Add fast path for bidirectional mask creation to fix regression

src/transformers/masking_utils.py

@@ -405,6 +405,16 @@ def sdpa_mask_recent_torch(
     if allow_is_bidirectional_skip and _ignore_bidirectional_mask_sdpa(padding_mask):
         return None
 
+    # vmap can incur performance issues as reported in #41566 for bidirectional mask as we only need to expand the
+    # padding mask. Thus, we allow early exit here if we do not detect any modification to the base mask function
+    if mask_function is bidirectional_mask_function:
+        if padding_mask is not None:
+            # used for slicing without data-dependent slicing
+            mask_indices = torch.arange(kv_length, device=cache_position.device) + kv_offset
+            return padding_mask[:, None, None, mask_indices].expand(-1, -1, q_length, -1)
+        else:
+            return torch.ones(batch_size, 1, q_length, kv_length, dtype=torch.bool, device=cache_position.device)
+
     # Similar to `kv_arange = torch.arange(start=kv_offset, end=kv_offset + kv_length, device=cache_position.device)`
     # but without data-dependent slicing (i.e. torch.compile friendly)
     kv_arange = torch.arange(kv_length, device=cache_position.device)
@@ -485,6 +495,14 @@ def sdpa_mask_older_torch(
     if allow_is_bidirectional_skip and _ignore_bidirectional_mask_sdpa(padding_mask):
         return None
 
+    # vmap can incur performance issues as reported in #41566 for bidirectional mask as we only need to expand the
+    # padding mask. Thus, we allow early exit here if we do not detect any modification to the base mask function
+    if mask_function is bidirectional_mask_function:
+        if padding_mask is not None:
+            return padding_mask[:, None, None, :].expand(-1, -1, q_length, -1)
+        else:
+            return torch.ones(batch_size, 1, q_length, kv_length, dtype=torch.bool, device=cache_position.device)
+
     # Similar to `kv_arange = torch.arange(start=kv_offset, end=kv_offset + kv_length, device=cache_position.device)`
     # but without data-dependent slicing (i.e. torch.compile friendly)
     kv_arange = torch.arange(kv_length, device=cache_position.device)