// 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,
      absl::Span<AbstractTensorHandle* const> sources,
      absl::Span<AbstractTensorHandle* const> output_gradients,

class ImmediateExecutionTensorHandle : public AbstractTensorHandle {
 public:
  // Returns number of dimensions.
  virtual Status NumDims(int* num_dims) const = 0;
  // Returns number of elements across all dimensions.
  virtual Status NumElements(int64_t* num_elements) const = 0;
  // Returns size of specified dimension
  //
  // -1 indicates an unknown axis length; this is unreachable for most standard