When I do fineweb-edu for classifier scoring, how do I overlap the tokenizer with the process of model infer?

[HuaYZhao]

trafficstars

I find that when tokenize, gpu utilization is always zero