get_alg_op->erase();
  return success();
}

void TFXLADeviceSpecificTransforms::runOnOperation() {
  if (!device_type_.hasValue()) return;
  auto func_op = getOperation();

  auto walk_result = func_op->walk([&](TF::StatelessRandomGetAlgOp op) {
    if (failed(ConvertGetAlgOp(op))) {
      op->emitOpError(
          "Could not convert and remove Device specific information");

 public:
  void runOnOperation() override;
};

void InputMetricsLoweringPass::runOnOperation() {
  bool has_dynamic_op = false;
  Operation* func_op = getOperation();

  func_op->walk([&](Operation* op) {
    auto abstractOp = op->getRegisteredInfo();
    if (!abstractOp) return WalkResult::advance();

    if (mlir::mhlo::IsDynamicPadderOp(abstractOp->getTypeID())) {

  return op.getAttrOfType<StringAttr>(kXlaOutsideCompilationAttr) != nullptr;
}

void VerifyNoOutsideCompilationMarkersPass::runOnOperation() {
  Operation* func_op = getOperation();

  auto walk_result = func_op->walk([&](Operation* op) {
    if (IsDeviceClusterOp(*op) && HasChildLaunchDeviceOp(*op)) {
      std::string launch_error =
          absl::StrCat("Node `", op->getName().getStringRef().str(), "` ",

  }

  return init_func_ops;
}

func::FuncOp GetInitializerFunction(ModuleOp module_op,
                                    const StringRef initializer_type) {
  SmallVector<func::FuncOp, 2> init_func_ops =
      GetInitializerFunctions(module_op);

  auto init_func_itr = absl::c_find_if(
      init_func_ops, [initializer_type](const func::FuncOp init_func_op) {

}

bool IsMhloUniformQuantizedOp(Operation& op) {
  return llvm::isa<mhlo::UniformQuantizeOp, mhlo::UniformDequantizeOp>(op);
}

void VerifyQuantLegalization::runOnOperation() {
  Operation* func_op = getOperation();

  auto walk_result = func_op->walk([&](Operation* op) {
    // Verify all uq and qint types are lowered.
    if (llvm::any_of(op->getOperandTypes(), IsQuantType) ||
        llvm::any_of(op->getResultTypes(), IsQuantType) ||

  auto func_op_quant_scale_spec = GetStableHloQuantConstraints;

  for (auto func_op : module_op.getOps<func::FuncOp>()) {
    // The function might contain more stats ops than required, and it will
    // introduce requantize if the calibration stats have conflicts. This tries
    // to remove all the redundant stats ops.
    RemoveRedundantStatsOps(func_op, func_op_quant_spec,
                            func_op_quant_scale_spec);

  Value input = func_op.getArgument(0);
  Value output_init_state = func_op.getArgument(1);
  Value hidden_init_state = func_op.getArgument(2);
  Value weight_kernel = func_op.getArgument(3);
  Value recurrent_kernel = func_op.getArgument(4);
  Value bias = func_op.getArgument(5);

  // The func op should have 5 outputs.
  if (func_op.getNumResults() != 5) return failure();

  std::string function_name = absl::StrCat("func_", subgraph.subgraph_id_);

  func::FuncOp new_func = func::FuncOp::create(builder.getUnknownLoc(),
                                               function_name, function_type);
  new_func.setVisibility(func::FuncOp::Visibility::Private);
  new_func.addEntryBlock();

  // To form the body of the new function we need to clone each

  // Quantizer via the tf_quant.composite_function UnitAttr attached to
  // func ops. We remove this attribute as this interferes with VHLO conversion.
  // Remove this temporary hack.
  for (auto func_op : module_op.getOps<func::FuncOp>()) {
    func_op->removeAttr(kFusedFunctionAttr);
  }
}

}  // namespace

  absl::flat_hash_set<FunctionName> existing_func_names;
  module_op->walk([&](func::FuncOp func_op) {
    FunctionName func_name = func_op.getSymName().str();
    existing_func_names.insert(func_name);
    // We may retrieve the original function's name from the attribute.
    // Functions without this attribute are ignored.
    auto original_func_name =
        func_op->getAttrOfType<StringAttr>("tf._original_func_name");
    if (original_func_name) {