Mamba: add generative tests (#31478)

src/transformers/generation/utils.py

@@ -1830,6 +1830,12 @@ class GenerationMixin:
                 raise ValueError("assisted generate requires `use_cache=True`")
             if generation_config.cache_implementation == "static":
                 raise ValueError("assisted generate is not supported with `static_cache`")
+            if self._is_stateful:
+                # In assisted generation we need the ability to confirm whether the model would pick certain tokens,
+                # which is not possible with stateful models (they can't reset to a previous subset of generated text)
+                raise ValueError(
+                    f"assisted generation is not supported with stateful models, such as {self.__class__.__name__}"
+                )
 
             # 11. Get the candidate generator, given the parameterization
             candidate_generator = self._get_candidate_generator(
@@ -1867,6 +1873,11 @@ class GenerationMixin:
         elif generation_mode == GenerationMode.CONTRASTIVE_SEARCH:
             if not model_kwargs["use_cache"]:
                 raise ValueError("Contrastive search requires `use_cache=True`")
+            if self._is_stateful:
+                # Just like assisted generation, we need to be able to rollback to a previous state (see comment above)
+                raise ValueError(
+                    f"contrastive search is not supported with stateful models, such as {self.__class__.__name__}"
+                )
 
             result = self._contrastive_search(
                 input_ids,

src/transformers/modeling_utils.py

@@ -1281,6 +1281,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
     is_parallelizable = False
     supports_gradient_checkpointing = False
+    _is_stateful = False
 
     # Flash Attention 2 support
     _supports_flash_attn_2 = False

src/transformers/models/jamba/modeling_jamba.py

@@ -1266,6 +1266,7 @@ class JambaPreTrainedModel(PreTrainedModel):
     _supports_flash_attn_2 = True
     _supports_sdpa = True
     _supports_cache_class = True  # Note: only supports HybridMambaAttentionDynamicCache
+    _is_stateful = True
 
     def _init_weights(self, module):
         std = self.config.initializer_range

src/transformers/models/mamba/modeling_mamba.py

@@ -354,6 +354,7 @@ class MambaPreTrainedModel(PreTrainedModel):
     base_model_prefix = "backbone"
     _no_split_modules = ["MambaBlock"]
     supports_gradient_checkpointing = True
+    _is_stateful = True
 
     def _init_weights(self, module):
         """Initialize the weights."""

src/transformers/models/rwkv/modeling_rwkv.py

@@ -394,6 +394,7 @@ class RwkvPreTrainedModel(PreTrainedModel):
     _no_split_modules = ["RwkvBlock"]
     _keep_in_fp32_modules = ["time_decay", "time_first"]
     supports_gradient_checkpointing = True
+    _is_stateful = True
 
     def _init_weights(self, module):
         """Initialize the weights."""

tests/generation/test_utils.py

@@ -102,7 +102,11 @@ class GenerationTesterMixin:
             if isinstance(config.eos_token_id, int):
                 config.eos_token_id = [config.eos_token_id]
             config.pad_token_id = config.eos_token_id[0]
+
+        if self.has_attentions:
             attention_mask = torch.ones_like(input_ids, dtype=torch.long)
+        else:
+            attention_mask = None
 
         # It is important set set the eos_token_id to None to ensure that no sequences
         # shorter than `max_length` can be generated
@@ -437,7 +441,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -471,7 +475,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -529,7 +533,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -595,7 +599,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
             if model.config.is_encoder_decoder:
@@ -642,7 +646,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -733,7 +737,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -834,7 +838,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
             if model.config.is_encoder_decoder:
@@ -952,7 +956,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -973,6 +977,9 @@ class GenerationTesterMixin:
 
     def test_contrastive_generate(self):
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support contrastive search generation")
+
             # won't fix: FSMT and Reformer have a different cache variable type (and format).
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
@@ -997,6 +1004,9 @@ class GenerationTesterMixin:
 
     def test_contrastive_generate_dict_outputs_use_cache(self):
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support contrastive search generation")
+
             # won't fix: FSMT and Reformer have a different cache variable type (and format).
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
@@ -1017,7 +1027,7 @@ class GenerationTesterMixin:
                 output_scores=True,
                 output_logits=True,
                 output_hidden_states=True,
-                output_attentions=True,
+                output_attentions=self.has_attentions,
                 return_dict_in_generate=True,
             )
 
@@ -1030,9 +1040,12 @@ class GenerationTesterMixin:
     def test_contrastive_generate_low_memory(self):
         # Check that choosing 'low_memory' does not change the model output
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support contrastive search generation")
+
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer", "speech2text"]):
                 self.skipTest("Won't fix: old model with different cache format")
-            if any(model_name in model_class.__name__.lower() for model_name in ["gptbigcode", "jamba"]):
+            if any(model_name in model_class.__name__.lower() for model_name in ["gptbigcode"]):
                 self.skipTest("TODO: fix me")
 
             config, input_ids, attention_mask = self._get_input_ids_and_config(batch_size=1)
@@ -1069,6 +1082,8 @@ class GenerationTesterMixin:
     def test_beam_search_low_memory(self):
         # Check that choosing 'low_memory' does not change the model output
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("May fix in the future: need custom cache handling")
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
             if any(
@@ -1115,6 +1130,8 @@ class GenerationTesterMixin:
         # - assisted_decoding does not support `batch_size > 1`
 
