} else {
        EXPECT_EQ(dltensor_out->strides[i],
                  dltensor_out->shape[i + 1] * dltensor_out->strides[i + 1]);
      }
    }
  }
  const float* data_in = static_cast<const float*>(dltensor_in->data);
  const float* data_out = static_cast<const float*>(dltensor_out->data);
  for (size_t j = 0; j < num_elements; ++j) {
    EXPECT_EQ(data_out[j], data_in[j]);
  }

  TFE_CallDLManagedTensorDeleter(dlm_out);

                           &out_components, status.get());
    ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

    ExpectScalarEq<float>(out_components[0].get(), 9.);
    ExpectScalarEq<float>(out_components[1].get(), 4.);
  }

  worker_server1.release();
  worker_server2.release();
}
}  // namespace parallel_device

                     float value) {
  desc->node_builder.Attr(attr_name, value);
}

void TF_SetAttrFloatList(TF_OperationDescription* desc, const char* attr_name,
                         const float* values, int num_values) {
  desc->node_builder.Attr(attr_name,
                          ArraySlice<const float>(values, num_values));
}

  return TF_NewTensor(TF_DOUBLE, nullptr, 0, values, num_bytes,
                      &DoubleDeallocator, nullptr);
}

TF_Tensor* FloatTensor(float v) {
  const int num_bytes = sizeof(float);
  float* values = new float[1];
  values[0] = v;
  return TF_NewTensor(TF_FLOAT, nullptr, 0, values, num_bytes,
                      &FloatDeallocator, nullptr);
}

}

// Passed to `TF_NewTensor` to indicate how an array of floats should be
// deleted.
static void FloatDeallocator(void* data, size_t, void* arg) {
  delete[] static_cast<float*>(data);
}

// Creates a TFE_TensorHandle with value `v`.
TensorHandlePtr FloatTensorHandle(float v, TF_Status* status) {
  const int num_bytes = sizeof(float);
  float* values = new float[1];
  values[0] = v;

    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 2.0f, &x_raw);
    ASSERT_EQ(errors::OK, status_.code()) << status_.message();
    x.reset(x_raw);
  }

  AbstractTensorHandlePtr y;
  {
    AbstractTensorHandle* y_raw = nullptr;
    status_ = TestScalarTensorHandle<float, TF_FLOAT>(
        immediate_execution_ctx_.get(), 2.0f, &y_raw);

  {
    std::unique_ptr<ParallelTensor> float_tensors =
        parallel_device.ScalarsFromSequence<float>({10.0, 11.0}, context.get(),
                                                   status.get());
    ASSERT_EQ(TF_GetCode(status.get()), TF_OK) << TF_Message(status.get());
    ExpectScalarEq<float>(float_tensors->tensor(0), 10.0);
    ExpectScalarEq<float>(float_tensors->tensor(1), 11.0);
  }

  {

    ASSERT_EQ(num_elem_numerical, num_elem_analytical);

    float* dnumerical = new float[num_elem_numerical]{0};
    memcpy(&dnumerical[0], TF_TensorData(numerical_tensor),
           TF_TensorByteSize(numerical_tensor));
    float* danalytical = new float[num_elem_analytical]{0};
    memcpy(&danalytical[0], TF_TensorData(analytical_tensor),
           TF_TensorByteSize(analytical_tensor));

    ASSERT_EQ(errors::OK, s.code()) << s.message();
    ctx.reset(ctx_raw);
  }

  AbstractTensorHandlePtr x;
  {
    AbstractTensorHandle* x_raw = nullptr;
    Status s = TestScalarTensorHandle<float, TF_FLOAT>(ctx.get(), 1.0f, &x_raw);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    x.reset(x_raw);
  }

  // Pseudo-code:
  //
  // tape.watch(x)
  // y = exp(x)
  // outputs = tape.gradient(y, x)

  std::unique_ptr<TF_Status, decltype(&TF_DeleteStatus)> status(
      TF_NewStatus(), TF_DeleteStatus);
  TF_Tensor* t = TFE_TensorHandleResolve(handle, status.get());
  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());
  std::unique_ptr<float[]> actual_values(new float[expected_values.size()]);
  EXPECT_EQ(sizeof(float) * expected_values.size(), TF_TensorByteSize(t));