public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {
    for (auto output : forward_outputs_) {
      if (output) {
        output->Ref();
      }
    }
  }

  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {

  tape.RecordOperation(inputs, {exp_output}, gradient_function.release());
  TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx,
                                          /*targets*/ {exp_output},
                                          /*sources=*/inputs,
                                          /*output_gradients=*/{},
                                          /*result=*/outputs));
  exp_output->Unref();
  return absl::OkStatus();
}

    TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx, /*targets=*/temp_outputs,
                                            /*sources=*/inputs,
                                            /*output_gradients=*/{}, outputs));
    for (auto temp_output : temp_outputs) {
      temp_output->Unref();
    }
    return absl::OkStatus();
  };
}

}  // namespace internal

void TF_OutputListSetNumOutputs(TF_OutputList* o, int num_outputs,
                                TF_Status* s) {
  unwrap(o)->expected_num_outputs = num_outputs;
  unwrap(o)->outputs.clear();
  unwrap(o)->outputs.resize(num_outputs);
}
int TF_OutputListNumOutputs(TF_OutputList* o) {
  return unwrap(o)->outputs.size();
}
TF_AbstractTensor* TF_OutputListGet(TF_OutputList* o, int i) {

  }
  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()) {

  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();

 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) {

    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,

      for (int i = 0; i < model_outputs.size(); i++) {
        if (model_outputs[i]) {
          output_list.outputs.emplace_back(model_outputs[i]);
          output_list.expected_num_outputs += 1;
        } else {
          null_indices.insert(i);
        }
      }
      TF_RETURN_IF_ERROR(dyn_cast<tracing::TracingContext>(func_ctx.get())
                             ->Finalize(&output_list, &func));