QUASI_RANDOM repeats same classes in one batch (not random?)

[ArneNx]

I am trying to run some ImageNet training in my own setup. Since I am not getting the same result reported in the literature, I am currently investigating what's going wrong. I found that when I use QUASI_RANDOM as the order option for the FFCV loader, I get non-random composition of batches. A batch of 16 images has only between 10 and 13 unique classes. This happens consistently and is highly unlikely with 1000 classes. Choosing RANDOM as the ordering solves this issue, but is significantly slower. Is there any explanation for this or is this a bug? I haven't finished my training yet, but I can already see a significant difference in validation performance between the two orderings.