Usually for 25Gbps line-rate, channel insertion-loss will be high, so DFE will perform better than LPM on most cases. But, as mentioned in UG578 Chapter4 DFE needs a random data for its auto adaptation module to work properly. Without a random data sequence, BER may increased.
Scrambling the RAW data before sending into GTY TX may help a lot. ( You don't need to scramble comma/alignment codes )