diff --git a/include/onnxruntime/core/framework/execution_provider.h b/include/onnxruntime/core/framework/execution_provider.h
index 1de0217c7e1fa..31c988f500779 100644
--- a/include/onnxruntime/core/framework/execution_provider.h
+++ b/include/onnxruntime/core/framework/execution_provider.h
@@ -59,14 +59,11 @@ enum class DataLayout {
 
 class IExecutionProvider {
  protected:
-  IExecutionProvider(const std::string& type, bool use_metadef_id_creator = false)
-      : IExecutionProvider(type, OrtDevice(), use_metadef_id_creator) {}
+  IExecutionProvider(const std::string& type)
+      : IExecutionProvider(type, OrtDevice()) {}
 
-  IExecutionProvider(const std::string& type, OrtDevice device, bool use_metadef_id_creator = false)
+  IExecutionProvider(const std::string& type, OrtDevice device)
       : default_device_(device), type_{type} {
-    if (use_metadef_id_creator) {
-      metadef_id_generator_ = std::make_unique<ModelMetadefIdGenerator>();
-    }
   }
 
   /*
@@ -274,19 +271,6 @@ class IExecutionProvider {
     return logger_;
   }
 
-  /** Generate a unique id that can be used in a MetaDef name. Values are unique for a model instance.
-   The model hash is also returned if you wish to include that in the MetaDef name to ensure uniqueness across models.
-   @param graph_viewer[in] Graph viewer that GetCapability was called with. Can be for the main graph or nested graph.
-   @param model_hash[out] Returns the hash for the main (i.e. top level) graph in the model.
-                          This is created using the model path if available,
-                          or the model input names and the output names from all nodes in the main graph.
-   @remarks e.g. the TensorRT Execution Provider is used in multiple sessions and the underlying infrastructure caches
-            compiled kernels, so the name must be unique and deterministic across models and sessions.
-            NOTE: Ideally this would be a protected method, but to work across the EP bridge it has to be public and
-                  virtual, and ModelMetadefIdGenerator but be defined in the header as well.
-   */
-  virtual int GenerateMetaDefId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) const;
-
   virtual std::unique_ptr<profiling::EpProfiler> GetProfiler() {
     return {};
   }
@@ -340,18 +324,5 @@ class IExecutionProvider {
 
   // It will be set when this object is registered to a session
   const logging::Logger* logger_ = nullptr;
-
-  // helper to generate ids that are unique to model and deterministic, even if the execution provider is shared across
-  // multiple sessions.
-  class ModelMetadefIdGenerator {
-   public:
-    int GenerateId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash);
-
-   private:
-    std::unordered_map<HashValue, HashValue> main_graph_hash_;  // map graph instance hash to model contents hash
-    std::unordered_map<HashValue, int> model_metadef_id_;       // current unique id for model
-  };
-
-  std::unique_ptr<ModelMetadefIdGenerator> metadef_id_generator_;
 };
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/framework/execution_provider.cc b/onnxruntime/core/framework/execution_provider.cc
index 7f8009216ce3a..b39924d4c3ff9 100644
--- a/onnxruntime/core/framework/execution_provider.cc
+++ b/onnxruntime/core/framework/execution_provider.cc
@@ -35,77 +35,4 @@ common::Status IExecutionProvider::Compile(const std::vector<FusedNodeAndGraph>&
 }
 
