From 1d92cb3fc7d05720a4f2d2ee288772d5aff85032 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Wed, 29 Oct 2025 16:39:11 -0700
Subject: [PATCH 01/12] update

---
 .../tensorrt/tensorrt_execution_provider.cc   | 64 +++++++++++++------
 1 file changed, 43 insertions(+), 21 deletions(-)

diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
index cd0c0e4bffdb5..932253ff1b634 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -2035,9 +2035,25 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
   }
 
   // Find inputs and outputs of the subgraph
+
   std::unique_ptr<IndexedSubGraph> sub_graph = onnxruntime::IndexedSubGraph::Create();
-  std::unordered_map<const NodeArg*, int> original_inputs, fused_inputs, fused_outputs, fused_outputs_to_add, graph_outputs_to_add;
+  std::unordered_map<const NodeArg*, int> original_inputs;
+
+  // These maps store the inputs and outpus of the subgraph.
+  // Please note that the inputs and outputs of the maps will be dynamically updated during node iteration
+  // to determine the final inputs and outputs of the subgraph.
+  std::unordered_map<const NodeArg*, int> fused_inputs, fused_outputs;
+
+  // This map stores the node's output that will be consumed by another node outside of this subgraph.
+  // So the node's output should be put into the subgraph's output list.
+  std::unordered_map<const NodeArg*, int> fused_outputs_to_add;
+
+  // This map stores the node's output that is original graph's output.
+  // So the node's output should be put into the subgraph's output list.
+  std::unordered_map<const NodeArg*, int> graph_outputs_to_add;
+
   std::unordered_set<const NodeArg*> erased;
+
   int input_order = 0;
   int output_order = 0;
 
@@ -2056,7 +2072,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
         erased.insert(input);
       } else if (erased.find(input) == erased.end()) {
         // Only when input is neither in output list nor erased list, add the input to input list
-        fused_inputs[input] = input_order++;
+        fused_inputs.insert({input, input_order++});
       }
     }
 
@@ -2071,7 +2087,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
         erased.insert(input);
       } else if (erased.find(input) == erased.end()) {
         // Only when input is neither in output list nor erased list, add the input to input list
-        fused_inputs[input] = input_order++;
+        fused_inputs.insert({input, input_order++});
       }
     }
 
@@ -2099,32 +2115,38 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
             erased.insert(output);
           } else if (erased.find(output) == erased.end()) {
             if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-              graph_outputs_to_add[output] = output_order;
+              // This output is the graph's output.
+              // So the output should be put into the subgraph's output list.
+              graph_outputs_to_add.insert({output, output_order});
             }
-            fused_outputs[output] = output_order++;
+            fused_outputs.insert({output, output_order++});
           }
         } else {
-          fused_outputs_to_add[output] = output_order++;
+          // This output will be consumed by another node outside of this subgraph.
+          // So the output should be put into the subgraph's output list.
+          fused_outputs_to_add.insert({output, output_order++});
         }
       }
-    } else {
-      for (const auto& output : node->OutputDefs()) {
-        const auto& it = fused_inputs.find(output);
-        if (it != fused_inputs.end()) {
-          fused_inputs.erase(it);
-          erased.insert(output);
-        }
-        // Only when output is neither in input list nor erased list, and the output is consumed by another node, add the output to output list
-        else if (erased.find(output) == erased.end()) {
-          if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-            graph_outputs_to_add[output] = output_order;
-          }
+    }
 
-          if (graph.GetGraph().GetConsumerNodes(output->Name()).size() > 0) {
-            fused_outputs[output] = output_order++;
-          }
+    for (const auto& output : node->OutputDefs()) {
+      const auto& it = fused_inputs.find(output);
+      if (it != fused_inputs.end()) {
+        fused_inputs.erase(it);
+        erased.insert(output);
+      } else if (erased.find(output) == erased.end()) {
+        if (graph.GetGraph().GetConsumerNodes(output->Name()).size() > 0) {
+          // Only when output is neither in input list nor erased list,
+          // and the output is consumed by another node, add the output to output list
+          fused_outputs.insert({output, output_order++});
         }
       }
+
+      if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
+        // This output is the graph's output.
+        // So the output should be put into the subgraph's output list.
+        graph_outputs_to_add.insert({output, output_order++});
+      }
     }
   }
 

From 3f00973323c34628a054f05d449ad7c2829fb005 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 30 Oct 2025 14:17:31 -0700
Subject: [PATCH 02/12] Add unit tests

---
 .../providers/tensorrt/tensorrt_basic_test.cc |  56 +++-
 .../providers/tensorrt/tensorrt_test_utils.cc | 284 ++++++++++++++++++
 .../providers/tensorrt/tensorrt_test_utils.h  |  10 +
 3 files changed, 349 insertions(+), 1 deletion(-)
 create mode 100644 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
 create mode 100644 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h

diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
index a746493d779f8..a0149613c21c3 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
@@ -1,5 +1,7 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
+// Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
+#include "onnxruntime_cxx_api.h"
+#include "tensorrt_test_utils.h"
 #include "core/graph/onnx_protobuf.h"
 #include "core/session/inference_session.h"
 #include "test/providers/provider_test_utils.h"
@@ -1358,5 +1360,57 @@ TEST(TensorrtExecutionProviderTest, RemoveCycleTest) {
   ASSERT_STATUS_OK(session_object.Run(run_options, feeds, output_names, &fetches));
   VerifyOutputs(fetches, expected_dims_mul_m, expected_values_mul_m);
 }
+
+TEST(TensorrtExecutionProviderTest, TestSessionOutputs) {
+  /*
+   * Model #1:
+   *
+   * "input" ---> TopK ---
+   *                     |---> "scores"
+   *                     |--- Less ---> "Less_output_0"
+   *                     |--- Div ---> "Div_output_0"
+   *                     |--- Mod ---> "labels"
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    OrtTensorRTProviderOptionsV2 provider_options;
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
+
+    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
+    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 4);
+  }
+
+  /*
+   * Model #2:
+   *
+   * "X" ---> Dropout ---> MatMul ---> "Y"
+   *          ^     |
+   *          |     |
+   * "W" ------     ----> Can't be graph's output
+   *
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    OrtTensorRTProviderOptionsV2 provider_options;
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
+
+    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
+    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 1);
+  }
+}
 }  // namespace test
 }  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
new file mode 100644
index 0000000000000..1456ea1a4e138
--- /dev/null
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
@@ -0,0 +1,284 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+#pragma once
+
+#include "onnxruntime_cxx_api.h"
+#include "core/common/path_string.h"
+
+#include <iostream>
+#include <fstream>
+
+#include <onnx/onnx_pb.h>
+#include <onnx/defs/schema.h>
+#include <onnx/checker.h>
+
+using namespace ONNX_NAMESPACE;
+
+#define ONNX_IR_VERSION 11
+#define OPSET_VERSION 23
+
+namespace onnxruntime {
+namespace test {
+
+// Helper: make TensorProto
+TensorProto MakeTensor(const std::string& name,
+                       TensorProto::DataType dtype,
+                       const std::vector<int64_t>& dims,
+                       const std::vector<int64_t>& vals) {
+  TensorProto t;
+  t.set_name(name);
+  t.set_data_type(dtype);
+  for (auto d : dims) t.add_dims(d);
+  for (auto v : vals) t.add_int64_data(v);
+  return t;
+}
+
+TensorProto MakeTensorFloat(const std::string& name,
+                            const std::vector<int64_t>& dims,
+                            const std::vector<float>& vals) {
+  TensorProto t;
+  t.set_name(name);
+  t.set_data_type(TensorProto::FLOAT);
+  for (auto d : dims) t.add_dims(d);
+  for (auto v : vals) t.add_float_data(v);
+  return t;
+}
+
+OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
+  // --------------------
+  // Create Model
+  // --------------------
+  ModelProto model;
+  model.set_ir_version(ONNX_IR_VERSION);
+  auto* opset = model.add_opset_import();
+  opset->set_domain("");  // empty = default ONNX domain
+  opset->set_version(OPSET_VERSION);
+
+  GraphProto* graph = model.mutable_graph();
+  graph->set_name("DropoutMatMulGraph");
+
+  // --------------------
+  // Create Inputs
+  // X: [3, 2]
+  // W: [2, 3]
+  // --------------------
+  {
+    auto* x = graph->add_input();
+    x->set_name("X");
+
+    auto* type = x->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(3);
+    shape->add_dim()->set_dim_value(2);
+  }
+
+  {
+    auto* w = graph->add_input();
+    w->set_name("W");
+
+    auto* type = w->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(2);
+    shape->add_dim()->set_dim_value(3);
+  }
+
+  // --------------------
+  // Output Y: [2, 3]
+  // --------------------
+  {
+    auto* x = graph->add_output();
+    x->set_name("Y");
+
+    auto* type = x->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(2);
+    shape->add_dim()->set_dim_value(3);
+  }
+
+  // --------------------
+  // Dropout Node
+  // --------------------
+  {
+    NodeProto* node = graph->add_node();
+    node->set_name("DropoutNode");
+    node->set_op_type("Dropout");
+
+    node->add_input("X");
+    node->add_output("dropout_out");
+    node->add_output("dropout_mask");
+  }
+
+  // --------------------
+  // MatMul Node
+  // --------------------
+  {
+    NodeProto* node = graph->add_node();
+    node->set_name("MatMulNode");
+    node->set_op_type("MatMul");
+
+    node->add_input("dropout_out");
+    node->add_input("W");
+    node->add_output("Y");
+  }
+
+  // --------------------
+  // Validate
+  // --------------------
+  try {
+    onnx::checker::check_model(model);
+  } catch (const std::exception& ex) {
+    std::string error_msg = "Model validation failed: " + std::string(ex.what());
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
+  }
+
+  std::ofstream ofs(model_name, std::ios::binary);
+  if (!model.SerializeToOstream(&ofs)) {
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
+  }
+
+  return nullptr;
+}
+
+OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
+  ModelProto model;
+  model.set_ir_version(ONNX_IR_VERSION);
+  auto* opset = model.add_opset_import();
+  opset->set_domain("");   // empty = default ONNX domain
+  opset->set_version(OPSET_VERSION);
+
+  auto* graph = model.mutable_graph();
+  graph->set_name("TopKGraph");
+
+  // ======================
+  // ---- Model Input ----
+  // ======================
+  {
+    auto* inp = graph->add_input();
+    inp->set_name("input");
+
+    auto* type = inp->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    // Shape: ["N"]
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_param("N");
+  }
+
+  // ======================
+  // ---- Initializers ----
+  // ======================
+  {
+    // K = [300]
+    TensorProto K = MakeTensor("K", TensorProto::INT64, {1}, {300});
+    *graph->add_initializer() = K;
+
+    // zero = 0.0 (scalar)
+    TensorProto zero = MakeTensor("zero", TensorProto::INT64, {}, {0});
+    *graph->add_initializer() = zero;
+
+    // twenty_six = 26 (scalar)
+    TensorProto ts = MakeTensor("twenty_six", TensorProto::INT64, {}, {26});
+    *graph->add_initializer() = ts;
+  }
+
+  // ======================
+  // ---- TopK ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("TopK");
+    n->add_input("input");
+    n->add_input("K");
+    n->add_output("scores");
+    n->add_output("topk_indices");
+    n->set_name("TopK");
+  }
+
+  // ======================
+  // ---- Less ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Less");
+    n->add_input("topk_indices");
+    n->add_input("zero");
+    n->add_output("Less_output_0");
+    n->set_name("Less");
+  }
+
+  // ======================
+  // ---- Div ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Div");
+    n->add_input("topk_indices");
+    n->add_input("twenty_six");
+    n->add_output("Div_17_output_0");
+    n->set_name("Div");
+  }
+
+  // ======================
+  // ---- Mod ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Mod");
+    n->add_input("topk_indices");
+    n->add_input("twenty_six");
+    n->add_output("labels");
+    n->set_name("Mod");
+  }
+
+  // =========================
+  // ---- Graph Outputs ----
+  // =========================
+  auto add_output = [&](const std::string& name, TensorProto::DataType type, const std::string& dim) {
+    auto* out = graph->add_output();
+    out->set_name(name);
+
+    auto* tt = out->mutable_type()->mutable_tensor_type();
+    tt->set_elem_type(type);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_param(dim);
+  };
+
+  add_output("scores", TensorProto::FLOAT, "K");
+  add_output("Less_output_0", TensorProto::BOOL, "K");
+  add_output("Div_17_output_0", TensorProto::INT64, "K");
+  add_output("labels", TensorProto::INT64, "K");
+
+  // ======================
+  // Validate + Save
+  // ======================
+  try {
+    onnx::checker::check_model(model);
+  } catch (const std::exception& e) {
+    std::string error_msg = "Model validation failed: " + std::string(e.what());
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
+  }
+
+  std::ofstream ofs(model_name, std::ios::binary);
+  if (!model.SerializeToOstream(&ofs)) {
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
+  }
+
+  return nullptr;
+}
+}
+}
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
new file mode 100644
index 0000000000000..9f98b1f396469
--- /dev/null
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
@@ -0,0 +1,10 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+#include "onnxruntime_cxx_api.h"
+
+namespace onnxruntime {
+namespace test {
+OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name);
+OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name);
+}
+}

