fix transpose on attn mask

Author: NeoLegends

i6_models/parts/conformer/mhsa_rel_pos.py

@@ -224,7 +224,7 @@ def forward(self, input_tensor: torch.Tensor, sequence_mask: torch.Tensor) -> to
             q_with_bias_u.transpose(-3, -2),  # [B, #heads, T, F']
             k.transpose(-3, -2),  # [B, #heads, T', F']
             v.transpose(-3, -2),  # [B, #heads, T, F']
-            attn_mask=attn_bd_mask_scaled.transpose(-3, -2),  # [B, #heads, T, T']
+            attn_mask=attn_bd_mask_scaled,  # [B, #heads, T, T']
             dropout_p=self.att_weights_dropout.p if self.training else 0.0,
             scale=scale,
         )  # [B, #heads, T, F']