Gradient* grad_tensor = grad[grad_index];
    if (grad_tensor != nullptr) {
      int64_t tensor_id = vspace_.TensorId(grad_tensor);
      targets.push_back(tensor_id);
      if (sources_set.find(tensor_id) != sources_set.end()) {
        sources_that_are_targets.emplace(
            tensor_id, vspace_.TapeTensorFromGradient(grad_tensor));
      }
      Gradient* in_grad = in_grads[grad_index];

    for (const int itensor : operations_[ioperation].tensors) {
      if (const auto& tensor = tensors_[itensor];
          tensor.first_use() >= begin && tensor.first_use() < end &&
          tensor.last_use() > peak_loc) {
        for (const int ioperation : tensor.operations) {
          // We return the earliest dependence on any output tensor.
          if (ioperation > peak_loc && ioperation < best) {

// Tensor types
def TF_BoolTensor : TensorOf<[TF_Bool]>;

def TF_IntTensor : TensorOf<[TF_Int]>;
def TF_Int4Tensor : TensorOf<[TF_Int4]>;
def TF_Int8Tensor : TensorOf<[TF_Int8]>;
def TF_Int16Tensor : TensorOf<[TF_Int16]>;
def TF_Int32Tensor : TensorOf<[TF_Int32]>;
def TF_Int64Tensor : TensorOf<[TF_Int64]>;
def TF_I32OrI64Tensor : TensorOf<[TF_I32OrI64]>;

    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),

                                                       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
// from a placeholder node "arg_tensor_enqueue_<tensor_id>".
//

// model.
table SubGraph {
  // A list of all tensors used in this subgraph.
  tensors:[Tensor];

  // Indices of the tensors that are inputs into this subgraph. Note this is
  // the list of non-static tensors that feed into the subgraph for inference.
  inputs:[int];

  // Indices of the tensors that are outputs out of this subgraph. Note this is
  // the list of output tensors that are considered the product of the

    auto* ret = emptyTensor->tensor;
    delete emptyTensor;
    return ret;
  }

  Tensor tensor;
  if (!tensor.CopyFrom(src, src.shape())) {
    return nullptr;
  }
  return new tensorflow::TensorInterface(std::move(tensor));
}

// Non-static for testing.
TF_Tensor* TF_TensorFromTensor(const tensorflow::Tensor& src, Status* status) {

  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());

  DataType data_type() const { return data_type_; }
  Status status() const { return status_; }
  const Tensor& tensor() const { return tensor_; }

 private:
  Status status_;
  Output output_ = Output(Operation(nullptr), 0);
  Tensor tensor_;
  const std::string node_name_ = "";
  int32 index_ = 0;
  DataType data_type_ = DT_INVALID;
};

          (TF_SubOp $max_range, $min_range),
          (TF_ConstOp (GetF32Scalar<255>)))),
      $min_range),
    /*truncate=*/ConstBoolAttrFalse),
  [(TensorOf<[TF_Qint8, TF_Quint8]> $input),
   (TF_Float32Tensor $min_range),
   (TF_Float32Tensor $max_range),
   (TensorOf<[TF_Float32, TF_Bfloat16]> $result)]>;

//===----------------------------------------------------------------------===//
// Difference op patterns.