The objectives of this study were to evaluate the prediction performance of the single-step genomic BLUP method using a multi-trait random regression model in genomic evaluation for milk production traits of Chinese Holsteins, and investigate how parameters w, τ, and ω used in the construction of the combined relationship matrix (H) affected prediction accuracy and bias. A total of 2.8 million test-day records from 0.2 million cows were available for milk, protein, and fat yields. Pedigree information included 0.3 million animals and 7,577 of them were genotyped with medium-density single nucleotide polymorphism marker panels. Genotypes were imputed into Geneseek Genomic Profiler HD (GeneSeek, Lincoln, NE) including 77K markers. A reduced data set for evaluating models was extracted from the full data set by removing test-day records from the last 4 yr. Bull and cow validation populations were constructed for each trait. We evaluated the prediction performance of the multiple-trait multiple-lactation random regression single-step genomic BLUP (RR-ssGBLUP) models with different values of parameters w, τ, and ω in the H matrix, taking consideration of inbreeding. We compared RR-ssGBLUP with the multiple-trait multiple-lactation random regression model based on pedigree and genomic BLUP. De-regressed proofs for 305-d milk, protein, and fat yields averaged over 3 lactations, which were calculated from the full data set, were used for posteriori validations. The results showed that RR-ssGBLUP was feasible for implementation in breeding practice, and its prediction performance was superior to the other 2 methods in the comparison, including prediction accuracy and unbiasedness. For bulls, RR-ssGBLUP models with $w_{0.05}{\tau }_{2.0}{\omega }_{1.0},w_{0.05}{\tau }_{2.5}{\omega }_{1.0}$, and $w_{0.1}{\tau }_{1.6}{\omega }_{0.4}$ achieved the best performance for milk, protein, and fat yields, respectively. For cows, the RR-ssGBLUP with $w_{0.2}{\tau }_{1.6}{\omega }_{0.4}$ performed the best for all 3 traits. The H matrix constructed with larger τ and smaller ω gave better convergence in solving mixed model equations. Among different RR-ssGBLUP models, the differences in validation accuracy were small. However, the regression coefficient indicating prediction bias varied substantially. The increase of w and τ, and decrease of ω, led to an increase in the regression coefficient. The results demonstrated RR-ssGBLUP is a good alternative to the multi-step approach, but the optimal choice of parameters should be found via preliminary validation study to achieve the best performance.