If you run HAProxy in multiprocess mode, your HTTP requests are probably load-balanced between the different processes. The stats are not shared between processes. So the value will differ between processes. Try to read the value of haproxy_process_relative_process_id to know where a request was processed.
In nbproc, you need to get stats from each process individually and aggregate the results. Prometheus can handle the aggregation, so you need to update your HAProxy configuration to have a listen socket bound on each process. Another solution is to use threads instead of processes. Note that the same is true for the stats socket.