int output_index,
                                     PatternRewriter& rewriter) const {
    assert(output_index == 0);
    if (!op.getResult().hasOneUse()) {
      op.emitError()
          << "output " << output_index
          << " should have only one use, which should be quant.stats.";
      return failure();
    }

      : OpRewritePattern<DequantizeOp>(context) {}

  LogicalResult matchAndRewrite(DequantizeOp dequantize_op,
                                PatternRewriter& rewriter) const override {
    if (!dequantize_op->hasOneUse()) return failure();

    auto use = dequantize_op->use_begin();
    Operation* passthrough_op = use->getOwner();
    unsigned operand_index = use->getOperandNumber();

        // Don't remove leading and trailing QDQ for PTQ workflow, so the io
        // modifying lib can work correctly.
        if (!q.getArg().getDefiningOp()) return failure();
        if (op->hasOneUse() &&
            op->user_begin()->hasTrait<OpTrait::IsTerminator>())
          return failure();
      }
      // If the quantize op is a requantize op, it is being used in other scale

void QuantizationDriver::RequantizeArg(const BlockArgument arg,
                                       RequantizeStates& states) {
  Value value = arg;
  builder_.setInsertionPointToStart(arg.getOwner());
  if (value.hasOneUse()) {
    Operation* user = value.use_begin().getUser();
    if (auto q = dyn_cast<quantfork::QuantizeCastOp>(user)) {
      value = q.getResult();
      builder_.setInsertionPoint(arg.getOwner(), ++Block::iterator(user));

ResourceValueAndSubtype GetResourceWriteResult(
    tf_device::ClusterFuncOp cluster_func, Value result) {
  ResourceValueAndSubtype resource;
  if (!result.hasOneUse()) return resource;
  Operation* result_user = *result.getUsers().begin();
  auto assign_var = dyn_cast<TF::AssignVariableOp>(result_user);
  if (!assign_var) return resource;

  auto handle = assign_var.getResource();

                                PatternRewriter &rewriter) const override {
    // We don't want to insert quantize/dequantize if the quantize op exists.
    auto res = tf_op.getOutputs();
    if (!res.hasOneUse() || isa<quantfork::QuantizeCastOp>(*res.user_begin()))
      return failure();

    // Extract the min/max constant values from the operands. We also consider

  // If activation exists, omit clipping op.
  // Since out_scale and out_zp are computed based on clipped range,
  // explicit activation clipping op is not required.
  if (isa<AddOp>(next_op) && gemm_style_op->hasOneUse()) {
    // bias fusion
    CreateAndReturnQuantizedBiasPattern(
        next_op, rewriter, entry_func_op, func_result_type,
        accumulation_quantized_element_type, gemm_style_op);

// before or above the parallel_execute.
TF::AssignVariableOp GetSingleUseResourceWrite(
    tf_device::ParallelExecuteOp parallel_execute, Value result) {
  if (!result.hasOneUse()) return nullptr;

  OpOperand& use = *result.getUses().begin();
  auto assign_var = dyn_cast<TF::AssignVariableOp>(use.getOwner());
  if (!assign_var) return nullptr;

                                OpBuilder &rewriter) const {
    // We don't want to insert quantize/dequantize if the quantize op exists.
    auto res = tf_op.getOutputs();
    if (!res.hasOneUse() || isa<QuantizeOp>(*res.user_begin())) {
      return failure();
    }

    // Extract the min/max constant values from the operands. We also consider

        constant_op.getResult().hasOneUse())
      continue;
    // Do not duplicate constant op if the size is too large.
    // 32 is chosen to be larger than all constants useful for shape references,
    // while not too large to possibly significantly increase model size.
    if (constant_op.getValue().getNumElements() > 32) continue;
    while (!constant_op.getResult().hasOneUse()) {