aaand remove tests after all green!!

Author: manueldeprada

src/transformers/configuration_utils.py

@@ -1139,6 +1139,9 @@ def _get_global_generation_defaults() -> dict[str, Any]:
             "exponential_decay_length_penalty": None,
             "suppress_tokens": None,
             "begin_suppress_tokens": None,
+            # Deprecated arguments (moved to the Hub). TODO joao, manuel: remove in v4.62.0
+            "num_beam_groups": 1,
+            "diversity_penalty": 0.0,
         }
 
     def _get_non_default_generation_parameters(self) -> dict[str, Any]:

tests/generation/test_utils.py

@@ -209,18 +209,6 @@ def _get_beam_kwargs(self, num_return_sequences=1):
         }
         return beam_kwargs
 
-    def _get_diverse_beam_kwargs(self, num_return_sequences=1):
-        beam_kwargs = {
-            "early_stopping": False,
-            "length_penalty": 2.0,
-            "num_beams": 2,
-            "num_return_sequences": num_return_sequences,
-            "num_beam_groups": 2,  # one beam per group
-            "diversity_penalty": 2.0,
-            "trust_remote_code": True,
-        }
-        return beam_kwargs
-
     def _get_constrained_beam_kwargs(self, num_return_sequences=1):
         beam_kwargs = {
             "early_stopping": False,
@@ -352,36 +340,6 @@ def _beam_sample_generate(
 
         return output_generate
 
-    def _group_beam_search_generate(
-        self,
-        model,
-        inputs_dict,
-        beam_kwargs,
-        output_scores=False,
-        output_logits=False,
-        output_attentions=False,
-        output_hidden_states=False,
-        return_dict_in_generate=False,
-        use_cache=True,
-    ):
-        logits_processor_kwargs = self._get_logits_processor_kwargs(do_sample=False, config=model.config)
-        output_generate = model.generate(
-            do_sample=False,
-            max_new_tokens=self.max_new_tokens,
-            min_new_tokens=self.max_new_tokens,
-            output_scores=output_scores,
-            output_logits=output_logits,
-            output_attentions=output_attentions,
-            output_hidden_states=output_hidden_states,
-            return_dict_in_generate=return_dict_in_generate,
-            use_cache=use_cache,
-            **beam_kwargs,
-            **logits_processor_kwargs,
-            **inputs_dict,
-        )
-
-        return output_generate
-
     def _constrained_beam_search_generate(
         self,
         model,
@@ -748,77 +706,6 @@ def test_generate_without_input_ids(self):
             )
             self.assertIsNotNone(output_ids_generate)
 
