From b2cddc5337d4e4c058b2f57fe4a65e539c3c491b Mon Sep 17 00:00:00 2001
From: Edward Chen <18449977+edgchen1@users.noreply.github.com>
Date: Fri, 12 Feb 2021 13:26:56 -0800
Subject: [PATCH] Consolidate MLTypeCallDispatcher classes (#6651)

---
 .../core/framework/data_types_internal.h      | 206 ++++++++++--------
 onnxruntime/contrib_ops/cpu/inverse.cc        |   4 +-
 onnxruntime/contrib_ops/cuda/inverse.cc       |   5 +-
 .../contrib_ops/cuda/math/bias_softmax.cc     |  10 +-
 .../contrib_ops/rocm/math/bias_softmax.cc     |  11 +-
 .../core/framework/tensorprotoutils.cc        |  10 +-
 .../core/optimizer/matmul_scale_fusion.cc     |   6 +-
 .../core/providers/cpu/generator/range.cc     |   5 +-
 onnxruntime/core/providers/cpu/math/clip.cc   |   4 +-
 .../providers/cpu/math/element_wise_ops.cc    |  14 +-
 onnxruntime/core/providers/cpu/math/sign.cc   |   4 +-
 .../core/providers/cpu/ml/normalizer.cc       |   4 +-
 onnxruntime/core/providers/cpu/nn/pool.cc     |   4 +-
 onnxruntime/core/providers/cpu/nn/shrink.cc   |   6 +-
 .../core/providers/cuda/generator/range.cc    |   5 +-
 onnxruntime/core/providers/cuda/math/clip.cc  |   4 +-
 .../cuda/math/variadic_elementwise_ops.cc     |  29 +--
 onnxruntime/core/providers/cuda/nn/dropout.h  |  14 +-
 .../core/providers/cuda/tensor/gather_nd.cc   |   6 +-
 .../providers/cuda/tensor/scatter_elements.cc |  11 +-
 onnxruntime/core/session/onnxruntime_c_api.cc |  14 +-
 .../cpu/cat_imputer_transformer.cc            |   4 +-
 .../cpu/forecasting_pivot_transformer.cc      |  13 +-
 .../cpu/from_string_transformer.cc            |   4 +-
 .../hash_one_hot_vectorizer_transformer.cc    |   4 +-
 .../cpu/imputation_marker_transformer.cc      |   4 +-
 .../cpu/label_encoder_transformer.cc          |   4 +-
 .../cpu/laglead_operator_transformer.cc       |   5 +-
 .../cpu/max_abs_scaler_transformer.cc         |   4 +-
 .../cpu/mean_imputer_transformer.cc           |   4 +-
 .../cpu/median_imputer_transformer.cc         |   4 +-
 .../cpu/min_max_imputer_transformer.cc        |   4 +-
 .../cpu/min_max_scaler_transformer.cc         |   4 +-
 .../cpu/missing_dummies_transformer.cc        |   4 +-
 .../cpu/mode_imputer_transformer.cc           |   4 +-
 .../cpu/normalize_transformer.cc              |   4 +-
 .../cpu/numericalize_transformer.cc           |   7 +-
 .../cpu/one_hot_encoder_transformer.cc        |   4 +-
 .../featurizers_ops/cpu/pca_transformer.cc    |   5 +-
 .../cpu/robust_scaler_transformer.cc          |   4 +-
 .../cpu/rolling_window_transformer.cc         |  10 +-
 .../cpu/short_grain_dropper_transformer.cc    |   8 +-
 .../cpu/standard_scale_wrapper_transformer.cc |   4 +-
 .../featurizers_ops/cpu/string_transformer.cc |   4 +-
 .../cpu/time_series_imputer_transformer.cc    |   8 +-
 .../cpu/truncated_svd_transformer.cc          |   5 +-
 .../training_ops/cpu/tensor/gather_nd_grad.cc |   4 +-
 .../cuda/activation/bias_gelu_grad.cc         |   7 +-
 .../training_ops/cuda/math/scale.cc           |   4 +-
 .../training_ops/cuda/nn/dropout.cc           |  17 +-
 .../cuda/tensor/gather_elements_grad.cc       |  10 +-
 .../cuda/tensor/gather_nd_grad.cc             |   6 +-
 52 files changed, 277 insertions(+), 276 deletions(-)

diff --git a/include/onnxruntime/core/framework/data_types_internal.h b/include/onnxruntime/core/framework/data_types_internal.h
index 23e4166e20..8e1649c552 100644
--- a/include/onnxruntime/core/framework/data_types_internal.h
+++ b/include/onnxruntime/core/framework/data_types_internal.h
@@ -7,6 +7,7 @@
 #include <cassert>
 #include <cstdint>
 #include <string>
+#include <type_traits>
 #include <vector>
 
 #include "boost/mp11.hpp"
@@ -16,11 +17,6 @@
 #include "core/framework/data_types.h"
 #include "core/graph/onnx_protobuf.h"
 
-#ifdef _MSC_VER
-#pragma warning(push)
-//TODO: fix the warning in CallableDispatchableRetHelper
-#pragma warning(disable : 4702)
-#endif
 namespace onnxruntime {
 namespace utils {
 
@@ -223,6 +219,7 @@ inline bool IsPrimitiveDataType(const PrimitiveDataTypeBase* prim_type) {
 // This implementation contains a workaround for GCC bug https://gcc.gnu.org/bugzilla/show_bug.cgi?id=47226
 // GCC until very recently does not support template parameter pack expansion within lambda context.
 namespace mltype_dispatcher_internal {
+
 // T - type handled by this helper
 class CallableDispatchableHelper {
   int32_t dt_type_;  // Type currently dispatched
@@ -242,7 +239,6 @@ class CallableDispatchableHelper {
   }
 
   void CheckCalledOnce() {
-    ORT_ENFORCE(called_ < 2, "Check for duplicate types in MLTypeCallDispatcher");
     ORT_ENFORCE(called_ == 1, "Unsupported data type: ", dt_type_);
   }
 };
@@ -256,7 +252,7 @@ struct UnsupportedTypeDefaultPolicy {
 };
 
 // Helper with the result type
-template <class Ret, class UnsupportedPolicy = UnsupportedTypeDefaultPolicy<Ret>>
+template <class Ret, class UnsupportedPolicy>
 class CallableDispatchableRetHelper {
   int32_t dt_type_;  // Type currently dispatched
   size_t called_;
@@ -266,8 +262,6 @@ class CallableDispatchableRetHelper {
   explicit CallableDispatchableRetHelper(int32_t dt_type) noexcept : dt_type_(dt_type), called_(0), result_() {}
 
   Ret Get() {
-    // See if there were multiple invocations.It is a bug.
-    ORT_ENFORCE(called_ < 2, "Check for duplicate types in MLTypeCallDispatcherRet");
     // No type was invoked
     if (called_ == 0) {
       result_ = UnsupportedPolicy()(dt_type_);
@@ -286,103 +280,89 @@ class CallableDispatchableRetHelper {
   }
 };
 
+template <typename T>
+using TensorProtoElementTypeConstant =
+    std::integral_constant<ONNX_NAMESPACE::TensorProto_DataType, ToTensorProtoElementType<T>()>;
+
+using UndefinedTensorProtoElementTypeConstant =
+    std::integral_constant<ONNX_NAMESPACE::TensorProto_DataType, ONNX_NAMESPACE::TensorProto_DataType_UNDEFINED>;
+
 }  // namespace mltype_dispatcher_internal
 
-// This class helps to efficiently dispatch calls for templated
-// kernel implementation functions that has no return value.
-// If your implementation function must return a value such as Status
-// Use MLTypeCallDispatcherRet class.
-//
-// The first template parameter is a template<T> struct/class functor
-// that must implement operator() with arbitrary number of arguments
-// and void return turn. It must return Ret type if you are using MLTypeCallDispatcherRet.
-// Fn must be default constructible.
-//
-// Types is a type list that are supported by this kernel implementation.
-// There should be no duplicate types. An exception will be thrown if there
-// a duplicate.
-//
-// The constructor accepts an enum that is obtained from
-// input_tensor->DataType()->AsPrimitiveType()->GetDataType().
-// Fn will be called only once the type designated by dt_type value.
-// If current dt_type is not handled, the Dispatcher will throw an exception.
-//
-template <template <typename> class Fn, typename... Types>
+/**
+ * This class helps to efficiently dispatch calls to implementation function
+ * objects with a tensor element type template argument.
+ *
+ * The constructor accepts a value corresponding to a tensor element type.
+ * For example, it can be obtained from:
+ *   input_tensor->GetElementType()
+ *
+ * The Invoke member functions will instantiate and invoke the provided
+ * function object template, Fn. Fn must be default constructible. Fn must also
+ * have a tensor element type template argument. This type template argument
+ * will be the type that corresponds to the value given in the constructor.
+ * These functions accept and forward arbitrary function arguments. They ensure
+ * that Fn is called once with the type specified in the constructor.
+ *
+ * @tparam Types The types supported by the implementation. This should be a
+ *         set of ONNX tensor element types that are supported by ORT.
+ */
+template <typename... Types>
 class MLTypeCallDispatcher {
+  using SupportedTypeList = TypeList<Types...>;
+  using SupportedTensorProtoElementTypeList =
+      boost::mp11::mp_transform<
+          mltype_dispatcher_internal::TensorProtoElementTypeConstant, SupportedTypeList>;
+
+  static_assert(
+      boost::mp11::mp_and<
+          boost::mp11::mp_is_set<SupportedTensorProtoElementTypeList>,
+          boost::mp11::mp_not<
+              boost::mp11::mp_set_contains<
+                  SupportedTensorProtoElementTypeList,
+                  mltype_dispatcher_internal::UndefinedTensorProtoElementTypeConstant>>>::value,
+      "Types must map to a unique set of ONNX tensor element data types supported by ORT.");
+
   int32_t dt_type_;
 
  public:
+  /**
+   * Constructor.
+   * @param dt_type The value corresponding to the tensor element type to be
+   *        dispatched to. This can be obtained from
+   *        input_tensor->GetElementType() or
+   *        utils::ToTensorProtoElementType<T>().
+   */
   explicit MLTypeCallDispatcher(int32_t dt_type) noexcept : dt_type_(dt_type) {}
 
