[CI] Check test if the GenerationTesterMixin inheritance is correct 🐛 🔫

src/transformers/modeling_utils.py

@@ -1685,17 +1685,14 @@ def base_model(self) -> nn.Module:
     @classmethod
     def can_generate(cls) -> bool:
         """
-        Returns whether this model can generate sequences with `.generate()`.
+        Returns whether this model can generate sequences with `.generate()` from the `GenerationMixin`.
 
         Returns:
             `bool`: Whether this model can generate sequences with `.generate()`.
         """
         # Directly inherits `GenerationMixin` -> can generate
         if "GenerationMixin" in str(cls.__bases__):
             return True
-        # Model class overwrites `generate` (e.g. time series models) -> can generate
-        if str(cls.__name__) in str(cls.generate):
-            return True
         # The class inherits from a class that can generate (recursive check) -> can generate
         for base in cls.__bases__:
             if not hasattr(base, "can_generate"):

src/transformers/models/albert/modeling_albert.py

@@ -24,7 +24,6 @@
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 
 from ...activations import ACT2FN
-from ...generation import GenerationMixin
 from ...modeling_attn_mask_utils import _prepare_4d_attention_mask_for_sdpa
 from ...modeling_outputs import (
     BaseModelOutput,
@@ -984,7 +983,7 @@ def forward(self, pooled_output: torch.Tensor) -> torch.Tensor:
     "Albert Model with a `language modeling` head on top.",
     ALBERT_START_DOCSTRING,
 )
-class AlbertForMaskedLM(AlbertPreTrainedModel, GenerationMixin):
+class AlbertForMaskedLM(AlbertPreTrainedModel):
     _tied_weights_keys = ["predictions.decoder.bias", "predictions.decoder.weight"]
 
     def __init__(self, config):

src/transformers/models/seamless_m4t/modeling_seamless_m4t.py

@@ -2912,7 +2912,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The speech-to-text SeamlessM4T Model transformer which can be used for S2TT.",
     SEAMLESS_M4T_START_DOCSTRING,
 )
-class SeamlessM4TForSpeechToText(SeamlessM4TPreTrainedModel):
+class SeamlessM4TForSpeechToText(SeamlessM4TPreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["text_decoder", "t2u_model", "vocoder"]
     main_input_name = "input_features"
 
@@ -3182,7 +3182,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The text-to-speech SeamlessM4T Model transformer which can be used for T2ST.",
     SEAMLESS_M4T_START_DOCSTRING,
 )
-class SeamlessM4TForTextToSpeech(SeamlessM4TPreTrainedModel):
+class SeamlessM4TForTextToSpeech(SeamlessM4TPreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["speech_encoder"]
     main_input_name = "input_ids"
 
@@ -3511,7 +3511,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The speech-to-speech SeamlessM4T Model transformer which can be used for S2ST.",
     SEAMLESS_M4T_START_DOCSTRING,
 )
-class SeamlessM4TForSpeechToSpeech(SeamlessM4TPreTrainedModel):
+class SeamlessM4TForSpeechToSpeech(SeamlessM4TPreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["text_encoder"]
     main_input_name = "input_features"
 
@@ -3854,7 +3854,7 @@ def _reorder_cache(past_key_values, beam_idx):
             Default modality. Used to initialize the model.
     """,
 )
-class SeamlessM4TModel(SeamlessM4TPreTrainedModel):
+class SeamlessM4TModel(SeamlessM4TPreTrainedModel, GenerationMixin):
     _tied_weights_keys = [
         "lm_head.weight",
         "text_encoder.embed_tokens.weight",

src/transformers/models/seamless_m4t_v2/modeling_seamless_m4t_v2.py

@@ -3192,7 +3192,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The speech-to-text SeamlessM4Tv2 Model transformer which can be used for S2TT.",
     SEAMLESS_M4T_V2_START_DOCSTRING,
 )
-class SeamlessM4Tv2ForSpeechToText(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2ForSpeechToText(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["text_decoder", "t2u_model", "vocoder"]
     main_input_name = "input_features"
 
@@ -3473,7 +3473,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The text-to-speech SeamlessM4Tv2 Model transformer which can be used for T2ST.",
     SEAMLESS_M4T_V2_START_DOCSTRING,
 )
-class SeamlessM4Tv2ForTextToSpeech(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2ForTextToSpeech(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["speech_encoder"]
     main_input_name = "input_ids"
 
@@ -3844,7 +3844,7 @@ def _reorder_cache(past_key_values, beam_idx):
     "The speech-to-speech SeamlessM4Tv2 Model transformer which can be used for S2ST.",
     SEAMLESS_M4T_V2_START_DOCSTRING,
 )
-class SeamlessM4Tv2ForSpeechToSpeech(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2ForSpeechToSpeech(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
     _keys_to_ignore_on_load_missing = ["text_encoder"]
     main_input_name = "input_features"
 