 #endif
-
-int IExecutionProvider::ModelMetadefIdGenerator::GenerateId(const onnxruntime::GraphViewer& graph_viewer,
-                                                            HashValue& model_hash) {
-  model_hash = 0;
-
-  // find the top level graph
-  const Graph* cur_graph = &graph_viewer.GetGraph();
-  while (cur_graph->IsSubgraph()) {
-    cur_graph = cur_graph->ParentGraph();
-  }
-
-  uint32_t instance_hash[4] = {0, 0, 0, 0};
-
-  const Graph& main_graph = *cur_graph;
-
-  // hash the bytes in the Graph instance. we can't just use the address as a new Graph instance may use
-  // the same memory (unit tests prove this can occur). the raw bytes of the Graph instance should be a unique
-  // fingerprint for the instance that can use used as the key to the hash of the model path/contents.
-  MurmurHash3::x86_128(&main_graph, gsl::narrow_cast<int32_t>(sizeof(Graph)), instance_hash[0], &instance_hash);
-  HashValue graph_instance_hash = instance_hash[0] | (uint64_t(instance_hash[1]) << 32);
-
-  // if we've already hashed this main graph instance use the cached value
-  auto entry = main_graph_hash_.find(graph_instance_hash);
-  if (entry != main_graph_hash_.cend()) {
-    model_hash = entry->second;
-  } else {
-    uint32_t hash[4] = {0, 0, 0, 0};
-
-    // prefer path the model was loaded from
-    // this may not be available if the model was loaded from a stream or in-memory bytes
-    const auto& model_path_str = main_graph.ModelPath().ToPathString();
-    if (!model_path_str.empty()) {
-      MurmurHash3::x86_128(model_path_str.data(), gsl::narrow_cast<int32_t>(model_path_str.size()), hash[0], &hash);
-    } else {
-      auto hash_str = [&hash](const std::string& str) {
-        MurmurHash3::x86_128(str.data(), gsl::narrow_cast<int32_t>(str.size()), hash[0], &hash);
-      };
-
-      // fingerprint the main graph by hashing graph inputs and the ordered outputs from each node
-      for (const auto* node_arg : main_graph.GetInputsIncludingInitializers()) {
-        hash_str(node_arg->Name());
-      }
-
-      // note: process nodes in order defined in model to be deterministic
-      for (const auto& node : main_graph.Nodes()) {
-        for (const auto* node_arg : node.OutputDefs()) {
-          if (node_arg->Exists()) {
-            hash_str(node_arg->Name());
-          }
-        }
-      }
-    }
-
-    model_hash = hash[0] | (uint64_t(hash[1]) << 32);
-
-    main_graph_hash_[graph_instance_hash] = model_hash;
-  }
-
-  // return the current unique id, and increment to update
-  return model_metadef_id_[model_hash]++;
-}
-
-int IExecutionProvider::GenerateMetaDefId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) const {
-  ORT_ENFORCE(metadef_id_generator_,
-              "IExecutionProvider constructor must be called with true for use_metadef_id_creator");
-
-  // if the EP is shared across multiple sessions there's a very small potential for concurrency issues.
-  // use a lock when generating an id to be paranoid
-  static OrtMutex mutex;
-  std::lock_guard<OrtMutex> lock(mutex);
-  return metadef_id_generator_->GenerateId(graph_viewer, model_hash);
-}
-
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/framework/model_metadef_id_generator.cc b/onnxruntime/core/framework/model_metadef_id_generator.cc
new file mode 100644
index 0000000000000..e51c6ebc29975
--- /dev/null
+++ b/onnxruntime/core/framework/model_metadef_id_generator.cc
@@ -0,0 +1,75 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+#include <unordered_map>
+#include "model_metadef_id_generator.h"
+#include "core/platform/ort_mutex.h"
+#include "core/graph/graph_viewer.h"
+#include "core/framework/murmurhash3.h"
+
+namespace onnxruntime {
+int ModelMetadefIdGenerator::GenerateId(const onnxruntime::GraphViewer& graph_viewer,
+                                        HashValue& model_hash) const {
+  // if the EP is shared across multiple sessions there's a very small potential for concurrency issues.
+  // use a lock when generating an id to be paranoid
+  static OrtMutex mutex;
+  std::lock_guard<OrtMutex> lock(mutex);
+  model_hash = 0;
+
+  // find the top level graph
+  const Graph* cur_graph = &graph_viewer.GetGraph();
+  while (cur_graph->IsSubgraph()) {
+    cur_graph = cur_graph->ParentGraph();
+  }
+
+  uint32_t instance_hash[4] = {0, 0, 0, 0};
+
+  const Graph& main_graph = *cur_graph;
+
+  // hash the bytes in the Graph instance. we can't just use the address as a new Graph instance may use
+  // the same memory (unit tests prove this can occur). the raw bytes of the Graph instance should be a unique
+  // fingerprint for the instance that can use used as the key to the hash of the model path/contents.
+  MurmurHash3::x86_128(&main_graph, gsl::narrow_cast<int32_t>(sizeof(Graph)), instance_hash[0], &instance_hash);
+  HashValue graph_instance_hash = instance_hash[0] | (uint64_t(instance_hash[1]) << 32);
+
+  // if we've already hashed this main graph instance use the cached value
+  auto entry = main_graph_hash_.find(graph_instance_hash);
+  if (entry != main_graph_hash_.cend()) {
+    model_hash = entry->second;
+  } else {
+    uint32_t hash[4] = {0, 0, 0, 0};
+
+    // prefer path the model was loaded from
+    // this may not be available if the model was loaded from a stream or in-memory bytes
+    const auto& model_path_str = main_graph.ModelPath().ToPathString();
+    if (!model_path_str.empty()) {
+      MurmurHash3::x86_128(model_path_str.data(), gsl::narrow_cast<int32_t>(model_path_str.size()), hash[0], &hash);
+    } else {
+      auto hash_str = [&hash](const std::string& str) {
+        MurmurHash3::x86_128(str.data(), gsl::narrow_cast<int32_t>(str.size()), hash[0], &hash);
+      };
+
+      // fingerprint the main graph by hashing graph inputs and the ordered outputs from each node
+      for (const auto* node_arg : main_graph.GetInputsIncludingInitializers()) {
+        hash_str(node_arg->Name());
+      }
+
+      // note: process nodes in order defined in model to be deterministic
+      for (const auto& node : main_graph.Nodes()) {
+        for (const auto* node_arg : node.OutputDefs()) {
+          if (node_arg->Exists()) {
+            hash_str(node_arg->Name());
+          }
+        }
+      }
+    }
+
+    model_hash = hash[0] | (uint64_t(hash[1]) << 32);
+
+    main_graph_hash_[graph_instance_hash] = model_hash;
+  }
+
+  // return the current unique id, and increment to update
+  return model_metadef_id_[model_hash]++;
+}
+
+}  // namespace onnxruntime
diff --git a/onnxruntime/core/framework/model_metadef_id_generator.h b/onnxruntime/core/framework/model_metadef_id_generator.h
new file mode 100644
index 0000000000000..82f68c42b5c35
--- /dev/null
+++ b/onnxruntime/core/framework/model_metadef_id_generator.h
@@ -0,0 +1,31 @@
+// Copyright (c) Microsoft Corporation. All rights reserved.
+// Licensed under the MIT License.
+
+#pragma once
+#include <unordered_map>
+#include "core/common/basic_types.h"
+namespace onnxruntime {
+class GraphViewer;
+
+/// <summary>
+/// helper to generate ids that are unique to model and deterministic, even if the execution provider is shared across
+/// multiple sessions.
+/// </summary>
+class ModelMetadefIdGenerator {
+ public:
+  /** Generate a unique id that can be used in a MetaDef name. Values are unique for a model instance.
+   The model hash is also returned if you wish to include that in the MetaDef name to ensure uniqueness across models.
+   @param graph_viewer[in] Graph viewer that GetCapability was called with. Can be for the main graph or nested graph.
+   @param model_hash[out] Returns the hash for the main (i.e. top level) graph in the model.
+                          This is created using the model path if available,
+                          or the model input names and the output names from all nodes in the main graph.
+   */
+  int GenerateId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) const;
+
+ private:
+  // mutable as these are caches so we can minimize the hashing required on each usage of GenerateId
+  mutable std::unordered_map<HashValue, HashValue> main_graph_hash_;  // map graph instance hash to model contents hash
+  mutable std::unordered_map<HashValue, int> model_metadef_id_;       // current unique id for model
+};
+
+}  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cann/cann_execution_provider.cc b/onnxruntime/core/providers/cann/cann_execution_provider.cc
index 127c37bd84d0f..752b742805a7c 100644
--- a/onnxruntime/core/providers/cann/cann_execution_provider.cc
+++ b/onnxruntime/core/providers/cann/cann_execution_provider.cc
@@ -9,7 +9,6 @@
 #include <map>
 #include <unordered_set>
 
