We consider multiple-input multiple-output decode-and-forward relay systems in Rayleigh fading channels under the partial channel state information (CSI) that the channel statistics of the source-relay (SR) link and the instantaneous CSI of the source-destination and relay-destination links are known at the destination. In this paper, we propose a new near maximum likelihood (near-ML) decoder with two-level pairwise error probability (near-ML-2PEP) which uses the average PEP instead of the exact PEP. Then, we theoretically prove that the near-ML and near-ML-2PEP decoders achieve the maximum diversity, which is confirmed by Monte Carlo simulations. Moreover, we show that the near-ML-2PEP decoder can also achieve the maximum diversity by substituting the average PEP with the values that represent the error performance of the SR link.