Estimate the floating-point operations of a single attention layer and of the whole model from sequence length, head count and hidden size.
注意力和序列長度平方相關,長序列代價高。