-#include "core/providers/shared_library/provider_api.h"
 #define ORT_API_MANUAL_INIT
 #include "core/session/onnxruntime_cxx_api.h"
 #include "core/providers/cann/cann_execution_provider.h"
@@ -1029,13 +1028,14 @@ Status RegisterCANNKernels(KernelRegistry& kernel_registry) {
 }  // namespace cann
 
 CANNExecutionProvider::CANNExecutionProvider(const CANNExecutionProviderInfo& info)
-    : IExecutionProvider{onnxruntime::kCannExecutionProvider, OrtDevice(OrtDevice::NPU, OrtDevice::MemType::DEFAULT, info.device_id), true}, info_{info} {
+    : IExecutionProvider{onnxruntime::kCannExecutionProvider, OrtDevice(OrtDevice::NPU, OrtDevice::MemType::DEFAULT, info.device_id)}, info_{info} {
   InitProviderOrtApi();
 
   CANN_CALL_THROW(aclrtSetDevice(info_.device_id));
 
   soc_name_ = aclrtGetSocName();
   ORT_ENFORCE(soc_name_ != nullptr, "aclrtGetSocName return nullptr");
+  metadef_id_generator_ = ModelMetadefIdGenerator::Create();
 }
 
 CANNExecutionProvider::~CANNExecutionProvider() {
@@ -1197,7 +1197,7 @@ std::unique_ptr<IndexedSubGraph> CANNExecutionProvider::GetSubGraph(
 
   // Generate unique kernel name for CANN subgraph
   HashValue model_hash = 0;
-  int id = GenerateMetaDefId(graph_viewer, model_hash);
+  int id = metadef_id_generator_->GenerateId(graph_viewer, model_hash);
   auto meta_def = IndexedSubGraph_MetaDef::Create();
   meta_def->name() = graph_viewer.Name() + "_" + std::to_string(model_hash) + "_" + std::to_string(id);
 
diff --git a/onnxruntime/core/providers/cann/cann_execution_provider.h b/onnxruntime/core/providers/cann/cann_execution_provider.h
index 76d3d9c331563..63ae980869c65 100644
--- a/onnxruntime/core/providers/cann/cann_execution_provider.h
+++ b/onnxruntime/core/providers/cann/cann_execution_provider.h
@@ -81,6 +81,7 @@ class CANNExecutionProvider : public IExecutionProvider {
   std::unordered_map<std::string, uint32_t> modelIDs_;
   std::unordered_map<std::string, std::string> models_;
   std::unordered_map<std::string, std::unordered_map<std::size_t, std::string>> names_;
+  std::unique_ptr<ModelMetadefIdGenerator> metadef_id_generator_;
 };
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/coreml/coreml_execution_provider.cc b/onnxruntime/core/providers/coreml/coreml_execution_provider.cc
index c9973671ffa28..c133f7b82aba4 100644
--- a/onnxruntime/core/providers/coreml/coreml_execution_provider.cc
+++ b/onnxruntime/core/providers/coreml/coreml_execution_provider.cc
@@ -24,7 +24,7 @@ namespace onnxruntime {
 constexpr const char* COREML = "CoreML";
 
 CoreMLExecutionProvider::CoreMLExecutionProvider(uint32_t coreml_flags)
-    : IExecutionProvider{onnxruntime::kCoreMLExecutionProvider, true},
+    : IExecutionProvider{onnxruntime::kCoreMLExecutionProvider},
       coreml_flags_(coreml_flags) {
 }
 
@@ -54,7 +54,7 @@ CoreMLExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_vie
 
   const auto gen_metadef_name = [&]() {
     HashValue model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
     return MakeString(COREML, "_", model_hash, "_", metadef_id);
   };
 
diff --git a/onnxruntime/core/providers/coreml/coreml_execution_provider.h b/onnxruntime/core/providers/coreml/coreml_execution_provider.h
index 67050e8079cf9..0201739547dd1 100644
--- a/onnxruntime/core/providers/coreml/coreml_execution_provider.h
+++ b/onnxruntime/core/providers/coreml/coreml_execution_provider.h
@@ -4,6 +4,7 @@
 #pragma once
 
 #include "core/framework/execution_provider.h"
+#include "core/framework/model_metadef_id_generator.h"
 #include "core/providers/coreml/coreml_provider_factory.h"
 
 namespace onnxruntime {
@@ -34,5 +35,6 @@ class CoreMLExecutionProvider : public IExecutionProvider {
 #ifdef __APPLE__
   std::unordered_map<std::string, std::unique_ptr<onnxruntime::coreml::Model>> coreml_models_;
 #endif
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/dnnl/dnnl_execution_provider.cc b/onnxruntime/core/providers/dnnl/dnnl_execution_provider.cc
index 05eb0091a8c83..3271dab13f675 100644
--- a/onnxruntime/core/providers/dnnl/dnnl_execution_provider.cc
+++ b/onnxruntime/core/providers/dnnl/dnnl_execution_provider.cc
@@ -5,8 +5,6 @@
 #pragma warning(disable : 4996)
 #endif
 
