TF_Status* status);

// Helper to un-pack `num_replicas` TFE_TensorHandles from one parallel handle.
template <std::size_t num_replicas>
void ExtractPerDeviceValues(
    TFE_Context* context, TFE_TensorHandle* input,
    std::array<TensorHandlePtr, num_replicas>* components, TF_Status* status);

// Helper to pack `num_replicas` TFE_TensorHandles into one parallel handle.
template <std::size_t num_replicas>

  std::vector<TFE_TensorHandle*> tensors() const {
    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,

// (for debugging/performance-analysis) by thread_name.
//
// The given param (which may be null) is passed to work_func when the thread
// starts. In this way, data may be passed from the thread back to the caller.
//
// Caller takes ownership of the result and must call TF_JoinThread on it
// eventually.
TF_CAPI_EXPORT extern TF_Thread* TF_StartThread(const TF_ThreadOptions* options,

  std::vector<const char*> raw_strs;
  std::vector<size_t> lengths;
  raw_strs.reserve(values.size());
  lengths.reserve(values.size());
  for (const auto& s : values) {
    raw_strs.emplace_back(s.data());
    lengths.emplace_back(s.size());
  }
  return SetAttrStringList(attr_name,
                           reinterpret_cast<const void**>(raw_strs.data()),
                           lengths.data(), values.size());
}

        // GradientTape which uses ones.
        Gradient* zero = input_tensors[target_index].ZerosLike();
        new_zeros.push_back(zero);
        in_grads.push_back(zero);
      } else {
        in_grads.push_back(nullptr);
      }
    } else {
      in_grads.push_back(current_grad->second);
    }
  }

  // Avoid infinite recursion. Whichever forward function we run, it'll end up

      lru_list_.erase(insert.first->second.lru_iterator);
      insert.first->second = entry;
    } else if (max_entries_ > 0 && cache_.size() > max_entries_) {
      cache_.erase(lru_list_.back());
      lru_list_.pop_back();
    }
  }

  bool DeleteLocked(const std::string& key) ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_) {
    auto it = cache_.find(key);
    if (it == cache_.end()) {
      return false;
    }

  // custom device.
  //
  // Many devices will want to simply return an "unimplemented" status
  // here. This is the default behavior if `pack` is null when passed to
  // TFE_RegisterCustomDevice.
  TFE_TensorHandle* (*pack)(TFE_Context* context, TFE_TensorHandle** handles,
                            int num_handles, TF_Status* s,
                            void* device_info) = nullptr;

///
/// The `TF_InitPlugin` function is used by plugins to set up the data
/// structures that implement this interface, as presented in Section 2. In
/// order to not have plugin shared objects call back symbols defined in core
/// TensorFlow, `TF_InitPlugin` has a `TF_FilesystemPluginInfo` argument which
/// the plugin must fill (using the `TF_SetFilesystemVersionMetadata` for the