Hi @Lennon_Lee,
The output predicted abundances are the result of multiplying the normalized # of 16S reads for each ASV by the number of KO copies per predicted genome for that ASV.
The output is non-integers because the the normalized 16S read counts are generated by dividing the raw 16S abundances by the predicted number of 16S copies in that ASV's genome.
I think it's reasonable to round the counts if you want to use DEseq2. I would be more worried about which model to use with DEseq2 because the default settings aren't appropriate for microbiome data in general (you are at risk of generating many false positives in my experience with either taxonomic or functional microbiome data with DEseq2).
Hopefully that helps!
Best,
Gavin