Yes, but I imagine traditional compiler optimisations like inlining, unrolling, etc. are easier to apply to serial code.

Well, to the compiler, the parallelism probably just looks like an opaque call to a higher-order function. The bulk of the program will still be serial code, which will be subject to all the normal optimisations. I talked a bit with one of the main authors of MPL, and he mentioned that the existing optimisation passes in MLton just kept working without any nontrivial modifications.