#include "tensorflow/core/platform/errors.h"

using std::vector;
using tensorflow::ops::BiasAddGrad;
using tensorflow::ops::Mul;
using tensorflow::ops::ReluGrad;

namespace tensorflow {
namespace gradients {
namespace {

class ReluGradientFunction : public GradientFunction {
 public:
  explicit ReluGradientFunction(vector<AbstractTensorHandle*> f_outputs)
      : forward_outputs_(f_outputs) {

#include "tensorflow/c/eager/dlpack.h"

#include <vector>

#include "absl/strings/str_join.h"
#include "include/dlpack/dlpack.h"  // from @dlpack
#include "tensorflow/c/eager/c_api.h"
#include "tensorflow/core/platform/test.h"

namespace tensorflow {
namespace {

void TestHandleFromDLPack(TF_Status* status, TFE_Context* ctx,
                          std::vector<int64_t> shape,

    // Input tensors
    const std::vector<std::pair<string, Tensor>>& input_pairs,
    // Output tensors
    const std::vector<string>& output_tensor_names, TF_Tensor** c_outputs,
    // Target nodes
    const std::vector<string>& target_oper_names, TF_Buffer* run_metadata,
    TF_Status* status) {
  const int noutputs = output_tensor_names.size();
  std::vector<Tensor> outputs(noutputs);
  Status result;

limitations under the License.
==============================================================================*/

#include "tensorflow/c/eager/c_api_unified_experimental.h"

#include <vector>

#include "absl/container/flat_hash_map.h"
#include "absl/strings/str_cat.h"
#include "tensorflow/c/eager/c_api_unified_experimental_internal.h"
#include "tensorflow/c/tf_datatype.h"
#include "tensorflow/c/tf_status.h"

  TensorHandlePtr three_vector =
      VectorFloatTensorHandle({5., 6., 7.}, status.get());
  ASSERT_TRUE(TF_GetCode(status.get()) == TF_OK) << TF_Message(status.get());

  std::vector<TensorHandlePtr> vector_handles;
  vector_handles.reserve(2);
  vector_handles.push_back(std::move(two_vector));
  vector_handles.push_back(std::move(three_vector));
  std::unique_ptr<ParallelTensor> unknown_length_vector =

  }
}

absl::optional<std::vector<std::unique_ptr<ParallelTensor>>>
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;

  device->delete_device = &DeleteParallelDevice;
  device->execute = &ParallelDeviceExecute;
  std::vector<std::string> underlying_devices_vector;
  underlying_devices_vector.reserve(num_underlying_devices);
  for (int device_index = 0; device_index < num_underlying_devices;
       ++device_index) {
    underlying_devices_vector.push_back(underlying_devices[device_index]);
  }
  std::unique_ptr<ParallelDevice> parallel_device(

  if (!status->status.ok()) return;

  // Initialize a input_tensor vector with `nullptr` values.
  std::vector<const Tensor*> input_tensors_vector(num_inputs, nullptr);
  // A vector to keep track of newly created `tf::Tensor` objects.
  std::vector<Tensor> all_input_tensors;
  // Update the vector with information from `input_tensors` if provided.
  if (input_tensors != nullptr) {

  int num_elems = TF_TensorElementCount(theta_tensor);
  vector<float> theta_data(num_elems);
  memcpy(theta_data.data(), TF_TensorData(theta_tensor),
         TF_TensorByteSize(theta_tensor));

  // Initialize space for the numerical gradient.
  vector<float> dtheta_approx(num_elems);

  // Get theta shape and store in theta_dims.
  int num_dims = TF_NumDims(theta_tensor);
  vector<int64_t> theta_dims(num_dims);

  // Process inputs.
  std::vector<tensorflow::OutputTensor> input_tensors;
  std::unordered_map<const Node*, std::vector<int>> input_nodes;
  status->status = tensorflow::ProcessInputs(fn_body, fn_name, ninputs, inputs,
                                             &input_tensors, &input_nodes);
  if (TF_GetCode(status) != TF_OK) return nullptr;

  // Process outputs.
  std::vector<tensorflow::OutputTensor> output_tensors;