-    @pytest.mark.generate
-    def test_group_beam_search_generate(self):
-        for model_class in self.all_generative_model_classes:
-            config, inputs_dict = self.prepare_config_and_inputs_for_generate()
-
-            model = model_class(config).to(torch_device).eval()
-            # check `generate()` and `group_beam_search()` are equal
-            beam_kwargs = self._get_diverse_beam_kwargs()
-            output_generate = self._group_beam_search_generate(
-                model=model,
-                inputs_dict=inputs_dict,
-                beam_kwargs=beam_kwargs,
-            )
-            if model.config.get_text_config(decoder=True).is_encoder_decoder:
-                self.assertTrue(output_generate.shape[1] == self.max_new_tokens + 1)
-            else:
-                self.assertTrue(output_generate.shape[1] == self.max_new_tokens + inputs_dict["input_ids"].shape[1])
-
-            # check `group_beam_search` for higher than 1 `num_return_sequences`
-            num_return_sequences = 2
-            beam_kwargs = self._get_diverse_beam_kwargs(num_return_sequences=num_return_sequences)
-            output_generate = self._group_beam_search_generate(
-                model=model,
-                inputs_dict=inputs_dict,
-                beam_kwargs=beam_kwargs,
-            )
-            if model.config.get_text_config(decoder=True).is_encoder_decoder:
-                self.assertTrue(output_generate.shape[1] == self.max_new_tokens + 1)
-            else:
-                self.assertTrue(output_generate.shape[1] == self.max_new_tokens + inputs_dict["input_ids"].shape[1])
-
-    @pytest.mark.generate
-    def test_group_beam_search_generate_dict_output(self):
-        for model_class in self.all_generative_model_classes:
-            config, inputs_dict = self.prepare_config_and_inputs_for_generate()
-            if self.has_attentions:
-                config._attn_implementation = "eager"  # can't output attentions otherwise
-
-            model = model_class(config).to(torch_device).eval()
-            beam_kwargs = self._get_diverse_beam_kwargs()
-            output_generate = self._group_beam_search_generate(
-                model=model,
-                inputs_dict=inputs_dict,
-                beam_kwargs=beam_kwargs,
-                output_scores=True,
-                output_logits=True,
-                output_hidden_states=True,
-                output_attentions=self.has_attentions,
-                return_dict_in_generate=True,
-                use_cache=False,
-            )
-            if model.config.get_text_config(decoder=True).is_encoder_decoder:
-                self.assertTrue(output_generate.sequences.shape[1] == self.max_new_tokens + 1)
-                self.assertIsInstance(output_generate, GenerateBeamEncoderDecoderOutput)
-                # Retrocompatibility check
-                self.assertIsInstance(output_generate, BeamSearchEncoderDecoderOutput)
-            else:
-                self.assertTrue(
-                    output_generate.sequences.shape[1] == self.max_new_tokens + inputs_dict["input_ids"].shape[1]
-                )
-                self.assertIsInstance(output_generate, GenerateBeamDecoderOnlyOutput)
-                # Retrocompatibility check
-                self.assertIsInstance(output_generate, BeamSearchDecoderOnlyOutput)
-
-            self._check_generate_outputs(
-                output_generate,
-                model.config,
-                num_return_sequences=beam_kwargs["num_return_sequences"],
-                num_beams=beam_kwargs["num_beams"],
-            )
-
     @is_flaky()  # Some models have position-specific tokens, this test may try to force them in an invalid position
     @pytest.mark.generate
     def test_constrained_beam_search_generate(self):
@@ -2672,6 +2559,7 @@ def test_diverse_beam_search(self):
             diversity_penalty=2.0,
             remove_invalid_values=True,
             trust_remote_code=True,
+            custom_generate="transformers-community/group-beam-search",
         )
 
         generated_text = bart_tokenizer.batch_decode(outputs, skip_special_tokens=True)
@@ -2831,6 +2719,7 @@ def test_generate_input_values_as_encoder_kwarg(self):
         self.assertListEqual(output_sequences.tolist(), output_sequences_kwargs.tolist())
         self.assertEqual(output_sequences.shape, (2, 5))
 
+    # TODO joao, manuel: remove in v4.62.0
     def test_transition_scores_group_beam_search_encoder_decoder(self):
         articles = [
             "Justin Timberlake and Jessica Biel, welcome to parenthood.",
@@ -2839,20 +2728,27 @@ def test_transition_scores_group_beam_search_encoder_decoder(self):
         tokenizer = BartTokenizer.from_pretrained("hf-internal-testing/tiny-random-bart")
         model = BartForConditionalGeneration.from_pretrained(
             "hf-internal-testing/tiny-random-bart",
+            eos_token_id=None,
+        )
+        generation_config = GenerationConfig(
             max_length=10,
             num_beams=2,
             num_beam_groups=2,
             num_return_sequences=2,
             diversity_penalty=1.0,
-            eos_token_id=None,
             return_dict_in_generate=True,
             output_scores=True,
             length_penalty=0.0,
         )
         model = model.to(torch_device)
 
         input_ids = tokenizer(articles, return_tensors="pt", padding=True).input_ids.to(torch_device)
-        outputs = model.generate(input_ids=input_ids, trust_remote_code=True)
+        outputs = model.generate(
+            input_ids=input_ids,
+            generation_config=generation_config,
+            trust_remote_code=True,
+            custom_generate="transformers-community/group-beam-search",
+        )
 
