Fix Causality Handling in Flash Attention to Support Bidirectional Attention (#39707)

Fix the is_causal logic to enable bidirectional attention

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

src/transformers/integrations/flash_attention.py

@@ -58,8 +58,10 @@ def flash_attention_forward(
         else:
             target_dtype = next(layer for layer in module.modules() if isinstance(layer, torch.nn.Linear)).weight.dtype
 
-    # FA2 always relies on the value set in the module, so remove it if present in kwargs to avoid passing it twice
-    kwargs.pop("is_causal", None)
+    # Instead of relying on the value set in the module directly, we use the is_causal passed in kwargs if it is presented
+    is_causal = kwargs.pop("is_causal", None)
+    if is_causal is None:
+        is_causal = module.is_causal
 
     attn_output = _flash_attention_forward(
         query,
@@ -67,7 +69,7 @@ def flash_attention_forward(
         value,
         attention_mask,
         query_length=seq_len,
-        is_causal=module.is_causal,
+        is_causal=is_causal,
         dropout=dropout,
         softmax_scale=scaling,
         sliding_window=sliding_window,