diff --git a/onnxruntime/core/framework/data_types.cc b/onnxruntime/core/framework/data_types.cc
index 7255d8eebc..2a5b8a3133 100644
--- a/onnxruntime/core/framework/data_types.cc
+++ b/onnxruntime/core/framework/data_types.cc
@@ -726,7 +726,7 @@ const std::vector<MLDataType>& DataTypeImpl::AllTensorTypes() {
 // helper to stream. expected to only be used for error output, so any typeid lookup
 // cost should be fine. alternative would be to add a static string field to DataTypeImpl
 // that we set in the register macro to the type name, and output that instead.
-std::ostream& operator<<(std::ostream& out, const MLDataType data_type) {
+std::ostream& operator<<(std::ostream& out, const DataTypeImpl* data_type) {
   if (data_type == nullptr)
     return out << "(null)";
 
diff --git a/onnxruntime/core/framework/execution_frame.cc b/onnxruntime/core/framework/execution_frame.cc
index b6993932ef..b5df11f7d0 100644
--- a/onnxruntime/core/framework/execution_frame.cc
+++ b/onnxruntime/core/framework/execution_frame.cc
@@ -62,7 +62,7 @@ ExecutionFrame::ExecutionFrame(const std::unordered_map<std::string, MLValue>& f
 ExecutionFrame::~ExecutionFrame() = default;
 
 Status ExecutionFrame::AllocateMLValueTensorSelfOwnBuffer(int mlvalue_index,
-                                                          const MLDataType element_type,
+                                                          const DataTypeImpl* element_type,
                                                           const ONNXRuntimeAllocatorInfo& location,
                                                           const TensorShape& shape,
                                                           bool create_fence) {
@@ -71,7 +71,7 @@ Status ExecutionFrame::AllocateMLValueTensorSelfOwnBuffer(int mlvalue_index,
 }
 
 Status ExecutionFrame::AllocateMLValueTensorSelfOwnBufferHelper(int mlvalue_index,
-                                                                const MLDataType element_type,
+                                                                const DataTypeImpl* element_type,
                                                                 const ONNXRuntimeAllocatorInfo& location,
                                                                 const TensorShape& shape,
                                                                 bool create_fence) {
@@ -163,7 +163,7 @@ void ExecutionFrame::TraceAllocate(int mlvalue_idx, size_t size) {
 }
 
 Status ExecutionFrame::AllocateTensorWithSelfOwnBuffer(const int index,
-                                                       const MLDataType element_type,
+                                                       const DataTypeImpl* element_type,
                                                        const ONNXRuntimeAllocatorInfo& location,
                                                        const TensorShape& shape,
                                                        bool create_fence) {
@@ -173,7 +173,7 @@ Status ExecutionFrame::AllocateTensorWithSelfOwnBuffer(const int index,
 
 Status ExecutionFrame::AllocateMLValueTensorPreAllocateBuffer(int mlvalue_index_to_allocate,
                                                               int mlvalue_index_reuse,
-                                                              const MLDataType element_type,
+                                                              const DataTypeImpl* element_type,
                                                               const ONNXRuntimeAllocatorInfo& location,
                                                               const TensorShape& shape,
                                                               bool create_fence) {
@@ -200,7 +200,7 @@ Status ExecutionFrame::AllocateMLValueTensorPreAllocateBuffer(int mlvalue_index_
 
 Status ExecutionFrame::AllocateTensorWithPreAllocateBufferHelper(MLValue* p_mlvalue,
                                                                  void* pBuffer,
-                                                                 const MLDataType element_type,
+                                                                 const DataTypeImpl* element_type,
                                                                  const ONNXRuntimeAllocatorInfo& location,
                                                                  const TensorShape& shape) {
   if (p_mlvalue->IsAllocated()) {
@@ -219,7 +219,7 @@ Status ExecutionFrame::AllocateTensorWithPreAllocateBufferHelper(MLValue* p_mlva
 
 Status ExecutionFrame::AllocateTensorWithPreAllocateBuffer(const int offset,
                                                            void* pBuffer,
-                                                           const MLDataType element_type,
+                                                           const DataTypeImpl* element_type,
                                                            const ONNXRuntimeAllocatorInfo& location,
                                                            const TensorShape& shape) {
   ONNXRUNTIME_ENFORCE(offset >= 0 && offset < node_values_.size());
diff --git a/onnxruntime/core/framework/kernel_registry.cc b/onnxruntime/core/framework/kernel_registry.cc
index 30a0af85cc..81c6fae987 100644
--- a/onnxruntime/core/framework/kernel_registry.cc
+++ b/onnxruntime/core/framework/kernel_registry.cc
@@ -133,7 +133,7 @@ bool KernelRegistry::VerifyKernelDef(const onnxruntime::Node& node,
     // valid names (of types or parameters) at the time that kernels are registered.
     if ((nullptr != actual_type) &&
         !std::any_of(allowed_types.begin(), allowed_types.end(),
-                     [actual_type, &node, &error_str](const MLDataType& expected_type) {
+                     [actual_type, &node, &error_str](const DataTypeImpl* expected_type) {
                        bool rc = expected_type->IsCompatible(*actual_type);  // for easier debugging
                        if (!rc) {
                          // TODO print type information as well
diff --git a/onnxruntime/python/onnxruntime_pybind_mlvalue.cc b/onnxruntime/python/onnxruntime_pybind_mlvalue.cc
index 1725f9d403..d0ae5504b2 100644
--- a/onnxruntime/python/onnxruntime_pybind_mlvalue.cc
+++ b/onnxruntime/python/onnxruntime_pybind_mlvalue.cc
@@ -19,7 +19,7 @@ namespace python {
 namespace py = pybind11;
 using namespace onnxruntime::logging;
 
-int OnnxRuntimeTensorToNumpyType(const MLDataType& tensor_type) {
+int OnnxRuntimeTensorToNumpyType(const DataTypeImpl* tensor_type) {
   static std::map<MLDataType, int> type_map{
       {DataTypeImpl::GetType<bool>(), NPY_BOOL},
       {DataTypeImpl::GetType<float>(), NPY_FLOAT},
@@ -42,7 +42,7 @@ int OnnxRuntimeTensorToNumpyType(const MLDataType& tensor_type) {
   }
 }
 
-const MLDataType& NumpyToOnnxRuntimeTensorType(int numpy_type) {
+const DataTypeImpl* NumpyToOnnxRuntimeTensorType(int numpy_type) {
   static std::map<int, MLDataType> type_map{
       {NPY_BOOL, DataTypeImpl::GetType<bool>()},
       {NPY_FLOAT, DataTypeImpl::GetType<float>()},
diff --git a/onnxruntime/python/onnxruntime_pybind_mlvalue.h b/onnxruntime/python/onnxruntime_pybind_mlvalue.h
index 4673893d37..8b47883700 100644
--- a/onnxruntime/python/onnxruntime_pybind_mlvalue.h
+++ b/onnxruntime/python/onnxruntime_pybind_mlvalue.h
@@ -22,7 +22,7 @@ namespace python {
 
 namespace py = pybind11;
 
-int OnnxRuntimeTensorToNumpyType(const MLDataType& tensor_type);
+int OnnxRuntimeTensorToNumpyType(const DataTypeImpl* tensor_type);
 
 void CreateGenericMLValue(AllocatorPtr alloc, const std::string& name_input, py::object& value, MLValue* p_mlvalue);