Mentioned by James in theano-dev (thread: Inconsistent behavior in integer divisions):

'tensor.div(a,b,dtype='float64') might also be a good thing to have if you want to specify it. We could also make an optimization that replaces cast(div(a,b)) with a div with a built-in cast.'

I think such ideas may extend to other arithmetic operations as well, when we want the result to be of higher precision than the inputs.