I think it’s considered an implementation detail so it’s just not mentioned. A lot of models still work if you round the values to FP16, which personally I find surprising.