Many of the proposed multicast congestion avoidance algorithms are single-rate where heterogeneity is accommodated by adjusting the transmission rate as a response to the worst receiver in the group. Due to the Internet heterogeneity, a single-rate congestion co ntrol affects the overall satisfaction of the receivers in a multicast session. In this paper, we propose a multi-rate replicated scheme where some receivers (instead of the source) are designated to perform data replication for other receivers with lower capacity. To be more scalable and to minimize the bandwidth consumption due to data replication, the partitioning algorithm is per- formed on-the-fly by the routers depending on the feedback they receive. Neither a prior estimation of the receivers capacity is necessary nor a complex computation is required to execute our partitioning algorithm.