surprising singleton behavior

[yaroslavvb]

Consider following, np.split is always a list of objects, but after going through Ray, the result is either a single Ray object or a list of Ray objects depending on whether shards is equal to 1

@ray.method(num_return_vals=shards)
        def compute_gradients(self, *weights):
            return np.split(array, shards)

[richardliaw]

Yeah this was a little odd - I'm not totally sure what the right thing to do here is.

One possibility is to force the semantics of method to always return a list of objects, and also enforce that the wrapped function also returns a list? In which case we could consider renaming ray.method into ray.split.

Thoughts? @robertnishihara