-#include "core/providers/dnnl/dnnl_execution_provider.h"
-
 #include <fstream>
 #include <iomanip>
 #include <unordered_set>
@@ -16,6 +14,7 @@
 
 #include "core/platform/ort_mutex.h"
 #include "core/providers/shared_library/provider_api.h"
+#include "core/providers/dnnl/dnnl_execution_provider.h"
 
 #include "core/providers/dnnl/dnnl_fwd.h"
 #include "core/providers/dnnl/dnnl_node_capability.h"
@@ -30,7 +29,7 @@ constexpr const char* DNNL = "Dnnl";
 constexpr const char* DNNL_CPU = "DnnlCpu";
 
 DnnlExecutionProvider::DnnlExecutionProvider(const DnnlExecutionProviderInfo& info)
-    : IExecutionProvider{onnxruntime::kDnnlExecutionProvider, true},
+    : IExecutionProvider{onnxruntime::kDnnlExecutionProvider},
       info_(info) {
   InitProviderOrtApi();
 
@@ -77,8 +76,8 @@ DnnlExecutionProvider::DnnlExecutionProvider(const DnnlExecutionProviderInfo& in
   // Log the number of threads used
   LOGS_DEFAULT(INFO) << "Allocated " << omp_get_max_threads() << " OpenMP threads for oneDNN ep\n";
 #endif  // defined(DNNL_OPENMP)
-
-}  // namespace onnxruntime
+  metadef_id_generator_ = ModelMetadefIdGenerator::Create();
+}
 
 DnnlExecutionProvider::~DnnlExecutionProvider() {
 }
@@ -229,7 +228,7 @@ std::vector<std::unique_ptr<ComputeCapability>> DnnlExecutionProvider::GetCapabi
 
     // Assign inputs and outputs to subgraph's meta_def
     HashValue model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_->GenerateId(graph_viewer, model_hash);
     auto meta_def = ::onnxruntime::IndexedSubGraph_MetaDef::Create();
     meta_def->name() = "DNNL_" + std::to_string(model_hash) + "_" + std::to_string(metadef_id);
     meta_def->domain() = kMSDomain;
@@ -264,7 +263,7 @@ std::vector<std::unique_ptr<ComputeCapability>> DnnlExecutionProvider::GetCapabi
     graph_viewer.ToProto(*model_proto->mutable_graph(), false, true);
     model_proto->set_ir_version(ONNX_NAMESPACE::Version::IR_VERSION);
     HashValue model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_->GenerateId(graph_viewer, model_hash);
     std::fstream dump("DNNL_" + std::to_string(model_hash) + "_" + std::to_string(metadef_id) + ".onnx", std::ios::out | std::ios::trunc | std::ios::binary);
     model_proto->SerializeToOstream(dump);
   }
diff --git a/onnxruntime/core/providers/dnnl/dnnl_execution_provider.h b/onnxruntime/core/providers/dnnl/dnnl_execution_provider.h
index 41062ccb4bc1b..b7fcbb7765180 100644
--- a/onnxruntime/core/providers/dnnl/dnnl_execution_provider.h
+++ b/onnxruntime/core/providers/dnnl/dnnl_execution_provider.h
@@ -41,6 +41,7 @@ class DnnlExecutionProvider : public IExecutionProvider {
   bool debug_log_ = false;
   // enable fusion by default
   bool enable_fusion_ = true;
+  std::unique_ptr<ModelMetadefIdGenerator> metadef_id_generator_;
 };
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/js/js_execution_provider.cc b/onnxruntime/core/providers/js/js_execution_provider.cc
index c2ff2ebc39e13..21bf09b274aff 100644
--- a/onnxruntime/core/providers/js/js_execution_provider.cc
+++ b/onnxruntime/core/providers/js/js_execution_provider.cc
@@ -680,7 +680,7 @@ std::unique_ptr<KernelRegistry> RegisterKernels() {
 using namespace js;
 
 JsExecutionProvider::JsExecutionProvider(const JsExecutionProviderInfo& info)
-    : IExecutionProvider{kJsExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0), true},
+    : IExecutionProvider{kJsExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, 0)},
       preferred_data_layout_{info.data_layout} {
 }
 
diff --git a/onnxruntime/core/providers/migraphx/migraphx_execution_provider.cc b/onnxruntime/core/providers/migraphx/migraphx_execution_provider.cc
index 8bfa66710e2fc..40e76a0a67782 100644
--- a/onnxruntime/core/providers/migraphx/migraphx_execution_provider.cc
+++ b/onnxruntime/core/providers/migraphx/migraphx_execution_provider.cc
@@ -102,7 +102,7 @@ std::shared_ptr<KernelRegistry> MIGraphXExecutionProvider::GetKernelRegistry() c
 }
 
 MIGraphXExecutionProvider::MIGraphXExecutionProvider(const MIGraphXExecutionProviderInfo& info)
-    : IExecutionProvider{onnxruntime::kMIGraphXExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, info.device_id), true}, device_id_(info.device_id) {
+    : IExecutionProvider{onnxruntime::kMIGraphXExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, info.device_id)}, device_id_(info.device_id) {
   InitProviderOrtApi();
   // Set GPU device to be used
   HIP_CALL_THROW(hipSetDevice(device_id_));
@@ -165,6 +165,8 @@ MIGraphXExecutionProvider::MIGraphXExecutionProvider(const MIGraphXExecutionProv
   MIOPEN_CALL_THROW(miopenCreate(&external_miopen_handle_));
   MIOPEN_CALL_THROW(miopenSetStream(external_miopen_handle_, stream_));
 
+  metadef_id_generator_ = ModelMetadefIdGenerator::Create();
+
   LOGS_DEFAULT(VERBOSE) << "[MIGraphX EP] MIGraphX provider options: "
                         << "device_id: " << device_id_
                         << ", migraphx_fp16_enable: " << fp16_enable_
@@ -757,7 +759,7 @@ std::unique_ptr<IndexedSubGraph> MIGraphXExecutionProvider::GetSubGraph(const st
 
