return 0;
    } else if (to_inference_type == QUANTIZED_INT8 ||
               to_inference_type == QUANTIZED_UINT8) {
      // QUANT path.
      return kQuantCost * total_element_count_transferred;
    }
  }

  if (from_inference_type == QUANTIZED_INT8 ||
      from_inference_type == QUANTIZED_UINT8) {
    // Dequant path.

  %0 = "tfl.pseudo_qconst"() {qtype = tensor<128x!quant.uniform<i32:f32, 0.7>>, value = dense<0> : tensor<128xi32>} : () -> tensor<128x!quant.uniform<i32:f32, 0.7>>

    %0 = tfl.add %arg0, %arg1 {fused_activation_function = "RELU6", tac.device = "CPU", tac.inference_type = "QUANTIZED_INT8"} : tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>
    %1 = tfl.mul %0, %arg2 {fused_activation_function = "RELU6", tac.device = "CPU", tac.inference_type = "QUANTIZED_INT8"} : tensor<100x!quant.uniform<i8:f32, 2.000000e-01:-3>>

#include "tensorflow/lite/c/c_api_types.h"
#include "tensorflow/lite/model.h"

namespace mlir {
namespace lite {

// Supported resulting types from quantization process.
enum class BufferType { QUANTIZED_INT8, QUANTIZED_FLOAT16 };

// Stores information about how to quantize a user-specified custom operation.
// CustomOpInfo contains info of its corresponding CustomOp registered in the

  %2 = "tfl.mul"(%arg0, %0) {tac.device = "CPU", tac.inference_type = "QUANTIZED_INT8", fused_activation_function = "NONE"} : (tensor<1x!quant.uniform<i8:f32, 0.003:-128>>, tensor<1x!quant.uniform<i8:f32, 0.003:-128>>) -> tensor<1x!quant.uniform<i8:f32, 0.003:-128>>

  if (float_type_observed) {
    if (int8_type_observed || uint8_type_observed) {
      return HYBRID;
    } else {
      return FLOAT;
    }
  }

  if (int8_type_observed) {
    return QUANTIZED_INT8;
  }

  if (uint8_type_observed) {
    return QUANTIZED_UINT8;
  }

  // Default to float inference.
  return FLOAT;
}

}  // namespace tac
}  // namespace TFL

    case toco::IODataType::QUANTIZED_UINT8:
      return tflite::TensorType_UINT8;
    case toco::IODataType::UINT8:
      return tflite::TensorType_UINT8;
    case toco::IODataType::QUANTIZED_INT8:
      return tflite::TensorType_INT8;
    case toco::IODataType::INT8:
      return tflite::TensorType_INT8;
    default:
      return tflite::TensorType_FLOAT32;
  }
}

    case toco::IODataType::FLOAT16:
      return DT_HALF;
    case toco::IODataType::FLOAT64:
      return DT_DOUBLE;
    case toco::IODataType::QUANTIZED_UINT8:
      return DT_QUINT8;
    case toco::IODataType::QUANTIZED_INT8:
      return DT_QINT8;
    case toco::IODataType::QUANTIZED_INT16:
      return DT_QINT16;
    case toco::IODataType::INT8:
      return DT_INT8;
    case toco::IODataType::INT16:
      return DT_INT16;

  ::tflite::optimize::BufferType quantized_type;
  switch (quant_specs.inference_type) {
    case DT_QINT8:
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_INT8;
      break;
    case DT_HALF:
      quantized_type = ::tflite::optimize::BufferType::QUANTIZED_FLOAT16;
      break;
    default:
      return absl::InvalidArgumentError("Quantized type not supported");