// nodes in parallel.
// in_flight_nodes_limit: when is_async is true, this value controls the
// maximum number of in flight async nodes. Enqueuing of additional async ops
// after the limit is reached blocks until some inflight nodes finishes.
// The effect is bounding the memory held by inflight TensorHandles that are
// referenced by the inflight nodes.
// A recommended value has not been established.

  // `StartExecute`s will deadlock).
  //
  // If `is_async=false` (constructor argument), `cancellation_manager` must
  // live until `Join` finishes. If `is_async=true` it must live until `Join` is
  // followed by `TFE_ContextAsyncWait` to clear pending operations. It will be
  // used to cancel all other operations if any fails.
  //

//
// Keeps references to Tensors watched via Watch and computed in Accumulate
// corresponding to output_tensors, and releases these references in its
// destructor. However, waiting until the destructor runs loses the memory
// efficiency of forward-mode autodiff. Instead, language bindings should call
// DeleteGradient as soon as a Tensor which was `Watch`ed or was an output

  // 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

            (IsCacheEnabled() ? "enabled" : "disabled"));
  }

  ~RamFileBlockCache() {
    if (pruning_thread_) {
      stop_pruning_thread_.Notify();
      // Destroying pruning_thread_ will block until Prune() receives the above
      // notification and returns.
      pruning_thread_.reset();
    }
  }

  /// Read `n` bytes from `filename` starting at `offset` into `buffer`. It