In compute fep you need to use a variable giving you the perturbation (the delta) on the parameters you are changing,
so you need to define a depsilon which in this case would be is -epsilon/20 at each stage.
Since you are not using a soft-core pair style (e.g. lj/cut/soft), lambda has no effect.
Maybe you should consider a soft repulsive potential to avoid spurious effects when epsilon approaches 0.
Agilio