Generation: deprecate PreTrainedModel inheriting from GenerationMixin (#33203)

tests/generation/test_utils.py

@@ -2099,6 +2099,15 @@ class GenerationTesterMixin:
             )
             self.assertEqual(with_all_logits.tolist(), without_all_logits.tolist())
 
+    @pytest.mark.generate
+    def test_inherits_generation_mixin(self):
+        """
+        Tests that the model class directly inherits `GenerationMixin`, as opposed to relying on `PreTrainedModel`
+        to inherit it.
+        """
+        for model_class in self.all_generative_model_classes:
+            self.assertTrue("GenerationMixin" in str(model_class.__bases__))
+
     def _check_outputs(self, output, input_ids, config, use_cache=False, num_return_sequences=1):
         batch_size, seq_length = input_ids.shape
         config = config.text_config if hasattr(config, "text_config") else config

tests/models/auto/test_modeling_auto.py

@@ -67,6 +67,7 @@ if is_torch_available():
         BertModel,
         FunnelBaseModel,
         FunnelModel,
+        GenerationMixin,
         GPT2Config,
         GPT2LMHeadModel,
         ResNetBackbone,
@@ -571,3 +572,20 @@ class AutoModelTest(unittest.TestCase):
             _ = AutoModelForCausalLM.from_pretrained(tmp_dir_out, trust_remote_code=True)
             self.assertTrue((Path(tmp_dir_out) / "modeling_fake_custom.py").is_file())
             self.assertTrue((Path(tmp_dir_out) / "configuration_fake_custom.py").is_file())
+
+    def test_custom_model_patched_generation_inheritance(self):
+        """
+        Tests that our inheritance patching for generate-compatible models works as expected. Without this feature,
+        old Hub models lose the ability to call `generate`.
+        """
+        model = AutoModelForCausalLM.from_pretrained(
+            "hf-internal-testing/test_dynamic_model_generation", trust_remote_code=True
+        )
+        self.assertTrue(model.__class__.__name__ == "NewModelForCausalLM")
+
+        # It inherits from GenerationMixin. This means it can `generate`. Because `PreTrainedModel` is scheduled to
+        # stop inheriting from `GenerationMixin` in v4.50, this check will fail if patching is not present.
+        self.assertTrue(isinstance(model, GenerationMixin))
+        # More precisely, it directly inherits from GenerationMixin. This check would fail prior to v4.45 (inheritance
+        # patching was added in v4.45)
+        self.assertTrue("GenerationMixin" in str(model.__class__.__bases__))

tests/utils/test_modeling_utils.py

@@ -90,6 +90,7 @@ if is_torch_available():
         BertConfig,
         BertModel,
         CLIPTextModel,
+        GenerationMixin,
         PreTrainedModel,
         T5Config,
         T5ForConditionalGeneration,
@@ -1715,6 +1716,32 @@ class ModelUtilsTest(TestCasePlus):
             torch.equal(torch.isin(random_ids, random_test_tensor), isin_mps_friendly(random_ids, random_test_tensor))
         )
 
+    def test_can_generate(self):
+        """Tests the behavior of `PreTrainedModel.can_generate` method."""
+        # 1 - By default, a model CAN'T generate
+        self.assertFalse(BertModel.can_generate())
+
+        # 2 - The most common case for a model to be able to generate is to inherit from `GenerationMixin` directly
+        class DummyBertWithMixin(BertModel, GenerationMixin):
+            pass
+
+        self.assertTrue(DummyBertWithMixin.can_generate())
+
+        # 3 - Alternatively, a model can implement a `generate` method
+        class DummyBertWithGenerate(BertModel):
+            def generate(self):
+                pass
+
+        self.assertTrue(DummyBertWithGenerate.can_generate())
+
+        # 4 - BC: models with a custom `prepare_inputs_for_generation` can generate (it was assumed they inherited
+        # `GenerationMixin`)
+        class DummyBertWithPrepareInputs(BertModel):
+            def prepare_inputs_for_generation(self):
+                pass
+
+        self.assertTrue(DummyBertWithPrepareInputs.can_generate())
+
     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