void GradientTape<Gradient, BackwardFunction, TapeTensor>::DeleteTrace(
    int64_t tensor_id) {
  auto it = tensor_usage_.find(tensor_id);
  if (it == tensor_usage_.end()) {
    return;
  }
  it->second--;
  if (it->second != 0) {
    return;
  }
  tensor_usage_.erase(it);
  auto tensor_op_it = tensor_tape_.find(tensor_id);
  if (tensor_op_it == tensor_tape_.end()) {
    return;
  }

      TF_Status* status);

  size_t num_tensors() const { return tensors_.size(); }
  TFE_TensorHandle* tensor(size_t index) const { return tensors_[index].get(); }

  // If the `shape` argument to `FromTensorHandles` is specified, returns that.
  //
  // Otherwise if all of the tensors have the same shape, returns that via the
  // `shape` output argument. This blocks waiting for async tensors, may return

  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),

    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);
}

}

// Converts `noutputs` and `outputs` into `outputs_tensors` and does various
// checks while doing so.
Status ProcessOutputs(const TF_Graph* fn_body, const char* fn_name,
                      int noutputs, const TF_Output* outputs,
                      std::vector<OutputTensor>* output_tensors)
    TF_EXCLUSIVE_LOCKS_REQUIRED(fn_body->mu) {
  output_tensors->reserve(noutputs);
  for (int i = 0; i < noutputs; ++i) {

  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());

    // below. Allocate enough space so that no reallocation happens, which will
    // make the pointers invalid.
    all_input_tensors.reserve(num_inputs);
    for (int i = 0; i < num_inputs; ++i) {
      if (input_tensors[i] == nullptr) continue;
      all_input_tensors.emplace_back();
      Tensor& input_tensor = all_input_tensors.back();
      status->status = TF_TensorToTensor(input_tensors[i], &input_tensor);

// an error.
//
// All component tensors must have the same dtype. Currently they must also have
// the same shape, although this requirement may be relaxed in the future.
//
// `device_name` must not name an existing physical or custom device (see
// the documentation for TFE_RegisterCustomDevice for more information).
//
// Tensors may be copied on or off the device explicitly using

                                                   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;

// from a placeholder node "arg_tensor_enqueue_<tensor_id>".
//
// `tensor` is still owned by the caller. This call will be blocked if the queue
// has reached its capacity, and will be unblocked when the queued tensors again
// 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,