Does MuLan require attention masking for the audio part of the tower?

[14leeyuchieh]

I was taking a look at the forward implementation of AudioSpectrogramTransformer and notice it does not intake an attention mask or pass it to the underlying transformer

def forward(
        self,
        x,
        force_no_patch_dropout = False,
        return_all_layers = False
    ):

Is this a bug? Or does Mulan handle attention masking when batching audio of arbitrary length via just padding?