Yeah, the problem is that distributions cannot infer the resulting shape of a given sample from their tensor parameter values. That means that a normal distribution cannot infer the shape it should have from its mu and sigma tensor parameters. This is why, in order to help sampling, we must usually declare the distribution’s shape by hand.
The reason is that distribution instances are not tensors, nor Ops, so there is no inplace mechanism that allows them to do the symbolic math required to infer shapes.