Address some issues and fix VS 2026 build

include/onnxruntime/core/graph/graph.h

@@ -1726,7 +1726,8 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
 #if !defined(ORT_MINIMAL_BUILD)
   // Build and verify node connection (edges).
   // Verify NodeArg name/type/shape matching correctly.
-  common::Status BuildConnections(std::unordered_set<std::string>& outer_scope_node_args_consumed);
+  common::Status BuildConnections(std::unordered_set<std::string>& outer_scope_node_args_consumed,
+                                  bool& removed_node_with_subgraph);
 
   common::Status VerifyNoDuplicateName();
 

onnxruntime/core/graph/graph.cc

@@ -1344,7 +1344,12 @@ Graph::Graph(const Model& owning_model,
       ORT_THROW("This is an invalid model. Tensor does not have type information.");
     }
 
-    if (utils::HasDataType(tensor) && (tensor.data_type() < TensorProto_DataType_DataType_ARRAYSIZE)) {
+    // all initializers must have a valid data type
+    if (!utils::HasDataType(tensor) || !tensor.DataType_IsValid(tensor.data_type())) {
+      ORT_THROW("This is an invalid model. Tensor '", tensor.name(), "' does not have valid data type.");
+    }
+
+    if ((tensor.data_type() < TensorProto_DataType_DataType_ARRAYSIZE)) {
       weight_data_type_freq_[tensor.data_type()]++;
     }
 
@@ -1669,13 +1674,14 @@ void Graph::RemoveEdge(NodeIndex src_node_index, NodeIndex dst_node_index, int s
 
 #if !defined(ORT_MINIMAL_BUILD)
 GSL_SUPPRESS(es .84)  // ignoring return value from unordered_map::insert causes noisy complaint
-Status Graph::BuildConnections(std::unordered_set<std::string>& outer_scope_node_args_consumed) {
+Status Graph::BuildConnections(std::unordered_set<std::string>& outer_scope_node_args_consumed,
+                               bool& removed_node_with_subgraph) {
   // recurse into subgraphs first so we can update any nodes in this graph that are used by those subgraphs
   if (!resolve_context_.nodes_with_subgraphs.empty()) {
     for (auto* node : resolve_context_.nodes_with_subgraphs) {
       for (auto& subgraph : node->MutableSubgraphs()) {
         std::unordered_set<std::string> node_args_consumed;
-        ORT_RETURN_IF_ERROR(subgraph->BuildConnections(node_args_consumed));
+        ORT_RETURN_IF_ERROR(subgraph->BuildConnections(node_args_consumed, removed_node_with_subgraph));
 
         for (auto& node_arg_name : node_args_consumed) {
           auto node_arg = GetNodeArg(node_arg_name);
@@ -1805,6 +1811,10 @@ Status Graph::BuildConnections(std::unordered_set<std::string>& outer_scope_node
     } else if (node.OutputDefs().empty()) {
       // This is a useless node.
       // It has no input/output.
+      if (node.ContainsSubgraph()) {
+        removed_node_with_subgraph = true;
+      }
+
       RemoveNode(node.Index());
     }
   }
@@ -3683,10 +3693,17 @@ Status Graph::Resolve(const ResolveOptions& options) {
   std::unordered_set<std::string> outer_scope_node_args_consumed;
 
   // recursively build connections between nodes in this graph and all subgraphs
-  ORT_RETURN_IF_ERROR(BuildConnections(outer_scope_node_args_consumed));
+  bool removed_node_with_subgraph = false;
+  ORT_RETURN_IF_ERROR(BuildConnections(outer_scope_node_args_consumed, removed_node_with_subgraph));
   ORT_ENFORCE(outer_scope_node_args_consumed.empty(),
               "Shouldn't be possible to have NodeArgs that haven't been handled already.");
 
+  // if we removed any nodes with subgraphs, we need to refresh the list of subgraphs.
+  if (removed_node_with_subgraph) {
+    all_subgraphs.clear();
+    FindAllSubgraphs(all_subgraphs);
+  }
+
   // topological sort of this and any subgraphs is non-recursive
   auto topo_sort_func = [](Graph& graph) { return graph.PerformTopologicalSortAndCheckIsAcyclic(); };
   ORT_RETURN_IF_ERROR(ForThisAndAllSubgraphs(all_subgraphs, topo_sort_func));

onnxruntime/core/session/custom_ops.cc

@@ -1257,12 +1257,8 @@ common::Status CreateCustomRegistry(gsl::span<OrtCustomOpDomain* const> op_domai
     std::unordered_map<std::string, std::vector<const OrtCustomOp*>> domain_kernels;
     for (const auto* op : domain->custom_ops_) {
       // define kernel
-      auto it = domain_kernels.find(op->GetName(op));
-      if (it == domain_kernels.end()) {
-        domain_kernels[op->GetName(op)] = {op};
-      } else {
-        domain_kernels[op->GetName(op)].push_back(op);
-      }
+      const auto* name = op->GetName(op);
+      domain_kernels[name].push_back(op);
     }
 
