Merge branch 'main' into patch-1

Author: saattrupdan

src/transformers/modeling_flash_attention_utils.py

@@ -280,7 +280,7 @@ def _flash_attention_forward(
     query_states: torch.Tensor,
     key_states: torch.Tensor,
     value_states: torch.Tensor,
-    attention_mask: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
     query_length: int,
     is_causal: bool,
     dropout: float = 0.0,
@@ -308,7 +308,7 @@ def _flash_attention_forward(
             Input key states to be passed to Flash Attention API
         value_states (`torch.Tensor`):
             Input value states to be passed to Flash Attention API
-        attention_mask (`torch.Tensor`):
+        attention_mask (`torch.Tensor`, *optional*):
             The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
             position of padding tokens and 1 for the position of non-padding tokens.
         dropout (`float`):