From 4308033f331da6e02b54cec8b9dd2121224a177c Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 30 Oct 2025 14:30:47 -0700
Subject: [PATCH 03/12] lintrunner -a

---
 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc | 6 +++---
 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h  | 4 ++--
 2 files changed, 5 insertions(+), 5 deletions(-)

diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
index 1456ea1a4e138..18b66f85ec34f 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
@@ -154,7 +154,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   ModelProto model;
   model.set_ir_version(ONNX_IR_VERSION);
   auto* opset = model.add_opset_import();
-  opset->set_domain("");   // empty = default ONNX domain
+  opset->set_domain("");  // empty = default ONNX domain
   opset->set_version(OPSET_VERSION);
 
   auto* graph = model.mutable_graph();
@@ -280,5 +280,5 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
 
   return nullptr;
 }
-}
-}
+}  // namespace test
+}  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
index 9f98b1f396469..f13ccae413d31 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
@@ -6,5 +6,5 @@ namespace onnxruntime {
 namespace test {
 OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name);
 OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name);
-}
-}
+}  // namespace test
+}  // namespace onnxruntime

From fb44919859e686fd6c38bddf7cbb6fad994862b8 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 30 Oct 2025 14:31:24 -0700
Subject: [PATCH 04/12] update

---
 .../nv_tensorrt_rtx/nv_execution_provider.cc  |  64 +++--
 .../nv_tensorrt_rtx/nv_basic_test.cc          |  51 ++++
 .../test_nv_trt_rtx_ep_util.cc                | 261 ++++++++++++++++++
 .../nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h |   3 +
 4 files changed, 358 insertions(+), 21 deletions(-)

diff --git a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
index 62210d65848d1..3c6f6c9b82c83 100644
--- a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
+++ b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
@@ -1387,9 +1387,25 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
   }
 
   // Find inputs and outputs of the subgraph
+
   std::unique_ptr<IndexedSubGraph> sub_graph = onnxruntime::IndexedSubGraph::Create();
-  std::unordered_map<const NodeArg*, int> original_inputs, fused_inputs, fused_outputs, fused_outputs_to_add, graph_outputs_to_add;
+  std::unordered_map<const NodeArg*, int> original_inputs;
+
+  // These maps store the inputs and outpus of the subgraph.
+  // Please note that the inputs and outputs of the maps will be dynamically updated during node iteration
+  // to determine the final inputs and outputs of the subgraph.
+  std::unordered_map<const NodeArg*, int> fused_inputs, fused_outputs;
+
+  // This map stores the node's output that will be consumed by another node outside of this subgraph.
+  // So the node's output should be put into the subgraph's output list.
+  std::unordered_map<const NodeArg*, int> fused_outputs_to_add;
+
+  // This map stores the node's output that is original graph's output.
+  // So the node's output should be put into the subgraph's output list.
+  std::unordered_map<const NodeArg*, int> graph_outputs_to_add;
+
   std::unordered_set<const NodeArg*> erased;
+
   int input_order = 0;
   int output_order = 0;
 
@@ -1408,7 +1424,7 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
         erased.insert(input);
       } else if (erased.find(input) == erased.end()) {
         // Only when input is neither in output list nor erased list, add the input to input list
-        fused_inputs[input] = input_order++;
+        fused_inputs.insert({input, input_order++});
       }
     }
 
@@ -1423,7 +1439,7 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
         erased.insert(input);
       } else if (erased.find(input) == erased.end()) {
         // Only when input is neither in output list nor erased list, add the input to input list
-        fused_inputs[input] = input_order++;
+        fused_inputs.insert({input, input_order++});
       }
     }
 
@@ -1451,32 +1467,38 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
             erased.insert(output);
           } else if (erased.find(output) == erased.end()) {
             if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-              graph_outputs_to_add[output] = output_order;
+              // This output is the graph's output.
+              // So the output should be put into the subgraph's output list.
+              graph_outputs_to_add.insert({output, output_order});
             }
-            fused_outputs[output] = output_order++;
+            fused_outputs.insert({output, output_order++});
           }
         } else {
-          fused_outputs_to_add[output] = output_order++;
+          // This output will be consumed by another node outside of this subgraph.
+          // So the output should be put into the subgraph's output list.
+          fused_outputs_to_add.insert({output, output_order++});
         }
       }
-    } else {
-      for (const auto& output : node->OutputDefs()) {
-        const auto& it = fused_inputs.find(output);
-        if (it != fused_inputs.end()) {
-          fused_inputs.erase(it);
-          erased.insert(output);
-        }
-        // Only when output is neither in input list nor erased list, and the output is consumed by another node, add the output to output list
-        else if (erased.find(output) == erased.end()) {
-          if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-            graph_outputs_to_add[output] = output_order;
-          }
+    }
 
-          if (graph.GetGraph().GetConsumerNodes(output->Name()).size() > 0) {
-            fused_outputs[output] = output_order++;
-          }
+    for (const auto& output : node->OutputDefs()) {
+      const auto& it = fused_inputs.find(output);
+      if (it != fused_inputs.end()) {
+        fused_inputs.erase(it);
+        erased.insert(output);
+      } else if (erased.find(output) == erased.end()) {
+        if (graph.GetGraph().GetConsumerNodes(output->Name()).size() > 0) {
+          // Only when output is neither in input list nor erased list,
+          // and the output is consumed by another node, add the output to output list
+          fused_outputs.insert({output, output_order++});
         }
       }
+
+      if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
+        // This output is the graph's output.
+        // So the output should be put into the subgraph's output list.
+        graph_outputs_to_add.insert({output, output_order++});
+      }
     }
   }
 
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
index d8cc56d738175..1e14ec61d638a 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
@@ -398,6 +398,57 @@ TEST(NvExecutionProviderTest, DataTransfer) {
   device_tensor = Ort::Value();
 }
 
+TEST(NvExecutionProviderTest, TestSessionOutputs) {
+  /*
+   * Model #1:
+   *
+   * "input" ---> TopK ---
+   *                     |---> "scores"
+   *                     |--- Less ---> "Less_output_0"
+   *                     |--- Div ---> "Div_output_0"
+   *                     |--- Mod ---> "labels"
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    OrtTensorRTProviderOptionsV2 provider_options;
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
+
+    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
+    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 4);
+  }
+
+  /*
+   * Model #2:
+   *
+   * "X" ---> Dropout ---> MatMul ---> "Y"
+   *          ^     |
+   *          |     |
+   * "W" ------     ----> Can't be graph's output
+   *
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    OrtTensorRTProviderOptionsV2 provider_options;
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
+
+    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
+    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 1);
+  }
+
 #endif
 
 }  // namespace test
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
index 3a91fc1ba09bb..e2aaec7f097fb 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
@@ -464,5 +464,266 @@ Ort::IoBinding generate_io_binding(
   return binding;
 }
 