   // Generate unique kernel name for MIGraphX subgraph
   uint64_t model_hash = 0;
-  int id = GenerateMetaDefId(graph, model_hash);
+  int id = metadef_id_generator_->GenerateId(graph, model_hash);
   std::string subgraph_id = std::to_string(model_hash) + "_" + std::to_string(id);
   auto meta_def = IndexedSubGraph_MetaDef::Create();
   const std::string graph_type = graph.IsSubgraph() ? "subgraph" : "graph";
diff --git a/onnxruntime/core/providers/migraphx/migraphx_execution_provider.h b/onnxruntime/core/providers/migraphx/migraphx_execution_provider.h
index c094be51012e4..d582338c7e067 100644
--- a/onnxruntime/core/providers/migraphx/migraphx_execution_provider.h
+++ b/onnxruntime/core/providers/migraphx/migraphx_execution_provider.h
@@ -98,6 +98,7 @@ class MIGraphXExecutionProvider : public IExecutionProvider {
   AllocatorPtr allocator_;
   miopenHandle_t external_miopen_handle_ = nullptr;
   rocblas_handle external_rocblas_handle_ = nullptr;
+  std::unique_ptr<ModelMetadefIdGenerator> metadef_id_generator_;
 };
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.cc b/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.cc
index 727917ad9232e..b04703d7611ee 100644
--- a/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.cc
+++ b/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.cc
@@ -50,7 +50,7 @@ std::unordered_set<std::string> GetPartitioningStopOps(const optional<std::strin
 
 NnapiExecutionProvider::NnapiExecutionProvider(uint32_t nnapi_flags,
                                                const optional<std::string>& partitioning_stop_ops_list)
-    : IExecutionProvider{onnxruntime::kNnapiExecutionProvider, true},
+    : IExecutionProvider{onnxruntime::kNnapiExecutionProvider},
       nnapi_flags_(nnapi_flags),
       partitioning_stop_ops_(GetPartitioningStopOps(partitioning_stop_ops_list)) {
   nnapi_handle_ = NnApiImplementation();
@@ -176,7 +176,7 @@ NnapiExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_view
 
   const auto gen_metadef_name = [&]() {
     HashValue model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
     return MakeString(NNAPI, "_", model_hash, "_", metadef_id);
   };
 
diff --git a/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.h b/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.h
index e4911511e6db0..460616c41991f 100644
--- a/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.h
+++ b/onnxruntime/core/providers/nnapi/nnapi_builtin/nnapi_execution_provider.h
@@ -6,6 +6,7 @@
 #include "core/common/inlined_containers_fwd.h"
 #include "core/common/optional.h"
 #include "core/framework/execution_provider.h"
+#include "core/framework/model_metadef_id_generator.h"
 #include "core/providers/nnapi/nnapi_builtin/nnapi_api_helper.h"
 #include "core/providers/nnapi/nnapi_provider_factory.h"
 
@@ -48,5 +49,6 @@ class NnapiExecutionProvider : public IExecutionProvider {
   const NnApi* nnapi_handle_ = nullptr;
   nnapi::DeviceWrapperVector nnapi_target_devices_;
   nnapi::TargetDeviceOption target_device_option_;
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/partitioning_utils.h b/onnxruntime/core/providers/partitioning_utils.h
index f9d5f7403f17b..136725c2f7250 100644
--- a/onnxruntime/core/providers/partitioning_utils.h
+++ b/onnxruntime/core/providers/partitioning_utils.h
@@ -40,7 +40,7 @@ using OnGroupClosedFn = std::function<bool(const std::vector<const Node*>& group
 
 /**
 Called to create a metadef name.
-Most likely should call IExecutionProvider::GenerateMetaDefId.
+Most likely should call ModelMetadefIdGenerator.GenerateId.
 See onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.cc for example usage.
 
 @return The metadef name.
diff --git a/onnxruntime/core/providers/qnn/qnn_execution_provider.cc b/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
index 0310cc2bc8f26..5f4e2e62f063e 100644
--- a/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
+++ b/onnxruntime/core/providers/qnn/qnn_execution_provider.cc
@@ -129,7 +129,7 @@ static void ParseHtpArchitecture(const std::string& htp_arch_string, QnnHtpDevic
 
 QNNExecutionProvider::QNNExecutionProvider(const ProviderOptions& provider_options_map,
                                            const SessionOptions* session_options)
-    : IExecutionProvider{onnxruntime::kQnnExecutionProvider, true} {
+    : IExecutionProvider{onnxruntime::kQnnExecutionProvider} {
   if (session_options) {
     disable_cpu_ep_fallback_ = session_options->config_options.GetConfigOrDefault(
                                    kOrtSessionOptionsDisableCPUEPFallback, "0") == "1";
@@ -472,7 +472,7 @@ QNNExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_viewer
 
   const auto gen_metadef_name = [&]() {
     uint64_t model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
     return MakeString(QNN, "_", model_hash, "_", metadef_id);
   };
 
diff --git a/onnxruntime/core/providers/qnn/qnn_execution_provider.h b/onnxruntime/core/providers/qnn/qnn_execution_provider.h
index 3f75be0efebcd..09bcb24db4dc2 100644
--- a/onnxruntime/core/providers/qnn/qnn_execution_provider.h
+++ b/onnxruntime/core/providers/qnn/qnn_execution_provider.h
@@ -5,6 +5,7 @@
 
