// "Jacobian-vector product"; `Watch`ing a new Tensor and immediately calling
  // FetchJVP for it would return `tangent`.
  void Watch(int64_t tensor_id, Gradient* tangent);

  // Removes the gradient associated with tensor_id. Should be called when the
  // Tensor associated with `tensor_id` is deleted.
  void DeleteGradient(int64_t tensor_id);

      # to remove it.
      key = re.sub(r"0x\w+", "", p.getparent().get("name", "")) + p.text
      if key in seen:
        testsuite._elem.remove(p.getparent())
      seen[key] += 1
    # Remove this testsuite if it doesn't have anything in it any more
    if len(testsuite) == 0:  # pylint: disable=g-explicit-length-test
      r._elem.remove(testsuite._elem)

  EXPECT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Pack 3 variable handles into one TFE_TensorHandle.
  // When remote is false, function device is placed on task0. Handle types are
  // REMOTE, REMOTE, LOCAL on task0. When remote is true, function device is
  // placed on task1, Handle types are LOCAL, REMOTE, LOCAL on task1.
  int num_replicas = 3;
  std::vector<TFE_TensorHandle*> handles = {h0, h1, h2};

    const TFE_OpAttrs* attributes, int expected_max_outputs,
    TF_Status* status) {
  absl::optional<std::vector<MaybeParallelTensorOwned>> result;
  // TODO(allenl): We should remove "TPU" from these op names at the very least,
  // or consider other ways of packing/unpacking parallel tensors.
  if (operation_name == std::string("TPUReplicatedInput")) {

class DeviceThread;

// Forwards operations to `devices`, maintaining ParallelTensor with components
// placed on each underlying device.
class ParallelDevice {
 public:
  // Eager async execution is only supported when remote eager is not in use
  // (b/157523095).
  explicit ParallelDevice(const std::vector<std::string>& devices,
                          bool is_async = false, int in_flight_nodes_limit = 0);

  ~ParallelDevice();

    per_device_output_tensors.push_back(device_thread->Join(status));
    // We will run every Join even if there are bad statuses in case the user
    // wants to recover and continue running ops on the parallel device (which
    // would otherwise deadlock).
    if (TF_GetCode(status) != TF_OK &&
        (first_bad_status == nullptr