+// Helper: make TensorProto
+TensorProto MakeTensor(const std::string& name,
+                       TensorProto::DataType dtype,
+                       const std::vector<int64_t>& dims,
+                       const std::vector<int64_t>& vals) {
+  TensorProto t;
+  t.set_name(name);
+  t.set_data_type(dtype);
+  for (auto d : dims) t.add_dims(d);
+  for (auto v : vals) t.add_int64_data(v);
+  return t;
+}
+
+TensorProto MakeTensorFloat(const std::string& name,
+                            const std::vector<int64_t>& dims,
+                            const std::vector<float>& vals) {
+  TensorProto t;
+  t.set_name(name);
+  t.set_data_type(TensorProto::FLOAT);
+  for (auto d : dims) t.add_dims(d);
+  for (auto v : vals) t.add_float_data(v);
+  return t;
+}
+
+OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
+  // --------------------
+  // Create Model
+  // --------------------
+  ModelProto model;
+  model.set_ir_version(ONNX_IR_VERSION);
+  auto* opset = model.add_opset_import();
+  opset->set_domain("");  // empty = default ONNX domain
+  opset->set_version(OPSET_VERSION);
+
+  GraphProto* graph = model.mutable_graph();
+  graph->set_name("DropoutMatMulGraph");
+
+  // --------------------
+  // Create Inputs
+  // X: [3, 2]
+  // W: [2, 3]
+  // --------------------
+  {
+    auto* x = graph->add_input();
+    x->set_name("X");
+
+    auto* type = x->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(3);
+    shape->add_dim()->set_dim_value(2);
+  }
+
+  {
+    auto* w = graph->add_input();
+    w->set_name("W");
+
+    auto* type = w->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(2);
+    shape->add_dim()->set_dim_value(3);
+  }
+
+  // --------------------
+  // Output Y: [2, 3]
+  // --------------------
+  {
+    auto* x = graph->add_output();
+    x->set_name("Y");
+
+    auto* type = x->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_value(2);
+    shape->add_dim()->set_dim_value(3);
+  }
+
+  // --------------------
+  // Dropout Node
+  // --------------------
+  {
+    NodeProto* node = graph->add_node();
+    node->set_name("DropoutNode");
+    node->set_op_type("Dropout");
+
+    node->add_input("X");
+    node->add_output("dropout_out");
+    node->add_output("dropout_mask");
+  }
+
+  // --------------------
+  // MatMul Node
+  // --------------------
+  {
+    NodeProto* node = graph->add_node();
+    node->set_name("MatMulNode");
+    node->set_op_type("MatMul");
+
+    node->add_input("dropout_out");
+    node->add_input("W");
+    node->add_output("Y");
+  }
+
+  // --------------------
+  // Validate
+  // --------------------
+  try {
+    onnx::checker::check_model(model);
+  } catch (const std::exception& ex) {
+    std::string error_msg = "Model validation failed: " + std::string(ex.what());
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
+  }
+
+  std::ofstream ofs(model_name, std::ios::binary);
+  if (!model.SerializeToOstream(&ofs)) {
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
+  }
+
+  return nullptr;
+}
+
+OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
+  ModelProto model;
+  model.set_ir_version(ONNX_IR_VERSION);
+  auto* opset = model.add_opset_import();
+  opset->set_domain("");  // empty = default ONNX domain
+  opset->set_version(OPSET_VERSION);
+
+  auto* graph = model.mutable_graph();
+  graph->set_name("TopKGraph");
+
+  // ======================
+  // ---- Model Input ----
+  // ======================
+  {
+    auto* inp = graph->add_input();
+    inp->set_name("input");
+
+    auto* type = inp->mutable_type();
+    auto* tt = type->mutable_tensor_type();
+    tt->set_elem_type(TensorProto::FLOAT);
+
+    // Shape: ["N"]
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_param("N");
+  }
+
+  // ======================
+  // ---- Initializers ----
+  // ======================
+  {
+    // K = [300]
+    TensorProto K = MakeTensor("K", TensorProto::INT64, {1}, {300});
+    *graph->add_initializer() = K;
+
+    // zero = 0.0 (scalar)
+    TensorProto zero = MakeTensor("zero", TensorProto::INT64, {}, {0});
+    *graph->add_initializer() = zero;
+
+    // twenty_six = 26 (scalar)
+    TensorProto ts = MakeTensor("twenty_six", TensorProto::INT64, {}, {26});
+    *graph->add_initializer() = ts;
+  }
+
+  // ======================
+  // ---- TopK ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("TopK");
+    n->add_input("input");
+    n->add_input("K");
+    n->add_output("scores");
+    n->add_output("topk_indices");
+    n->set_name("TopK");
+  }
+
+  // ======================
+  // ---- Less ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Less");
+    n->add_input("topk_indices");
+    n->add_input("zero");
+    n->add_output("Less_output_0");
+    n->set_name("Less");
+  }
+
+  // ======================
+  // ---- Div ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Div");
+    n->add_input("topk_indices");
+    n->add_input("twenty_six");
+    n->add_output("Div_17_output_0");
+    n->set_name("Div");
+  }
+
+  // ======================
+  // ---- Mod ----
+  // ======================
+  {
+    NodeProto* n = graph->add_node();
+    n->set_op_type("Mod");
+    n->add_input("topk_indices");
+    n->add_input("twenty_six");
+    n->add_output("labels");
+    n->set_name("Mod");
+  }
+
+  // =========================
+  // ---- Graph Outputs ----
+  // =========================
+  auto add_output = [&](const std::string& name, TensorProto::DataType type, const std::string& dim) {
+    auto* out = graph->add_output();
+    out->set_name(name);
+
+    auto* tt = out->mutable_type()->mutable_tensor_type();
+    tt->set_elem_type(type);
+
+    auto* shape = tt->mutable_shape();
+    shape->add_dim()->set_dim_param(dim);
+  };
+
+  add_output("scores", TensorProto::FLOAT, "K");
+  add_output("Less_output_0", TensorProto::BOOL, "K");
+  add_output("Div_17_output_0", TensorProto::INT64, "K");
+  add_output("labels", TensorProto::INT64, "K");
+
+  // ======================
+  // Validate + Save
+  // ======================
+  try {
+    onnx::checker::check_model(model);
+  } catch (const std::exception& e) {
+    std::string error_msg = "Model validation failed: " + std::string(e.what());
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
+  }
+
+  std::ofstream ofs(model_name, std::ios::binary);
+  if (!model.SerializeToOstream(&ofs)) {
+    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
+    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
+  }
+
+  return nullptr;
+}
+
 }  // namespace test
 }  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
index 0f011af8211ca..11000a76ad2cd 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
@@ -119,6 +119,9 @@ void CreateBaseModel(const PathString& model_name,
 
 void CreateLargeLLMModel(const PathString& model_path, const PathString& external_data_path);
 
+OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name);
+OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name);
+
 Ort::IoBinding generate_io_binding(
     Ort::Session& session,
     std::map<std::string, std::vector<int64_t>> shape_overwrites = {},

From f763515ae67e23312c0c0b4b66cad8b3d63c76ae Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 30 Oct 2025 17:41:15 -0700
Subject: [PATCH 05/12] fix build issues

---
 .../nv_tensorrt_rtx/nv_basic_test.cc          | 101 +++++++++---------
 .../test_nv_trt_rtx_ep_util.cc                |  72 +++++++------
 .../providers/tensorrt/tensorrt_test_utils.cc |  66 ++++++------
 .../providers/tensorrt/tensorrt_test_utils.h  |   1 +
 4 files changed, 123 insertions(+), 117 deletions(-)

diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
index 1e14ec61d638a..72b275c14ea3b 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
@@ -203,6 +203,56 @@ TEST_P(TypeTests, IOTypes) {
   }
 }
 
+TEST(NvExecutionProviderTest, TestSessionOutputs) {
+  /*
+   * Model #1:
+   *
+   * "input" ---> TopK ---
+   *                     |---> "scores"
+   *                     |--- Less ---> "Less_output_0"
+   *                     |--- Div ---> "Div_output_0"
+   *                     |--- Mod ---> "labels"
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider(kNvTensorRTRTXExecutionProvider, {});
+
+    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
+    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 4);
+  }
+
+  /*
+   * Model #2:
+   *
+   * "X" ---> Dropout ---> MatMul ---> "Y"
+   *          ^     |
+   *          |     |
+   * "W" ------     ----> Can't be graph's output
+   *
+   */
+  {
+    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
+    Ort::SessionOptions session_options;
+    session_options.AppendExecutionProvider(kNvTensorRTRTXExecutionProvider, {});
+
+    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
+    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
+    ASSERT_TRUE(status.IsOK());
+
+    Ort::Session session(env, model_path, session_options);
+
+    size_t output_count = session.GetOutputCount();
+    ASSERT_TRUE(output_count == 1);
+  }
+}
+
 INSTANTIATE_TEST_SUITE_P(NvExecutionProviderTest, TypeTests,
                          ::testing::Values(ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT,
                                            ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT16,
@@ -398,57 +448,6 @@ TEST(NvExecutionProviderTest, DataTransfer) {
   device_tensor = Ort::Value();
 }
 
-TEST(NvExecutionProviderTest, TestSessionOutputs) {
-  /*
-   * Model #1:
-   *
-   * "input" ---> TopK ---
-   *                     |---> "scores"
-   *                     |--- Less ---> "Less_output_0"
-   *                     |--- Div ---> "Div_output_0"
-   *                     |--- Mod ---> "labels"
-   */
-  {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
-    OrtTensorRTProviderOptionsV2 provider_options;
-    Ort::SessionOptions session_options;
-    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
-
-    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
-    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
-
-    size_t output_count = session.GetOutputCount();
-    ASSERT_TRUE(output_count == 4);
-  }
-
-  /*
-   * Model #2:
-   *
-   * "X" ---> Dropout ---> MatMul ---> "Y"
-   *          ^     |
-   *          |     |
-   * "W" ------     ----> Can't be graph's output
-   *
-   */
-  {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
-    OrtTensorRTProviderOptionsV2 provider_options;
-    Ort::SessionOptions session_options;
-    session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
-
-    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
-    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
-
-    size_t output_count = session.GetOutputCount();
-    ASSERT_TRUE(output_count == 1);
-  }
-
 #endif
 
 }  // namespace test
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
index e2aaec7f097fb..79cc2fef42666 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
@@ -19,6 +19,14 @@
 #include "test/common/trt_op_test_utils.h"
 #include "test/providers/provider_test_utils.h"
 #include "test/unittest_util/framework_test_utils.h"
+#include <iostream>
+#include <fstream>
+#include <onnx/onnx_pb.h>
+#include <onnx/defs/schema.h>
+#include <onnx/checker.h>
+
+#define ONNX_IR_VERSION 11
+#define OPSET_VERSION 23
 
 namespace onnxruntime {
 namespace test {
@@ -464,12 +472,12 @@ Ort::IoBinding generate_io_binding(
   return binding;
 }
 
