// Causes the calling thread to block till all ops dispatched in this executor
// have been executed. Note that "execution" here refers to kernel execution /
// scheduling of copies, etc. Similar to sync execution, it doesn't guarantee
// that lower level device queues (like GPU streams) have been flushed.
//
// This call may not block for execution of ops enqueued concurrently with this
// call.

                    int expected_max_outputs,
                    CancellationManager& cancellation_manager,
                    std::optional<int64_t> step_id = std::nullopt) const;

  // Blocks until the previous `StartExecute` has run `TFE_Execute` on each
  // device. If is_async=false (constructor argument) this means the ops have
  // run and have results. If is_async=true it means that all of the

// you want the latter, use TFE_TensorHandleBackingDeviceName. This function
// will block till the operation that produces `h` has completed.
TF_CAPI_EXPORT extern const char* TFE_TensorHandleDeviceName(
    TFE_TensorHandle* h, TF_Status* status);

// Returns the name of the device in whose memory `h` resides.
//
// This function will block till the operation that produces `h` has completed.

// "fifo_queue_dequeue_<tensor_id>", to be executed by this API call.

// Caller must call TF_DeleteTensor() over the returned tensor. If the queue is
// empty, this call is blocked.
//
// Tensors are enqueued via the corresponding TF enqueue op.
// TODO(hongm): Add support for `timeout_ms`.
TF_CAPI_EXPORT extern TF_Tensor* TF_DequeueNamedTensor(TF_Session* session,

// (also known as adjoint function) to compute, given downstream gradients,
// upstream gradients.
//
// TODO(apassos) provide concrete template instantiations for TFE_TensorHandle
// specialization, which is blocked by quite a few things needing to loop back
// into python now.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class VSpace {
 public:
  virtual ~VSpace() {}

  };

  /// \brief The block map type for the file block cache.
  ///
  /// The block map is an ordered map from Key to Block.
  typedef std::map<Key, std::shared_ptr<Block>> BlockMap;

  /// Prune the cache by removing files with expired blocks.
  void Prune() ABSL_LOCKS_EXCLUDED(mu_);

  bool BlockNotStale(const std::shared_ptr<Block>& block)
      ABSL_EXCLUSIVE_LOCKS_REQUIRED(mu_);

  // Add `devices` into context's device manager. Context's device manager
  // will take ownership and maintain devices' lifetime.
  virtual Status AddDevices(std::vector<std::unique_ptr<Device>> devices) = 0;

  // Block until all pending nodes are finished.
  virtual Status AsyncWait() = 0;

  // Add a function (serialized FunctionDef protocol buffer) so that it can
  // be executed as an op. Return error if the function with the same name