if bias:
            self.b_q = np.random.rand(num_hiddens)
            self.b_k = np.random.rand(num_hiddens)
            self.b_v = np.random.rand(num_hiddens)
            self.b_o = np.random.rand(num_hiddens)
        else:
            self.b_q = self.b_k = self.b_v = self.b_o = np.zeros(num_hiddens)