Variants with very low allele count (<5) seem to cause problems with tensorqtl, producing dozens of independent QTLs for some genes.
There could be a step that inspects the genotypes and reports issues. Maybe a standalone script whose output is a non-functioning input for QTL mapping.
Or, tensorqtl does have a MAF filtering option, so the default could be a threshold computed from the number of samples to ensure an allele count of, e.g., >=10.