-  template <typename... Args>
-  void Invoke(Args&&... args) const {
-    mltype_dispatcher_internal::CallableDispatchableHelper helper(dt_type_);
-    int results[] = {0, helper.template Invoke<Types>(Fn<Types>(), std::forward<Args>(args)...)...};
-    ORT_UNUSED_PARAMETER(results);
-    helper.CheckCalledOnce();
-  }
-};
-
-// Version of the MLTypeDispatcher with a return type.
-// Return type of Fn must return type convertible to Ret
-// The value of the return type will be the return value
-// of the function for type T which was specified for execution.
-template <class Ret, template <typename> class Fn, typename... Types>
-class MLTypeCallDispatcherRet {
-  int32_t dt_type_;
-
- public:
-  explicit MLTypeCallDispatcherRet(int32_t dt_type) noexcept : dt_type_(dt_type) {}
-
-  template <typename... Args>
-  Ret Invoke(Args&&... args) const {
-    mltype_dispatcher_internal::CallableDispatchableRetHelper<Ret> helper(dt_type_);
-    int results[] = {0, helper.template Invoke<Types>(Fn<Types>(), std::forward<Args>(args)...)...};
-    ORT_UNUSED_PARAMETER(results);
-    return helper.Get();
-  }
-
-  template <class UnsupportedPolicy, typename... Args>
-  Ret InvokeWithUnsupportedPolicy(Args&&... args) const {
-    mltype_dispatcher_internal::CallableDispatchableRetHelper<Ret, UnsupportedPolicy> helper(dt_type_);
-    int results[] = {0, helper.template Invoke<Types>(Fn<Types>(), std::forward<Args>(args)...)...};
-    ORT_UNUSED_PARAMETER(results);
-    return helper.Get();
-  }
-};
-
-// Version of MLTypeCallDispatcher that takes supported types as class-level template parameters.
-// This enables easier use with type list representations of the supported types.
-// The invocation-related template parameters like Fn move to the individual Invoke() methods.
-// TODO consolidate this with the other MLTypeCallDispatcher classes
-// can add additional methods to cover their usages, but need to update call sites
-template <typename... Types>
-class MLTypeCallDispatcher2 {
-  static_assert(boost::mp11::mp_is_set<TypeList<Types...>>::value,
-                "MLTypeCallDispatcher requires a set of unique types.");
-
-  int32_t dt_type_;
-
- public:
-  explicit MLTypeCallDispatcher2(int32_t dt_type) noexcept : dt_type_(dt_type) {}
-
+  /**
+   * Invokes Fn<T> with the specified arguments.
+   *
+   * @tparam Fn The function object template.
+   * @tparam Args The argument types.
+   */
   template <template <typename> class Fn, typename... Args>
   void Invoke(Args&&... args) const {
-    mltype_dispatcher_internal::CallableDispatchableHelper helper(dt_type_);
-
-    static_cast<void>(std::array<int, sizeof...(Types)>{
-        helper.template Invoke<Types>(Fn<Types>(), std::forward<Args>(args)...)...});
-
-    // avoid "unused parameter" warning for the case where Types is empty
-    static_cast<void>(std::array<int, sizeof...(Args)>{(ORT_UNUSED_PARAMETER(args), 0)...});
-
-    helper.CheckCalledOnce();
+    InvokeWithLeadingTemplateArgs<Fn, TypeList<>>(std::forward<Args>(args)...);
   }
 
+  /**
+   * Invokes Fn<..., T> with leading template arguments and the specified arguments.
+   *
+   * @tparam Fn The function object template.
+   * @tparam LeadingTemplateArgTypeList A type list of the leading template arguments.
+   * @tparam Args The argument types.
+   */
   template <template <typename...> class Fn, typename LeadingTemplateArgTypeList, typename... Args>
   void InvokeWithLeadingTemplateArgs(Args&&... args) const {
+    static_assert(
+        boost::mp11::mp_is_list<LeadingTemplateArgTypeList>::value,
+        "LeadingTemplateArgTypeList must be a type list (e.g., onnxruntime::TypeList<T1, T2, ...>).");
+
     mltype_dispatcher_internal::CallableDispatchableHelper helper(dt_type_);
 
+    // given LeadingTemplateArgTypeList is a type list L<U1, U2, ...>,
+    //   call helper.Invoke() with Fn<U1, U2, ..., T> for each T in Types
     static_cast<void>(std::array<int, sizeof...(Types)>{
         helper.template Invoke<Types>(
             boost::mp11::mp_apply<Fn, boost::mp11::mp_push_back<LeadingTemplateArgTypeList, Types>>(),
@@ -393,11 +373,49 @@ class MLTypeCallDispatcher2 {
 
     helper.CheckCalledOnce();
   }
+
+  /**
+   * Invokes Fn<T> with the specified arguments and returns the result.
+   *
+   * @tparam Ret The return type. Fn should return a type convertible to Ret.
+   * @tparam Fn The function object template.
+   * @tparam Args The argument types.
+   */
+  template <class Ret, template <typename> class Fn, typename... Args>
+  Ret InvokeRet(Args&&... args) const {
+    return InvokeRetWithUnsupportedPolicy<
+        Ret, Fn, mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy<Ret>>(
+        std::forward<Args>(args)...);
+  }
+
+  /**
+   * Invokes Fn<T> with the specified arguments and returns the result.
+   *
+   * @tparam Ret The return type. Fn should return a type convertible to Ret.
+   * @tparam Fn The function object template.
+   * @tparam UnsupportedPolicy The policy used to handle unsupported types.
+   *         See mltype_dispatcher_internal::UnsupportedTypeDefaultPolicy
+   *         for an example.
+   * @tparam Args The argument types.
+   */
+  template <class Ret, template <typename> class Fn, class UnsupportedPolicy, typename... Args>
+  Ret InvokeRetWithUnsupportedPolicy(Args&&... args) const {
+    mltype_dispatcher_internal::CallableDispatchableRetHelper<Ret, UnsupportedPolicy> helper(dt_type_);
+
+    // call helper.Invoke() with Fn<T> for each T in Types
+    static_cast<void>(std::array<int, sizeof...(Types)>{
+        helper.template Invoke<Types>(Fn<Types>(), std::forward<Args>(args)...)...});
+
+    // avoid "unused parameter" warning for the case where Types is empty
+    static_cast<void>(std::array<int, sizeof...(Args)>{(ORT_UNUSED_PARAMETER(args), 0)...});
+
+    return helper.Get();
+  }
 };
 
-// the type MLTypeCallDispatcher2<T...> given a type list L<T...>
+// the type MLTypeCallDispatcher<T...> given a type list L<T...>
 template <typename L>
-using MLTypeCallDispatcherFromTypeList = boost::mp11::mp_apply<MLTypeCallDispatcher2, L>;
+using MLTypeCallDispatcherFromTypeList = boost::mp11::mp_apply<MLTypeCallDispatcher, L>;
 
 namespace data_types_internal {
 
@@ -553,7 +571,3 @@ bool IsOpaqueType(MLDataType ml_type, const char* domain, const char* name);
 
