Merge branch 'main' into pad

Author: xhcao

.gitmodules

@@ -7,4 +7,4 @@
 [submodule "cmake/external/emsdk"]
 	path = cmake/external/emsdk
 	url = https://github.com/emscripten-core/emsdk.git
-	branch = 4.0.3
+	branch = 4.0.4

cmake/CMakeLists.txt

@@ -1512,6 +1512,11 @@ if (onnxruntime_USE_OPENVINO)
 
 endif()
 
+if (onnxruntime_USE_OPENVINO_INTERFACE AND (NOT onnxruntime_USE_OPENVINO))
+    add_definitions(-DUSE_OPENVINO=1)
+    add_definitions(-DOPENVINO_CONFIG_NPU=1)
+endif()
+
 if (onnxruntime_USE_VITISAI)
     set(CMAKE_MODULE_PATH "${CMAKE_MODULE_PATH};${CMAKE_CURRENT_LIST_DIR}")
 endif()

cmake/deps.txt

@@ -37,7 +37,7 @@ mimalloc;https://github.com/microsoft/mimalloc/archive/refs/tags/v2.1.1.zip;d5ee
 mp11;https://github.com/boostorg/mp11/archive/refs/tags/boost-1.82.0.zip;9bc9e01dffb64d9e0773b2e44d2f22c51aace063
 onnx;https://github.com/onnx/onnx/archive/refs/tags/v1.17.0.zip;13a60ac5217c104139ce0fd024f48628e7bcf5bc
 # Use the latest commit of 10.8-GA
-onnx_tensorrt;https://github.com/onnx/onnx-tensorrt/archive/118ed0aea197fa9a7d3ea66180a1d5ddb9deecc3.zip;b78aed3728ad4daf6dc47ea10c1d243dee1d95b1
+onnx_tensorrt;https://github.com/onnx/onnx-tensorrt/archive/c5ca8912f30e9ad630a0ef565e3d5f4bd5e91563.zip;588b294aaa9e84679ed5815cea1d399210ac98c2
 protobuf;https://github.com/protocolbuffers/protobuf/archive/refs/tags/v21.12.zip;7cf2733949036c7d52fda017badcab093fe73bfa
 protoc_win64;https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protoc-21.12-win64.zip;b4521f7ada5b260380f94c4bd7f1b7684c76969a
 protoc_win32;https://github.com/protocolbuffers/protobuf/releases/download/v21.12/protoc-21.12-win32.zip;3688010318192c46ce73213cdfb6b3e5656da874

cmake/external/emsdk

@@ -301,6 +301,7 @@ if(NOT TARGET Boost::mp11)
     onnxruntime_fetchcontent_declare(
      mp11
      URL ${DEP_URL_mp11}
+     EXCLUDE_FROM_ALL
      FIND_PACKAGE_ARGS NAMES Boost
     )
     onnxruntime_fetchcontent_makeavailable(mp11)    

cmake/external/onnxruntime_external_deps.cmake

@@ -22,6 +22,7 @@ endif()
 if (onnxruntime_MINIMAL_BUILD)
   set(onnxruntime_session_src_exclude
     "${ONNXRUNTIME_ROOT}/core/session/provider_bridge_ort.cc"
+    "${ONNXRUNTIME_ROOT}/core/session/model_builder_c_api.cc"
   )
 
   list(REMOVE_ITEM onnxruntime_session_srcs ${onnxruntime_session_src_exclude})

cmake/onnxruntime_session.cmake

@@ -503,6 +503,7 @@ set (onnxruntime_shared_lib_test_SRC
 
 if (NOT onnxruntime_MINIMAL_BUILD)
   list(APPEND onnxruntime_shared_lib_test_SRC ${ONNXRUNTIME_SHARED_LIB_TEST_SRC_DIR}/test_inference.cc)
+  list(APPEND onnxruntime_shared_lib_test_SRC ${ONNXRUNTIME_SHARED_LIB_TEST_SRC_DIR}/test_model_builder_api.cc)
 endif()
 
