// If `dx` is nullptr, the implementation will use dx of `OnesLike` for all
// shapes in `y`.
// The partial derivatives are returned in `dy`. `dy` should be allocated to
// size `nx`.
// `prefix` names the scope into which all gradients operations are being added.
// `prefix` must be unique within the provided graph otherwise this operation
// will fail. If `prefix` is nullptr, the default prefixing behaviour takes

// efficiency of forward-mode autodiff. Instead, language bindings should call
// DeleteGradient as soon as a Tensor which was `Watch`ed or was an output
// Tensor passed to Accumulate goes out of scope.
//
// Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the

// need for `Env::Default()`. At that time, we might decide to remove the need
// for `Env::Default()` altogether, but that's a different project, not in
// scope for now. I'm just mentioning this here as that transition will mean
// removal of the registration part from `Env` and adding it here instead: we
// will need tables to hold for each scheme the function tables that implement