 }  // namespace utils
 }  // namespace onnxruntime
-
-#ifdef _MSC_VER
-#pragma warning(pop)
-#endif
diff --git a/onnxruntime/contrib_ops/cpu/inverse.cc b/onnxruntime/contrib_ops/cpu/inverse.cc
index 9691acbf02..a428946aa8 100644
--- a/onnxruntime/contrib_ops/cpu/inverse.cc
+++ b/onnxruntime/contrib_ops/cpu/inverse.cc
@@ -77,8 +77,8 @@ Status Inverse::Compute(OpKernelContext* ctx) const {
   }
 
   std::function<void(ptrdiff_t)> fn = [elem_type, input, output, rows, cols](ptrdiff_t batch_num) {
-    utils::MLTypeCallDispatcher<ComputeImpl, float, double, MLFloat16> t_disp(elem_type);
-    t_disp.Invoke(input, output, batch_num, rows, cols);
+    utils::MLTypeCallDispatcher<float, double, MLFloat16> t_disp(elem_type);
+    t_disp.Invoke<ComputeImpl>(input, output, batch_num, rows, cols);
   };
 
   concurrency::ThreadPool::TryBatchParallelFor(ctx->GetOperatorThreadPool(), num_batches, std::move(fn), 0);
diff --git a/onnxruntime/contrib_ops/cuda/inverse.cc b/onnxruntime/contrib_ops/cuda/inverse.cc
index f762b09d9f..f75ae8a1e6 100644
--- a/onnxruntime/contrib_ops/cuda/inverse.cc
+++ b/onnxruntime/contrib_ops/cuda/inverse.cc
@@ -153,8 +153,9 @@ Status Inverse::ComputeInternal(OpKernelContext* ctx) const {
   CUDA_RETURN_IF_ERROR(cudaMemsetAsync(info.get(), 0, num_batches, Stream()));
   IAllocatorUniquePtr<int> pivots = GetScratchBuffer<int>(rows * num_batches);
 
-  utils::MLTypeCallDispatcherRet<Status, ComputeImpl, float, double, MLFloat16> t_disp(input->GetElementType());
-  return t_disp.Invoke(Stream(), Base::CublasHandle(), this, *input, *output, info, pivots, num_batches, rows);
+  utils::MLTypeCallDispatcher<float, double, MLFloat16> t_disp(input->GetElementType());
+  return t_disp.InvokeRet<Status, ComputeImpl>(
+      Stream(), Base::CublasHandle(), this, *input, *output, info, pivots, num_batches, rows);
 }
 
 }  // namespace cuda
diff --git a/onnxruntime/contrib_ops/cuda/math/bias_softmax.cc b/onnxruntime/contrib_ops/cuda/math/bias_softmax.cc
index d9d30055dd..9a09761a03 100644
--- a/onnxruntime/contrib_ops/cuda/math/bias_softmax.cc
+++ b/onnxruntime/contrib_ops/cuda/math/bias_softmax.cc
@@ -62,16 +62,14 @@ Status BiasSoftmax::ComputeInternal(OpKernelContext* ctx) const {
   const int broadcast_size = N / static_cast<int>(X_shape.SizeToDimension(broadcast_axis));
 
   const size_t elem_size = X->DataType()->Size();
+  utils::MLTypeCallDispatcher<double, float, MLFloat16> t_disp(X->GetElementType());
+
   if (D <= 1024 && D * elem_size <= 4096) {
     // expect thread blocks can fill SM at high occupancy without overflowing registers
-    utils::MLTypeCallDispatcher<DispatchBiasSoftmaxForward, double, float, MLFloat16>
-        t_disp(X->GetElementType());
-    t_disp.Invoke(Stream(), Y, X, B, D, N, D, broadcast_size);
+    t_disp.Invoke<DispatchBiasSoftmaxForward>(Stream(), Y, X, B, D, N, D, broadcast_size);
   } else {
     // need to fallback to add kernel + CUDA DNN library softmax call :/
-    utils::MLTypeCallDispatcher<DispatchBiasSoftMaxForwardViaDnnLibrary, double, float, MLFloat16>
-        t_disp(X->GetElementType());
-    t_disp.Invoke(Stream(), CudnnHandle(), D, N, broadcast_axis, softmax_axis, X_shape, X, B_shape, B, Y);
+    t_disp.Invoke<DispatchBiasSoftMaxForwardViaDnnLibrary>(Stream(), CudnnHandle(), D, N, broadcast_axis, softmax_axis, X_shape, X, B_shape, B, Y);
   }
 
   return Status::OK();
diff --git a/onnxruntime/contrib_ops/rocm/math/bias_softmax.cc b/onnxruntime/contrib_ops/rocm/math/bias_softmax.cc
index 954474aa09..2612d7eb52 100644
--- a/onnxruntime/contrib_ops/rocm/math/bias_softmax.cc
+++ b/onnxruntime/contrib_ops/rocm/math/bias_softmax.cc
@@ -65,16 +65,15 @@ Status BiasSoftmax::ComputeInternal(OpKernelContext* ctx) const {
   const int broadcast_size = N / static_cast<int>(X_shape.SizeToDimension(broadcast_axis));
 
   const size_t elem_size = X->DataType()->Size();
+  utils::MLTypeCallDispatcher<float, MLFloat16> t_disp(X->GetElementType());
+
   if (D <= 1024 && D * elem_size <= 4096) {
     // expect thread blocks can fill SM at high occupancy without overflowing registers
-    utils::MLTypeCallDispatcher<DispatchBiasSoftmaxForward, float, MLFloat16>
-        t_disp(X->GetElementType());
-    t_disp.Invoke(Stream(), Y, X, B, D, N, D, broadcast_size);
+    t_disp.Invoke<DispatchBiasSoftmaxForward>(Stream(), Y, X, B, D, N, D, broadcast_size);
   } else {
     // need to fallback to add kernel + CUDA DNN library softmax call :/
-    utils::MLTypeCallDispatcher<DispatchBiasSoftMaxForwardViaDnnLibrary, float, MLFloat16>
-        t_disp(X->GetElementType());
-    t_disp.Invoke(Stream(), MiopenHandle(), D, N, broadcast_axis, softmax_axis, X_shape, X, B_shape, B, Y);
+    t_disp.Invoke<DispatchBiasSoftMaxForwardViaDnnLibrary>(
+        Stream(), MiopenHandle(), D, N, broadcast_axis, softmax_axis, X_shape, X, B_shape, B, Y);
   }
 
   return Status::OK();
diff --git a/onnxruntime/core/framework/tensorprotoutils.cc b/onnxruntime/core/framework/tensorprotoutils.cc
index 0e6f63cae9..79a19983ef 100644
--- a/onnxruntime/core/framework/tensorprotoutils.cc
+++ b/onnxruntime/core/framework/tensorprotoutils.cc
@@ -954,8 +954,9 @@ common::Status SparseTensorProtoToDenseTensorProto(const ONNX_NAMESPACE::SparseT
     void* sparse_data = sparse_data_storage.get();
     size_t element_size = 0;
     // We want to this list to match the one used below in DenseTensorToSparseTensorProto()
-    MLTypeCallDispatcherRet<Status, GetElementSize, float, int8_t, uint8_t> type_disp(type);
-    ORT_RETURN_IF_ERROR(type_disp.InvokeWithUnsupportedPolicy<UnsupportedSparseDataType>(element_size));
+    MLTypeCallDispatcher<float, int8_t, uint8_t> type_disp(type);
+    ORT_RETURN_IF_ERROR(
+        (type_disp.InvokeRetWithUnsupportedPolicy<Status, GetElementSize, UnsupportedSparseDataType>(element_size)));
 
     // by putting the data into a std::string we can avoid a copy as set_raw_data can do a std::move
     // into the TensorProto. however to actually write to the buffer we have created in the std::string we need
@@ -1076,8 +1077,9 @@ common::Status DenseTensorToSparseTensorProto(const ONNX_NAMESPACE::TensorProto&
   std::unique_ptr<uint8_t[]> dense_raw_data;
   ORT_RETURN_IF_ERROR(UnpackInitializerData(dense_proto, model_path, dense_raw_data, tensor_bytes_size));
   size_t element_size = 0;
-  MLTypeCallDispatcherRet<Status, GetElementSize, float, int8_t, uint8_t> type_disp(data_type);
-  ORT_RETURN_IF_ERROR(type_disp.InvokeWithUnsupportedPolicy<UnsupportedSparseDataType>(element_size));
+  MLTypeCallDispatcher<float, int8_t, uint8_t> type_disp(data_type);
+  ORT_RETURN_IF_ERROR(
+      (type_disp.InvokeRetWithUnsupportedPolicy<Status, GetElementSize, UnsupportedSparseDataType>(element_size)));
 
   switch (element_size) {
     case 1: {
diff --git a/onnxruntime/core/optimizer/matmul_scale_fusion.cc b/onnxruntime/core/optimizer/matmul_scale_fusion.cc
index bd577324d8..07f695659a 100644
--- a/onnxruntime/core/optimizer/matmul_scale_fusion.cc
+++ b/onnxruntime/core/optimizer/matmul_scale_fusion.cc
@@ -44,11 +44,11 @@ optional<float> GetScalarConstantInitializer(const Graph& graph, const NodeArg&
   }
 
   float scalar{};
-  utils::MLTypeCallDispatcherRet<
-      Status, ExtractScalarAsFloatDispatchTarget,
+  utils::MLTypeCallDispatcher<
       uint32_t, uint64_t, int32_t, int64_t, MLFloat16, float, double, BFloat16>
       dispatcher{initializer->data_type()};
-  ORT_THROW_IF_ERROR(dispatcher.Invoke(*initializer, graph.ModelPath(), scalar));
+  ORT_THROW_IF_ERROR(
+      (dispatcher.InvokeRet<Status, ExtractScalarAsFloatDispatchTarget>(*initializer, graph.ModelPath(), scalar)));
 
   return {scalar};
 }
diff --git a/onnxruntime/core/providers/cpu/generator/range.cc b/onnxruntime/core/providers/cpu/generator/range.cc
index 27dbbb28d6..5645b089db 100644
--- a/onnxruntime/core/providers/cpu/generator/range.cc
+++ b/onnxruntime/core/providers/cpu/generator/range.cc
@@ -95,10 +95,9 @@ struct CallRangeImpl {
 Status Range::Compute(OpKernelContext* ctx) const {
   const auto* input_tensor = ctx->Input<Tensor>(0);
   if (input_tensor == nullptr) return Status(common::ONNXRUNTIME, common::FAIL, "input count mismatch");
-  utils::MLTypeCallDispatcherRet<Status, range_internal::CallRangeImpl,
-                                 int32_t, float, int64_t, double, int16_t>
+  utils::MLTypeCallDispatcher<int32_t, float, int64_t, double, int16_t>
       t_disp(input_tensor->GetElementType());
-  return t_disp.Invoke(ctx);
+  return t_disp.InvokeRet<Status, range_internal::CallRangeImpl>(ctx);
 }
 