@@ -4229,7 +4229,7 @@ def _reorder_cache(past_key_values, beam_idx):
             This will be updated automatically according to the modality passed to the forward and generate passes (`input_ids` for text and `input_features` for audio).
     """,
 )
-class SeamlessM4Tv2Model(SeamlessM4Tv2PreTrainedModel):
+class SeamlessM4Tv2Model(SeamlessM4Tv2PreTrainedModel, GenerationMixin):
     _tied_weights_keys = [
         "lm_head.weight",
         "text_encoder.embed_tokens.weight",

tests/generation/test_utils.py

@@ -19,6 +19,7 @@
 import datetime
 import gc
 import inspect
+import random
 import tempfile
 import unittest
 import warnings
@@ -48,8 +49,6 @@
 )
 from transformers.utils import is_ipex_available
 
-from ..test_modeling_common import floats_tensor, ids_tensor
-
 
 if is_torch_available():
     import torch
@@ -2786,6 +2785,43 @@ def test_speculative_sampling_target_distribution(self):
         self.assertTrue(last_token_counts[8] > last_token_counts[3])
 
 
+global_rng = random.Random()
+
+
+# Copied from tests.test_modeling_common.ids_tensor
+def ids_tensor(shape, vocab_size, rng=None, name=None):
+    #  Creates a random int32 tensor of the shape within the vocab size
+    if rng is None:
+        rng = global_rng
+
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.randint(0, vocab_size - 1))
+
+    return torch.tensor(data=values, dtype=torch.long, device=torch_device).view(shape).contiguous()
+
+
+# Copied from tests.test_modeling_common.floats_tensor
+def floats_tensor(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = global_rng
+
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.random() * scale)
+
+    return torch.tensor(data=values, dtype=torch.float, device=torch_device).view(shape).contiguous()
+
+
 @pytest.mark.generate
 @require_torch
 class GenerationIntegrationTests(unittest.TestCase):

tests/models/big_bird/test_modeling_big_bird.py

@@ -451,6 +451,8 @@ class BigBirdModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
         if is_torch_available()
         else ()
     )
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {
             "feature-extraction": BigBirdModel,

tests/models/blip/test_modeling_blip.py

@@ -799,6 +799,8 @@ def prepare_config_and_inputs_for_common(self):
 @require_vision
 class BlipVQAModelTest(ModelTesterMixin, unittest.TestCase):
     all_model_classes = (BlipForQuestionAnswering,) if is_torch_available() else ()
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     fx_compatible = False
     test_head_masking = False
     test_pruning = False
@@ -1106,6 +1108,8 @@ def test_model_from_pretrained(self):
 @require_torch
 class BlipTextImageModelTest(ModelTesterMixin, unittest.TestCase):
     all_model_classes = (BlipForConditionalGeneration,) if is_torch_available() else ()
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     fx_compatible = False
     test_head_masking = False
     test_pruning = False

tests/models/blip_2/test_modeling_blip_2.py

@@ -885,7 +885,7 @@ def prepare_config_and_inputs_for_common(self):
 
 
 @require_torch
-class Blip2ModelTest(ModelTesterMixin, PipelineTesterMixin, GenerationTesterMixin, unittest.TestCase):
+class Blip2ModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (Blip2ForConditionalGeneration, Blip2Model) if is_torch_available() else ()
     # Doesn't run generation tests. TODO: fix generation tests for Blip2ForConditionalGeneration
     all_generative_model_classes = ()

tests/models/clvp/test_modeling_clvp.py

@@ -408,6 +408,8 @@ def prepare_config_and_inputs_for_common(self):
 @require_torch
 class ClvpModelForConditionalGenerationTest(ModelTesterMixin, unittest.TestCase):
     all_model_classes = (ClvpModelForConditionalGeneration,) if is_torch_available() else ()
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
 
     test_head_masking = False
     test_pruning = False

tests/models/conditional_detr/test_modeling_conditional_detr.py

@@ -22,7 +22,6 @@
 from transformers.testing_utils import require_timm, require_torch, require_vision, slow, torch_device
 from transformers.utils import cached_property
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -173,7 +172,7 @@ def create_and_check_conditional_detr_object_detection_head_model(self, config,
 
 
 @require_torch
-class ConditionalDetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class ConditionalDetrModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (
         (
             ConditionalDetrModel,

tests/models/cpmant/test_modeling_cpmant.py

@@ -136,6 +136,8 @@ def prepare_config_and_inputs_for_common(self):
 @require_torch
 class CpmAntModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (CpmAntModel, CpmAntForCausalLM) if is_torch_available() else ()
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {"feature-extraction": CpmAntModel, "text-generation": CpmAntForCausalLM} if is_torch_available() else {}
     )

tests/models/dab_detr/test_modeling_dab_detr.py

@@ -23,7 +23,6 @@
 from transformers.testing_utils import require_timm, require_torch, require_vision, slow, torch_device
 from transformers.utils import cached_property
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -174,15 +173,8 @@ def create_and_check_dab_detr_object_detection_head_model(self, config, pixel_va
 