 if(onnxruntime_RUN_ONNX_TESTS)
@@ -1359,14 +1360,19 @@ if (NOT onnxruntime_ENABLE_TRAINING_TORCH_INTEROP)
             LIBS ${onnxruntime_shared_lib_test_LIBS}
             DEPENDS ${all_dependencies}
     )
+
+    target_include_directories(onnxruntime_shared_lib_test PRIVATE ${ONNXRUNTIME_ROOT})
+
     if (onnxruntime_USE_CUDA)
       target_include_directories(onnxruntime_shared_lib_test PRIVATE ${CMAKE_CUDA_TOOLKIT_INCLUDE_DIRECTORIES})
       target_sources(onnxruntime_shared_lib_test PRIVATE ${ONNXRUNTIME_SHARED_LIB_TEST_SRC_DIR}/cuda_ops.cu)
     endif()
+
     if (onnxruntime_USE_ROCM)
       target_include_directories(onnxruntime_shared_lib_test PRIVATE ${onnxruntime_ROCM_HOME}/include)
       target_compile_definitions(onnxruntime_shared_lib_test PRIVATE __HIP_PLATFORM_AMD__)
     endif()
+
     if (CMAKE_SYSTEM_NAME STREQUAL "Android")
       target_sources(onnxruntime_shared_lib_test PRIVATE
         "${ONNXRUNTIME_ROOT}/core/platform/android/cxa_demangle.cc"

cmake/onnxruntime_unittests.cmake

@@ -25,7 +25,7 @@ internal class ManagedTypeProjection
         /// </summary>
         /// <param name="namedOnnxValue"></param>
         /// <param name="metadata"></param>
-        /// <returns>OrtValye created accoding to the metadata</returns>
+        /// <returns>OrtValue created according to the metadata</returns>
         internal static OrtValue CreateProjection(NamedOnnxValue namedOnnxValue, NodeMetadata metadata)
         {
             OrtValue result;
@@ -191,4 +191,3 @@ private static OrtValue CreateTensorProjection(NamedOnnxValue node, NodeMetadata
         }
     }
 }
-

csharp/src/Microsoft.ML.OnnxRuntime/ManagedProjections.shared.cs

@@ -847,7 +847,7 @@ internal class NativeLib
         /// Creates an instance of OrtSession with provided parameters
         /// </summary>
         /// <param name="environment">Native OrtEnv instance</param>
-        /// <param name="modelData">Byte array correspoonding to the model</param>
+        /// <param name="modelData">Byte array corresponding to the model</param>
         /// <param name="modelSize">Size of the model in bytes</param>
         /// <param name="sessionOptions">Native SessionOptions instance</param>
         /// <param name="prepackedWeightsContainer">Native OrtPrepackedWeightsContainer instance</param>
@@ -1258,7 +1258,7 @@ IntPtr[] outputValues /* An array of output value pointers. Array must be alloca
         /// </summary>
         /// <param name="options">Native SessionOptions instance</param>
         /// <param name="name">Name of the initializer</param>
-        /// <param name="ortValue">Native OrtValue instnce</param>
+        /// <param name="ortValue">Native OrtValue instance</param>
         [UnmanagedFunctionPointer(CallingConvention.Winapi)]
         public delegate IntPtr /*(OrtStatus*)*/ DOrtAddInitializer(IntPtr /*(OrtSessionOptions*)*/ options,
                                                                    byte[] /*(const char*)*/ name,

csharp/src/Microsoft.ML.OnnxRuntime/NativeMethods.shared.cs

@@ -27,6 +27,7 @@
 #include "core/common/span_utils.h"
 #include "core/common/status.h"
 #include "core/common/logging/logging.h"
+#include "core/framework/ort_value.h"
 #include "core/framework/prepacked_weights_container.h"
 #include "core/graph/onnx_protobuf.h"
 #include "core/graph/basic_types.h"
@@ -39,6 +40,9 @@
 #include "core/graph/node_arg.h"
 #include "core/graph/ort_format_load_options.h"
 
+// Type from Model Editor API in ORT C API so can't be in a namespace
+struct OrtGraph;
+
 namespace onnxruntime {
 class Graph;
 struct IndexedSubGraph;
@@ -763,6 +767,10 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
   */
   bool GetInitializedTensor(const std::string& tensor_name, const ONNX_NAMESPACE::TensorProto*& value) const;
 
+  /** Populate `value` if an externally allocated OrtValue exists for an initializer with the given name.
+   */
+  bool GetOrtValueInitializer(const std::string& name, OrtValue& value) const;
+
   /** Gets all the initializer tensors in this Graph. */
   const InitializedTensorSet& GetAllInitializedTensors() const noexcept { return name_to_initial_tensor_; }
 