-// Helper: make TensorProto
-TensorProto MakeTensor(const std::string& name,
-                       TensorProto::DataType dtype,
-                       const std::vector<int64_t>& dims,
-                       const std::vector<int64_t>& vals) {
-  TensorProto t;
+// Helper: make ONNX_NAMESPACE::TensorProto
+ONNX_NAMESPACE::TensorProto MakeTensor(const std::string& name,
+                                       ONNX_NAMESPACE::TensorProto::DataType dtype,
+                                       const std::vector<int64_t>& dims,
+                                       const std::vector<int64_t>& vals) {
+  ONNX_NAMESPACE::TensorProto t;
   t.set_name(name);
   t.set_data_type(dtype);
   for (auto d : dims) t.add_dims(d);
@@ -477,12 +485,12 @@ TensorProto MakeTensor(const std::string& name,
   return t;
 }
 
-TensorProto MakeTensorFloat(const std::string& name,
-                            const std::vector<int64_t>& dims,
-                            const std::vector<float>& vals) {
-  TensorProto t;
+ONNX_NAMESPACE::TensorProto MakeTensorFloat(const std::string& name,
+                                            const std::vector<int64_t>& dims,
+                                            const std::vector<float>& vals) {
+  ONNX_NAMESPACE::TensorProto t;
   t.set_name(name);
-  t.set_data_type(TensorProto::FLOAT);
+  t.set_data_type(ONNX_NAMESPACE::TensorProto::FLOAT);
   for (auto d : dims) t.add_dims(d);
   for (auto v : vals) t.add_float_data(v);
   return t;
@@ -492,13 +500,13 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // --------------------
   // Create Model
   // --------------------
-  ModelProto model;
+  ONNX_NAMESPACE::ModelProto model;
   model.set_ir_version(ONNX_IR_VERSION);
   auto* opset = model.add_opset_import();
   opset->set_domain("");  // empty = default ONNX domain
   opset->set_version(OPSET_VERSION);
 
-  GraphProto* graph = model.mutable_graph();
+  ONNX_NAMESPACE::GraphProto* graph = model.mutable_graph();
   graph->set_name("DropoutMatMulGraph");
 
   // --------------------
@@ -512,7 +520,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = x->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(3);
@@ -525,7 +533,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = w->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(2);
@@ -541,7 +549,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = x->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(2);
@@ -552,7 +560,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // Dropout Node
   // --------------------
   {
-    NodeProto* node = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
     node->set_name("DropoutNode");
     node->set_op_type("Dropout");
 
@@ -565,7 +573,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // MatMul Node
   // --------------------
   {
-    NodeProto* node = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
     node->set_name("MatMulNode");
     node->set_op_type("MatMul");
 
@@ -595,7 +603,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 }
 
 OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
-  ModelProto model;
+  ONNX_NAMESPACE::ModelProto model;
   model.set_ir_version(ONNX_IR_VERSION);
   auto* opset = model.add_opset_import();
   opset->set_domain("");  // empty = default ONNX domain
@@ -613,7 +621,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
 
     auto* type = inp->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     // Shape: ["N"]
     auto* shape = tt->mutable_shape();
@@ -625,15 +633,15 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ======================
   {
     // K = [300]
-    TensorProto K = MakeTensor("K", TensorProto::INT64, {1}, {300});
+    ONNX_NAMESPACE::TensorProto K = MakeTensor("K", ONNX_NAMESPACE::TensorProto::INT64, {1}, {300});
     *graph->add_initializer() = K;
 
     // zero = 0.0 (scalar)
-    TensorProto zero = MakeTensor("zero", TensorProto::INT64, {}, {0});
+    ONNX_NAMESPACE::TensorProto zero = MakeTensor("zero", ONNX_NAMESPACE::TensorProto::INT64, {}, {0});
     *graph->add_initializer() = zero;
 
     // twenty_six = 26 (scalar)
-    TensorProto ts = MakeTensor("twenty_six", TensorProto::INT64, {}, {26});
+    ONNX_NAMESPACE::TensorProto ts = MakeTensor("twenty_six", ONNX_NAMESPACE::TensorProto::INT64, {}, {26});
     *graph->add_initializer() = ts;
   }
 
@@ -641,7 +649,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- TopK ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("TopK");
     n->add_input("input");
     n->add_input("K");
@@ -654,7 +662,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Less ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Less");
     n->add_input("topk_indices");
     n->add_input("zero");
@@ -666,7 +674,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Div ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Div");
     n->add_input("topk_indices");
     n->add_input("twenty_six");
@@ -678,7 +686,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Mod ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Mod");
     n->add_input("topk_indices");
     n->add_input("twenty_six");
@@ -689,7 +697,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // =========================
   // ---- Graph Outputs ----
   // =========================
-  auto add_output = [&](const std::string& name, TensorProto::DataType type, const std::string& dim) {
+  auto add_output = [&](const std::string& name, ONNX_NAMESPACE::TensorProto::DataType type, const std::string& dim) {
     auto* out = graph->add_output();
     out->set_name(name);
 
@@ -700,10 +708,10 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
     shape->add_dim()->set_dim_param(dim);
   };
 
-  add_output("scores", TensorProto::FLOAT, "K");
-  add_output("Less_output_0", TensorProto::BOOL, "K");
-  add_output("Div_17_output_0", TensorProto::INT64, "K");
-  add_output("labels", TensorProto::INT64, "K");
+  add_output("scores", ONNX_NAMESPACE::TensorProto::FLOAT, "K");
+  add_output("Less_output_0", ONNX_NAMESPACE::TensorProto::BOOL, "K");
+  add_output("Div_17_output_0", ONNX_NAMESPACE::TensorProto::INT64, "K");
+  add_output("labels", ONNX_NAMESPACE::TensorProto::INT64, "K");
 
   // ======================
   // Validate + Save
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
index 18b66f85ec34f..43e68e1879d34 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
@@ -12,20 +12,18 @@
 #include <onnx/defs/schema.h>
 #include <onnx/checker.h>
 
-using namespace ONNX_NAMESPACE;
-
 #define ONNX_IR_VERSION 11
 #define OPSET_VERSION 23
 
 namespace onnxruntime {
 namespace test {
 
-// Helper: make TensorProto
-TensorProto MakeTensor(const std::string& name,
-                       TensorProto::DataType dtype,
-                       const std::vector<int64_t>& dims,
-                       const std::vector<int64_t>& vals) {
-  TensorProto t;
+// Helper: make ONNX_NAMESPACE::TensorProto
+ONNX_NAMESPACE::TensorProto MakeTensor(const std::string& name,
+                                       ONNX_NAMESPACE::TensorProto::DataType dtype,
+                                       const std::vector<int64_t>& dims,
+                                       const std::vector<int64_t>& vals) {
+  ONNX_NAMESPACE::TensorProto t;
   t.set_name(name);
   t.set_data_type(dtype);
   for (auto d : dims) t.add_dims(d);
@@ -33,12 +31,12 @@ TensorProto MakeTensor(const std::string& name,
   return t;
 }
 
-TensorProto MakeTensorFloat(const std::string& name,
-                            const std::vector<int64_t>& dims,
-                            const std::vector<float>& vals) {
-  TensorProto t;
+ONNX_NAMESPACE::TensorProto MakeTensorFloat(const std::string& name,
+                                            const std::vector<int64_t>& dims,
+                                            const std::vector<float>& vals) {
+  ONNX_NAMESPACE::TensorProto t;
   t.set_name(name);
-  t.set_data_type(TensorProto::FLOAT);
+  t.set_data_type(ONNX_NAMESPACE::TensorProto::FLOAT);
   for (auto d : dims) t.add_dims(d);
   for (auto v : vals) t.add_float_data(v);
   return t;
@@ -48,13 +46,13 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // --------------------
   // Create Model
   // --------------------
-  ModelProto model;
+  ONNX_NAMESPACE::ModelProto model;
   model.set_ir_version(ONNX_IR_VERSION);
   auto* opset = model.add_opset_import();
   opset->set_domain("");  // empty = default ONNX domain
   opset->set_version(OPSET_VERSION);
 
-  GraphProto* graph = model.mutable_graph();
+  ONNX_NAMESPACE::GraphProto* graph = model.mutable_graph();
   graph->set_name("DropoutMatMulGraph");
 
   // --------------------
@@ -68,7 +66,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = x->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(3);
@@ -81,7 +79,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = w->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(2);
@@ -97,7 +95,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 
     auto* type = x->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     auto* shape = tt->mutable_shape();
     shape->add_dim()->set_dim_value(2);
@@ -108,7 +106,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // Dropout Node
   // --------------------
   {
-    NodeProto* node = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
     node->set_name("DropoutNode");
     node->set_op_type("Dropout");
 
@@ -121,7 +119,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
   // MatMul Node
   // --------------------
   {
-    NodeProto* node = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
     node->set_name("MatMulNode");
     node->set_op_type("MatMul");
 
@@ -151,7 +149,7 @@ OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
 }
 
 OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
-  ModelProto model;
+  ONNX_NAMESPACE::ModelProto model;
   model.set_ir_version(ONNX_IR_VERSION);
   auto* opset = model.add_opset_import();
   opset->set_domain("");  // empty = default ONNX domain
@@ -169,7 +167,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
 
     auto* type = inp->mutable_type();
     auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(TensorProto::FLOAT);
+    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
 
     // Shape: ["N"]
     auto* shape = tt->mutable_shape();
@@ -181,15 +179,15 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ======================
   {
     // K = [300]
-    TensorProto K = MakeTensor("K", TensorProto::INT64, {1}, {300});
+    ONNX_NAMESPACE::TensorProto K = MakeTensor("K", ONNX_NAMESPACE::TensorProto::INT64, {1}, {300});
     *graph->add_initializer() = K;
 
     // zero = 0.0 (scalar)
-    TensorProto zero = MakeTensor("zero", TensorProto::INT64, {}, {0});
+    ONNX_NAMESPACE::TensorProto zero = MakeTensor("zero", ONNX_NAMESPACE::TensorProto::INT64, {}, {0});
     *graph->add_initializer() = zero;
 
     // twenty_six = 26 (scalar)
-    TensorProto ts = MakeTensor("twenty_six", TensorProto::INT64, {}, {26});
+    ONNX_NAMESPACE::TensorProto ts = MakeTensor("twenty_six", ONNX_NAMESPACE::TensorProto::INT64, {}, {26});
     *graph->add_initializer() = ts;
   }
 
