Question about mask used in TransformerModel.forward

[fredlumm]

def forward(self, src):
    mask = self._generate_square_subsequent_mask(len(src)).to(self.device)
    self.src_mask = mask

    src = self.input_emb(src) * math.sqrt(self.ninp)
    src = self.pos_encoder(src)
    output_enc = self.encoder(src, mask=self.src_mask)
    output_dec = self.decoder(output_enc)
    return F.log_softmax(output_dec, dim=-1), output_enc

it seems mask is always a causal mask, no matter it is an arm or llada model. According to the llada paper，it should not use it.
Correct me if I am wrong