 #include "core/framework/execution_provider.h"
 #include "core/framework/session_options.h"
+#include "core/framework/model_metadef_id_generator.h"
 #include "core/graph/model.h"
 #include <string>
 #include "core/providers/qnn/builder/qnn_backend_manager.h"
@@ -71,6 +72,7 @@ class QNNExecutionProvider : public IExecutionProvider {
   bool qnn_context_embed_mode_ = true;
   int32_t vtcm_size_in_mb_ = 0;
   std::unique_ptr<onnxruntime::Model> qnn_ep_context_model_;
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/shared_library/provider_api.h b/onnxruntime/core/providers/shared_library/provider_api.h
index 53ba4874c643c..1e3a528d87721 100644
--- a/onnxruntime/core/providers/shared_library/provider_api.h
+++ b/onnxruntime/core/providers/shared_library/provider_api.h
@@ -142,7 +142,7 @@ struct KernelDefBuilder;
 struct KernelRegistry;
 struct Function;
 struct Graph;
-struct GraphViewer;
+class GraphViewer;
 enum class DataLayout;
 struct Model;
 struct Path;
@@ -157,6 +157,7 @@ struct Tensor;
 struct SparseTensor;
 class TensorSeq;
 class SessionState;
+class ModelMetadefIdGenerator;
 
 class If;
 class Loop;
diff --git a/onnxruntime/core/providers/shared_library/provider_bridge_provider.cc b/onnxruntime/core/providers/shared_library/provider_bridge_provider.cc
index a3155fe6b86cf..a16d2890b47c2 100644
--- a/onnxruntime/core/providers/shared_library/provider_bridge_provider.cc
+++ b/onnxruntime/core/providers/shared_library/provider_bridge_provider.cc
@@ -329,10 +329,6 @@ common::Status IExecutionProvider::Compile(const std::vector<FusedNodeAndGraph>&
   return g_host->IExecutionProvider__Compile(this, fused_nodes_and_graphs, node_compute_funcs);
 }
 
-int IExecutionProvider::GenerateMetaDefId(const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) const {
-  return g_host->IExecutionProvider__GenerateMetaDefId(this, graph_viewer, model_hash);
-}
-
 #ifdef USE_TENSORRT
 std::unique_ptr<IAllocator> CreateCUDAAllocator(int16_t device_id, const char* name) {
   return g_host->CreateCUDAAllocator(device_id, name);
diff --git a/onnxruntime/core/providers/shared_library/provider_interfaces.h b/onnxruntime/core/providers/shared_library/provider_interfaces.h
index 21c14ce784a38..a216b2bfc6d04 100644
--- a/onnxruntime/core/providers/shared_library/provider_interfaces.h
+++ b/onnxruntime/core/providers/shared_library/provider_interfaces.h
@@ -229,8 +229,6 @@ struct ProviderHost {
 
   virtual common::Status IExecutionProvider__Compile(IExecutionProvider* p, const std::vector<IExecutionProvider::FusedNodeAndGraph>& fused_nodes_and_graphs, std::vector<NodeComputeInfo>& node_compute_funcs) = 0;
 
-  virtual int IExecutionProvider__GenerateMetaDefId(const IExecutionProvider* p, const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) = 0;
-
   // Status
   virtual std::string Status__ToString(const Status* p) = 0;
 
@@ -972,6 +970,11 @@ struct ProviderHost {
 #if !defined(ORT_MINIMAL_BUILD) || defined(ORT_MINIMAL_BUILD_CUSTOM_OPS)
   virtual Status LoadDynamicLibrary(onnxruntime::PathString library_name) = 0;
 #endif
+
+  // ModelMetadefIdGenerator
+  virtual std::unique_ptr<ModelMetadefIdGenerator> ModelMetadefIdGenerator__construct() = 0;
+  virtual void ModelMetadefIdGenerator__operator_delete(ModelMetadefIdGenerator* p) = 0;
+  virtual int ModelMetadefIdGenerator__GenerateId(const ModelMetadefIdGenerator* p, const GraphViewer& graph_viewer, HashValue& model_hash) = 0;
 };
 
 #if defined(_MSC_VER) && !defined(__clang__)
diff --git a/onnxruntime/core/providers/shared_library/provider_wrappedtypes.h b/onnxruntime/core/providers/shared_library/provider_wrappedtypes.h
index eaf8ef459cf00..f46c76fd3421b 100644
--- a/onnxruntime/core/providers/shared_library/provider_wrappedtypes.h
+++ b/onnxruntime/core/providers/shared_library/provider_wrappedtypes.h
@@ -750,7 +750,8 @@ struct Graph final {
   PROVIDER_DISALLOW_ALL(Graph)
 };
 
-struct GraphViewer final {
+class GraphViewer final {
+ public:
   static void operator delete(void* p) { g_host->GraphViewer__operator_delete(reinterpret_cast<GraphViewer*>(p)); }
 
   std::unique_ptr<Model> CreateModel(const logging::Logger& logger) const { return g_host->GraphViewer__CreateModel(this, logger); }
@@ -1152,6 +1153,13 @@ class TensorSeq final {
   void Reserve(size_t capacity) { g_host->TensorSeq__Reserve(this, capacity); }
 };
 
+class ModelMetadefIdGenerator {
+ public:
+  static std::unique_ptr<ModelMetadefIdGenerator> Create() { return g_host->ModelMetadefIdGenerator__construct(); }
+  static void operator delete(void* p) { g_host->ModelMetadefIdGenerator__operator_delete(reinterpret_cast<ModelMetadefIdGenerator*>(p)); }
+  int GenerateId(const GraphViewer& graph_viewer, HashValue& model_hash) const { return g_host->ModelMetadefIdGenerator__GenerateId(this, graph_viewer, model_hash); }
+};
+
 template <>
 inline gsl::span<const int64_t> Tensor::DataAsSpan() const { return g_host->Tensor__DataAsSpan_int64(this); }
 
diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
index fe6b959b962de..c709ca31f8643 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider.cc
@@ -1310,7 +1310,7 @@ TensorrtExecutionProvider::PerThreadContext& TensorrtExecutionProvider::GetPerTh
 }
 