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cpu/math/clip.cc b/onnxruntime/core/providers/cpu/math/clip.cc
index 04656f870a..c8f7bc8bb7 100644
--- a/onnxruntime/core/providers/cpu/math/clip.cc
+++ b/onnxruntime/core/providers/cpu/math/clip.cc
@@ -74,10 +74,10 @@ Status Clip::Compute(OpKernelContext* ctx) const {
   const auto* max = ctx->Input<Tensor>(2);
   Tensor* Y = ctx->Output(0, X->Shape());
 
-  utils::MLTypeCallDispatcher<ComputeImpl, float, double, int8_t, uint8_t, int64_t, uint64_t>
+  utils::MLTypeCallDispatcher<float, double, int8_t, uint8_t, int64_t, uint64_t>
       t_disp(X->GetElementType());
 
-  t_disp.Invoke(X, min, max, Y);
+  t_disp.Invoke<ComputeImpl>(X, min, max, Y);
 
   return Status::OK();
 }
diff --git a/onnxruntime/core/providers/cpu/math/element_wise_ops.cc b/onnxruntime/core/providers/cpu/math/element_wise_ops.cc
index 781603d505..fa7fa56b0d 100644
--- a/onnxruntime/core/providers/cpu/math/element_wise_ops.cc
+++ b/onnxruntime/core/providers/cpu/math/element_wise_ops.cc
@@ -724,9 +724,9 @@ Status Min_8::Compute(OpKernelContext* context) const {
       return MinMaxMLFloat16<true>(*this, context);
       break;
     default:
-      utils::MLTypeCallDispatcherRet<Status, ComputeImpl, float, double, int32_t, uint32_t, int64_t, uint64_t>
+      utils::MLTypeCallDispatcher<float, double, int32_t, uint32_t, int64_t, uint64_t>
           t_disp(dt_type);
-      return t_disp.Invoke(*this, context);
+      return t_disp.InvokeRet<Status, ComputeImpl>(*this, context);
   }
 }
 
@@ -784,9 +784,9 @@ Status Max_8::Compute(OpKernelContext* context) const {
       return MinMaxMLFloat16<false>(*this, context);
       break;
     default:
-      utils::MLTypeCallDispatcherRet<Status, ComputeImpl, float, double, int32_t, uint32_t, int64_t, uint64_t>
+      utils::MLTypeCallDispatcher<float, double, int32_t, uint32_t, int64_t, uint64_t>
           t_disp(dt_type);
-      return t_disp.Invoke(*this, context);
+      return t_disp.InvokeRet<Status, ComputeImpl>(*this, context);
   }
 }
 
@@ -1671,10 +1671,10 @@ Status Mod::Compute(OpKernelContext* context) const {
       mod_internal::BroadCastMFloat16FMod(context);
       break;
     default:
-      utils::MLTypeCallDispatcher<mod_internal::CallModImpl, uint8_t, int8_t, uint16_t, int16_t,
-                                  uint32_t, int32_t, uint64_t, int64_t>
+      utils::MLTypeCallDispatcher<uint8_t, int8_t, uint16_t, int16_t,
+                                   uint32_t, int32_t, uint64_t, int64_t>
           t_disp(dt_type);
-      t_disp.Invoke(fmod_, context);
+      t_disp.Invoke<mod_internal::CallModImpl>(fmod_, context);
       break;
   }
 
diff --git a/onnxruntime/core/providers/cpu/math/sign.cc b/onnxruntime/core/providers/cpu/math/sign.cc
index 0a7ced3eb0..06a27cc776 100644
--- a/onnxruntime/core/providers/cpu/math/sign.cc
+++ b/onnxruntime/core/providers/cpu/math/sign.cc
@@ -115,10 +115,10 @@ Status Sign::Compute(OpKernelContext* ctx) const {
       SignMLFloat16(input, output);
       break;
     default:
-      utils::MLTypeCallDispatcher<CallSignImpl, float, double, int8_t, uint8_t,
+      utils::MLTypeCallDispatcher<float, double, int8_t, uint8_t,
                                   int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
           t_disp(dtype);
-      t_disp.Invoke(input, output);
+      t_disp.Invoke<CallSignImpl>(input, output);
       break;
   }
   return Status::OK();
diff --git a/onnxruntime/core/providers/cpu/ml/normalizer.cc b/onnxruntime/core/providers/cpu/ml/normalizer.cc
index 5652291962..72dd395ea2 100644
--- a/onnxruntime/core/providers/cpu/ml/normalizer.cc
+++ b/onnxruntime/core/providers/cpu/ml/normalizer.cc
@@ -169,10 +169,10 @@ struct Normalizer::CallNormalizerImpl {
 Status Normalizer::Compute(OpKernelContext* context) const {
   const auto& input_tensor_ptr = *context->Input<Tensor>(0);
 
-  utils::MLTypeCallDispatcherRet<Status, CallNormalizerImpl, float, double, int64_t, int32_t>
+  utils::MLTypeCallDispatcher<float, double, int64_t, int32_t>
       t_disp(input_tensor_ptr.GetElementType());
 
-  auto status = t_disp.Invoke(this, context);
+  auto status = t_disp.InvokeRet<Status, CallNormalizerImpl>(this, context);
   return status;
 }
 
diff --git a/onnxruntime/core/providers/cpu/nn/pool.cc b/onnxruntime/core/providers/cpu/nn/pool.cc
index e94b072b8c..6e8eb1b336 100644
--- a/onnxruntime/core/providers/cpu/nn/pool.cc
+++ b/onnxruntime/core/providers/cpu/nn/pool.cc
@@ -132,9 +132,9 @@ Status Pool<float, AveragePool>::Compute(OpKernelContext* context) const {
 
 
 Status MaxPoolV8::Compute(OpKernelContext* context) const {
-  utils::MLTypeCallDispatcherRet<Status, ComputeHelper, float, double, int8_t, uint8_t>
+  utils::MLTypeCallDispatcher<float, double, int8_t, uint8_t>
       t_disp(context->Input<Tensor>(0)->GetElementType());
-  return t_disp.Invoke(this, context);
+  return t_disp.InvokeRet<Status, ComputeHelper>(this, context);
 }
 
 template <typename T>
diff --git a/onnxruntime/core/providers/cpu/nn/shrink.cc b/onnxruntime/core/providers/cpu/nn/shrink.cc
index b95878d1fd..46e6e00878 100644
--- a/onnxruntime/core/providers/cpu/nn/shrink.cc
+++ b/onnxruntime/core/providers/cpu/nn/shrink.cc
@@ -74,9 +74,9 @@ Status Shrink::Compute(OpKernelContext* p_op_kernel_context) const {
   const auto* input = p_op_kernel_context->Input<Tensor>(0);
   auto* output = p_op_kernel_context->Output(0, input->Shape());
   // bool, std::string are not supported.
-  utils::MLTypeCallDispatcherRet<Status, shrink_internal::CallShrinkImpl, float, double, MLFloat16, BFloat16, int8_t, uint8_t,
-                                 int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
+  utils::MLTypeCallDispatcher<float, double, MLFloat16, BFloat16, int8_t, uint8_t,
+                              int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
       t_disp(input->GetElementType());
-  return t_disp.Invoke(input, output, bias_, lambd_);
+  return t_disp.InvokeRet<Status, shrink_internal::CallShrinkImpl>(input, output, bias_, lambd_);
 }
 }  // namespace onnxruntime
diff --git a/onnxruntime/core/providers/cuda/generator/range.cc b/onnxruntime/core/providers/cuda/generator/range.cc
index bc9c94c120..df1fda338e 100644
--- a/onnxruntime/core/providers/cuda/generator/range.cc
+++ b/onnxruntime/core/providers/cuda/generator/range.cc
@@ -97,10 +97,9 @@ Status Range::ComputeInternal(OpKernelContext* ctx) const {
     return Status(common::ONNXRUNTIME, common::FAIL, "input count mismatch");
   }
 
