TF_Status* status);

// A handle to a tensor on a device.
//
// Like a TF_Tensor, a TFE_TensorHandle refers to a tensor with a value, shape,
// type etc. Unlike a TF_Tensor, a TFE_TensorHandle may refer to such tensors
// placed in the memory of different devices or remote address spaces.
typedef struct TFE_TensorHandle TFE_TensorHandle;

    absl::Span<const AbstractTensorHandle* const> tensors) const {
  std::vector<int64_t> tensor_ids(tensors.size());
  std::vector<tensorflow::DataType> tensor_dtypes(tensors.size());
  for (int i = 0; i < tensors.size(); i++) {
    tensor_ids[i] = ToId(tensors[i]);
    tensor_dtypes[i] = tensors[i]->DataType();
  }
  return GradientTape::ShouldRecord(tensor_ids, tensor_dtypes);
}

  if (VLOG_IS_ON(1) && status->status.ok()) {
    tensorflow::Tensor tensor;
    if (tensorflow::TF_TensorToTensor(ret, &tensor).ok()) {
      VLOG(1) << "Dequeued tensor content: " << tensor.DebugString();
    }
  }
  return ret;
}

void TF_EnqueueNamedTensor(TF_Session* session, int tensor_id,
                           TF_Tensor* tensor, TF_Status* status) {
  assert(session);
  {

  ParallelTensor* parallel_tensor = reinterpret_cast<ParallelTensor*>(
      TFE_TensorHandleDevicePointer(tensor, status));
  if (TF_GetCode(status) != TF_OK) return nullptr;
  if (parallel_tensor->num_tensors() == 1) {
    // Copy-off for single-device tensors is allowed to make debugging dynamic
    // control flow easier.
    return TFE_TensorHandleCopySharingTensor(parallel_tensor->tensor(0),

TF_CAPI_EXPORT extern TF_Tensor* TF_DequeueNamedTensor(TF_Session* session,
                                                       int tensor_id,
                                                       TF_Status* status);

// On success, enqueues `tensor` into a TF-managed FifoQueue given by
// `tensor_id`, associated with `session`. There must be a graph node named
// "fifo_queue_enqueue_<tensor_id>", to be executed by this API call. It reads

      numerical_grad.reset(numerical_grad_raw);
    }

    TF_Tensor* numerical_tensor;
    s = GetValue(numerical_grad.get(), &numerical_tensor);
    ASSERT_EQ(errors::OK, s.code()) << s.message();
    auto num_elem_numerical = TF_TensorElementCount(numerical_tensor);

    TF_Tensor* analytical_tensor;
    s = GetValue(outputs[i], &analytical_tensor);
    ASSERT_EQ(errors::OK, s.code()) << s.message();

        ],
    }),
)

tf_cuda_library(
    name = "tf_tensor_internal",
    hdrs = [
        "tf_tensor.h",
        "tf_tensor_helper.h",
        "tf_tensor_internal.h",
    ],
    visibility = ["//tensorflow:internal"],
    deps = [
        ":c_api_macros",
        ":tensor_interface",
        ":tf_datatype",
        ":tf_status",

  // Convert from AbstractTensor to TF_Tensor.
  TF_Tensor* theta_tensor;
  TF_RETURN_IF_ERROR(GetValue(theta, &theta_tensor));

  // Get number of elements and fill data.
  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.

    std::vector<TFE_TensorHandle*> result;
    result.reserve(tensors_.size());
    for (const TensorHandlePtr& tensor : tensors_) {
      result.emplace_back(tensor.get());
    }
    return result;
  }

 private:
  ParallelTensor(const ParallelDevice& device,
                 std::vector<TensorHandlePtr> tensors,
                 absl::Span<const int64_t> shape, const TF_DataType dtype)
      : device_(device),

    status->status = cc_status;
    return nullptr;
  }

  tensorflow::Tensor* tensor = var_info->var_info.var()->tensor();
  TF_Tensor* result_tensor =
      tensorflow::TF_TensorFromTensor(*tensor, &cc_status);
  status->status = cc_status;
  return result_tensor;
}

void TF_DeleteVariableInfo(TF_VariableInfo* var_info) {
  if (var_info != nullptr) {
    delete var_info;