 TensorrtExecutionProvider::TensorrtExecutionProvider(const TensorrtExecutionProviderInfo& info)
-    : IExecutionProvider{onnxruntime::kTensorrtExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, info.device_id), true}, info_(info), device_id_(info.device_id) {
+    : IExecutionProvider{onnxruntime::kTensorrtExecutionProvider, OrtDevice(OrtDevice::GPU, OrtDevice::MemType::DEFAULT, info.device_id)}, info_(info), device_id_(info.device_id) {
   InitProviderOrtApi();
 
   CUDA_CALL_THROW(cudaSetDevice(device_id_));
diff --git a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider_utils.h b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider_utils.h
index a8e3ae3ddf6ec..92cce0c203927 100644
--- a/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider_utils.h
+++ b/onnxruntime/core/providers/tensorrt/tensorrt_execution_provider_utils.h
@@ -497,7 +497,15 @@ void RemoveCachesByType(const std::string& root, std::string file_extension) {
   }
 }
 
-// Helper class to generate engine id via model name/model content/env metadata
+/**
+ * <summary>
+ * Helper class to generate engine id via model name/model content/env metadata
+ * </summary>
+ * <remarks>
+ * The TensorRT Execution Provider is used in multiple sessions and the underlying infrastructure caches
+ * compiled kernels, so the name must be unique and deterministic across models and sessions.
+ * </remarks>
+ */
 HashValue TRTGenerateId(const GraphViewer& graph_viewer) {
   HashValue model_hash = 0;
 
diff --git a/onnxruntime/core/providers/webnn/webnn_execution_provider.cc b/onnxruntime/core/providers/webnn/webnn_execution_provider.cc
index 2922cf9540a8e..257a8fa74d930 100644
--- a/onnxruntime/core/providers/webnn/webnn_execution_provider.cc
+++ b/onnxruntime/core/providers/webnn/webnn_execution_provider.cc
@@ -19,7 +19,7 @@ namespace onnxruntime {
 
 WebNNExecutionProvider::WebNNExecutionProvider(const std::string& webnn_device_flags,
                                                const std::string& webnn_threads_number, const std::string& webnn_power_flags)
-    : IExecutionProvider{onnxruntime::kWebNNExecutionProvider, true} {
+    : IExecutionProvider{onnxruntime::kWebNNExecutionProvider} {
   // Create WebNN context and graph builder.
   const emscripten::val ml = emscripten::val::global("navigator")["ml"];
   if (!ml.as<bool>()) {
@@ -168,7 +168,7 @@ WebNNExecutionProvider::GetCapability(const onnxruntime::GraphViewer& graph_view
 
     // Assign inputs and outputs to subgraph's meta_def.
     uint64_t model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
     auto meta_def = std::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>();
     meta_def->name = "WEBNN_" + std::to_string(model_hash) + "_" + std::to_string(metadef_id);
     meta_def->domain = kMSDomain;
diff --git a/onnxruntime/core/providers/webnn/webnn_execution_provider.h b/onnxruntime/core/providers/webnn/webnn_execution_provider.h
index 13a475327dc0c..d9cfa5f17c0d4 100644
--- a/onnxruntime/core/providers/webnn/webnn_execution_provider.h
+++ b/onnxruntime/core/providers/webnn/webnn_execution_provider.h
@@ -6,6 +6,7 @@
 
 #include "core/common/inlined_containers.h"
 #include "core/framework/execution_provider.h"
+#include "core/framework/model_metadef_id_generator.h"
 #include "core/providers/webnn/builders/helper.h"
 
 #include <emscripten.h>
@@ -48,5 +49,6 @@ class WebNNExecutionProvider : public IExecutionProvider {
   DataLayout preferred_layout_;
   webnn::WebnnDeviceType wnn_device_type_;
   InlinedHashMap<std::string, std::unique_ptr<onnxruntime::webnn::Model>> models_;
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/xnnpack/xnnpack_execution_provider.cc b/onnxruntime/core/providers/xnnpack/xnnpack_execution_provider.cc
index a2a776df439e4..eafbfae6f01e1 100644
--- a/onnxruntime/core/providers/xnnpack/xnnpack_execution_provider.cc
+++ b/onnxruntime/core/providers/xnnpack/xnnpack_execution_provider.cc
@@ -155,7 +155,7 @@ std::unique_ptr<KernelRegistry> RegisterKernels() {
 using namespace xnnpack;
 
 XnnpackExecutionProvider::XnnpackExecutionProvider(const XnnpackExecutionProviderInfo& info)
-    : IExecutionProvider{kXnnpackExecutionProvider, true} {
+    : IExecutionProvider{kXnnpackExecutionProvider} {
   int xnn_thread_pool_size = info.xnn_thread_pool_size;
   int ort_thread_pool_size = info.session_options ? info.session_options->intra_op_param.thread_pool_size : 1;
   bool allow_intra_op_spinning = (info.session_options == nullptr) ||
diff --git a/onnxruntime/core/session/provider_bridge_ort.cc b/onnxruntime/core/session/provider_bridge_ort.cc
index 3269c9f0f4e4b..f8bd7f4aa208c 100644
--- a/onnxruntime/core/session/provider_bridge_ort.cc
+++ b/onnxruntime/core/session/provider_bridge_ort.cc
@@ -30,6 +30,7 @@
 #include "core/framework/sparse_utils.h"
 #include "core/graph/graph_proto_serializer.h"
 #include "core/framework/murmurhash3.h"
+#include "core/framework/model_metadef_id_generator.h"
 
 #include "core/session/onnxruntime_c_api.h"
 #include "core/common/string_helper.h"
@@ -317,10 +318,6 @@ struct ProviderHostImpl : ProviderHost {
     return p->IExecutionProvider::Compile(fused_nodes_and_graphs, node_compute_funcs);
   }
 