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support assisted generation")
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
             if any(
@@ -1156,7 +1173,7 @@ class GenerationTesterMixin:
                 "output_scores": True,
                 "output_logits": True,
                 "output_hidden_states": True,
-                "output_attentions": True,
+                "output_attentions": self.has_attentions,
                 "return_dict_in_generate": True,
             }
             output_greedy = model.generate(input_ids, attention_mask=attention_mask, **generation_kwargs)
@@ -1184,6 +1201,8 @@ class GenerationTesterMixin:
         # This test is mostly a copy of test_assisted_decoding_matches_greedy_search
 
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support assisted generation")
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
             if any(
@@ -1225,7 +1244,7 @@ class GenerationTesterMixin:
                 "output_scores": True,
                 "output_logits": True,
                 "output_hidden_states": True,
-                "output_attentions": True,
+                "output_attentions": self.has_attentions,
                 "return_dict_in_generate": True,
             }
 
@@ -1244,6 +1263,8 @@ class GenerationTesterMixin:
         # match sample for the same seed, as the forward pass does not return the exact same logits (due to matmul with
         # different shapes, see https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535).
         for model_class in self.all_generative_model_classes:
+            if model_class._is_stateful:
+                self.skipTest("Stateful models don't support assisted generation")
             if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer"]):
                 self.skipTest("Won't fix: old model with different cache format")
             if any(
@@ -1289,7 +1310,7 @@ class GenerationTesterMixin:
                 "output_scores": True,
                 "output_logits": True,
                 "output_hidden_states": True,
-                "output_attentions": True,
+                "output_attentions": self.has_attentions,
                 "return_dict_in_generate": True,
             }
             output_assisted = model.generate(input_ids, attention_mask=attention_mask, **generation_kwargs)
@@ -1326,7 +1347,7 @@ class GenerationTesterMixin:
                     input_ids,
                     attention_mask=attention_mask,
                     num_beams=1,
-                    output_attentions=True,
+                    output_attentions=self.has_attentions,
                     return_dict_in_generate=True,
                     remove_invalid_values=True,
                     **{name: mask},
@@ -1344,6 +1365,10 @@ class GenerationTesterMixin:
         if len(self.all_generative_model_classes) == 0:
             self.skipTest(reason="No generative architecture available for this model.")
 
+        # - The model must support padding
+        if not self.has_attentions:
+            self.skipTest(reason="This model doesn't support padding.")
+
         # - The model must be a decoder-only architecture (encoder-based architectures use right-padding)
         decoder_only_classes = []
         for model_class in self.all_generative_model_classes:
@@ -1704,6 +1729,7 @@ class GenerationTesterMixin:
         self._check_logits(num_sequences_in_output, output.logits, config=config)
 
         # Attentions
+        if self.has_attentions:
             if config.is_encoder_decoder:
                 # encoder
                 self._check_encoder_attention_for_generate(output.encoder_attentions, batch_size, config, seq_length)
@@ -1763,7 +1789,7 @@ class GenerationTesterMixin:
         # 2. Some old models still return `output.past_key_values` even without `use_cache=True`
         # 3. TODO (joao): A few models have different formats/types, skipping those until the cache refactor is
         # complete
-        models_without_standard_cache = ("bloom", "ctrl", "fsmt", "gptbigcode", "mega", "reformer", "jamba")
+        models_without_standard_cache = ("bloom", "ctrl", "fsmt", "gptbigcode", "mega", "reformer", "jamba", "mamba")
         has_standard_cache = not any(
             model_name in config.__class__.__name__.lower() for model_name in models_without_standard_cache
         )

tests/models/jamba/test_modeling_jamba.py

@@ -503,10 +503,6 @@ class JambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
             # They should result in very similar logits
             self.assertTrue(torch.allclose(next_logits_wo_padding, next_logits_with_padding, atol=3e-3))
 
-    @unittest.skip("Jamba has its own special cache type")  # FIXME: @gante
-    def test_assisted_decoding_matches_greedy_search_0_random(self):
-        pass
-
     @require_flash_attn
     @require_torch_gpu
     @require_bitsandbytes

tests/models/mamba/test_modeling_mamba.py

@@ -250,6 +250,8 @@ class MambaModelTester:
 @require_torch
 class MambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
     all_model_classes = (MambaModel, MambaForCausalLM) if is_torch_available() else ()
+    all_generative_model_classes = (MambaForCausalLM,) if is_torch_available() else ()
+    has_attentions = False  # Mamba does not support attentions
     fx_compatible = False  # FIXME let's try to support this @ArthurZucker
     test_torchscript = False  # FIXME let's try to support this @ArthurZucker
     test_missing_keys = False
@@ -292,10 +294,6 @@ class MambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
     def test_config(self):
         self.config_tester.run_common_tests()
 
-    @unittest.skip("No attention in mamba")
-    def test_retain_grad_hidden_states_attentions(self):
-        pass
-
     @require_torch_multi_gpu
     def test_multi_gpu_data_parallel_forward(self):
         config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
@@ -364,14 +362,6 @@ class MambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixi
                         # check if it's a ones like
                         self.assertTrue(torch.allclose(param.data, torch.ones_like(param.data), atol=1e-5, rtol=1e-5))
 
-    @unittest.skip("Mamba does not use attention")
-    def test_attention_outputs(self):
-        r"""
-        Overriding the test_attention_outputs test as the attention outputs of Mamba are different from other models
-        it has a shape `batch_size, seq_len, hidden_size`.
-        """
-        pass
-
     @slow
     def test_model_from_pretrained(self):
         model = MambaModel.from_pretrained("hf-internal-testing/mamba-130m")