components.reserve(underlying_devices_.size());
    for (int j = 0; j < underlying_devices_.size(); ++j) {
      components.push_back(std::move(per_device_output_tensors[j][i]));
    }
    if (expected_output_shapes[i].IsFullyDefined()) {
      per_device_outputs.push_back(ParallelTensor::FromTensorHandles(
          *this, std::move(components),
          absl::Span<const int64_t>(expected_output_shapes[i].dim_sizes()),
          status));

    // Skip adding shape op for operands that have static shapes.
    tensorflow::PartialTensorShape shape(
        metadata.args(operand_and_idx.index()).shape());
    if (shape.IsFullyDefined()) continue;

    auto shape_op = builder->create<TF::ShapeOp>(
        cluster_func.getLoc(),
        tensorflow::GetTypeFromTFTensorShape({-1}, builder->getIntegerType(64)),
        operand_and_idx.value());