         transition_scores = model.compute_transition_scores(outputs.sequences, outputs.scores, outputs.beam_indices)
         transition_scores_sum = transition_scores.sum(-1)
@@ -4823,6 +4719,16 @@ def test_generate_custom_cache_position(self):
         [
             ("transformers-community/dola", {"dola_layers": "low"}),
             ("transformers-community/contrastive-search", {"penalty_alpha": 0.6, "top_k": 4}),
+            (
+                "transformers-community/group-beam-search",
+                {
+                    "do_sample": False,
+                    "num_beams": 2,
+                    "num_beam_groups": 2,
+                    "diversity_penalty": 2.0,
+                    "length_penalty": 2.0,
+                },
+            ),
         ]
     )
     def test_hub_gen_strategies(self, custom_generate, extra_kwargs):

tests/models/csm/test_modeling_csm.py

@@ -272,16 +272,6 @@ def test_beam_search_generate_dict_outputs_use_cache(self):
     def test_beam_sample_generate_dict_output(self):
         pass
 
-    @pytest.mark.generate
-    @unittest.skip(reason="CSM does not support group beam search.")
-    def test_group_beam_search_generate(self):
-        pass
-
-    @pytest.mark.generate
-    @unittest.skip(reason="CSM does not support group beam search.")
-    def test_group_beam_search_generate_dict_output(self):
-        pass
-
     @pytest.mark.generate
     @unittest.skip(reason="CSM does not support constrained beam search.")
     def test_constrained_beam_search_generate(self):

tests/models/dia/test_modeling_dia.py

@@ -237,7 +237,6 @@ def skip_non_greedy_generate(self):
         skippable_tests = [
             "test_sample_generate_dict_output",  # return sequences > 1
             "test_beam",
-            "test_group_beam",
             "test_constrained_beam",
             "test_contrastive",
             "test_assisted",

tests/models/recurrent_gemma/test_modeling_recurrent_gemma.py

@@ -138,16 +138,6 @@ def test_constrained_beam_search_generate_dict_output(self):
     def test_generate_without_input_ids(self):
         pass
 
-    @unittest.skip(reason="RecurrentGemma is unusual and fails a lot of generation tests")
-    @pytest.mark.generate
-    def test_group_beam_search_generate(self):
-        pass
-
-    @unittest.skip(reason="RecurrentGemma is unusual and fails a lot of generation tests")
-    @pytest.mark.generate
-    def test_group_beam_search_generate_dict_output(self):
-        pass
-
     @unittest.skip(reason="RecurrentGemma is unusual and fails a lot of generation tests")
     @pytest.mark.generate
     def test_constrained_beam_search_generate(self):

tests/models/rwkv/test_modeling_rwkv.py

@@ -401,13 +401,6 @@ def test_greedy_generate_dict_outputs(self):
         super().test_greedy_generate_dict_outputs()
         self.has_attentions = old_has_attentions
 
-    def test_group_beam_search_generate_dict_output(self):
-        # This model has a custom attention output shape AND config flags, let's skip those checks
-        old_has_attentions = self.has_attentions
-        self.has_attentions = False
-        super().test_group_beam_search_generate_dict_output()
-        self.has_attentions = old_has_attentions
-
     def test_sample_generate_dict_output(self):
         # This model has a custom attention output shape AND config flags, let's skip those checks
         old_has_attentions = self.has_attentions

tests/models/whisper/test_modeling_whisper.py

@@ -403,12 +403,6 @@ def _get_beam_kwargs(self, num_return_sequences=1):
         beam_kwargs["num_return_sequences"] = beam_kwargs["num_beams"]
         return beam_kwargs
 
-    def _get_diverse_beam_kwargs(self, num_return_sequences=1):
-        # Overwritten from `GenerationTesterMixin`, Whisper's `num_return_sequences` differs from the core `generate`
-        beam_kwargs = super()._get_diverse_beam_kwargs(num_return_sequences=num_return_sequences)
-        beam_kwargs["num_return_sequences"] = beam_kwargs["num_beams"]
-        return beam_kwargs
-
     def _get_constrained_beam_kwargs(self, num_return_sequences=1):
         # Overwritten from `GenerationTesterMixin`, Whisper's `num_return_sequences` differs from the core `generate`
         beam_kwargs = super()._get_constrained_beam_kwargs(num_return_sequences=num_return_sequences)