BasicTestsForTwoDevices(context.get(),
                          "/job:worker/replica:0/task:1/device:CPU:0",
                          "/job:worker/replica:0/task:2/device:CPU:0");

  worker_server1.release();
  worker_server2.release();
}

TEST(PARALLEL_DEVICE, TestAsyncCopyOff) {
  std::unique_ptr<TFE_ContextOptions, decltype(&TFE_DeleteContextOptions)> opts(
      TFE_NewContextOptions(), TFE_DeleteContextOptions);

  return wrap(t);
}

void TF_DeleteExecutionContext(TF_ExecutionContext* c) { unwrap(c)->Release(); }

TF_AbstractOp* TF_NewAbstractOp(TF_ExecutionContext* c) {
  return wrap((unwrap(c)->CreateOperation()));
}

void TF_DeleteAbstractOp(TF_AbstractOp* op) { unwrap(op)->Release(); }

void TF_DeleteAbstractTensor(TF_AbstractTensor* t) { unwrap(t)->Unref(); }

    tensorflow::shape_inference::InferenceContext* ic, int num_dims,
    const int64_t* dims) {
  if (num_dims != -1) {
    std::vector<tensorflow::shape_inference::DimensionHandle> dim_vec;
    dim_vec.reserve(num_dims);
    for (int i = 0; i < num_dims; ++i) {
      dim_vec.push_back(ic->MakeDim(dims[i]));
    }
    return ic->MakeShape(dim_vec);
  } else {
    return ic->UnknownShape();
  }
}

  TFE_ExecutorWaitForAllPendingNodes(executor, status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // TODO(b/136478427): Figure out how to correctly shut the server down.
  worker_server.release();

  // Update the server def with a new set of names (worker instead of
  // localhost).
  tensorflow::ServerDef updated_server_def = GetServerDef("worker", 2);
  serialized = updated_server_def.SerializeAsString();

  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());
  const std::vector<std::unique_ptr<ParallelTensor>>& handles = *outputs;
  std::vector<ParallelTensor*> handle_inputs;
  handle_inputs.reserve(handles.size());
  for (auto& handle : handles) {
    handle_inputs.push_back(handle.get());
  }
  std::unique_ptr<TFE_Op, decltype(&TFE_DeleteOp)> read_op(

  TF_RETURN_IF_ERROR(ops::Exp(ctx, inputs[0], &exp_output, "Exp"));
  std::unique_ptr<GradientFunction> gradient_function(
      new PassThroughGradientFunction);
  tape.RecordOperation(inputs, {exp_output}, gradient_function.release());
  TF_RETURN_IF_ERROR(tape.ComputeGradient(ctx,
                                          /*targets*/ {exp_output},
                                          /*sources=*/inputs,

  TFE_DeleteExecutor(executor);
  TFE_DeleteContext(ctx);

  TF_DeleteStatus(status);

  // TODO(b/136478427): Figure out how to correctly shut the server down.
  worker_server.release();
}

TEST(CAPI, RemoteExecute) { TestRemoteExecute(false); }
TEST(CAPI, RemoteExecuteAsync) { TestRemoteExecute(true); }

void TestRemoteExecuteSilentCopiesOp(bool async, bool remote,

    // Note that we take the address of the elements in `all_input_tensors`
    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();

      parent_op_(parent_op),
      tape_(tape),
      registry_(registry) {
  // TODO(b/172003047): Consider making AbstractOperation RefCounted.
  // parent_op_->Ref();
}
void TapeOperation::Release() {
  // TODO(srbs): Change to Unref().
  delete this;
}
TapeOperation::~TapeOperation() {
  // TODO(b/172003047): Consider making AbstractOperation RefCounted.
  // parent_op->Unref();
}

                                         tensorflow::Status* status) {
  std::vector<int64_t> shape;
  int rank = -1;
  *status = handle.NumDims(&rank);
  if (!status->ok()) {
    return shape;
  }
  shape.reserve(rank);
  for (int i = 0; i < rank; ++i) {
    int64_t dim;
    *status = handle.Dim(i, &dim);
    if (!status->ok()) {
      return shape;
    }
    shape.push_back(dim);
  }
  return shape;
}