 
 @require_torch
-class DabDetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
-    all_model_classes = (
-        (
-            DabDetrModel,
-            DabDetrForObjectDetection,
-        )
-        if is_torch_available()
-        else ()
-    )
+class DabDetrModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
+    all_model_classes = (DabDetrModel, DabDetrForObjectDetection) if is_torch_available() else ()
     pipeline_model_mapping = (
         {
             "image-feature-extraction": DabDetrModel,

tests/models/decision_transformer/test_modeling_decision_transformer.py

@@ -20,7 +20,6 @@
 from transformers import DecisionTransformerConfig, is_torch_available
 from transformers.testing_utils import require_torch, slow, torch_device
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -125,7 +124,7 @@ def prepare_config_and_inputs_for_common(self):
 
 
 @require_torch
-class DecisionTransformerModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class DecisionTransformerModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (DecisionTransformerModel,) if is_torch_available() else ()
     pipeline_model_mapping = {"feature-extraction": DecisionTransformerModel} if is_torch_available() else {}
 

tests/models/deformable_detr/test_modeling_deformable_detr.py

@@ -31,7 +31,6 @@
     torch_device,
 )
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -188,7 +187,7 @@ def create_and_check_deformable_detr_object_detection_head_model(self, config, p
 
 
 @require_torch
-class DeformableDetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class DeformableDetrModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (DeformableDetrModel, DeformableDetrForObjectDetection) if is_torch_available() else ()
     pipeline_model_mapping = (
         {"image-feature-extraction": DeformableDetrModel, "object-detection": DeformableDetrForObjectDetection}

tests/models/detr/test_modeling_detr.py

@@ -22,7 +22,6 @@
 from transformers.testing_utils import require_timm, require_torch, require_vision, slow, torch_device
 from transformers.utils import cached_property
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, floats_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -169,7 +168,7 @@ def create_and_check_detr_object_detection_head_model(self, config, pixel_values
 
 
 @require_torch
-class DetrModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class DetrModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (
         (
             DetrModel,

tests/models/electra/test_modeling_electra.py

@@ -389,6 +389,8 @@ class ElectraModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase)
         if is_torch_available()
         else ()
     )
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {
             "feature-extraction": ElectraModel,

tests/models/flaubert/test_modeling_flaubert.py

@@ -377,6 +377,8 @@ class FlaubertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase
         if is_torch_available()
         else ()
     )
+    # Doesn't run generation tests. Outdated custom `prepare_inputs_for_generation` -- TODO @gante
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {
             "feature-extraction": FlaubertModel,

tests/models/idefics/test_modeling_idefics.py

@@ -326,6 +326,8 @@ def test_eager_matches_sdpa_generate(self):
 @require_torch
 class IdeficsModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (IdeficsModel, IdeficsForVisionText2Text) if is_torch_available() else ()
+    # Doesn't run generation tests here -- idefics has a dedicated tester for generation tests below
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {"feature-extraction": IdeficsModel, "image-text-to-text": IdeficsForVisionText2Text}
         if is_torch_available()
@@ -868,6 +870,12 @@ def test_training_gradient_checkpointing_use_reentrant_false(self):
     def test_sdpa_can_dispatch_non_composite_models(self):
         pass
 
+    @unittest.skip(
+        "Idefics has a separate test runner for generation tests with complex inheritance, causing this check to fail"
+    )
+    def test_generation_tester_mixin_inheritance(self):
+        pass
+
 
 @require_torch
 @require_vision

tests/models/lilt/test_modeling_lilt.py

@@ -19,7 +19,6 @@
 from transformers import LiltConfig, is_torch_available
 from transformers.testing_utils import require_torch, slow, torch_device
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, ids_tensor
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -218,7 +217,7 @@ def prepare_config_and_inputs_for_common(self):
 
 
 @require_torch
-class LiltModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class LiltModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (
         (
             LiltModel,

tests/models/megatron_bert/test_modeling_megatron_bert.py

@@ -282,6 +282,8 @@ class MegatronBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.Test
         if is_torch_available()
         else ()
     )
+    # Doesn't run generation tests. There are interface mismatches when using `generate` -- TODO @gante
+    all_generative_model_classes = ()
     pipeline_model_mapping = (
         {
             "feature-extraction": MegatronBertModel,

tests/models/modernbert/test_modeling_modernbert.py

@@ -29,7 +29,6 @@
     torch_device,
 )
 
-from ...generation.test_utils import GenerationTesterMixin
 from ...test_configuration_common import ConfigTester
 from ...test_modeling_common import ModelTesterMixin, _config_zero_init, ids_tensor, random_attention_mask
 from ...test_pipeline_mixin import PipelineTesterMixin
@@ -216,7 +215,7 @@ def prepare_config_and_inputs_for_common(self):
 
 
 @require_torch
-class ModernBertModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
+class ModernBertModelTest(ModelTesterMixin, PipelineTesterMixin, unittest.TestCase):
     test_torchscript = False
 
     all_model_classes = (