@@ -2962,12 +2962,6 @@ def __call__(
         # perturbed path (identity attention)
         batch_size, sequence_length, _ = hidden_states_ptb.shape
 
-        if attention_mask is not None:
-            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-            # scaled_dot_product_attention expects attention_mask shape to be
-            # (batch, heads, source_length, target_length)
-            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-
         if attn.group_norm is not None:
             hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)
 
@@ -3070,12 +3064,6 @@ def __call__(
         # perturbed path (identity attention)
         batch_size, sequence_length, _ = hidden_states_ptb.shape
 
-        if attention_mask is not None:
-            attention_mask = attn.prepare_attention_mask(attention_mask, sequence_length, batch_size)
-            # scaled_dot_product_attention expects attention_mask shape to be
-            # (batch, heads, source_length, target_length)
-            attention_mask = attention_mask.view(batch_size, attn.heads, -1, attention_mask.shape[-1])
-
         if attn.group_norm is not None:
             hidden_states_ptb = attn.group_norm(hidden_states_ptb.transpose(1, 2)).transpose(1, 2)