-  utils::MLTypeCallDispatcherRet<Status, cuda_range_internal::CallCudaRangeImpl, int32_t,
-                                 float, int64_t, double, int16_t>
+  utils::MLTypeCallDispatcher<int32_t, float, int64_t, double, int16_t>
       t_disp(input_tensor->GetElementType());
-  return t_disp.Invoke(Stream(), ctx);
+  return t_disp.InvokeRet<Status, cuda_range_internal::CallCudaRangeImpl>(Stream(), ctx);
 }
 
 }  // namespace cuda
diff --git a/onnxruntime/core/providers/cuda/math/clip.cc b/onnxruntime/core/providers/cuda/math/clip.cc
index c6d41f4cb7..6c17ecc523 100644
--- a/onnxruntime/core/providers/cuda/math/clip.cc
+++ b/onnxruntime/core/providers/cuda/math/clip.cc
@@ -121,10 +121,10 @@ Status Clip::ComputeInternal(OpKernelContext* ctx) const {
   const auto* max = ctx->Input<Tensor>(2);
   Tensor* Y = ctx->Output(0, X->Shape());
 
-  utils::MLTypeCallDispatcher<ComputeImpl, float, double, int8_t, uint8_t, int64_t, uint64_t>
+  utils::MLTypeCallDispatcher<float, double, int8_t, uint8_t, int64_t, uint64_t>
       t_disp(X->GetElementType());
 
-  t_disp.Invoke(Stream(), X, min, max, Y);
+  t_disp.Invoke<ComputeImpl>(Stream(), X, min, max, Y);
 
   return Status::OK();
 }
diff --git a/onnxruntime/core/providers/cuda/math/variadic_elementwise_ops.cc b/onnxruntime/core/providers/cuda/math/variadic_elementwise_ops.cc
index 7db97d0ed1..96223da2f9 100644
--- a/onnxruntime/core/providers/cuda/math/variadic_elementwise_ops.cc
+++ b/onnxruntime/core/providers/cuda/math/variadic_elementwise_ops.cc
@@ -142,6 +142,7 @@ Status VariadicElementwiseOp<VariadicElementwiseOpTag, SupportedElementTypes...>
   }
 
   const auto element_type = first_input_tensor.GetElementType();
+  utils::MLTypeCallDispatcher<SupportedElementTypes...> dispatcher(element_type);
 
   // special case for no broadcasting and few enough inputs
   if (input_count <= k_max_input_batch_size &&
@@ -154,17 +155,12 @@ Status VariadicElementwiseOp<VariadicElementwiseOpTag, SupportedElementTypes...>
 
     // special case for no broadcasting and 2 inputs
     if (input_count == 2) {
-      utils::MLTypeCallDispatcherRet<Status, BinaryImplDispatchTarget, SupportedElementTypes...> dispatcher(element_type);
-      ORT_RETURN_IF_ERROR(dispatcher.Invoke(Stream(), input_tensors[0], input_tensors[1], output_tensor));
-
-      return Status::OK();
+      return dispatcher.template InvokeRet<Status, BinaryImplDispatchTarget>(
+          Stream(), input_tensors[0], input_tensors[1], output_tensor);
     }
 
-    utils::MLTypeCallDispatcherRet<Status, NoBroadcastBatchImplDispatchTarget, SupportedElementTypes...> dispatcher(
-        element_type);
-    ORT_RETURN_IF_ERROR(dispatcher.Invoke(Stream(), input_tensors, output_tensor));
-
-    return Status::OK();
+    return dispatcher.template InvokeRet<Status, NoBroadcastBatchImplDispatchTarget>(
+        Stream(), input_tensors, output_tensor);
   }
 
   // compute output shape first, using broadcast rule
@@ -179,20 +175,13 @@ Status VariadicElementwiseOp<VariadicElementwiseOpTag, SupportedElementTypes...>
 
   // special case for 2 inputs
   if (input_count == 2) {
-    utils::MLTypeCallDispatcherRet<Status, BinaryImplDispatchTarget, SupportedElementTypes...> dispatcher(element_type);
-    ORT_RETURN_IF_ERROR(dispatcher.Invoke(Stream(), input_tensors[0], input_tensors[1], output_tensor));
-
-    return Status::OK();
+    return dispatcher.template InvokeRet<Status, BinaryImplDispatchTarget>(
+        Stream(), input_tensors[0], input_tensors[1], output_tensor);
   }
 
   // general case for more than 2 inputs
-  {
-    utils::MLTypeCallDispatcherRet<Status, GeneralImplDispatchTarget, SupportedElementTypes...> dispatcher(
-        element_type);
-    ORT_RETURN_IF_ERROR(dispatcher.Invoke(Stream(), input_tensors, output_tensor));
-  }
-
-  return Status::OK();
+  return dispatcher.template InvokeRet<Status, GeneralImplDispatchTarget>(
+      Stream(), input_tensors, output_tensor);
 }
 
 namespace {
diff --git a/onnxruntime/core/providers/cuda/nn/dropout.h b/onnxruntime/core/providers/cuda/nn/dropout.h
index 5e38a587df..bea0bfbb42 100644
--- a/onnxruntime/core/providers/cuda/nn/dropout.h
+++ b/onnxruntime/core/providers/cuda/nn/dropout.h
@@ -72,8 +72,8 @@ Status Dropout::ComputeInternal(OpKernelContext* context) const {
   float ratio_data = default_ratio_;
   auto ratio = context->Input<Tensor>(1);
   if (ratio) {
-    utils::MLTypeCallDispatcher<GetRatioDataImpl, float, MLFloat16, double> t_disp(ratio->GetElementType());
-    t_disp.Invoke(ratio, ratio_data);
+    utils::MLTypeCallDispatcher<float, MLFloat16, double> t_disp(ratio->GetElementType());
+    t_disp.Invoke<GetRatioDataImpl>(ratio, ratio_data);
   }
 
   const Tensor* training_mode = context->Input<Tensor>(2);
@@ -102,12 +102,16 @@ Status Dropout::ComputeInternal(OpKernelContext* context) const {
 
   PhiloxGenerator& generator = generator_ ? *generator_ : PhiloxGenerator::Default();
 
+  using SupportedTypes = onnxruntime::TypeList<
 #if defined(CUDA_VERSION) && CUDA_VERSION >= 11000
-  utils::MLTypeCallDispatcher<DropoutComputeImpl, float, MLFloat16, double, BFloat16> t_disp(X->GetElementType());
+      float, MLFloat16, double, BFloat16
 #else
-  utils::MLTypeCallDispatcher<DropoutComputeImpl, float, MLFloat16, double> t_disp(X->GetElementType());
+      float, MLFloat16, double
 #endif
-  t_disp.Invoke(GetDeviceProp(), Stream(), N, ratio_data, generator, *X, *Y, mask_data);
+      >;
+
+  utils::MLTypeCallDispatcherFromTypeList<SupportedTypes> t_disp(X->GetElementType());
+  t_disp.Invoke<DropoutComputeImpl>(GetDeviceProp(), Stream(), N, ratio_data, generator, *X, *Y, mask_data);
 
   return Status::OK();
 }
diff --git a/onnxruntime/core/providers/cuda/tensor/gather_nd.cc b/onnxruntime/core/providers/cuda/tensor/gather_nd.cc
index fd452d37e3..03b65164ff 100644
--- a/onnxruntime/core/providers/cuda/tensor/gather_nd.cc
+++ b/onnxruntime/core/providers/cuda/tensor/gather_nd.cc
@@ -203,9 +203,9 @@ Status GatherND<TIndex>::ComputeInternal(OpKernelContext* context) const {
 
   const void* const kernel_input_data = input_tensor->DataRaw();
   void* const kernel_output_data = output_tensor->MutableDataRaw();
-  utils::MLTypeCallDispatcher<GatherNDComputeImpl, GATHER_ND_T_DATA_TYPES>
-      t_disp(input_tensor->GetElementType());
-  t_disp.Invoke(Stream(), num_slices, slice_size, kernel_input_data, kernel_output_data, input_slice_offsets_buffer.get());
+  utils::MLTypeCallDispatcher<GATHER_ND_T_DATA_TYPES> t_disp(input_tensor->GetElementType());
+  t_disp.Invoke<GatherNDComputeImpl>(
+      Stream(), num_slices, slice_size, kernel_input_data, kernel_output_data, input_slice_offsets_buffer.get());
 
   return Status::OK();
 }
diff --git a/onnxruntime/core/providers/cuda/tensor/scatter_elements.cc b/onnxruntime/core/providers/cuda/tensor/scatter_elements.cc
index ab04fb150a..1536efe168 100755
--- a/onnxruntime/core/providers/cuda/tensor/scatter_elements.cc
+++ b/onnxruntime/core/providers/cuda/tensor/scatter_elements.cc
@@ -163,12 +163,13 @@ Status ScatterElements::ComputeInternal(OpKernelContext* context) const {
     fdm_indices_strides[i] = fast_divmod(static_cast<int>(indices_strides[i]));
   }
 
