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)

      });
}
bool Tape::ShouldRecord(
    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);
}

  for (int component_index = 0; component_index < tensors_.size();
       ++component_index) {
    // TODO(allenl): Add a C API for summarizing tensors. Currently custom
    // devices limiting themselves to a C API (for ABI compatibility) would need
    // to implement summarization for component tensors themselves.
    ImmediateExecutionTensorHandle* component =
        tensorflow::unwrap(tensors_[component_index].get());

                                       TFE_TensorHandle* tensor,
                                       TF_Status* status, void* device_info) {
  TF_SetStatus(
      status, TF_UNIMPLEMENTED,
      absl::StrCat("Trying to copy a tensor ",
                   tensorflow::unwrap(tensor)->DebugString(),
                   " on to a parallel device. Pack non-parallel "
                   "tensors for each device into a parallel tensor explicitly.")

// some input tensors are not on that device.
// LINT.IfChange
// Note: Keep in sync with internal copy of enum in eager/context.h.
typedef enum TFE_ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  TFE_DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  TFE_DEVICE_PLACEMENT_WARN = 1,

// Note: Keep in sync with exported copy of enum in eager/c_api.h.
enum ContextDevicePlacementPolicy {
  // Running operations with input tensors on the wrong device will fail.
  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.

// drop below the capacity due to dequeuing.
//
// Tensors are dequeued via the corresponding TF dequeue op.
// TODO(hongm): Add support for `timeout_ms`.
TF_CAPI_EXPORT extern void TF_EnqueueNamedTensor(TF_Session* session,
                                                 int tensor_id,
                                                 TF_Tensor* tensor,

  TFE_DeleteContextOptions(opts);

  ASSERT_EQ(TF_OK, TF_GetCode(status.get())) << TF_Message(status.get());

  /* Want to test simple MatMul example with abstract tensors:
    [[1,2],   *  [[5,6],   =   [[19,22],
     [3,4]]       [7,8]]        [43,50]]
  */

  // Build 1st Matrix.
  int64_t dims[] = {2, 2};  // Matrices will be 2 x 2

                                             TF_Status* status,
                                             void* device_info);

  // Method to copy a tensor from the custom device to a target device.
  TFE_TensorHandle* (*copy_tensor_from_device)(TFE_Context* context,
                                               TFE_TensorHandle* tensor,

    status->status = tensorflow::errors::InvalidArgument(
        "DLPack doesn't support ", handle->TypeString(), " tensor");
    return nullptr;
  }
  const tensorflow::Tensor* tensor;
  status->status = handle->Tensor(&tensor);
  if (!status->status.ok()) {
    return nullptr;
  }
  return tensor;
}

// Deleter for DLManagedTensor
void DLManagedTensorDeleter(DLManagedTensor* arg) {
  TfDlManagedTensorCtx* owner =