// Consumes the internal state of the tape (so cannot be called more than
  // once unless the tape is persistent) and produces the gradient of the target
  // tensors with respect to the source tensors. The output gradients are used
  // if not empty and not null. The result is populated with one tensor per
  // target element.
  Status ComputeGradient(
      AbstractContext* ctx, absl::Span<AbstractTensorHandle* const> targets,

  std::list<Key> lru_list_ ABSL_GUARDED_BY(mu_);

  /// The LRA (least recently added) list of block keys. The front of the list
  /// identifies the most recently added block.
  ///
  /// Note: blocks are added to lra_list_ only after they have successfully been
  /// fetched from the underlying block store.
  std::list<Key> lra_list_ ABSL_GUARDED_BY(mu_);

  /// The combined number of bytes in all of the cached blocks.