-  utils::MLTypeCallDispatcherRet<Status, ComputeImpl, float, MLFloat16, int16_t, int8_t, int32_t,
-                                 int64_t, uint8_t, uint16_t, uint32_t, uint64_t, double, bool>
+  utils::MLTypeCallDispatcher<float, MLFloat16, int16_t, int8_t, int32_t,
+                              int64_t, uint8_t, uint16_t, uint32_t, uint64_t, double, bool>
       t_disp(data_tensor->GetElementType());
-  return t_disp.Invoke(Stream(), data_tensor, updates_tensor, indices_tensor, output_tensor, rank,
-                       input_data_size, buffer_input_dims, buffer_input_strides, indices_size,
-                       buffer_indices_dims, fdm_indices_strides, axis);
+  return t_disp.InvokeRet<Status, ComputeImpl>(
+      Stream(), data_tensor, updates_tensor, indices_tensor, output_tensor, rank,
+      input_data_size, buffer_input_dims, buffer_input_strides, indices_size,
+      buffer_indices_dims, fdm_indices_strides, axis);
 }
 
 }  // namespace cuda
diff --git a/onnxruntime/core/session/onnxruntime_c_api.cc b/onnxruntime/core/session/onnxruntime_c_api.cc
index 3fa6b01567..be59468022 100644
--- a/onnxruntime/core/session/onnxruntime_c_api.cc
+++ b/onnxruntime/core/session/onnxruntime_c_api.cc
@@ -1283,10 +1283,11 @@ ORT_STATUS_PTR OrtGetValueImplSeqOfTensors(_In_ const OrtValue* p_ml_value, int
   auto& one_tensor = data.Get(index);
 
   using namespace c_api_internal;
-  utils::MLTypeCallDispatcherRet<OrtStatusPtr, CallGetValueImpl, float, double, MLFloat16, BFloat16, bool, std::string,
-                                 int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
+  utils::MLTypeCallDispatcher<float, double, MLFloat16, BFloat16, bool, std::string,
+                              int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
       t_disp(one_tensor.GetElementType());
-  return t_disp.template InvokeWithUnsupportedPolicy<UnsupportedReturnFailStatus>(allocator, one_tensor, out);
+  return t_disp.template InvokeRetWithUnsupportedPolicy<OrtStatusPtr, CallGetValueImpl, UnsupportedReturnFailStatus>(
+      allocator, one_tensor, out);
 }
 
 #ifdef _MSC_VER
@@ -1489,11 +1490,12 @@ static ORT_STATUS_PTR OrtCreateValueImplSeqHelper(const OrtValue* const* in, siz
     }
 
     OrtStatus* st{};
-    utils::MLTypeCallDispatcherRet<OrtStatus*, CallCreateValueImpl, bool, float, double, std::string,
-                                   MLFloat16, BFloat16, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
+    utils::MLTypeCallDispatcher<bool, float, double, std::string,
+                                MLFloat16, BFloat16, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t>
         t_disp(one_tensor.GetElementType());
 
-    st = t_disp.InvokeWithUnsupportedPolicy<UnsupportedReturnFailStatus>(one_tensor, tensors[idx]);
+    st = t_disp.InvokeRetWithUnsupportedPolicy<OrtStatus*, CallCreateValueImpl, UnsupportedReturnFailStatus>(
+        one_tensor, tensors[idx]);
 
     if (st) {
       return st;
diff --git a/onnxruntime/featurizers_ops/cpu/cat_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/cat_imputer_transformer.cc
index f57c6c5ae7..890ca95343 100644
--- a/onnxruntime/featurizers_ops/cpu/cat_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/cat_imputer_transformer.cc
@@ -63,8 +63,8 @@ class CatImputerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<CatImputerTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<CatImputerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/forecasting_pivot_transformer.cc b/onnxruntime/featurizers_ops/cpu/forecasting_pivot_transformer.cc
index 03daa40158..e1bed0fc2c 100644
--- a/onnxruntime/featurizers_ops/cpu/forecasting_pivot_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/forecasting_pivot_transformer.cc
@@ -155,10 +155,10 @@ struct ForecastingPivotTransformerImpl {
 
       const auto elem_type = input_tensor->GetElementType();
 
-      utils::MLTypeCallDispatcher<CopyImputedColumnsImpl,
-                                  int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
-                                  float, double, bool, std::string> t_disp(elem_type);
-      t_disp.Invoke(input_tensor, output_tensor_imputed, row_idx_record, input_matrix_size, num_output_rows);
+      utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
+                                  float, double, bool, std::string>
+          t_disp(elem_type);
+      t_disp.Invoke<CopyImputedColumnsImpl>(input_tensor, output_tensor_imputed, row_idx_record, input_matrix_size, num_output_rows);
     }
 
     // Prepare the horizon Output(uint32)
@@ -177,9 +177,8 @@ class ForecastingPivotTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<ForecastingPivotTransformerImpl, float, double>
-        t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx, _num_pivot_columns);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<ForecastingPivotTransformerImpl>(ctx, _num_pivot_columns);
 
     return Status::OK();
   }
diff --git a/onnxruntime/featurizers_ops/cpu/from_string_transformer.cc b/onnxruntime/featurizers_ops/cpu/from_string_transformer.cc
index 0a237b8409..364c1afc76 100644
--- a/onnxruntime/featurizers_ops/cpu/from_string_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/from_string_transformer.cc
@@ -65,10 +65,10 @@ class FromStringTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<FromStringTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double, bool, std::string>
         t_disp(result_type_);
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<FromStringTransformerImpl>(ctx);
     return Status::OK();
   }
 
