// If coordination service is enabled, the key-value will be stored on the
// leader and become accessible to all workers in the cluster.
// Currently, a config key can only be set with one value, and subsequently
// setting the same key will lead to errors.
//
// Note that the key-values are only expected to be used for cluster
// configuration data, and should not be used for storing a large amount of data

  // followed by `TFE_ContextAsyncWait` to clear pending operations. It will be
  // used to cancel all other operations if any fails.
  //
  // Set step_id to configure the step id used for rendezvous creation. step id
  // of value -1 is reserved for global rendezvous and should not be set here.
  void StartExecute(TFE_Context* context,
                    const std::vector<ParallelTensor*>& inputs,

// `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>".
//
// `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.

// adding gradients, getting zeroes, etc. Currently cannot be implemented
// without using tensorflow python code, hence left unspecified here.
//
// Gradient is the type returned by gradient functions. In Python TF it's either
// Tensor or IndexedSlices or None, which here we map to nullptr. Gradients need
// to allow their size to be computed and they need to be passable to a backward

// that shares the underlying buffer. Otherwise, it currently requires at least
// one of the source or destination devices to be CPU (i.e., for the source or
// destination tensor to be placed in host memory).
// If async execution is enabled, the copy may be enqueued and the call will
// return "non-ready" handle. Else, this function returns after the copy has
// been done.

  ///
  /// Thread safety:
  /// The iterator and timestamp fields should only be accessed while holding
  /// the block-cache-wide mu_ instance variable. The state variable should only
  /// be accessed while holding the Block's mu lock. The data vector should only
  /// be accessed after state == FINISHED, and it should never be modified.
  ///
  /// In order to prevent deadlocks, never grab the block-cache-wide mu_ lock

  DEVICE_PLACEMENT_EXPLICIT = 0,
  // Copy the tensor to the right device but log a warning.
  DEVICE_PLACEMENT_WARN = 1,
  // Silently copy the tensor, which has a performance cost since the operation
  // will be blocked till the copy completes. This is the default policy.
  DEVICE_PLACEMENT_SILENT = 2,
  // Placement policy which silently copies int32 tensors but not other dtypes.
  DEVICE_PLACEMENT_SILENT_FOR_INT32 = 3,
};