// memory consumption is over kMinAggregateBytes, do an early aggregation
// so as to release the gradient tensor to save memory.
constexpr int kMinAggregateCount = 4;
constexpr int kMinAggregateBytes = 128 * 1024 * 1024;

template <typename Gradient, typename BackwardFunction, typename TapeTensor>
Status GradientTape<Gradient, BackwardFunction, TapeTensor>::ComputeGradient(