Use max/min

src/transformers/models/deprecated/gptsan_japanese/tokenization_gptsan_japanese.py

@@ -495,7 +495,7 @@ def checku2e(x):
                         candidates.append((self.vocab[wd], wd, e))
             if len(candidates) > 0:
                 # the smallest token_id is adopted
-                _, wd, e = sorted(candidates, key=lambda x: x[0])[0]
+                _, wd, e = min(candidates, key=lambda x: x[0])
                 result.append(wd)
                 pos = e
             else:

src/transformers/models/gpt_neox_japanese/tokenization_gpt_neox_japanese.py

@@ -318,7 +318,7 @@ def checku2e(x):
                         candidates.append((self.vocab[wd], wd, e))
             if len(candidates) > 0:
                 # the smallest token_id is adopted
-                _, wd, e = sorted(candidates, key=lambda x: x[0])[0]
+                _, wd, e = min(candidates, key=lambda x: x[0])
                 result.append(wd)
                 pos = e
             else:

src/transformers/models/ovis2/image_processing_ovis2.py

@@ -169,10 +169,10 @@ def get_min_tile_covering_grid(
 
     if sufficient_covering_grids:
         # Prefer fewer tiles and higher covering ratio
-        return sorted(sufficient_covering_grids, key=lambda x: (x[0][0] * x[0][1], -x[1]))[0][0]
+        return min(sufficient_covering_grids, key=lambda x: (x[0][0] * x[0][1], -x[1]))[0]
     else:
         # Fallback: prefer higher covering even if below threshold
-        return sorted(evaluated_grids, key=lambda x: (-x[1], x[0][0] * x[0][1]))[0][0]
+        return min(evaluated_grids, key=lambda x: (-x[1], x[0][0] * x[0][1]))[0]
 
 
 class Ovis2ImageProcessor(BaseImageProcessor):

src/transformers/tokenization_mistral_common.py

@@ -1789,7 +1789,7 @@ def from_pretrained(
                 if "tekken.json" in valid_tokenizer_files:
                     tokenizer_file = "tekken.json"
                 else:
-                    tokenizer_file = sorted(valid_tokenizer_files)[-1]
+                    tokenizer_file = max(valid_tokenizer_files)
                 logger.warning(
                     f"Multiple tokenizer files found in directory: {pretrained_model_name_or_path}. Using {tokenizer_file}."
                 )

tests/trainer/test_trainer.py

@@ -3478,7 +3478,7 @@ def test_resume_training_with_randomness(self):
             checkpoints = [d for d in os.listdir(tmp_dir) if d.startswith("checkpoint-")]
             # There should be one checkpoint per epoch.
             self.assertEqual(len(checkpoints), 3)
-            checkpoint_dir = sorted(checkpoints, key=lambda x: int(x.replace("checkpoint-", "")))[0]
+            checkpoint_dir = min(checkpoints, key=lambda x: int(x.replace("checkpoint-", "")))
 
             trainer.train(resume_from_checkpoint=os.path.join(tmp_dir, checkpoint_dir))
             (a1, b1) = trainer.model.a.item(), trainer.model.b.item()

utils/add_pipeline_model_mapping_to_test.py

@@ -134,7 +134,7 @@ def find_test_class(test_file):
             break
     # Take the test class with the shortest name (just a heuristic)
     if target_test_class is None and len(test_classes) > 0:
-        target_test_class = sorted(test_classes, key=lambda x: (len(x.__name__), x.__name__))[0]
+        target_test_class = min(test_classes, key=lambda x: (len(x.__name__), x.__name__))
 
     return target_test_class
 

utils/create_dummy_models.py

@@ -389,7 +389,7 @@ def get_tiny_config(config_class, model_class=None, **model_tester_kwargs):
             # This is to avoid `T5EncoderOnlyModelTest` is used instead of `T5ModelTest`, which has
             # `is_encoder_decoder=False` and causes some pipeline tests failing (also failures in `Optimum` CI).
             # TODO: More fine grained control of the desired tester class.
-            model_tester_class = sorted(tester_classes, key=lambda x: (len(x.__name__), x.__name__))[0]
+            model_tester_class = min(tester_classes, key=lambda x: (len(x.__name__), x.__name__))
     except ModuleNotFoundError:
         error = f"Tiny config not created for {model_type} - cannot find the testing module from the model name."
         raise ValueError(error)

utils/deprecate_models.py

@@ -37,7 +37,7 @@ def get_last_stable_minor_release():
     last_stable_minor_releases = [
         release for release in release_data["releases"] if release.startswith(last_major_minor)
     ]
-    last_stable_release = sorted(last_stable_minor_releases, key=version.parse)[-1]
+    last_stable_release = max(last_stable_minor_releases, key=version.parse)
 
     return last_stable_release