Sample multiplexing reduces cost and batch effects in population based large-scale single-cell genomics studies but requires accurate and scalable computational demultiplexing. Existing genotype-based methods, such as demuxlet, provide high accuracy but can be computationally slow and memory intensive as the number of cells, donors, and informative variants increases. Here, we introduce fastdemux, a scalable genotype-based demultiplexing framework based on a diagonal linear discriminant analysis (DLDA) model that substantially improves computational efficiency while maintaining accurate donor assignment. Using a pooled single-cell RNA-seq dataset from unrelated donors, we benchmarked fastdemux against demuxlet, vireo, and demuxalot. fastdemux achieved comparable or improved demultiplexing accuracy while reducing runtime and peak memory usage by orders of magnitude relative to alternative methods. Performance remained robust across varying sequencing depths and genotype SNP filtering thresholds. In addition, the DLDA framework naturally extends to doublet and higher-order multiplet detection. We also show that fastdemux works well with scATAC-seq data where genetic variants are more sparsely covered. Together, these results establish fastdemux as an efficient and scalable solution for genetic demultiplexing of pooled single-cell datasets.