inputs.reserve(op_with_region->getNumOperands());
      for (Value operand : op_with_region->getOperands()) {
        const Type operand_type = operand.getType();
        if (mlir::isa<NoneType>(operand_type)) {
          inputs.push_back(operand);
          continue;
        }

        const Type element_type =
            mlir::cast<TensorType>(operand.getType()).getElementType();

      SmallVector<Value, 4> inputs;
      inputs.reserve(quantizing_op->getNumOperands());
      for (auto operand : quantizing_op->getOperands()) {
        Type operand_type = operand.getType();
        if (operand_type.isa<NoneType>()) {
          inputs.push_back(operand);
          continue;
        }

        auto ele_type = operand.getType().cast<TensorType>().getElementType();
        if (static_cast<const ConcreteT*>(this)

TF_UNION_ACCESS_INSTANCE(impl::String, s);
TF_UNION_ACCESS_INSTANCE(impl::TaggedValueTensor, tensor);
TF_UNION_ACCESS_INSTANCE(impl::TensorSpec, tensor_spec);
#undef TF_UNION_ACCESS_INSTANCE

/// The union accessor for `NoneType`.
template <>
struct TaggedValue::UnionAccess<impl::None> {
  static impl::None& unsafe_reference(TaggedValue& t) {
    return None::GetInstance();
  }

}

LogicalResult EnsureBias(Operation* op, int bias_idx,
                         PatternRewriter& rewriter) {
  auto bias = op->getOperand(bias_idx);

  if (!mlir::isa<NoneType>(bias.getType())) return failure();

  // Proceed to create a zero bias.
  auto output = op->getResult(0);
  auto output_type = mlir::dyn_cast_or_null<RankedTensorType>(output.getType());

  // on failure.
  auto build_tensor_and_buffer = [&](Value value, const int subgraph_index,
                                     const std::string& tensor_name) {
    // NoneType represents optional and may be skipped here.
    if (mlir::isa<NoneType>(value.getType())) {
      return true;
    }

    tensor_index_map.insert({value, tensors.size()});
    tensor_index_map_[subgraph_index][tensor_name] = tensors.size();

  if (op_infos.empty())
    return parser.emitError(loc) << " expects at least one data operand";

  Attribute frame;
  if (parser.parseKeyword("frame") ||
      parser.parseAttribute(frame, NoneType::get(context), "frame_name",
                            result.attributes))
    return failure();

  Type i64 = parser.getBuilder().getIntegerType(64);
  if (parser.parseOptionalKeyword("parallel_iterations")) {

  if (getFusedActivationFunction() != "NONE") return failure();
  if (getWeightsFormat() != "DEFAULT") return failure();

  // Bias tensor is optional.
  const bool has_bias = !(!getBias() || getBias().getType().isa<NoneType>());

  // Get the tensors.
  DenseElementsAttr input_tensor, weights_tensor, bias_tensor;
  if (!matchPattern(getInput(), m_Constant(&input_tensor)) ||

            builder.classifierBuilder.buildClassLikeSymbolByLookupTag(coneType.lookupTag)
                ?: errorWithFirSpecificEntries("Class was not found", coneType = coneType)
        }

    override val qualifiers by cached {
        UsualClassTypeQualifierBuilder.buildQualifiers(coneType, builder)
    }

    override val typeArguments: List<KaTypeProjection> get() = withValidityAssertion { qualifiers.last().typeArguments }

    override fun approximateToSuperPublicDenotableType(type: KaType, approximateLocalTypes: Boolean): KaType? {
        require(type is KaFirType)
        val coneType = type.coneType
        val approximatedConeType = PublicTypeApproximator.approximateTypeToPublicDenotable(
            coneType,
            rootModuleSession,
            approximateLocalTypes = approximateLocalTypes,
        )

        return approximatedConeType?.asKtType()

            ?: errorWithFirSpecificEntries("Type parameter was not found", coneType = coneType)
    }

    override val annotations: KaAnnotationList by cached {
        KaFirAnnotationListForType.create(coneType, builder)
    }

    override val nullability: KaTypeNullability get() = withValidityAssertion { coneType.nullability.asKtNullability() }