}
  std::string value_string;
  if (!SummarizeValue(value_string).ok()) {
    value_string = "<error computing value>";
  }
  if (value_string.length() > 100) {
    // The default NumPy-style output can be distractingly long in error
    // messages.
    value_string = absl::StrCat(value_string.substr(0, 100), " [...]");
  }
  Status s;
  const char* device_name = DeviceName(&s);
  if (!s.ok()) {

ParallelDevice::Join(
    const std::vector<PartialTensorShape>& expected_output_shapes,
    TF_Status* status) const {
  absl::optional<std::vector<std::unique_ptr<ParallelTensor>>> result;
  // Compute per-device per-output tensors
  std::vector<std::vector<TensorHandlePtr>> per_device_output_tensors;
  per_device_output_tensors.reserve(underlying_devices_.size());
  int first_op_output_count = 0;
  StatusPtr first_bad_status(nullptr);

  AbstractTensorHandlePtr model_out(model_outputs[0]);

  TF_Tensor* model_out_tensor;
  TF_RETURN_IF_ERROR(GetValue(model_out.get(), &model_out_tensor));
  int num_dims_out = TF_NumDims(model_out_tensor);
  TF_DeleteTensor(model_out_tensor);

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();

                        TF_Bool assume_valid_feeds,
                        TF_Bool aggressive_shape_inference,
                        TF_Bool include_input_tensor_values,
                        TF_Bool include_output_tensor_values,
                        TF_Status* status) {
  TF_SetStatus(status, TF_OK, "");
  absl::Status s =
      reinterpret_cast<tensorflow::grappler::GraphProperties*>(graph_properties)

                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::AddV2(ctx, inputs[0], inputs[1], &outputs[0], "Add");
}

Status ExpModel(AbstractContext* ctx,
                absl::Span<AbstractTensorHandle* const> inputs,
                absl::Span<AbstractTensorHandle*> outputs) {
  return ops::Exp(ctx, inputs[0], &outputs[0], "Exp");
}

Status SqrtModel(AbstractContext* ctx,

    absl::Span<AbstractTensorHandle*> outputs) {
  Tape tape(/*persistent=*/false);
  tape.Watch(inputs[0]);
  AbstractTensorHandle* neg_output;
  TF_RETURN_IF_ERROR(ops::Neg(ctx, inputs[0], &neg_output, "Neg"));
  tape.RecordOperation(inputs, {neg_output}, nullptr, "Neg");
  return tape.ComputeGradient(ctx,
                              /*targets=*/{neg_output},
                              /*sources=*/inputs,

 public:
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {
    for (int i = 0; i < grad_outputs.size(); i++) {
      auto grad_input = grad_outputs[i];
      // TODO(srbs): Should we add a copy contructor to AbstractTensorHandle
      // that takes care of this similar to `Tensor`?

void TestRemoteExecuteSilentCopiesOp(bool async, bool remote,
                                     bool remote_func_outputs = false) {
  return TestRemoteExecuteSilentCopies(async, remote, /*func=*/false,
                                       /*heavy_load_on_streaming_rpc=*/false,
                                       remote_func_outputs);
}

TEST(CAPI, RemoteExecuteSilentCopies) {

  std::vector<MaybeParallelTensorOwned> typed_outputs(
      std::move(maybe_typed_outputs.value()));

  if (typed_outputs.size() > *num_outputs) {
    TF_SetStatus(status, TF_INTERNAL,
                 "The allocated output buffer was too small.");
    return;
  }

  for (int i = 0; i < typed_outputs.size(); ++i) {
    MaybeParallelTensorOwned typed_output(std::move(typed_outputs[i]));

          /* opers */ operations, /* ninputs */ 1, /* inputs */ &x,
          /* noutputs */ 1, /* outputs */ &y, /* output_names */ &output_name,
          /* opts */ nullptr, /* description */ "", /* status */ status),
      TF_DeleteFunction);
  if (TF_GetCode(status) != TF_OK) return;
  TFE_ContextAddFunction(context, function.get(), status);
}

TEST(PARALLEL_DEVICE, TestFunction) {