@@ -1430,6 +1438,16 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
                                   const OrtFormatLoadOptions& load_options,
                                   const logging::Logger& logger, std::unique_ptr<Graph>& graph);
 
+  static Status LoadFromModelEditorApiModel(const OrtGraph& api_graph,
+                                            const Model& owning_model,
+                                            const std::unordered_map<std::string, int>& domain_to_version,
+                                            IOnnxRuntimeOpSchemaCollectionPtr schema_registry,
+                                            bool strict_shape_type_inference,
+                                            const logging::Logger& logger,
+                                            std::unique_ptr<Graph>& graph);
+
+  Status UpdateUsingModelEditorApiModel(const OrtModel& api_model);
+
 #if !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
   const RuntimeOptimizationRecordContainer& RuntimeOptimizations() const {
     return runtime_optimizations_;
@@ -1630,7 +1648,8 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
   // Implementation for initializer replacement
   Status ReplaceInitializedTensorImpl(ONNX_NAMESPACE::TensorProto new_initializer, bool is_external);
 
-  std::vector<NodeArg*> CreateNodeArgs(const google::protobuf::RepeatedPtrField<std::string>& names,
+  template <typename StringRange>  // range-initializer returning std::string
+  std::vector<NodeArg*> CreateNodeArgs(const StringRange& names,
                                        const ArgNameToTypeMap& name_to_type_map);
 
   void ToGraphProtoInternal(ONNX_NAMESPACE::GraphProto& graph_proto) const;
@@ -1694,6 +1713,8 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
     return nodes_[node_index].get();
   }
 
+  Status LoadFromModelEditorApiModel(const OrtGraph& api_graph, bool updating_existing_graph = false);
+
   const Model& owning_model_;
 
   // GraphProto to store name, version, initializer.
@@ -1708,6 +1729,12 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
 
   InitializedTensorSet name_to_initial_tensor_;
 
+  // Initializers that are external to the Graph.
+  // e.g. created from existing memory using CreateTensorWithDataAndDeleterAsOrtValue in the ORT API.
+  // As we need to convert to TensorProto for the optimizers to work and keep the deleter information we store them
+  // in the Graph instance and retrieve during session state finalization.
+  std::unordered_map<std::string, OrtValue> ortvalue_initializers_;
+
   std::unordered_set<std::reference_wrapper<const std::string>,
                      std::hash<std::string>, std::equal_to<std::string>>
       sparse_tensor_names_;
@@ -1744,6 +1771,7 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
   // in some case, a fused sub-graph will happens multiple times in one model, we use a map
   // to store reusable-schema in lookup.
   InlinedHashMap<std::string, std::reference_wrapper<ONNX_NAMESPACE::OpSchema>> reusable_fused_schema_map_;
+
 #endif  // !defined(ORT_MINIMAL_BUILD)
 
   // Graph nodes.
@@ -1806,7 +1834,7 @@ class Graph {  // NOLINT(clang-analyzer-optin.performance.Padding): preserve exi
   std::unordered_map<std::string, std::unordered_set<NodeIndex>> node_arg_to_consumer_nodes_;
 #endif  // !defined(ORT_MINIMAL_BUILD) || defined(ORT_EXTENDED_MINIMAL_BUILD)
 
-  const std::unordered_map<std::string, int> domain_to_version_;
+  std::unordered_map<std::string, int> domain_to_version_;
 
   // Model IR version.
   Version ir_version_{ONNX_NAMESPACE::Version::IR_VERSION};

include/onnxruntime/core/graph/graph.h

@@ -193,6 +193,12 @@ class GraphViewer {
   IOnnxRuntimeOpSchemaCollectionPtr GetSchemaRegistry() const { return graph_->GetSchemaRegistry(); }
 #endif
 
+  /** Populate `value` if an externally allocated OrtValue exists for an initializer with the given name.
+   */
+  bool GetOrtValueInitializer(const std::string& name, OrtValue& value) const {
+    return graph_->GetOrtValueInitializer(name, value);
+  }
+
  private:
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(GraphViewer);
   GraphViewer(const Graph& graph, const IndexedSubGraph* filter_info);