@@ -197,7 +195,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- TopK ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("TopK");
     n->add_input("input");
     n->add_input("K");
@@ -210,7 +208,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Less ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Less");
     n->add_input("topk_indices");
     n->add_input("zero");
@@ -222,7 +220,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Div ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Div");
     n->add_input("topk_indices");
     n->add_input("twenty_six");
@@ -234,7 +232,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // ---- Mod ----
   // ======================
   {
-    NodeProto* n = graph->add_node();
+    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
     n->set_op_type("Mod");
     n->add_input("topk_indices");
     n->add_input("twenty_six");
@@ -245,7 +243,7 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
   // =========================
   // ---- Graph Outputs ----
   // =========================
-  auto add_output = [&](const std::string& name, TensorProto::DataType type, const std::string& dim) {
+  auto add_output = [&](const std::string& name, ONNX_NAMESPACE::TensorProto::DataType type, const std::string& dim) {
     auto* out = graph->add_output();
     out->set_name(name);
 
@@ -256,10 +254,10 @@ OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_n
     shape->add_dim()->set_dim_param(dim);
   };
 
-  add_output("scores", TensorProto::FLOAT, "K");
-  add_output("Less_output_0", TensorProto::BOOL, "K");
-  add_output("Div_17_output_0", TensorProto::INT64, "K");
-  add_output("labels", TensorProto::INT64, "K");
+  add_output("scores", ONNX_NAMESPACE::TensorProto::FLOAT, "K");
+  add_output("Less_output_0", ONNX_NAMESPACE::TensorProto::BOOL, "K");
+  add_output("Div_17_output_0", ONNX_NAMESPACE::TensorProto::INT64, "K");
+  add_output("labels", ONNX_NAMESPACE::TensorProto::INT64, "K");
 
   // ======================
   // Validate + Save
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
index f13ccae413d31..3857d8ee80131 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
@@ -1,6 +1,7 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 #include "onnxruntime_cxx_api.h"
+#include "core/common/path_string.h"
 
 namespace onnxruntime {
 namespace test {

From 186c938c014019b2ca5d80320631db6419bc687c Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Fri, 31 Oct 2025 12:30:11 -0700
Subject: [PATCH 06/12] fix build warning

---
 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc | 1 -
 1 file changed, 1 deletion(-)

diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
index 43e68e1879d34..51738342c6cb0 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
@@ -1,6 +1,5 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
-#pragma once
 
 #include "onnxruntime_cxx_api.h"
 #include "core/common/path_string.h"

From fe795311303eafe872be26e7d1eee81b143fee07 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 20 Nov 2025 10:36:12 -0800
Subject: [PATCH 07/12] add comments and refactor the logic of adding graph's
 outputs

---
 .../nv_tensorrt_rtx/nv_execution_provider.cc  | 21 +++++++------------
 .../tensorrt/tensorrt_execution_provider.cc   | 21 +++++++------------
 2 files changed, 14 insertions(+), 28 deletions(-)

diff --git a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
index 3c6f6c9b82c83..d49621bfe700e 100644
--- a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
+++ b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
@@ -1406,6 +1406,11 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
 
   std::unordered_set<const NodeArg*> erased;
 
+  // This is the relative ordering that ensures node's input or output being added to the 'fused_inputs',
+  // 'fused_outputs', 'fused_outputs_to_add' and 'graph_outputs_to_add' maps is associated with a relative order index.
+  // Items added earlier receive a smaller order index than items added later.
+  // When constructing the final sub_graph's input or output lists, entries with smaller
+  // order indices will appear before those with larger indices.
   int input_order = 0;
   int output_order = 0;
 
@@ -1460,20 +1465,8 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
         } else {
           output = (it->GetNode()).ImplicitInputDefs()[it->GetDstArgIndex() - static_cast<int>(it->GetNode().InputDefs().size())];
         }
-        if (node_set.find(node_idx) != node_set.end()) {
-          const auto& iter = fused_inputs.find(output);
-          if (iter != fused_inputs.end()) {
-            fused_inputs.erase(iter);
-            erased.insert(output);
-          } else if (erased.find(output) == erased.end()) {
-            if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-              // This output is the graph's output.
-              // So the output should be put into the subgraph's output list.
-              graph_outputs_to_add.insert({output, output_order});
-            }
-            fused_outputs.insert({output, output_order++});
-          }
-        } else {
+
+        if (node_set.find(node_idx) == node_set.end()) {
           // This output will be consumed by another node outside of this subgraph.
           // So the output should be put into the subgraph's output list.
           fused_outputs_to_add.insert({output, output_order++});
diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
index 932253ff1b634..0b92a4f648c80 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -2054,6 +2054,11 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
 
   std::unordered_set<const NodeArg*> erased;
 
+  // This is the relative ordering that ensures node's input or output being added to the 'fused_inputs',
+  // 'fused_outputs', 'fused_outputs_to_add' and 'graph_outputs_to_add' maps is associated with a relative order index.
+  // Items added earlier receive a smaller order index than items added later.
+  // When constructing the final sub_graph's input or output lists, entries with smaller
+  // order indices will appear before those with larger indices.
   int input_order = 0;
   int output_order = 0;
 
@@ -2108,20 +2113,8 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
         } else {
           output = (it->GetNode()).ImplicitInputDefs()[it->GetDstArgIndex() - static_cast<int>(it->GetNode().InputDefs().size())];
         }
-        if (node_set.find(node_idx) != node_set.end()) {
-          const auto& iter = fused_inputs.find(output);
-          if (iter != fused_inputs.end()) {
-            fused_inputs.erase(iter);
-            erased.insert(output);
-          } else if (erased.find(output) == erased.end()) {
-            if (graph_output_names.find(output->Name()) != graph_output_names.end()) {
-              // This output is the graph's output.
-              // So the output should be put into the subgraph's output list.
-              graph_outputs_to_add.insert({output, output_order});
-            }
-            fused_outputs.insert({output, output_order++});
-          }
-        } else {
+
+        if (node_set.find(node_idx) == node_set.end()) {
           // This output will be consumed by another node outside of this subgraph.
           // So the output should be put into the subgraph's output list.
           fused_outputs_to_add.insert({output, output_order++});

From aef06028e034c16dc7f01ca82cdc74b5ce3c7097 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 20 Nov 2025 10:47:58 -0800
Subject: [PATCH 08/12] fix typo

---
 .../core/providers/nv_tensorrt_rtx/nv_execution_provider.cc     | 2 +-
 .../core/providers/tensorrt/tensorrt_execution_provider.cc      | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
index d49621bfe700e..0e683db1c0d53 100644
--- a/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
+++ b/onnxruntime/core/providers/nv_tensorrt_rtx/nv_execution_provider.cc
@@ -1391,7 +1391,7 @@ std::unique_ptr<IndexedSubGraph> NvExecutionProvider::GetSubGraph(SubGraph_t gra
   std::unique_ptr<IndexedSubGraph> sub_graph = onnxruntime::IndexedSubGraph::Create();
   std::unordered_map<const NodeArg*, int> original_inputs;
 
-  // These maps store the inputs and outpus of the subgraph.
+  // These maps store the inputs and outputs of the subgraph.
   // Please note that the inputs and outputs of the maps will be dynamically updated during node iteration
   // to determine the final inputs and outputs of the subgraph.
   std::unordered_map<const NodeArg*, int> fused_inputs, fused_outputs;
diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
index 0b92a4f648c80..e5b48da33fbc3 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -2039,7 +2039,7 @@ std::unique_ptr<IndexedSubGraph> TensorrtExecutionProvider::GetSubGraph(SubGraph
   std::unique_ptr<IndexedSubGraph> sub_graph = onnxruntime::IndexedSubGraph::Create();
   std::unordered_map<const NodeArg*, int> original_inputs;
 
-  // These maps store the inputs and outpus of the subgraph.
+  // These maps store the inputs and outputs of the subgraph.
   // Please note that the inputs and outputs of the maps will be dynamically updated during node iteration
   // to determine the final inputs and outputs of the subgraph.
   std::unordered_map<const NodeArg*, int> fused_inputs, fused_outputs;

From f169747c9d37e11a620f3bcfd3a6e5f5408b630a Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 20 Nov 2025 11:27:23 -0800
Subject: [PATCH 09/12] address reviewer's comments

---
 .../nv_tensorrt_rtx/nv_basic_test.cc          |  16 +-
 .../test_nv_trt_rtx_ep_util.cc                | 269 -----------------
 .../nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h |   3 -
 .../providers/tensorrt/tensorrt_basic_test.cc |  18 +-
 .../providers/tensorrt/tensorrt_test_utils.cc | 281 ------------------
 .../providers/tensorrt/tensorrt_test_utils.h  |  11 -
 .../test/testdata/node_output_not_used.onnx   | Bin 0 -> 189 bytes
 .../test/testdata/node_output_not_used.py     |  42 +++
 .../topk_and_multiple_graph_outputs.onnx      | Bin 0 -> 393 bytes
 .../topk_and_multiple_graph_outputs.py        |  86 ++++++
 10 files changed, 138 insertions(+), 588 deletions(-)
 delete mode 100644 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
 delete mode 100644 onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
 create mode 100644 onnxruntime/test/testdata/node_output_not_used.onnx
 create mode 100644 onnxruntime/test/testdata/node_output_not_used.py
 create mode 100644 onnxruntime/test/testdata/topk_and_multiple_graph_outputs.onnx
 create mode 100644 onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py

diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
index 72b275c14ea3b..af9706855ee3c 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/nv_basic_test.cc
@@ -214,15 +214,11 @@ TEST(NvExecutionProviderTest, TestSessionOutputs) {
    *                     |--- Mod ---> "labels"
    */
   {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
     Ort::SessionOptions session_options;
     session_options.AppendExecutionProvider(kNvTensorRTRTXExecutionProvider, {});
 
-    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
-    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
+    auto model_path = ORT_TSTR("testdata/topk_and_multiple_graph_outputs.onnx");
+    Ort::Session session(*ort_env, model_path, session_options);
 
     size_t output_count = session.GetOutputCount();
     ASSERT_TRUE(output_count == 4);
@@ -238,15 +234,11 @@ TEST(NvExecutionProviderTest, TestSessionOutputs) {
    *
    */
   {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
     Ort::SessionOptions session_options;
     session_options.AppendExecutionProvider(kNvTensorRTRTXExecutionProvider, {});
 