     // Creation of the schemas, one per unique name.
@@ -1276,7 +1272,8 @@ common::Status CreateCustomRegistry(gsl::span<OrtCustomOpDomain* const> op_domai
       for (const auto* op : ops) {
         // define kernel
         auto kernel_create_info = CreateKernelCreateInfo(domain->domain_, op);
-        kernel_def_map[op->GetName(op)].push_back(kernel_create_info.kernel_def.get());
+        const auto* op_name = op->GetName(op);
+        kernel_def_map[op_name].push_back(kernel_create_info.kernel_def.get());
         ORT_RETURN_IF_ERROR(output->RegisterCustomKernel(kernel_create_info));
         // If IsCompatible returns false, then all custom operators named
         // 'op->GetName(op)' are not compatible among themselves.

onnxruntime/test/framework/inference_session_test.cc

@@ -255,8 +255,7 @@ void RunModel(InferenceSession& session_object,
   if (is_preallocate_output_vec) {
     fetches.resize(output_names.size());
     for (auto& elem : fetches) {
-      CreateMLValue<float>(TestCPUExecutionProvider()->CreatePreferredAllocators()[0], dims_mul_x, values_mul_x,
-                           &elem);
+      AllocateMLValue<float>(TestCPUExecutionProvider()->CreatePreferredAllocators()[0], dims_mul_x, &elem);
     }
   }
 
@@ -3069,5 +3068,26 @@ TEST(InferenceSessionTests, InterThreadPoolWithDenormalAsZero) {
 }
 #endif
 
+TEST(InferenceSessionTests, BadDataTypeInInitializerIsHandled) {
+  // model has an initializer with a bogus data type. Graph ctor should detect and throw.
+  auto model_uri = ORT_TSTR("testdata/icm-31000000518082.onnx");
+
+  SessionOptions so;
+  so.session_logid = "TempTest.LoadModel";
+  InferenceSession session{so, GetEnvironment()};
+  ASSERT_STATUS_NOT_OK_AND_HAS_SUBSTR(session.Load(model_uri), "does not have valid data type");
+}
+
+TEST(InferenceSessionTests, GraphResolveHandlesNodeWithSubgraphBeingRemoved) {
+  // model has a subgraph with output that is not consumed. the node with the subgraph should get removed in
+  // Graph::BuildConnections and Graph::Resolve should adjust its list of subgraphs to not access the removed subgraph.
+  auto model_uri = ORT_TSTR("testdata/icm-31000000518483.onnx");
+
+  SessionOptions so;
+  so.session_logid = "TempTest.LoadModel";
+  InferenceSession session{so, GetEnvironment()};
+  ASSERT_STATUS_OK(session.Load(model_uri));
+}
+
 }  // namespace test
 }  // namespace onnxruntime

onnxruntime/test/testdata/icm-31000000518483.py

@@ -0,0 +1,53 @@
+from onnx import TensorProto, helper, save_model
+
+# Add node with a subgraph that has no inputs or outputs.
+# Graph::BuildConnections should remove and the list of subgraphs in Graph::Resolve should be updated.
+# Other details here don't matter. Copied from ort_github_issue_10305.py
+if_body = helper.make_graph(
+    [
+        # need to use main_graph_initializer in a way that can't be constant folded
+        helper.make_node("Constant", inputs=[], outputs=["zero"], name="Constant", value_int=0),
+    ],
+    "if_branch_body",
+    [
+        # no explicit inputs
+    ],
+    [
+        helper.make_tensor_value_info("zero", TensorProto.BOOL, [1]),
+    ],
+)
+
+# Create the main graph
+graph_proto = helper.make_graph(
+    [
+        # add a Transpose that can be moved past the Slice
+        helper.make_node(
+            "Transpose",
+            inputs=["input:0"],
+            outputs=["transpose:0"],
+            name="transpose0",
+            perm=[1, 0, 2],
+        ),
+        helper.make_node(
+            "If",
+            [],
+            [],
+            "If1",
+            then_branch=if_body,
+            else_branch=if_body,
+        ),
+    ],
+    "Main_graph",
+    [
+        helper.make_tensor_value_info("input:0", TensorProto.FLOAT, [2, 2, 2]),
+        helper.make_tensor_value_info("state_var_in", TensorProto.FLOAT, [1]),
+    ],
+    [
+        helper.make_tensor_value_info("transpose:0", TensorProto.FLOAT, [2, 2]),
+    ],
+)
+
+model = helper.make_model(graph_proto)
+# model to repro issue is invalid. can't run checker.
+# onnx.checker.check_model(model, True)
+save_model(model, "icm-31000000518483.onnx")