To represent "abstract" graphs (those were the
equality for the nodes type is not used to check if two nodes of a graph
are the same), the library uses an internal counter. This counter must
be serialised along with the graphs and then it must be updated
correctly when graphs are unserialised to avoid creating a node with the
same identifier than o node in the unserialised graphs.

This is explained in the FAQ :
http://ocamlgraph.lri.fr/FAQ

But I bielieve that the FAQ got it wrong in case multiple graphs are
unserialised, or nodes are created before the unserialisation of other
nodes. In these situation using concrete graphs, which do not suffer
from this problem, is certainly easier (as advertised by the manual).
And I manually add identifiers to the nodes if I need many nodes with
the same label.