// * 3rd output is the outer offset.
  auto input_type = GetInputType(func, 0);
  if (!input_type || !mlir::isa<StringType>(input_type.getElementType()) ||
      !input_type.hasRank()) {
    return func.emitError() << "Input should be a string tensor";
  }

  const std::vector<int> kValidNumOfOutput = {1, 2, 3};
  if (input_type.getRank() >= kValidNumOfOutput.size()) {
    return func.emitError()

  auto input_type = input.getType().cast<ShapedType>();
  if (permutation_array.size() != input_type.getRank()) {
    return nullptr;
  }
  llvm::SmallVector<int64_t> transposed_shape(permutation_array.size());
  for (int64_t i = 0; i < permutation_array.size(); ++i) {
    transposed_shape[i] = input_type.getDimSize(permutation_array[i]);
  }
  auto transposed_type =

// accumulation over the given input type.
Type GetSumAccumulationType(Type input_type) {
  MLIRContext *ctx = input_type.getContext();
  if (input_type.isBF16() || input_type.isF16()) return FloatType::getF32(ctx);
  if (input_type.isSignlessInteger(8) || input_type.isSignlessInteger(16))
    return IntegerType::get(ctx, 32);
  return input_type;
}

// Returns axis in HLO format from TF elements attr with exactly one element or

    llvm::ArrayRef<int64_t> permutation_array, ShapedType input_type,
    ConversionPatternRewriter& rewriter) {
  assert(permutation_array.size() == input_type.getRank());
  llvm::SmallVector<int64_t> transposed_shape(permutation_array.size());
  for (int64_t i = 0; i < permutation_array.size(); ++i) {
    transposed_shape[i] = input_type.getDimSize(permutation_array[i]);
  }
  auto transposed_type =

    // can have smaller memory usage.
    auto input_type =
        mlir::dyn_cast<RankedTensorType>(dequantize_op.getOutput().getType());
    auto output_type = mlir::dyn_cast<RankedTensorType>(
        passthrough_op->getResult(0).getType());
    if (!input_type || !output_type ||
        get_num_elements(input_type) <= get_num_elements(output_type)) {
      return failure();
    }

  if (!dims_type) return success();
  if (dims_type.getRank() > 1)
    return emitError(loc, "dimensions can only be 0D or 1D tensor");

  auto input_type = mlir::dyn_cast<RankedTensorType>(input.getType());
  if (!input_type) return success();
  int64_t rank = input_type.getRank();

  DenseIntElementsAttr dims_attr;
  if (!matchPattern(dims, m_Constant(&dims_attr))) return success();

  // known.
  mlir::Type output_type;
  auto input_type = mlir::cast<mlir::TensorType>(src_input.getType());

  if (input_type.hasRank()) {
    if (input_type.getShape()[split_dimension] == mlir::ShapedType::kDynamic) {
      output_type = input_type;
    } else {
      auto shape = llvm::to_vector<4>(input_type.getShape());
      if (shape[split_dimension] % num_split != 0) {

    // Create a tfl.transpose op that performs ZX transpose on `input`.
    auto create_z_x_transpose_op = [&](Value input) -> Value {
      RankedTensorType input_type =
          mlir::cast<RankedTensorType>(input.getType());
      const int input_rank = input_type.getRank();

      // Create a 1D I32 tensor for representing the dimension permutation.
      auto permuation_tensor_type =

  for (auto input_type : func_type.getInputs()) {
    tensorflow::TensorShape tensor_shape;
    xla::Shape xla_shape = xla::TypeToShape(input_type);
    TF_RETURN_IF_ERROR(tensorflow::TensorShape::BuildTensorShape(
        xla_shape.dimensions(), &tensor_shape));
    arg_shapes.emplace_back(tensor_shape);

    DataType dtype;
    TF_RETURN_IF_ERROR(ConvertToDataType(input_type, &dtype));

    int quant_dim = -1;
    auto input_type = mlir::cast<ShapedType>(input.getType());
    if (PerAxis) {
      if (!input_type.hasRank()) {
        tf_op.emitError("The input should have known rank for per-channel op.");
        return failure();
      }
      // This is a special case that the quant_dim is the last dimensions.
      quant_dim = input_type.getRank() - 1;
    }