diff --git a/onnxruntime/featurizers_ops/cpu/hash_one_hot_vectorizer_transformer.cc b/onnxruntime/featurizers_ops/cpu/hash_one_hot_vectorizer_transformer.cc
index 44f880728c..fe99290911 100644
--- a/onnxruntime/featurizers_ops/cpu/hash_one_hot_vectorizer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/hash_one_hot_vectorizer_transformer.cc
@@ -57,10 +57,10 @@ class HashOneHotVectorizerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<HashOneHotVectorizerTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t,
                                 uint32_t, int64_t, uint64_t, float, double, bool, std::string>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<HashOneHotVectorizerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/imputation_marker_transformer.cc b/onnxruntime/featurizers_ops/cpu/imputation_marker_transformer.cc
index 478dbf20f5..365ad304ac 100644
--- a/onnxruntime/featurizers_ops/cpu/imputation_marker_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/imputation_marker_transformer.cc
@@ -54,8 +54,8 @@ class ImputationMarkerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<ImputationMarkerTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<ImputationMarkerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/label_encoder_transformer.cc b/onnxruntime/featurizers_ops/cpu/label_encoder_transformer.cc
index a6843f0780..011c116e3f 100644
--- a/onnxruntime/featurizers_ops/cpu/label_encoder_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/label_encoder_transformer.cc
@@ -48,10 +48,10 @@ class LabelEncoderTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<LabelEncoderTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double, bool, std::string>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<LabelEncoderTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/laglead_operator_transformer.cc b/onnxruntime/featurizers_ops/cpu/laglead_operator_transformer.cc
index 703c6068e7..5c1689a203 100644
--- a/onnxruntime/featurizers_ops/cpu/laglead_operator_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/laglead_operator_transformer.cc
@@ -93,9 +93,8 @@ class LagLeadOperatorTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<LagLeadOperatorTransformerImpl, float, double>
-        t_disp(ctx->Input<Tensor>(2)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(2)->GetElementType());
+    t_disp.Invoke<LagLeadOperatorTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/max_abs_scaler_transformer.cc b/onnxruntime/featurizers_ops/cpu/max_abs_scaler_transformer.cc
index 8cb9ddd781..4216faeb96 100644
--- a/onnxruntime/featurizers_ops/cpu/max_abs_scaler_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/max_abs_scaler_transformer.cc
@@ -71,10 +71,10 @@ class MaxAbsScalerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MaxAbsScalerTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<MaxAbsScalerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/mean_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/mean_imputer_transformer.cc
index 073f21608d..45e7b9f31d 100644
--- a/onnxruntime/featurizers_ops/cpu/mean_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/mean_imputer_transformer.cc
@@ -63,8 +63,8 @@ class MeanImputerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MeanImputerTransformerImpl, float, double> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<MeanImputerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/median_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/median_imputer_transformer.cc
index bcd5602dfb..772989155c 100644
--- a/onnxruntime/featurizers_ops/cpu/median_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/median_imputer_transformer.cc
@@ -77,8 +77,8 @@ class MedianImputerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MedianImputerTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<MedianImputerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/min_max_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/min_max_imputer_transformer.cc
index 94a6e399ce..8086ff9819 100644
--- a/onnxruntime/featurizers_ops/cpu/min_max_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/min_max_imputer_transformer.cc
@@ -72,8 +72,8 @@ class MinMaxImputerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MinMaxImputerTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<MinMaxImputerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/min_max_scaler_transformer.cc b/onnxruntime/featurizers_ops/cpu/min_max_scaler_transformer.cc
index 25d1e6c74f..c6e677bf2e 100644
--- a/onnxruntime/featurizers_ops/cpu/min_max_scaler_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/min_max_scaler_transformer.cc
@@ -48,10 +48,10 @@ class MinMaxScalerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MinMaxScalerTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<MinMaxScalerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/missing_dummies_transformer.cc b/onnxruntime/featurizers_ops/cpu/missing_dummies_transformer.cc
index 465b58e9f0..ba4c9a7e63 100644
--- a/onnxruntime/featurizers_ops/cpu/missing_dummies_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/missing_dummies_transformer.cc
@@ -54,8 +54,8 @@ class MissingDummiesTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<MissingDummiesTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<MissingDummiesTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/mode_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/mode_imputer_transformer.cc
index 6b8e39ca63..5e4dfbb09a 100644
--- a/onnxruntime/featurizers_ops/cpu/mode_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/mode_imputer_transformer.cc
@@ -72,8 +72,8 @@ class ModeImputerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<ModeImputerTransformerImpl, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<ModeImputerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/normalize_transformer.cc b/onnxruntime/featurizers_ops/cpu/normalize_transformer.cc
index 41c809aec7..a9160eb678 100644
--- a/onnxruntime/featurizers_ops/cpu/normalize_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/normalize_transformer.cc
@@ -72,10 +72,10 @@ class NormalizeTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<NormalizeTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<NormalizeTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/numericalize_transformer.cc b/onnxruntime/featurizers_ops/cpu/numericalize_transformer.cc
index 710654fbe0..7af8dac3cd 100644
--- a/onnxruntime/featurizers_ops/cpu/numericalize_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/numericalize_transformer.cc
@@ -48,9 +48,10 @@ class NumericalizeTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<NumericalizeTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
-        int64_t, uint64_t, float, double, std::string> t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+                                int64_t, uint64_t, float, double, std::string>
+        t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<NumericalizeTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/one_hot_encoder_transformer.cc b/onnxruntime/featurizers_ops/cpu/one_hot_encoder_transformer.cc
index 54c94622b0..13d3b6a67c 100644
--- a/onnxruntime/featurizers_ops/cpu/one_hot_encoder_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/one_hot_encoder_transformer.cc
@@ -57,10 +57,10 @@ class OneHotEncoderTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<OneHotEncoderTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double, bool, std::string>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<OneHotEncoderTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/pca_transformer.cc b/onnxruntime/featurizers_ops/cpu/pca_transformer.cc
index 2dd644fa35..8eed4f3c51 100644
--- a/onnxruntime/featurizers_ops/cpu/pca_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/pca_transformer.cc
@@ -65,9 +65,8 @@ class PCATransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<PCATransformerImpl, float, double>
-        t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<PCATransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/robust_scaler_transformer.cc b/onnxruntime/featurizers_ops/cpu/robust_scaler_transformer.cc
index 4d67b86c3e..e70f58e143 100644
--- a/onnxruntime/featurizers_ops/cpu/robust_scaler_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/robust_scaler_transformer.cc
@@ -71,10 +71,10 @@ class RobustScalerTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<RobustScalerTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<RobustScalerTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/rolling_window_transformer.cc b/onnxruntime/featurizers_ops/cpu/rolling_window_transformer.cc
index 46bfc9960b..4596edeb96 100644
--- a/onnxruntime/featurizers_ops/cpu/rolling_window_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/rolling_window_transformer.cc
@@ -91,9 +91,8 @@ class AnalyticalRollingWindowTransformer final : public OpKernel {
 }
 
   Status Compute(OpKernelContext* ctx) const override {
-
-    utils::MLTypeCallDispatcher<AnalyticalRollingWindowTransformerImpl, float, double> t_disp(ctx->Input<Tensor>(2)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(2)->GetElementType());
+    t_disp.Invoke<AnalyticalRollingWindowTransformerImpl>(ctx);
     return Status::OK();
   }
 };
@@ -104,9 +103,8 @@ class SimpleRollingWindowTransformer final : public OpKernel {
 }
 
   Status Compute(OpKernelContext* ctx) const override {
-
-    utils::MLTypeCallDispatcher<SimpleRollingWindowTransformerImpl, float, double> t_disp(ctx->Input<Tensor>(2)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(2)->GetElementType());
+    t_disp.Invoke<SimpleRollingWindowTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/short_grain_dropper_transformer.cc b/onnxruntime/featurizers_ops/cpu/short_grain_dropper_transformer.cc
index 6a2eb233d3..d543c0cb61 100644
--- a/onnxruntime/featurizers_ops/cpu/short_grain_dropper_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/short_grain_dropper_transformer.cc
@@ -89,10 +89,10 @@ void ShortGrainDropperTransformerImpl(OpKernelContext* ctx) {
 
     const auto elem_type = variadic_input_tensor->GetElementType();
 
-    utils::MLTypeCallDispatcher<CopyNonDroppedColumnsImpl,
-                                int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
-                                float, double, bool, std::string> t_disp(elem_type);
-    t_disp.Invoke(variadic_input_tensor, output_after_drop_tensor, rows_to_drop, input_row_size);
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
+                                float, double, bool, std::string>
+        t_disp(elem_type);
+    t_disp.Invoke<CopyNonDroppedColumnsImpl>(variadic_input_tensor, output_after_drop_tensor, rows_to_drop, input_row_size);
   }
 };
 
diff --git a/onnxruntime/featurizers_ops/cpu/standard_scale_wrapper_transformer.cc b/onnxruntime/featurizers_ops/cpu/standard_scale_wrapper_transformer.cc
index ed0213d1e0..aae0de0e67 100644
--- a/onnxruntime/featurizers_ops/cpu/standard_scale_wrapper_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/standard_scale_wrapper_transformer.cc
@@ -48,10 +48,10 @@ class StandardScaleWrapperTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<StandardScaleWrapperTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t,
                                 uint32_t, int64_t, uint64_t, float, double>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<StandardScaleWrapperTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/string_transformer.cc b/onnxruntime/featurizers_ops/cpu/string_transformer.cc
index ac9342e0f7..7bd93a823a 100644
--- a/onnxruntime/featurizers_ops/cpu/string_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/string_transformer.cc
@@ -48,10 +48,10 @@ class StringTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<StringTransformerImpl, int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
+    utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t,
                                 int64_t, uint64_t, float, double, bool, std::string>
         t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    t_disp.Invoke<StringTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/onnxruntime/featurizers_ops/cpu/time_series_imputer_transformer.cc b/onnxruntime/featurizers_ops/cpu/time_series_imputer_transformer.cc
index 3b738336d1..2f9385b199 100644
--- a/onnxruntime/featurizers_ops/cpu/time_series_imputer_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/time_series_imputer_transformer.cc
@@ -265,10 +265,10 @@ struct TimeSeriesImputerTransformerImpl {
 
       const auto elem_type = variadic_input_tensor->GetElementType();
 
-      utils::MLTypeCallDispatcher<GenerateImputedColumnsImpl,
-                                  int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
-                                  float, double, bool, std::string> t_disp(elem_type);
-      t_disp.Invoke(variadic_input_tensor, output_after_impute_tensor, is_row_imputed, input_row_size);
+      utils::MLTypeCallDispatcher<int8_t, uint8_t, int16_t, uint16_t, int32_t, uint32_t, int64_t, uint64_t,
+                                  float, double, bool, std::string>
+          t_disp(elem_type);
+      t_disp.Invoke<GenerateImputedColumnsImpl>(variadic_input_tensor, output_after_impute_tensor, is_row_imputed, input_row_size);
   }
 
   }
diff --git a/onnxruntime/featurizers_ops/cpu/truncated_svd_transformer.cc b/onnxruntime/featurizers_ops/cpu/truncated_svd_transformer.cc
index f707bbdf91..d0072c9405 100644
--- a/onnxruntime/featurizers_ops/cpu/truncated_svd_transformer.cc
+++ b/onnxruntime/featurizers_ops/cpu/truncated_svd_transformer.cc
@@ -65,9 +65,8 @@ class TruncatedSVDTransformer final : public OpKernel {
   }
 
   Status Compute(OpKernelContext* ctx) const override {
-    utils::MLTypeCallDispatcher<TruncatedSVDTransformerImpl, float, double>
-        t_disp(ctx->Input<Tensor>(1)->GetElementType());
-    t_disp.Invoke(ctx);
+    utils::MLTypeCallDispatcher<float, double> t_disp(ctx->Input<Tensor>(1)->GetElementType());
+    t_disp.Invoke<TruncatedSVDTransformerImpl>(ctx);
     return Status::OK();
   }
 };
diff --git a/orttraining/orttraining/training_ops/cpu/tensor/gather_nd_grad.cc b/orttraining/orttraining/training_ops/cpu/tensor/gather_nd_grad.cc
index a186287bfa..f00e7a7551 100644
--- a/orttraining/orttraining/training_ops/cpu/tensor/gather_nd_grad.cc
+++ b/orttraining/orttraining/training_ops/cpu/tensor/gather_nd_grad.cc
@@ -74,8 +74,8 @@ Status GatherNDGrad::Compute(OpKernelContext* context) const {
   }
 