-    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
-    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
+    auto model_path = ORT_TSTR("testdata/node_output_not_used.onnx");
+    Ort::Session session(*ort_env, model_path, session_options);
 
     size_t output_count = session.GetOutputCount();
     ASSERT_TRUE(output_count == 1);
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
index 79cc2fef42666..3a91fc1ba09bb 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.cc
@@ -19,14 +19,6 @@
 #include "test/common/trt_op_test_utils.h"
 #include "test/providers/provider_test_utils.h"
 #include "test/unittest_util/framework_test_utils.h"
-#include <iostream>
-#include <fstream>
-#include <onnx/onnx_pb.h>
-#include <onnx/defs/schema.h>
-#include <onnx/checker.h>
-
-#define ONNX_IR_VERSION 11
-#define OPSET_VERSION 23
 
 namespace onnxruntime {
 namespace test {
@@ -472,266 +464,5 @@ Ort::IoBinding generate_io_binding(
   return binding;
 }
 
-// Helper: make ONNX_NAMESPACE::TensorProto
-ONNX_NAMESPACE::TensorProto MakeTensor(const std::string& name,
-                                       ONNX_NAMESPACE::TensorProto::DataType dtype,
-                                       const std::vector<int64_t>& dims,
-                                       const std::vector<int64_t>& vals) {
-  ONNX_NAMESPACE::TensorProto t;
-  t.set_name(name);
-  t.set_data_type(dtype);
-  for (auto d : dims) t.add_dims(d);
-  for (auto v : vals) t.add_int64_data(v);
-  return t;
-}
-
-ONNX_NAMESPACE::TensorProto MakeTensorFloat(const std::string& name,
-                                            const std::vector<int64_t>& dims,
-                                            const std::vector<float>& vals) {
-  ONNX_NAMESPACE::TensorProto t;
-  t.set_name(name);
-  t.set_data_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-  for (auto d : dims) t.add_dims(d);
-  for (auto v : vals) t.add_float_data(v);
-  return t;
-}
-
-OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
-  // --------------------
-  // Create Model
-  // --------------------
-  ONNX_NAMESPACE::ModelProto model;
-  model.set_ir_version(ONNX_IR_VERSION);
-  auto* opset = model.add_opset_import();
-  opset->set_domain("");  // empty = default ONNX domain
-  opset->set_version(OPSET_VERSION);
-
-  ONNX_NAMESPACE::GraphProto* graph = model.mutable_graph();
-  graph->set_name("DropoutMatMulGraph");
-
-  // --------------------
-  // Create Inputs
-  // X: [3, 2]
-  // W: [2, 3]
-  // --------------------
-  {
-    auto* x = graph->add_input();
-    x->set_name("X");
-
-    auto* type = x->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(3);
-    shape->add_dim()->set_dim_value(2);
-  }
-
-  {
-    auto* w = graph->add_input();
-    w->set_name("W");
-
-    auto* type = w->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(2);
-    shape->add_dim()->set_dim_value(3);
-  }
-
-  // --------------------
-  // Output Y: [2, 3]
-  // --------------------
-  {
-    auto* x = graph->add_output();
-    x->set_name("Y");
-
-    auto* type = x->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(2);
-    shape->add_dim()->set_dim_value(3);
-  }
-
-  // --------------------
-  // Dropout Node
-  // --------------------
-  {
-    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
-    node->set_name("DropoutNode");
-    node->set_op_type("Dropout");
-
-    node->add_input("X");
-    node->add_output("dropout_out");
-    node->add_output("dropout_mask");
-  }
-
-  // --------------------
-  // MatMul Node
-  // --------------------
-  {
-    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
-    node->set_name("MatMulNode");
-    node->set_op_type("MatMul");
-
-    node->add_input("dropout_out");
-    node->add_input("W");
-    node->add_output("Y");
-  }
-
-  // --------------------
-  // Validate
-  // --------------------
-  try {
-    onnx::checker::check_model(model);
-  } catch (const std::exception& ex) {
-    std::string error_msg = "Model validation failed: " + std::string(ex.what());
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
-  }
-
-  std::ofstream ofs(model_name, std::ios::binary);
-  if (!model.SerializeToOstream(&ofs)) {
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
-  }
-
-  return nullptr;
-}
-
-OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
-  ONNX_NAMESPACE::ModelProto model;
-  model.set_ir_version(ONNX_IR_VERSION);
-  auto* opset = model.add_opset_import();
-  opset->set_domain("");  // empty = default ONNX domain
-  opset->set_version(OPSET_VERSION);
-
-  auto* graph = model.mutable_graph();
-  graph->set_name("TopKGraph");
-
-  // ======================
-  // ---- Model Input ----
-  // ======================
-  {
-    auto* inp = graph->add_input();
-    inp->set_name("input");
-
-    auto* type = inp->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    // Shape: ["N"]
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_param("N");
-  }
-
-  // ======================
-  // ---- Initializers ----
-  // ======================
-  {
-    // K = [300]
-    ONNX_NAMESPACE::TensorProto K = MakeTensor("K", ONNX_NAMESPACE::TensorProto::INT64, {1}, {300});
-    *graph->add_initializer() = K;
-
-    // zero = 0.0 (scalar)
-    ONNX_NAMESPACE::TensorProto zero = MakeTensor("zero", ONNX_NAMESPACE::TensorProto::INT64, {}, {0});
-    *graph->add_initializer() = zero;
-
-    // twenty_six = 26 (scalar)
-    ONNX_NAMESPACE::TensorProto ts = MakeTensor("twenty_six", ONNX_NAMESPACE::TensorProto::INT64, {}, {26});
-    *graph->add_initializer() = ts;
-  }
-
-  // ======================
-  // ---- TopK ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("TopK");
-    n->add_input("input");
-    n->add_input("K");
-    n->add_output("scores");
-    n->add_output("topk_indices");
-    n->set_name("TopK");
-  }
-
-  // ======================
-  // ---- Less ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Less");
-    n->add_input("topk_indices");
-    n->add_input("zero");
-    n->add_output("Less_output_0");
-    n->set_name("Less");
-  }
-
-  // ======================
-  // ---- Div ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Div");
-    n->add_input("topk_indices");
-    n->add_input("twenty_six");
-    n->add_output("Div_17_output_0");
-    n->set_name("Div");
-  }
-
-  // ======================
-  // ---- Mod ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Mod");
-    n->add_input("topk_indices");
-    n->add_input("twenty_six");
-    n->add_output("labels");
-    n->set_name("Mod");
-  }
-
-  // =========================
-  // ---- Graph Outputs ----
-  // =========================
-  auto add_output = [&](const std::string& name, ONNX_NAMESPACE::TensorProto::DataType type, const std::string& dim) {
-    auto* out = graph->add_output();
-    out->set_name(name);
-
-    auto* tt = out->mutable_type()->mutable_tensor_type();
-    tt->set_elem_type(type);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_param(dim);
-  };
-
-  add_output("scores", ONNX_NAMESPACE::TensorProto::FLOAT, "K");
-  add_output("Less_output_0", ONNX_NAMESPACE::TensorProto::BOOL, "K");
-  add_output("Div_17_output_0", ONNX_NAMESPACE::TensorProto::INT64, "K");
-  add_output("labels", ONNX_NAMESPACE::TensorProto::INT64, "K");
-
-  // ======================
-  // Validate + Save
-  // ======================
-  try {
-    onnx::checker::check_model(model);
-  } catch (const std::exception& e) {
-    std::string error_msg = "Model validation failed: " + std::string(e.what());
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
-  }
-
-  std::ofstream ofs(model_name, std::ios::binary);
-  if (!model.SerializeToOstream(&ofs)) {
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
-  }
-
-  return nullptr;
-}
-
 }  // namespace test
 }  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
index 11000a76ad2cd..0f011af8211ca 100644
--- a/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
+++ b/onnxruntime/test/providers/nv_tensorrt_rtx/test_nv_trt_rtx_ep_util.h
@@ -119,9 +119,6 @@ void CreateBaseModel(const PathString& model_name,
 
 void CreateLargeLLMModel(const PathString& model_path, const PathString& external_data_path);
 
-OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name);
-OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name);
-
 Ort::IoBinding generate_io_binding(
     Ort::Session& session,
     std::map<std::string, std::vector<int64_t>> shape_overwrites = {},
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
index a0149613c21c3..ac5f974c4a14f 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
@@ -20,6 +20,8 @@ using namespace std;
 using namespace ONNX_NAMESPACE;
 using namespace ::onnxruntime::logging;
 
+extern std::unique_ptr<Ort::Env> ort_env;
+
 namespace onnxruntime {
 
 namespace test {
@@ -1372,16 +1374,12 @@ TEST(TensorrtExecutionProviderTest, TestSessionOutputs) {
    *                     |--- Mod ---> "labels"
    */
   {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
     OrtTensorRTProviderOptionsV2 provider_options;
     Ort::SessionOptions session_options;
     session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
 
-    auto model_path = ORT_TSTR("model_with_topk_and_multiple_graph_outputs.onnx");
-    Ort::Status status(CreateModelWithTopKWhichContainsGraphOutput(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
+    auto model_path = ORT_TSTR("testdata/topk_and_multiple_graph_outputs.onnx");
+    Ort::Session session(*ort_env, model_path, session_options);
 
     size_t output_count = session.GetOutputCount();
     ASSERT_TRUE(output_count == 4);
@@ -1397,16 +1395,12 @@ TEST(TensorrtExecutionProviderTest, TestSessionOutputs) {
    *
    */
   {
-    Ort::Env env{ORT_LOGGING_LEVEL_WARNING, "test"};
     OrtTensorRTProviderOptionsV2 provider_options;
     Ort::SessionOptions session_options;
     session_options.AppendExecutionProvider_TensorRT_V2(provider_options);
 
