synchronized (lock) {
            List<WorkerDaemonClient> sortedClients = CollectionUtils.sort(idleClients, comparingInt(WorkerDaemonClient::getUses));
            List<WorkerDaemonClient> clientsToStop = selectionFunction.transform(new ArrayList<>(sortedClients));
            if (!clientsToStop.isEmpty()) {
                stopWorkers(clientsToStop);
            }

    // determine whether we have taken care of all uses and can remove
    // the arg.
    for (BlockArgument arg : func.getArguments()) {
      auto uses = arg.getUses();
      use_count[arg.getArgNumber()] = std::distance(uses.begin(), uses.end());
      argument_to_index.insert({arg, arg.getArgNumber()});
    }

    // We're only considering return values that are the same (at this

    llvm::SmallVector<Operation*, 4> candidates;
    for (Value value : values) {
      if (incoming) {
        candidates = {value.getDefiningOp()};
      } else {
        candidates.assign(value.getUsers().begin(), value.getUsers().end());
      }
      for (Operation* candidate_op : candidates) {
        if (cluster_ops.contains(candidate_op) ||
            cluster_dependent_ops.contains(candidate_op)) {

    if (!type || !type.getElementType().isF32()) {
      return failure();
    }
    return success(
        op->hasOneUse() &&
        IsWeightQuantizableFunction(*op->getUses().begin(), type.getRank()));
  }

  // Checks if the operand is second operand of `tf.XlaCallModule` op for
  // `stablehlo.convolution` or `stablehlo.dot_general` with fully_quantizable
  // trait.

  // bias is used immediately by the user. This assumption is always correct
  // after constant folding.
  bool UsedAsBias(Value value) {
    for (auto &use : value.getUses()) {
      auto biases = TFL::GetOpQuantSpec(use.getOwner())->biases_params;
      if (biases.find(use.getOperandNumber()) != biases.end()) return true;
    }
    return false;
  }

    if (!type || !type.getElementType().isF32()) return false;

    Value value = op.getResult();

    // Check whether dynamic range quantization can be applied.
    for (auto& use : value.getUses()) {
      Operation* user = use.getOwner();
      int operand_num = use.getOperandNumber();
      std::unique_ptr<OpQuantSpec> spec = GetTFOpQuantSpec(user);

          values.getUses().begin().getUser());
      values_slice_op.getResult().replaceAllUsesWith(top_k_op.getValues());
      values_slice_op.erase();
    }
    if (!indices.use_empty()) {
      auto indices_slice_op = llvm::dyn_cast_or_null<TFL::SliceOp>(
          indices.getUses().begin().getUser());
      indices_slice_op.getResult().replaceAllUsesWith(top_k_op.getIndices());

// if there is only one.
Operation* GetUserIfOnlyOne(Operation* op) {
  if (op->getNumResults() != 1) return nullptr;
  auto result = op->getResult(0);
  if (!result.hasOneUse()) return nullptr;
  return (*result.getUses().begin()).getOwner();
}

// Gets operation providing value for the given operand of given operation
// if the given operation is the only user.
Operation* GetInputOpWithOneUse(Operation* op, int opr_num) {

    // Unsure if this check is required.
    if (state.pos != pos) {
      return;
    }
  }
  if (pos == RequantizeState::ON_OUTPUT) {
    Operation* user = value.getUses().begin().getUser();
    if (isa<quantfork::QuantizeCastOp>(user)) {
      // The requantize op is inserted between `quantize` and `dequantize` ops.
      value = user->getResult(0);
      builder_.setInsertionPointAfter(user);

}

std::optional<RankedTensorType> GetElementTypeFromAccess(
    Value collection, ModuleOp module,
    llvm::function_ref<std::optional<Type>(Operation*)> infer_from_op) {
  for (auto& use : collection.getUses()) {
    if (auto while_op = llvm::dyn_cast<TF::WhileOp>(use.getOwner())) {
      auto body = while_op.body_function();
      assert(body);
      auto type_from_body = GetElementTypeFromAccess(