Limit output types to f32/f16, add const_cast

Author: mdvoretc-intel

onnxruntime/core/providers/openvino/qdq_transformations/qdq_scales_fix.cpp

@@ -466,34 +466,13 @@ struct CustomGraph {
         if (prev.node_ptr->OutputDefs()[0]->Type() != dq_node_ref.OutputDefs()[0]->Type()) {
           NodeArg& output = original_graph.GetOrCreateNodeArg(prev.node_name + "_cast_0", dq_node_ref.OutputDefs()[0]->TypeAsProto());
           std::string cast_node_name = prev.node_ptr->OutputDefs()[0]->Name() + "_cast";
-          InlinedVector<NodeArg*> input_args = {(NodeArg*)(prev.node_ptr->OutputDefs()[0])};
+          InlinedVector<NodeArg*> input_args = {const_cast<NodeArg*>(prev.node_ptr->OutputDefs()[0])};
           InlinedVector<NodeArg*> output_args = {&output};
-          std::unordered_map<std::string, int> type_str_to_tensor_data_type_;
-          type_str_to_tensor_data_type_["tensor(float)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT;
-          type_str_to_tensor_data_type_["tensor(float16)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT16;
-          type_str_to_tensor_data_type_["tensor(bfloat16)"] = ONNX_NAMESPACE::TensorProto_DataType_BFLOAT16;
-          type_str_to_tensor_data_type_["tensor(double)"] = ONNX_NAMESPACE::TensorProto_DataType_DOUBLE;
-          type_str_to_tensor_data_type_["tensor(int8)"] = ONNX_NAMESPACE::TensorProto_DataType_INT8;
-          type_str_to_tensor_data_type_["tensor(int16)"] = ONNX_NAMESPACE::TensorProto_DataType_INT16;
-          type_str_to_tensor_data_type_["tensor(int32)"] = ONNX_NAMESPACE::TensorProto_DataType_INT32;
-          type_str_to_tensor_data_type_["tensor(int64)"] = ONNX_NAMESPACE::TensorProto_DataType_INT64;
-          type_str_to_tensor_data_type_["tensor(uint8)"] = ONNX_NAMESPACE::TensorProto_DataType_UINT8;
-          type_str_to_tensor_data_type_["tensor(uint16)"] = ONNX_NAMESPACE::TensorProto_DataType_UINT16;
-          type_str_to_tensor_data_type_["tensor(uint32)"] = ONNX_NAMESPACE::TensorProto_DataType_UINT32;
-          type_str_to_tensor_data_type_["tensor(uint64)"] = ONNX_NAMESPACE::TensorProto_DataType_UINT64;
-          type_str_to_tensor_data_type_["tensor(complex64)"] = ONNX_NAMESPACE::TensorProto_DataType_COMPLEX64;
-          type_str_to_tensor_data_type_["tensor(complex128)"] = ONNX_NAMESPACE::TensorProto_DataType_COMPLEX128;
-          type_str_to_tensor_data_type_["tensor(string)"] = ONNX_NAMESPACE::TensorProto_DataType_STRING;
-          type_str_to_tensor_data_type_["tensor(bool)"] = ONNX_NAMESPACE::TensorProto_DataType_BOOL;
-          type_str_to_tensor_data_type_["tensor(float8e4m3fn)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT8E4M3FN;
-          type_str_to_tensor_data_type_["tensor(float8e4m3fnuz)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT8E4M3FNUZ;
-          type_str_to_tensor_data_type_["tensor(float8e5m2)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT8E5M2;
-          type_str_to_tensor_data_type_["tensor(float8e5m2fnuz)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT8E5M2FNUZ;
-          type_str_to_tensor_data_type_["tensor(uint4)"] = ONNX_NAMESPACE::TensorProto_DataType_UINT4;
-          type_str_to_tensor_data_type_["tensor(int4)"] = ONNX_NAMESPACE::TensorProto_DataType_INT4;
-          type_str_to_tensor_data_type_["tensor(float4e2m1)"] = ONNX_NAMESPACE::TensorProto_DataType_FLOAT4E2M1;
           Node& cast_node = original_graph.AddNode(cast_node_name, "Cast", "", input_args, output_args, nullptr, "");
-          auto type_cast = type_str_to_tensor_data_type_[*dq_node_ref.OutputDefs()[0]->Type()];
+          auto type_str = dq_node_ref.OutputDefs()[0]->Type();
+          auto type_cast = type_str->find("tensor(float)") != std::string::npos ? onnx::TensorProto_DataType_FLOAT : onnx::TensorProto_DataType_FLOAT16;
+          ORT_ENFORCE((type_cast == onnx::TensorProto_DataType_FLOAT) || (type_str->find("tensor(float16)") != std::string::npos),
+              "QDQ type misalignment, expected float32 or float16 output");
           cast_node.AddAttribute("to", static_cast<int64_t>(type_cast));
           original_graph.AddEdge(prev.node_ptr->Index(),
                                  cast_node.Index(),