Imputation of missing genotypes is essential for genomic studies, yet its application to genotyping-by-sequencing (GBS) datasets is constrained by high missingness and a lack of suitable reference panels. While recent reference-free tools address some of these issues, they often rely on shallow models or assumptions about linkage disequilibrium (LD) that limit their generalizability. Here, we present DeepGBSImpute, a reference-free genotype imputation framework based on a multi-head self-attention transformer architecture with LD-aware attention and positional encoding. Unlike existing approaches, DeepGBSImpute is optimized to model complex dependencies in sparse, uneven GBS data without relying on external haplotype references or phasing. On a real dataset with over 65% missing genotypes, DeepGBSImpute achieved > 99.9% accuracy and balanced F1 scores across different genotype classes. The pipeline’s modular, GPU-accelerated design supports scalable, window-based inference with interpretability and reproducibility features. DeepGBSImpute represents a significant advance in reference-free genotype imputation, enabling broader application of GBS data in diverse populations and non-model organisms.