[`FA2`] Fix flash attention 2 fine-tuning with Falcon (#26852)

fix fa2 + dropout issue

src/transformers/models/falcon/modeling_falcon.py

@@ -606,7 +606,7 @@ class FalconFlashAttention2(FalconAttention):
         if alibi is not None:
             raise ValueError("`alibi` is not supported when `use_flash_attn` is True")
 
-        attn_dropout = self.attention_dropout if self.training else 0.0
+        attn_dropout = self.config.attention_dropout if self.training else 0.0
 
         # In PEFT, usually we cast the layer norms in float32 for training stability reasons
         # therefore the input hidden states gets silently casted in float32. Hence, we need

tests/test_modeling_common.py

@@ -2810,6 +2810,10 @@ class ModelTesterMixin:
 
                 self.assertTrue(torch.allclose(logits_fa[1:], logits[1:], atol=4e-2, rtol=4e-2))
 
+                # check with inference + dropout
+                model.train()
+                _ = model_fa(dummy_input, attention_mask=dummy_attention_mask, output_hidden_states=True)
+
     @require_flash_attn
     @require_torch_gpu
     @mark.flash_attn_test