-  int IExecutionProvider__GenerateMetaDefId(const IExecutionProvider* p, const onnxruntime::GraphViewer& graph_viewer, HashValue& model_hash) override {
-    return p->IExecutionProvider::GenerateMetaDefId(graph_viewer, model_hash);
-  }
-
   // Status (direct)
   std::string Status__ToString(const Status* p) override { return p->Status::ToString(); }
 
@@ -1083,6 +1080,11 @@ struct ProviderHostImpl : ProviderHost {
   void TensorSeq__Add(TensorSeq* p, Tensor&& tensor) override { p->Add(std::move(tensor)); }
   void TensorSeq__Reserve(TensorSeq* p, size_t capacity) override { p->Reserve(capacity); }
 
+  // ModelMetadefIdGenerator(wrapped)
+  std::unique_ptr<ModelMetadefIdGenerator> ModelMetadefIdGenerator__construct() override { return std::make_unique<ModelMetadefIdGenerator>(); }
+  void ModelMetadefIdGenerator__operator_delete(ModelMetadefIdGenerator* p) override { delete p; }
+  int ModelMetadefIdGenerator__GenerateId(const ModelMetadefIdGenerator* p, const GraphViewer& graph_viewer, HashValue& model_hash) override { return p->GenerateId(graph_viewer, model_hash); }
+
 #if defined(ENABLE_TRAINING) && defined(ORT_USE_NCCL)
   training::DistributedRunContext& GetDistributedRunContextInstance() override { return training::DistributedRunContext::GetInstance(); }
 #endif
diff --git a/onnxruntime/test/framework/execution_provider_test.cc b/onnxruntime/test/framework/execution_provider_test.cc
index 5a7351a766fa3..390fda7bfc5ad 100644
--- a/onnxruntime/test/framework/execution_provider_test.cc
+++ b/onnxruntime/test/framework/execution_provider_test.cc
@@ -6,6 +6,7 @@
 #include "test_utils.h"
 #include "test/test_environment.h"
 #include "test/util/include/asserts.h"
+#include "core/framework/model_metadef_id_generator.h"
 
 #include "gtest/gtest.h"
 
@@ -18,11 +19,14 @@ class TestEP : public IExecutionProvider {
   static constexpr const char* kEPType = "TestEP";
 
  public:
-  TestEP() : IExecutionProvider{kEPType, true} {}
+  TestEP() : IExecutionProvider{kEPType} {}
 
   int GetId(const GraphViewer& viewer, HashValue& model_hash) {
-    return GenerateMetaDefId(viewer, model_hash);
+    return metadef_id_generator_.GenerateId(viewer, model_hash);
   }
+
+ private:
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 
 TEST(ExecutionProviderTest, MetadefIdGeneratorUsingModelPath) {
diff --git a/onnxruntime/test/framework/tunable_op_test.cc b/onnxruntime/test/framework/tunable_op_test.cc
index 19253e1a5bd2c..6fe0754db40d3 100644
--- a/onnxruntime/test/framework/tunable_op_test.cc
+++ b/onnxruntime/test/framework/tunable_op_test.cc
@@ -82,7 +82,7 @@ class TestEP : public IExecutionProvider {
   TestTuningContext tuning_ctx_{this};
 
  public:
-  TestEP() : IExecutionProvider{kEPType, true} {}
+  TestEP() : IExecutionProvider{kEPType} {}
 
   ITuningContext* GetTuningContext() const override {
     return const_cast<TestTuningContext*>(&tuning_ctx_);
diff --git a/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.cc b/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.cc
index 957443c23e7c3..0167f7a7718b1 100644
--- a/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.cc
+++ b/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.cc
@@ -85,7 +85,7 @@ constexpr const char* INTERNAL_TESTING_EP = "InternalTestingEP";
 InternalTestingExecutionProvider::InternalTestingExecutionProvider(const std::unordered_set<std::string>& ops,
                                                                    const std::unordered_set<std::string>& stop_ops,
                                                                    DataLayout preferred_layout)
-    : IExecutionProvider{utils::kInternalTestingExecutionProvider, true},
+    : IExecutionProvider{utils::kInternalTestingExecutionProvider},
       ep_name_{INTERNAL_TESTING_EP},
       ops_{ops},
       stop_ops_{stop_ops},
@@ -212,7 +212,7 @@ InternalTestingExecutionProvider::GetCapability(const onnxruntime::GraphViewer&
   // create functor to generate a guaranteed unique metadef id
   auto generate_metadef_name = [this, &graph_viewer]() {
     HashValue model_hash;
-    int metadef_id = GenerateMetaDefId(graph_viewer, model_hash);
+    int metadef_id = metadef_id_generator_.GenerateId(graph_viewer, model_hash);
     auto meta_def = std::make_unique<::onnxruntime::IndexedSubGraph::MetaDef>();
     return ep_name_ + "_" + std::to_string(model_hash) + "_" + std::to_string(metadef_id);
   };
diff --git a/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.h b/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.h
index 6103352627667..6615eb82f2b05 100644
--- a/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.h
+++ b/onnxruntime/test/providers/internal_testing/internal_testing_execution_provider.h
@@ -4,6 +4,7 @@
 #pragma once
 #include <set>
 #include "core/framework/execution_provider.h"
+#include "core/framework/model_metadef_id_generator.h"
 
 namespace onnxruntime {
 namespace internal_testing_ep {
@@ -82,6 +83,7 @@ class InternalTestingExecutionProvider : public IExecutionProvider {
   // per-instance kernel registry so tests using static kernels don't clash.
   // shared_ptr as required by IExecutionProvider::GetKernelRegistry
   std::shared_ptr<KernelRegistry> kernel_registry_;
+  ModelMetadefIdGenerator metadef_id_generator_;
 };
 
 }  // namespace internal_testing_ep