-    auto model_path = ORT_TSTR("model_with_node_output_not_used.onnx");
-    Ort::Status status(CreateModelWithNodeOutputNotUsed(model_path));
-    ASSERT_TRUE(status.IsOK());
-
-    Ort::Session session(env, model_path, session_options);
+    auto model_path = ORT_TSTR("testdata/node_output_not_used.onnx");
+    Ort::Session session(*ort_env, model_path, session_options);
 
     size_t output_count = session.GetOutputCount();
     ASSERT_TRUE(output_count == 1);
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
deleted file mode 100644
index 51738342c6cb0..0000000000000
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.cc
+++ /dev/null
@@ -1,281 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-
-#include "onnxruntime_cxx_api.h"
-#include "core/common/path_string.h"
-
-#include <iostream>
-#include <fstream>
-
-#include <onnx/onnx_pb.h>
-#include <onnx/defs/schema.h>
-#include <onnx/checker.h>
-
-#define ONNX_IR_VERSION 11
-#define OPSET_VERSION 23
-
-namespace onnxruntime {
-namespace test {
-
-// Helper: make ONNX_NAMESPACE::TensorProto
-ONNX_NAMESPACE::TensorProto MakeTensor(const std::string& name,
-                                       ONNX_NAMESPACE::TensorProto::DataType dtype,
-                                       const std::vector<int64_t>& dims,
-                                       const std::vector<int64_t>& vals) {
-  ONNX_NAMESPACE::TensorProto t;
-  t.set_name(name);
-  t.set_data_type(dtype);
-  for (auto d : dims) t.add_dims(d);
-  for (auto v : vals) t.add_int64_data(v);
-  return t;
-}
-
-ONNX_NAMESPACE::TensorProto MakeTensorFloat(const std::string& name,
-                                            const std::vector<int64_t>& dims,
-                                            const std::vector<float>& vals) {
-  ONNX_NAMESPACE::TensorProto t;
-  t.set_name(name);
-  t.set_data_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-  for (auto d : dims) t.add_dims(d);
-  for (auto v : vals) t.add_float_data(v);
-  return t;
-}
-
-OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name) {
-  // --------------------
-  // Create Model
-  // --------------------
-  ONNX_NAMESPACE::ModelProto model;
-  model.set_ir_version(ONNX_IR_VERSION);
-  auto* opset = model.add_opset_import();
-  opset->set_domain("");  // empty = default ONNX domain
-  opset->set_version(OPSET_VERSION);
-
-  ONNX_NAMESPACE::GraphProto* graph = model.mutable_graph();
-  graph->set_name("DropoutMatMulGraph");
-
-  // --------------------
-  // Create Inputs
-  // X: [3, 2]
-  // W: [2, 3]
-  // --------------------
-  {
-    auto* x = graph->add_input();
-    x->set_name("X");
-
-    auto* type = x->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(3);
-    shape->add_dim()->set_dim_value(2);
-  }
-
-  {
-    auto* w = graph->add_input();
-    w->set_name("W");
-
-    auto* type = w->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(2);
-    shape->add_dim()->set_dim_value(3);
-  }
-
-  // --------------------
-  // Output Y: [2, 3]
-  // --------------------
-  {
-    auto* x = graph->add_output();
-    x->set_name("Y");
-
-    auto* type = x->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_value(2);
-    shape->add_dim()->set_dim_value(3);
-  }
-
-  // --------------------
-  // Dropout Node
-  // --------------------
-  {
-    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
-    node->set_name("DropoutNode");
-    node->set_op_type("Dropout");
-
-    node->add_input("X");
-    node->add_output("dropout_out");
-    node->add_output("dropout_mask");
-  }
-
-  // --------------------
-  // MatMul Node
-  // --------------------
-  {
-    ONNX_NAMESPACE::NodeProto* node = graph->add_node();
-    node->set_name("MatMulNode");
-    node->set_op_type("MatMul");
-
-    node->add_input("dropout_out");
-    node->add_input("W");
-    node->add_output("Y");
-  }
-
-  // --------------------
-  // Validate
-  // --------------------
-  try {
-    onnx::checker::check_model(model);
-  } catch (const std::exception& ex) {
-    std::string error_msg = "Model validation failed: " + std::string(ex.what());
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
-  }
-
-  std::ofstream ofs(model_name, std::ios::binary);
-  if (!model.SerializeToOstream(&ofs)) {
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
-  }
-
-  return nullptr;
-}
-
-OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name) {
-  ONNX_NAMESPACE::ModelProto model;
-  model.set_ir_version(ONNX_IR_VERSION);
-  auto* opset = model.add_opset_import();
-  opset->set_domain("");  // empty = default ONNX domain
-  opset->set_version(OPSET_VERSION);
-
-  auto* graph = model.mutable_graph();
-  graph->set_name("TopKGraph");
-
-  // ======================
-  // ---- Model Input ----
-  // ======================
-  {
-    auto* inp = graph->add_input();
-    inp->set_name("input");
-
-    auto* type = inp->mutable_type();
-    auto* tt = type->mutable_tensor_type();
-    tt->set_elem_type(ONNX_NAMESPACE::TensorProto::FLOAT);
-
-    // Shape: ["N"]
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_param("N");
-  }
-
-  // ======================
-  // ---- Initializers ----
-  // ======================
-  {
-    // K = [300]
-    ONNX_NAMESPACE::TensorProto K = MakeTensor("K", ONNX_NAMESPACE::TensorProto::INT64, {1}, {300});
-    *graph->add_initializer() = K;
-
-    // zero = 0.0 (scalar)
-    ONNX_NAMESPACE::TensorProto zero = MakeTensor("zero", ONNX_NAMESPACE::TensorProto::INT64, {}, {0});
-    *graph->add_initializer() = zero;
-
-    // twenty_six = 26 (scalar)
-    ONNX_NAMESPACE::TensorProto ts = MakeTensor("twenty_six", ONNX_NAMESPACE::TensorProto::INT64, {}, {26});
-    *graph->add_initializer() = ts;
-  }
-
-  // ======================
-  // ---- TopK ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("TopK");
-    n->add_input("input");
-    n->add_input("K");
-    n->add_output("scores");
-    n->add_output("topk_indices");
-    n->set_name("TopK");
-  }
-
-  // ======================
-  // ---- Less ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Less");
-    n->add_input("topk_indices");
-    n->add_input("zero");
-    n->add_output("Less_output_0");
-    n->set_name("Less");
-  }
-
-  // ======================
-  // ---- Div ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Div");
-    n->add_input("topk_indices");
-    n->add_input("twenty_six");
-    n->add_output("Div_17_output_0");
-    n->set_name("Div");
-  }
-
-  // ======================
-  // ---- Mod ----
-  // ======================
-  {
-    ONNX_NAMESPACE::NodeProto* n = graph->add_node();
-    n->set_op_type("Mod");
-    n->add_input("topk_indices");
-    n->add_input("twenty_six");
-    n->add_output("labels");
-    n->set_name("Mod");
-  }
-
-  // =========================
-  // ---- Graph Outputs ----
-  // =========================
-  auto add_output = [&](const std::string& name, ONNX_NAMESPACE::TensorProto::DataType type, const std::string& dim) {
-    auto* out = graph->add_output();
-    out->set_name(name);
-
-    auto* tt = out->mutable_type()->mutable_tensor_type();
-    tt->set_elem_type(type);
-
-    auto* shape = tt->mutable_shape();
-    shape->add_dim()->set_dim_param(dim);
-  };
-
-  add_output("scores", ONNX_NAMESPACE::TensorProto::FLOAT, "K");
-  add_output("Less_output_0", ONNX_NAMESPACE::TensorProto::BOOL, "K");
-  add_output("Div_17_output_0", ONNX_NAMESPACE::TensorProto::INT64, "K");
-  add_output("labels", ONNX_NAMESPACE::TensorProto::INT64, "K");
-
-  // ======================
-  // Validate + Save
-  // ======================
-  try {
-    onnx::checker::check_model(model);
-  } catch (const std::exception& e) {
-    std::string error_msg = "Model validation failed: " + std::string(e.what());
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, error_msg.c_str());
-  }
-
-  std::ofstream ofs(model_name, std::ios::binary);
-  if (!model.SerializeToOstream(&ofs)) {
-    const OrtApi* ort_api = OrtGetApiBase()->GetApi(ORT_API_VERSION);
-    return ort_api->CreateStatus(OrtErrorCode::ORT_EP_FAIL, "Failed to write model");
-  }
-
-  return nullptr;
-}
-}  // namespace test
-}  // namespace onnxruntime
diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h b/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
deleted file mode 100644
index 3857d8ee80131..0000000000000
--- a/onnxruntime/test/providers/tensorrt/tensorrt_test_utils.h
+++ /dev/null
@@ -1,11 +0,0 @@
-// Copyright (c) Microsoft Corporation. All rights reserved.
-// Licensed under the MIT License.
-#include "onnxruntime_cxx_api.h"
-#include "core/common/path_string.h"
-
-namespace onnxruntime {
-namespace test {
-OrtStatus* CreateModelWithTopKWhichContainsGraphOutput(const PathString& model_name);
-OrtStatus* CreateModelWithNodeOutputNotUsed(const PathString& model_name);
-}  // namespace test
-}  // namespace onnxruntime
diff --git a/onnxruntime/test/testdata/node_output_not_used.onnx b/onnxruntime/test/testdata/node_output_not_used.onnx
new file mode 100644
index 0000000000000000000000000000000000000000..e2726182fddc2c265752e46346735c26e33add4b
GIT binary patch
literal 189
zcmd<!w%WwVWx~Z6A;g_hlwXivS`rV$LOd`|Zenq^6t@dR$}c}9Rf!$S;Zo;9*22XY
zF2ooq#pRn=;#-;nR>=lo3kgAWK)CKji3J%^!XPX8xOg}ig*dpFIGBN$2_zVfB*+Ak
RNCFB*q6<2)a4`t*0ss-ID|-L{

