generation/stopping_criteria: short-circuit StoppingCriteriaList when all sequences are done

src/transformers/generation/stopping_criteria.py

@@ -498,6 +498,8 @@ def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwa
         is_done = torch.full((input_ids.shape[0],), False, device=input_ids.device, dtype=torch.bool)
         for criteria in self:
             is_done = is_done | criteria(input_ids, scores, **kwargs)
+            if is_done.all():
+                break
         return is_done
 
     @property

tests/generation/test_stopping_criteria.py

@@ -25,14 +25,15 @@
     import torch
 
     from transformers.generation import (
-        ConfidenceCriteria,
-        EosTokenCriteria,
-        MaxLengthCriteria,
-        MaxTimeCriteria,
-        StoppingCriteriaList,
-        StopStringCriteria,
-        validate_stopping_criteria,
-    )
+    ConfidenceCriteria,
+    EosTokenCriteria,
+    MaxLengthCriteria,
+    MaxTimeCriteria,
+    StoppingCriteria,
+    StoppingCriteriaList,
+    StopStringCriteria,
+    validate_stopping_criteria,
+)
 
 
 @require_torch
@@ -63,6 +64,29 @@ def test_list_criteria(self):
         input_ids, scores = self._get_tensors(10)
         self.assertTrue(all(criteria(input_ids, scores)))
 
+    def test_list_criteria_short_circuits_when_all_done(self):
+        """Verify that StoppingCriteriaList stops evaluating remaining criteria once all
+        sequences are already marked done, avoiding unnecessary computation."""
+        call_count = 0
+
+        class CountingCriteria(StoppingCriteria):
+            def __call__(self, input_ids, scores, **kwargs):
+                nonlocal call_count
+                call_count += 1
+                return torch.zeros(input_ids.shape[0], dtype=torch.bool, device=input_ids.device)
+
+        # MaxLengthCriteria fires immediately (length == max_length), so CountingCriteria should be skipped
+        input_ids, scores = self._get_tensors(10)
+        criteria = StoppingCriteriaList(
+            [
+                MaxLengthCriteria(max_length=10),
+                CountingCriteria(),
+            ]
+        )
+        result = criteria(input_ids, scores)
+        self.assertTrue(all(result))
+        self.assertEqual(call_count, 0, "CountingCriteria should not be called when first criterion marks all done")
+
     def test_max_length_criteria(self):
         criteria = MaxLengthCriteria(max_length=10)