   ORT_RETURN_IF_NOT(nullptr == p.input_str_base, "nullptr != p.input_str_base");
-  utils::MLTypeCallDispatcher<GatherNDGradComputeImpl, float, double> t_disp(update_tensor->GetElementType());
-  t_disp.Invoke(p, update_tensor);
+  utils::MLTypeCallDispatcher<float, double> t_disp(update_tensor->GetElementType());
+  t_disp.Invoke<GatherNDGradComputeImpl>(p, update_tensor);
 
   return Status::OK();
 }
diff --git a/orttraining/orttraining/training_ops/cuda/activation/bias_gelu_grad.cc b/orttraining/orttraining/training_ops/cuda/activation/bias_gelu_grad.cc
index b948c5c77c..703623f16b 100644
--- a/orttraining/orttraining/training_ops/cuda/activation/bias_gelu_grad.cc
+++ b/orttraining/orttraining/training_ops/cuda/activation/bias_gelu_grad.cc
@@ -76,11 +76,8 @@ Status BiasGeluGrad_dX<GeluComputationMode>::ComputeInternal(OpKernelContext* co
 
   const auto input_size = input_shape.Size(), bias_size = bias_shape.Size();
 
-  utils::MLTypeCallDispatcher<
-      KernelLaunchDispatcher,
-      ALL_IEEE_FLOAT_DATA_TYPES>
-      dispatcher{X->GetElementType()};
-  dispatcher.Invoke(Stream(), input_size, bias_size, *dY, *X, *B, *dX);
+  utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> dispatcher{X->GetElementType()};
+  dispatcher.Invoke<KernelLaunchDispatcher>(Stream(), input_size, bias_size, *dY, *X, *B, *dX);
 
   return Status::OK();
 }
diff --git a/orttraining/orttraining/training_ops/cuda/math/scale.cc b/orttraining/orttraining/training_ops/cuda/math/scale.cc
index 7fe37d00f7..2345213d6b 100644
--- a/orttraining/orttraining/training_ops/cuda/math/scale.cc
+++ b/orttraining/orttraining/training_ops/cuda/math/scale.cc
@@ -37,8 +37,8 @@ Status Scale<T>::ComputeInternal(OpKernelContext* context) const {
   typedef typename ToCudaType<T>::MappedType CudaT;
   float scale_value;
   auto scale_tensor = context->Input<Tensor>(1);
-  utils::MLTypeCallDispatcher<GetScaleValueImpl, float, double, MLFloat16, int64_t, int32_t> t_disp(scale_tensor->GetElementType());
-  t_disp.Invoke(scale_tensor, scale_value);
+  utils::MLTypeCallDispatcher<float, double, MLFloat16, int64_t, int32_t> t_disp(scale_tensor->GetElementType());
+  t_disp.Invoke<GetScaleValueImpl>(scale_tensor, scale_value);
 
   if (scale_down_) {
     scale_value = 1.0f / scale_value;
diff --git a/orttraining/orttraining/training_ops/cuda/nn/dropout.cc b/orttraining/orttraining/training_ops/cuda/nn/dropout.cc
index b67eb70426..8d970bdd90 100644
--- a/orttraining/orttraining/training_ops/cuda/nn/dropout.cc
+++ b/orttraining/orttraining/training_ops/cuda/nn/dropout.cc
@@ -73,14 +73,14 @@ Status DropoutGrad::ComputeInternal(OpKernelContext* context) const {
   float ratio_data = default_ratio_;
   auto ratio = context->Input<Tensor>(2);
   if (ratio) {
-    utils::MLTypeCallDispatcher<GetRatioDataImpl, ALL_IEEE_FLOAT_DATA_TYPES> t_disp(ratio->GetElementType());
-    t_disp.Invoke(ratio, ratio_data);
+    utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(ratio->GetElementType());
+    t_disp.Invoke<GetRatioDataImpl>(ratio, ratio_data);
   }
 
   auto dX = context->Output(0, shape);
 
-  utils::MLTypeCallDispatcher<DropoutGradComputeImpl, ALL_IEEE_FLOAT_DATA_TYPES> t_disp(dY->GetElementType());
-  t_disp.Invoke(Stream(), N, *dY, mask_data, ratio_data, *dX);
+  utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(dY->GetElementType());
+  t_disp.Invoke<DropoutGradComputeImpl>(Stream(), N, *dY, mask_data, ratio_data, *dX);
 
   return Status::OK();
 }
@@ -165,8 +165,8 @@ Status BiasDropout::ComputeInternal(OpKernelContext* context) const {
   float ratio_data = default_ratio_;
   auto ratio = context->Input<Tensor>(3);
   if (ratio) {
-    utils::MLTypeCallDispatcher<GetRatioDataImpl, ALL_IEEE_FLOAT_DATA_TYPES> t_disp(ratio->GetElementType());
-    t_disp.Invoke(ratio, ratio_data);
+    utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(ratio->GetElementType());
+    t_disp.Invoke<GetRatioDataImpl>(ratio, ratio_data);
   }
 
   //Check for inference mode.
@@ -186,8 +186,9 @@ Status BiasDropout::ComputeInternal(OpKernelContext* context) const {
   const fast_divmod fdm_dim(gsl::narrow_cast<int>(dim));
   PhiloxGenerator& generator = generator_ ? *generator_ : PhiloxGenerator::Default();
 
-  utils::MLTypeCallDispatcherRet<Status, BiasDropoutComputeImpl, ALL_IEEE_FLOAT_DATA_TYPES> t_disp(X->GetElementType());
-  return t_disp.Invoke(GetDeviceProp(), Stream(), N, fdm_dim, ratio_data, generator, *X, *bias, residual, *Y, mask_data);
+  utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(X->GetElementType());
+  return t_disp.InvokeRet<Status, BiasDropoutComputeImpl>(
+      GetDeviceProp(), Stream(), N, fdm_dim, ratio_data, generator, *X, *bias, residual, *Y, mask_data);
 }
 
 }  // namespace cuda
diff --git a/orttraining/orttraining/training_ops/cuda/tensor/gather_elements_grad.cc b/orttraining/orttraining/training_ops/cuda/tensor/gather_elements_grad.cc
index 983960a2df..dae18fc35f 100644
--- a/orttraining/orttraining/training_ops/cuda/tensor/gather_elements_grad.cc
+++ b/orttraining/orttraining/training_ops/cuda/tensor/gather_elements_grad.cc
@@ -129,11 +129,11 @@ Status GatherElementsGrad::ComputeInternal(OpKernelContext* context) const {
     fdm_indices_strides[i] = fast_divmod(static_cast<int>(indices_strides[i]));
   }
 
-  utils::MLTypeCallDispatcherRet<Status, ComputeImpl, MLFloat16, float, double>
-      t_disp(dY->GetElementType());
-  return t_disp.Invoke(Stream(), dY, indices_tensor, dX, rank,
-                       buffer_output_dims, buffer_input_strides, indices_size,
-                       buffer_indices_dims, fdm_indices_strides, axis);
+  utils::MLTypeCallDispatcher<MLFloat16, float, double> t_disp(dY->GetElementType());
+  return t_disp.InvokeRet<Status, ComputeImpl>(
+      Stream(), dY, indices_tensor, dX, rank,
+      buffer_output_dims, buffer_input_strides, indices_size,
+      buffer_indices_dims, fdm_indices_strides, axis);
 }
 
 }  // namespace cuda
diff --git a/orttraining/orttraining/training_ops/cuda/tensor/gather_nd_grad.cc b/orttraining/orttraining/training_ops/cuda/tensor/gather_nd_grad.cc
index 7e7acd2974..6bfc0eb68f 100644
--- a/orttraining/orttraining/training_ops/cuda/tensor/gather_nd_grad.cc
+++ b/orttraining/orttraining/training_ops/cuda/tensor/gather_nd_grad.cc
@@ -96,9 +96,9 @@ Status GatherNDGrad<TIndex>::ComputeInternal(OpKernelContext* context) const {
 
   const void* const kernel_input_data = update_tensor->DataRaw();
   void* const kernel_output_data = output_tensor->MutableDataRaw();
-  utils::MLTypeCallDispatcher<GatherNDGradComputeImpl, ALL_IEEE_FLOAT_DATA_TYPES>
-      t_disp(update_tensor->GetElementType());
-  t_disp.Invoke(Stream(), num_slices, slice_size, kernel_input_data, kernel_output_data, input_slice_offsets_buffer.get());
+  utils::MLTypeCallDispatcher<ALL_IEEE_FLOAT_DATA_TYPES> t_disp(update_tensor->GetElementType());
+  t_disp.Invoke<GatherNDGradComputeImpl>(
+      Stream(), num_slices, slice_size, kernel_input_data, kernel_output_data, input_slice_offsets_buffer.get());
 
   return Status::OK();
 }