literal 0
HcmV?d00001

diff --git a/onnxruntime/test/testdata/node_output_not_used.py b/onnxruntime/test/testdata/node_output_not_used.py
new file mode 100644
index 0000000000000..3bacc1783c42d
--- /dev/null
+++ b/onnxruntime/test/testdata/node_output_not_used.py
@@ -0,0 +1,42 @@
+import onnx
+from onnx import helper, TensorProto
+
+def create_model_with_node_output_not_used(model_path):
+    # Create graph
+    X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [3, 2])
+    W = helper.make_tensor_value_info("W", TensorProto.FLOAT, [2, 3])
+    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 3])
+
+    # Dropout node (two outputs)
+    dropout_node = helper.make_node(
+        "Dropout",
+        inputs=["X"],
+        outputs=["dropout_out", "dropout_mask"],
+        name="DropoutNode",
+    )
+
+    # MatMul node
+    matmul_node = helper.make_node(
+        "MatMul",
+        inputs=["dropout_out", "W"],
+        outputs=["Y"],
+        name="MatMulNode",
+    )
+
+    graph = helper.make_graph(
+        nodes=[dropout_node, matmul_node],
+        name="DropoutMatMulGraph",
+        inputs=[X, W],
+        outputs=[Y],
+    )
+
+    model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])
+
+    onnx.checker.check_model(model)
+    onnx.save(model, model_path)
+
+    print(f"Model saved to: {model_path}")
+
+
+if __name__ == "__main__":
+    create_model_with_node_output_not_used("node_output_not_used.onnx")
diff --git a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.onnx b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.onnx
new file mode 100644
index 0000000000000000000000000000000000000000..340c3d420d5746844be0bd3769a174b4e69de801
GIT binary patch
literal 393
zcmd<!w))4!rNhOVnO9I+!o}z<#8#Y~UzA!b#8Z-AkR6|ymy(&BS}es9l3(Dh1SYxk
zxlrV|SgKNs@`ZSPQj3e@^Gi#B=ENIFv4FTrV3NxeO#@d+d1_usWqfgFg%H0>W?8(U
zIYJ{dP(TSp09}P@53)A4oW!KmoMI_v-~1FM5Fx|~a-n-sVnK!$HwU8tyA{(KCMQO3
zEp8x_k--V<NG(ASN6HDwq$m-n--WojI5`-FSh<*m82yq&xnSXdB<!6e!-eWiuqIB3
dChsIUE_C03<=J8KKtmzUMiF*m;bIWr1pvZ&Wx)Ud

literal 0
HcmV?d00001

diff --git a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
new file mode 100644
index 0000000000000..cca9d7bd63034
--- /dev/null
+++ b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
@@ -0,0 +1,86 @@
+import onnx
+from onnx import helper, TensorProto
+
+def create_model_with_topk_graph_output(model_path):
+
+    # ======================
+    # ---- Inputs ----
+    # ======================
+    input_tensor = helper.make_tensor_value_info(
+        "input", TensorProto.FLOAT, ["N"]
+    )
+
+    # ======================
+    # ---- Initializers ----
+    # ======================
+    K = helper.make_tensor(
+        "K", TensorProto.INT64, dims=[1], vals=[300]
+    )
+    zero = helper.make_tensor(
+        "zero", TensorProto.INT64, dims=[], vals=[0]
+    )
+    twenty_six = helper.make_tensor(
+        "twenty_six", TensorProto.INT64, dims=[], vals=[26]
+    )
+
+    # ======================
+    # ---- Nodes ----
+    # ======================
+    topk_node = helper.make_node(
+        "TopK",
+        inputs=["input", "K"],
+        outputs=["scores", "topk_indices"],
+        name="TopK",
+    )
+
+    less_node = helper.make_node(
+        "Less",
+        inputs=["topk_indices", "zero"],
+        outputs=["Less_output_0"],
+        name="Less",
+    )
+
+    div_node = helper.make_node(
+        "Div",
+        inputs=["topk_indices", "twenty_six"],
+        outputs=["Div_17_output_0"],
+        name="Div",
+    )
+
+    mod_node = helper.make_node(
+        "Mod",
+        inputs=["topk_indices", "twenty_six"],
+        outputs=["labels"],
+        name="Mod",
+    )
+
+    # =========================
+    # ---- Graph Outputs ----
+    # =========================
+    scores_out = helper.make_tensor_value_info("scores", TensorProto.FLOAT, ["K"])
+    less_out = helper.make_tensor_value_info("Less_output_0", TensorProto.BOOL, ["K"])
+    div_out = helper.make_tensor_value_info("Div_17_output_0", TensorProto.INT64, ["K"])
+    labels_out = helper.make_tensor_value_info("labels", TensorProto.INT64, ["K"])
+
+    # ======================
+    # ---- Graph ----
+    # ======================
+    graph = helper.make_graph(
+        nodes=[topk_node, less_node, div_node, mod_node],
+        name="TopKGraph",
+        inputs=[input_tensor],
+        outputs=[scores_out, less_out, div_out, labels_out],
+        initializer=[K, zero, twenty_six],
+    )
+
+    model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])
+
+    # Validate + Save
+    onnx.checker.check_model(model)
+    onnx.save(model, model_path)
+
+    print(f"Model saved to: {model_path}")
+
+
+if __name__ == "__main__":
+    create_model_with_topk_graph_output("topk_and_multiple_graph_outputs.onnx")

From 6c769c16928234e3656c0809c9cfd7b932699619 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 20 Nov 2025 11:30:38 -0800
Subject: [PATCH 10/12] lintrunner -a

---
 .../test/testdata/node_output_not_used.py     |  3 ++-
 .../topk_and_multiple_graph_outputs.py        | 20 ++++++-------------
 2 files changed, 8 insertions(+), 15 deletions(-)

diff --git a/onnxruntime/test/testdata/node_output_not_used.py b/onnxruntime/test/testdata/node_output_not_used.py
index 3bacc1783c42d..e1423ba8f38dc 100644
--- a/onnxruntime/test/testdata/node_output_not_used.py
+++ b/onnxruntime/test/testdata/node_output_not_used.py
@@ -1,5 +1,6 @@
 import onnx
-from onnx import helper, TensorProto
+from onnx import TensorProto, helper
+
 
 def create_model_with_node_output_not_used(model_path):
     # Create graph
diff --git a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
index cca9d7bd63034..caf2140d427c5 100644
--- a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
+++ b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
@@ -1,27 +1,19 @@
 import onnx
-from onnx import helper, TensorProto
+from onnx import TensorProto, helper
 
-def create_model_with_topk_graph_output(model_path):
 
+def create_model_with_topk_graph_output(model_path):
     # ======================
     # ---- Inputs ----
     # ======================
-    input_tensor = helper.make_tensor_value_info(
-        "input", TensorProto.FLOAT, ["N"]
-    )
+    input_tensor = helper.make_tensor_value_info("input", TensorProto.FLOAT, ["N"])
 
     # ======================
     # ---- Initializers ----
     # ======================
-    K = helper.make_tensor(
-        "K", TensorProto.INT64, dims=[1], vals=[300]
-    )
-    zero = helper.make_tensor(
-        "zero", TensorProto.INT64, dims=[], vals=[0]
-    )
-    twenty_six = helper.make_tensor(
-        "twenty_six", TensorProto.INT64, dims=[], vals=[26]
-    )
+    K = helper.make_tensor("K", TensorProto.INT64, dims=[1], vals=[300])
+    zero = helper.make_tensor("zero", TensorProto.INT64, dims=[], vals=[0])
+    twenty_six = helper.make_tensor("twenty_six", TensorProto.INT64, dims=[], vals=[26])
 
     # ======================
     # ---- Nodes ----

From accf540b2b94210378afc5340ba05b0a2fb505a3 Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Thu, 20 Nov 2025 11:53:00 -0800
Subject: [PATCH 11/12] update

---
 onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc | 1 -
 1 file changed, 1 deletion(-)

diff --git a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
index ac5f974c4a14f..c7062245ba746 100644
--- a/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
+++ b/onnxruntime/test/providers/tensorrt/tensorrt_basic_test.cc
@@ -1,7 +1,6 @@
 // Copyright (c) Microsoft Corporation. All rights reserved.
 // Licensed under the MIT License.
 #include "onnxruntime_cxx_api.h"
-#include "tensorrt_test_utils.h"
 #include "core/graph/onnx_protobuf.h"
 #include "core/session/inference_session.h"
 #include "test/providers/provider_test_utils.h"

From bf87f39ac4b59156718bb64110965ff42668be2c Mon Sep 17 00:00:00 2001
From: Chi Lo <Chi.Lo@microsoft.com>
Date: Mon, 1 Dec 2025 10:18:15 -0800
Subject: [PATCH 12/12] lintrunner -a

---
 onnxruntime/test/testdata/node_output_not_used.py      | 10 +++++-----
 .../test/testdata/topk_and_multiple_graph_outputs.py   |  4 ++--
 2 files changed, 7 insertions(+), 7 deletions(-)

diff --git a/onnxruntime/test/testdata/node_output_not_used.py b/onnxruntime/test/testdata/node_output_not_used.py
index e1423ba8f38dc..d36d5e9cfd2f8 100644
--- a/onnxruntime/test/testdata/node_output_not_used.py
+++ b/onnxruntime/test/testdata/node_output_not_used.py
@@ -4,9 +4,9 @@
 
 def create_model_with_node_output_not_used(model_path):
     # Create graph
-    X = helper.make_tensor_value_info("X", TensorProto.FLOAT, [3, 2])
-    W = helper.make_tensor_value_info("W", TensorProto.FLOAT, [2, 3])
-    Y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 3])
+    x = helper.make_tensor_value_info("X", TensorProto.FLOAT, [3, 2])
+    w = helper.make_tensor_value_info("W", TensorProto.FLOAT, [2, 3])
+    y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [2, 3])
 
     # Dropout node (two outputs)
     dropout_node = helper.make_node(
@@ -27,8 +27,8 @@ def create_model_with_node_output_not_used(model_path):
     graph = helper.make_graph(
         nodes=[dropout_node, matmul_node],
         name="DropoutMatMulGraph",
-        inputs=[X, W],
-        outputs=[Y],
+        inputs=[x, w],
+        outputs=[y],
     )
 
     model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])
diff --git a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
index caf2140d427c5..232abb2ed9163 100644
--- a/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
+++ b/onnxruntime/test/testdata/topk_and_multiple_graph_outputs.py
@@ -11,7 +11,7 @@ def create_model_with_topk_graph_output(model_path):
     # ======================
     # ---- Initializers ----
     # ======================
-    K = helper.make_tensor("K", TensorProto.INT64, dims=[1], vals=[300])
+    k = helper.make_tensor("K", TensorProto.INT64, dims=[1], vals=[300])
     zero = helper.make_tensor("zero", TensorProto.INT64, dims=[], vals=[0])
     twenty_six = helper.make_tensor("twenty_six", TensorProto.INT64, dims=[], vals=[26])
 
@@ -62,7 +62,7 @@ def create_model_with_topk_graph_output(model_path):
         name="TopKGraph",
         inputs=[input_tensor],
         outputs=[scores_out, less_out, div_out, labels_out],
-        initializer=[K, zero, twenty_six],
+        initializer=[k, zero, twenty_six],
     )
 
     model = helper.make_model(graph, opset_imports=[helper.make_operatorsetid("", 13)])