return strings::StrCat(forward_frame_name, "_backprop");
}

// Creates a loop that counts the number of iterations performed by the
// while loop associated with `while_ctx`. The returned output yields the
// iteration count.
Status AddForwardLoopCounter(WhileContext* while_ctx, const Scope& scope,
                             Output* count) {
  // Create while loop:
  //   i = 0
  //   while forward loop predicate is true:

      // Add dequantize.
      builder.setInsertionPointAfter(read_variable_op);
      auto new_read_variable_op =
          builder.create<ReadVariableOp>(read_variable_op.getLoc(), ref_qtype,
                                         read_variable_op.getResourceId());
      auto new_dq_op = builder.create<DequantizeOp>(
          read_variable_op.getLoc(), read_variable_op.getResult().getType(),
          new_read_variable_op.getResult());

    if (failed(applyPatternsAndFoldGreedily(func, frozen_patterns))) {
      func.emitError() << "quant-propagate-quantize-type failed.";
      signalPassFailure();
    }
  }
}

}  // namespace

// Creates an instance of the TensorFlow dialect PropagateQuantizeType pass.
std::unique_ptr<OperationPass<ModuleOp>> CreatePropagateQuantizeTypePass() {
  return std::make_unique<PropagateQuantizeType>();
}

  // available during compilation or compilation only device for on demand
  // execution which may create a recursion if used for constant folding.
  std::string host_cpu = tensorflow::DeviceNameUtils::FullName(
      /*job=*/"localhost", /*replica=*/0, /*task=*/0, /*type=*/"CPU", /*id=*/0);

  absl::StatusOr<OpKernelRunner> runner = OpKernelRunner::Create(
      node_def->get()->op(), node_def->get()->name(), host_cpu, operands.size(),

    if (constant.getType().getRank() != 2) return failure();

    // 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.

  // _Arg nodes, and let CompileGraph walk it. This could be optimized.
  std::unique_ptr<Graph> graph(new Graph(OpRegistry::Global()));

  // First create the actual node we care about computing.
  TF_ASSIGN_OR_RETURN(Node * main_node, graph->AddNode(node_def));

  // Create dummy _Arg nodes. Link these to `node` and also via a control
  // dependency edge to the _SOURCE node.
  for (int64_t i = 0, end = args.size(); i < end; ++i) {

    Value input = op->getOperand(0);
    const TensorType new_input_tensor_type = GetTransposedTensorType(
        mlir::cast<TensorType>(input.getType()), kNchwToNhwcPermutation);

    auto input_transpose_op = rewriter.create<mlir::stablehlo::TransposeOp>(
        op.getLoc(), /*resultType0=*/new_input_tensor_type, /*operand=*/input,
        rewriter.getDenseI64ArrayAttr(kNchwToNhwcPermutation));

      if (QuantizedType::getQuantizedElementType(operand.getType())) {
        auto newTy = QuantizedType::castToExpressedType(operand.getType());
        newOperands.push_back(
            rewriter.create<TFL::DequantizeOp>(loc, newTy, operand));
        continue;
      }

      newOperands.push_back(operand);
    }

    SmallVector<Type> newResultTys;
    for (auto result : op->getResults()) {

    if (globalTensor.getValue()) {
      initial_value = *globalTensor.getValue();
    } else {
      initial_value = mlir::Attribute();
    }
    opToName[globalTensor] = name;
    auto variableOp = globalBuilder.create<ml_program::GlobalOp>(
        globalTensor.getLoc(), name, globalTensor.getType(),
        globalTensor.getIsMutable(), initial_value,
        /*visibility=*/globalBuilder.getStringAttr("private"));

    b.setInsertionPoint(op);
    if (auto dq = llvm::dyn_cast<DequantizeOp>(op)) {
      auto dcast = b.create<quantfork::DequantizeCastOp>(
          dq.getLoc(), dq.getOutput().getType(), dq.getInput());
      dq.getOutput().replaceAllUsesWith(dcast);
      dq.erase();
    } else if (auto q = llvm::dyn_cast<QuantizeOp>(op)) {
      auto qcast = b.create<quantfork::QuantizeCastOp>(
          q.getLoc(), q.getOutput().getType(), q.getInput());