[TRT-EP] Add loadModelProto APIs (#25409)

### Description
This PR adds three new options for the TRT execution provider:

- trt_load_user_initializer
- trt_external_data_bytestream
- trt_external_data_bytestream_size

The idea is to use these options to leverage new TRT 10.13 APIs to give
the user more control on how the weights are loaded in the ONNX parser.

When `trt_load_user_initializer` is set to true, the EP will own the
weights instead of serializing the weights to ModelProto. This reduces
overhead in having to serialize large weights.

When `trt_external_data_bytestream / trt_external_data_bytestream_size`
is provided, the refitEngine() function will be able to read from this
bytestream directly to extract weights for the refitter.

Also fixes graph_proto_serializer to keep information about external
weights.

---------

Signed-off-by: Kevin Chen <[email protected]>

Author: kevinch-nv

include/onnxruntime/core/providers/tensorrt/tensorrt_provider_options.h

@@ -89,7 +89,12 @@ struct OrtTensorRTProviderOptionsV2 {
   size_t trt_onnx_bytestream_size{0};               // size of the byte stream provided as "trt_onnx_bytestream"
                                                     // can be updated using: UpdateTensorRTProviderOptionsWithValue
 
-  const char* trt_engine_cache_prefix{nullptr};  // specify engine cache prefix
-  int trt_engine_hw_compatible{0};               // Enable hardware compatibility. Default 0 = false, nonzero = true
-  const char* trt_op_types_to_exclude{};         // Exclude specific ops from running on TRT.
+  const void* trt_external_data_bytestream{nullptr};  // The byte stream containing the weights to override the ones provided in the ONNX model.
+                                                      // can be updated using: UpdateTensorRTProviderOptionsWithValue
+  size_t trt_external_data_bytestream_size{0};        // size of the byte stream provided as "trt_external_data_bytestream"
+                                                      // can be updated using: UpdateTensorRTProviderOptionsWithValue
+  const char* trt_engine_cache_prefix{nullptr};       // specify engine cache prefix
+  int trt_engine_hw_compatible{0};                    // Enable hardware compatibility. Default 0 = false, nonzero = true
+  const char* trt_op_types_to_exclude{};              // Exclude specific ops from running on TRT.
+  int trt_load_user_initializer{0};                   // Save initializers locally instead of to disk. Default 0 = false, nonzero = true
 };

onnxruntime/core/graph/graph_proto_serializer.cc

@@ -94,8 +94,8 @@ void GraphViewerToProto(const GraphViewer& graph_view,
       current_scope_initializer_set.insert(name);
       auto* p_initializer = graph_proto.add_initializer();
 
-      // Do not save raw or external data into the graph, only the metadata
-      if (!include_initializer_data && (init->has_raw_data() || init->has_data_location())) {
+      // Do not save raw into the graph, only the metadata
+      if (!include_initializer_data && init->has_raw_data()) {
         // Set datatype
         if (init->has_data_type()) {
           p_initializer->set_data_type(init->data_type());

onnxruntime/core/providers/tensorrt/onnx_ctx_model_helper.cc

@@ -298,6 +298,8 @@ Status TensorRTCacheModelHandler::GetEpContextFromGraph(const GraphViewer& graph
                                                            make_secure_path_checks,
                                                            onnx_model_bytestream_,
                                                            onnx_model_bytestream_size_,
+                                                           onnx_external_data_bytestream_,
+                                                           onnx_external_data_bytestream_size_,
                                                            (*trt_engine_).get(),
                                                            false /* serialize refitted engine to disk */,
                                                            detailed_build_log_);
@@ -367,6 +369,8 @@ Status TensorRTCacheModelHandler::GetEpContextFromGraph(const GraphViewer& graph
                                                            make_secure_path_checks,
                                                            onnx_model_bytestream_,
                                                            onnx_model_bytestream_size_,
+                                                           onnx_external_data_bytestream_,
+                                                           onnx_external_data_bytestream_size_,
                                                            (*trt_engine_).get(),
                                                            true /* serialize refitted engine to disk */,
                                                            detailed_build_log_);

onnxruntime/core/providers/tensorrt/onnx_ctx_model_helper.h

@@ -54,6 +54,8 @@ class TensorRTCacheModelHandler {
                             std::string onnx_model_folder_path,
                             const void* onnx_model_bytestream,
                             size_t onnx_model_bytestream_size,
+                            const void* onnx_external_data_bytestream,
+                            size_t onnx_external_data_bytestream_size,
                             bool detailed_build_log)
       : trt_engine_(trt_engine),
         trt_runtime_(trt_runtime),
@@ -63,6 +65,8 @@ class TensorRTCacheModelHandler {
         onnx_model_folder_path_(onnx_model_folder_path),
         onnx_model_bytestream_(onnx_model_bytestream),
         onnx_model_bytestream_size_(onnx_model_bytestream_size),
+        onnx_external_data_bytestream_(onnx_external_data_bytestream),
+        onnx_external_data_bytestream_size_(onnx_external_data_bytestream_size),
         detailed_build_log_(detailed_build_log) {
   }
   ORT_DISALLOW_COPY_ASSIGNMENT_AND_MOVE(TensorRTCacheModelHandler);
@@ -80,6 +84,8 @@ class TensorRTCacheModelHandler {
   std::string onnx_model_folder_path_;
   const void* onnx_model_bytestream_;
   size_t onnx_model_bytestream_size_;
+  const void* onnx_external_data_bytestream_;
+  size_t onnx_external_data_bytestream_size_;
   bool detailed_build_log_;
 };  // TRTCacheModelHandler
 }  // namespace onnxruntime