Fix falcon with SDPA, alibi but no passed mask (#30123)

* fix falcon without attention_mask & alibi

* add test

* Update tests/models/falcon/test_modeling_falcon.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

---------

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

tests/models/falcon/test_modeling_falcon.py

@@ -666,3 +666,27 @@ class FalconLanguageGenerationTest(unittest.TestCase):
         self.assertLess(unpadded_inputs.input_ids.shape[-1], padded_inputs.input_ids.shape[-1])  # left-padding exists
         self.assertEqual(unpadded_gen_text[0], expected_output)
         self.assertEqual(padded_gen_text[0], expected_output)
+
+    @slow
+    @require_torch_sdpa
+    def test_falcon_alibi_sdpa_matches_eager(self):
+        input_ids = torch.randint(0, 1000, (5, 20))
+
+        config = FalconConfig(
+            vocab_size=1000,
+            hidden_size=64,
+            num_hidden_layers=3,
+            num_attention_heads=4,
+            new_decoder_architecture=True,
+            alibi=True,
+        )
+
+        falcon = FalconForCausalLM(config)
+        falcon = falcon.eval()
+
+        with torch.no_grad():
+            # output_attentions=True dispatches to eager path
+            falcon_output_eager = falcon(input_ids, output_attentions=True)[0]
+            falcon_output_sdpa = falcon(input_ids)[0]
+
+        self.assertTrue(torch.allclose(falcon_output_eager, falcon_output_sdpa, atol=1e-3))