Pipeline: no side-effects on model.config and model.generation_config 🔫 (#33480)

tests/pipelines/test_pipelines_common.py

@@ -31,6 +31,7 @@ from transformers import (
     AutoTokenizer,
     DistilBertForSequenceClassification,
     MaskGenerationPipeline,
+    T5ForConditionalGeneration,
     TextClassificationPipeline,
     TextGenerationPipeline,
     TFAutoModelForSequenceClassification,
@@ -234,6 +235,31 @@ class CommonPipelineTest(unittest.TestCase):
 
             self.assertIsInstance(pipe, TextGenerationPipeline)  # Assert successful load
 
+    @require_torch
+    def test_pipeline_with_task_parameters_no_side_effects(self):
+        """
+        Regression test: certain pipeline flags, like `task`, modified the model configuration, causing unexpected
+        side-effects
+        """
+        # This checkpoint has task-specific parameters that will modify the behavior of the pipeline
+        model = T5ForConditionalGeneration.from_pretrained("t5-small")
+        self.assertTrue(model.config.num_beams == 1)
+
+        # The task-specific parameters used to cause side-effects on `model.config` -- not anymore
+        pipe = pipeline(model=model, tokenizer=AutoTokenizer.from_pretrained("t5-small"), task="translation_en_to_de")
+        self.assertTrue(model.config.num_beams == 1)
+        self.assertTrue(model.generation_config.num_beams == 1)
+
+        # Under the hood: we now store a generation config in the pipeline. This generation config stores the
+        # task-specific paremeters.
+        self.assertTrue(pipe.generation_config.num_beams == 4)
+
+        # We can confirm that the task-specific parameters have an effect. (In this case, the default is `num_beams=1`,
+        # which would crash when `num_return_sequences=4` is passed.)
+        pipe("Hugging Face doesn't sell hugs.", num_return_sequences=4)
+        with self.assertRaises(ValueError):
+            pipe("Hugging Face doesn't sell hugs.", num_return_sequences=4, num_beams=1)
+
 
 @is_pipeline_test
 class PipelineScikitCompatTest(unittest.TestCase):

tests/utils/test_modeling_utils.py

@@ -1715,6 +1715,38 @@ class ModelUtilsTest(TestCasePlus):
             torch.equal(torch.isin(random_ids, random_test_tensor), isin_mps_friendly(random_ids, random_test_tensor))
         )
 
+    def test_save_and_load_config_with_custom_generation(self):
+        """
+        Regression test for the ability to save and load a config with a custom generation kwarg (i.e. a parameter
+        that gets moved to the generation config and reset on the model config)
+        """
+        model = T5ForConditionalGeneration.from_pretrained(TINY_T5)
+
+        # The default for `num_beams` is 1 and `early_stopping` is False
+        self.assertTrue(model.config.num_beams == 1)
+        self.assertTrue(model.config.early_stopping is False)
+
+        # When we save the model, this custom parameter should be moved to the generation config AND the model
+        # config should contain `None`
+        model.config.num_beams = 2
+        model.config.early_stopping = True
+        self.assertTrue(model.generation_config.num_beams == 1)  # unmodified generation config
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(tmp_dir)
+            new_model = T5ForConditionalGeneration.from_pretrained(tmp_dir)
+            # moved to generation config
+            self.assertTrue(new_model.generation_config.num_beams == 2)
+            self.assertTrue(new_model.generation_config.early_stopping is True)
+            # reset in the model config
+            self.assertTrue(new_model.config.num_beams is None)
+            self.assertTrue(new_model.config.early_stopping is None)
+
+            # Sanity check: We can run `generate` with the new model without any warnings
+            random_ids = torch.randint(0, 100, (1, 5))
+            with warnings.catch_warnings(record=True) as w:
+                new_model.generate(random_ids, max_new_tokens=3)
+            self.assertTrue(len(w) == 0)
+
 
 @slow
 @require_torch