// is used as an input to an op (so we know when we're done computing gradients
// for that Tensor). We also count, for each tape entry, how many of its output
// Tensors need gradients to be computed (Tensors which are not used do not need
// any gradients to be computed).
//
// Finally, we start a backprop stack with a set of tape entries for which we
// have all gradients available. This set usually is a subset of the set of

// Struct to be filled in to define a custom device tensor handle. Fields are
// required except where indicated.
typedef struct TFE_CustomDeviceTensorHandleMethods {
  int version = TFE_CUSTOM_DEVICE_VERSION;

  // Computes the rank of the tensor handle.
  //
  // Shapes are specified via callbacks because retrieving the shape of a tensor
  // is a blocking operation for async eager; custom devices should avoid

        dtype_(dtype) {}

  const ParallelDevice& device_;
  std::vector<TensorHandlePtr> tensors_;
  // Parallel tensors are immutable but compute their shape lazily unless it is
  // provided on construction. The optional has a value if the lazy computation
  // has been completed or the shape was provided on construction.
  mutable absl::optional<std::vector<int64_t>> shape_;
  const TF_DataType dtype_;
};

// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleBackingDeviceName(

                                    TF_Status* status, TF_Output* dy);

// Adds operations to compute the partial derivatives of sum of `y`s w.r.t `x`s,
// i.e., d(y_1 + y_2 + ...)/dx_1, d(y_1 + y_2 + ...)/dx_2...
// This is a variant of TF_AddGradients that allows to caller to pass a custom
// name prefix to the operations added to a graph to compute the gradients.
//
// `dx` are used as initial gradients (which represent the symbolic partial

std::string getTF_OutputDebugString(TF_Output node);

// Set whether to propagate assigned device information when constructing a new
// Graph from a GraphDef. By default assigned device information is not copied
// and is re-computed by the runtime.
inline void TF_ImportGraphDefOptionsSetPropagateDeviceSpec(
    TF_ImportGraphDefOptions* opts, unsigned char propagate_device_spec) {
  opts->opts.propagate_device_spec = propagate_device_spec;
}

  TF_Status* s = TF_NewStatus();
  TF_OpKernelConstruction_GetAttrType(ctx, "foobar", &type, s);
  TF_DeleteStatus(s);
  return NULL;
}

// A compute function. This will never actually get called in this test, it's
// just nice to know that it compiles.
void compute(void* kernel, TF_OpKernelContext* ctx) {
  TF_Tensor* input;
  TF_Status* s = TF_NewStatus();
  TF_GetInput(ctx, 0, &input, s);
  TF_DeleteTensor(input);

namespace gradients {

// =============== Experimental C++ API for computing gradients ===============

// Sample gradient function:
//
// class AddGradientFunction : public GradientFunction {
//  public:
//   Status Compute(Context* ctx,
//                  absl::Span<AbstractTensorHandle* const> grad_inputs,
//                  absl::Span<AbstractTensorHandle*> grad_outputs) override {
//     grad_outputs[0] = grad_inputs[0];

  ///
  /// Replaces `dst` if it exists. In case of error, both `src` and `dst` keep
  /// the same state as before the call.
  ///
  /// Plugins:
  ///   * Must set `status` to `TF_OK` if rename was completed.
  ///   * Must set `status` to `TF_NOT_FOUND` if one of the parents entries in
  ///     either `src` or `dst` doesn't exist or if the specified `src` path
  ///     doesn't exist.