Add FlaxBartForCausalLM (#15995)

* add causal lm

* add CausalLM tests

* Add FlaxBartForCausalLM

* Add EncoderDecoder model tests

* change docstring

* make repo-consistency

* suggested changes

* remove jax ops

* correction

* rename pre-trained decoder model

docs/source/model_doc/bart.mdx

@@ -152,3 +152,8 @@ assert tok.batch_decode(generated_ids, skip_special_tokens=True) == [
     - __call__
     - encode
     - decode
+
+## FlaxBartForCausalLM
+
+[[autodoc]] FlaxBartForCausalLM
+    - __call__
\ No newline at end of file

src/transformers/__init__.py

@@ -2198,6 +2198,8 @@ if is_flax_available():
 
     _import_structure["models.bart"].extend(
         [
+            "FlaxBartDecoderPreTrainedModel",
+            "FlaxBartForCausalLM",
             "FlaxBartForConditionalGeneration",
             "FlaxBartForQuestionAnswering",
             "FlaxBartForSequenceClassification",
@@ -4170,6 +4172,8 @@ if TYPE_CHECKING:
             FlaxAutoModelForVision2Seq,
         )
         from .models.bart import (
+            FlaxBartDecoderPreTrainedModel,
+            FlaxBartForCausalLM,
             FlaxBartForConditionalGeneration,
             FlaxBartForQuestionAnswering,
             FlaxBartForSequenceClassification,

src/transformers/models/auto/modeling_flax_auto.py

@@ -126,6 +126,7 @@ FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gpt_neo", "FlaxGPTNeoForCausalLM"),
         ("gptj", "FlaxGPTJForCausalLM"),
         ("xglm", "FlaxXGLMForCausalLM"),
+        ("bart", "FlaxBartForCausalLM"),
     ]
 )
 

src/transformers/models/bart/__init__.py

@@ -45,6 +45,8 @@ if is_tf_available():
 
 if is_flax_available():
     _import_structure["modeling_flax_bart"] = [
+        "FlaxBartDecoderPreTrainedModel",
+        "FlaxBartForCausalLM",
         "FlaxBartForConditionalGeneration",
         "FlaxBartForQuestionAnswering",
         "FlaxBartForSequenceClassification",
@@ -76,6 +78,8 @@ if TYPE_CHECKING:
 
     if is_flax_available():
         from .modeling_flax_bart import (
+            FlaxBartDecoderPreTrainedModel,
+            FlaxBartForCausalLM,
             FlaxBartForConditionalGeneration,
             FlaxBartForQuestionAnswering,
             FlaxBartForSequenceClassification,

src/transformers/models/bart/modeling_flax_bart.py

@@ -1725,3 +1725,264 @@ append_call_sample_docstring(
     FlaxSeq2SeqQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+class FlaxBartDecoderPreTrainedModel(FlaxPreTrainedModel):
+    config_class = BartConfig
+    base_model_prefix: str = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: BartConfig,
+        input_shape: Tuple[int] = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs
+    ):
+        config.is_decoder = True
+        config.is_encoder_decoder = False
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+
+        batch_size, sequence_length = input_ids.shape
+        position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+        encoder_hidden_states = jnp.zeros(input_shape + (self.config.d_model,))
+        encoder_attention_mask = attention_mask
+        module_init_outputs = self.module.init(
+            rngs,
+            input_ids,
+            attention_mask,
+            position_ids,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            return_dict=False,
+        )
+        return module_init_outputs["params"]
+
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
+    @add_start_docstrings_to_model_forward(BART_DECODE_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_ids: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        past_key_values: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if encoder_hidden_states is not None and encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # prepare decoder inputs
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+        if position_ids is None:
+            batch_size, sequence_length = input_ids.shape
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+        # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+        # changed by FlaxBartAttention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        outputs = self.module.apply(
+            inputs,
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past_key_values = outputs
+            outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past_key_values = outputs
+            outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        return outputs
+
+
+class FlaxBartDecoderWrapper(nn.Module):
+    """
+    This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is
+    used in combination with the [`EncoderDecoderModel`] framework.
+    """
+
+    config: BartConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.d_model
+        embed_tokens = nn.Embed(
+            self.config.vocab_size,
+            embed_dim,
+            embedding_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+        self.decoder = FlaxBartDecoder(config=self.config, embed_tokens=embed_tokens, dtype=self.dtype)
+
+    def __call__(self, *args, **kwargs):
+        return self.decoder(*args, **kwargs)
+
+
+class FlaxBartForCausalLMModule(nn.Module):
+    config: BartConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.model = FlaxBartDecoderWrapper(config=self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(self.config.init_std),
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+
+        outputs = self.model(
+            input_ids,
+            attention_mask,
+            position_ids,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.model.variables["params"]["decoder"]["embed_tokens"]["embedding"]
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_embedding.T}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            return (lm_logits,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=lm_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bart Decoder Model with a language modeling head on top (linear layer with weights tied to the input embeddings)
+    e.g for autoregressive tasks.
+    """,
+    BART_START_DOCSTRING,
+)
+class FlaxBartForCausalLM(FlaxBartDecoderPreTrainedModel):
+    module_class = FlaxBartForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxBartForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)

src/transformers/utils/dummy_flax_objects.py

@@ -249,6 +249,20 @@ class FlaxAutoModelForVision2Seq(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxBartDecoderPreTrainedModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxBartForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxBartForConditionalGeneration(metaclass=DummyObject):
     _backends = ["flax"]
 

tests/bart/test_modeling_flax_bart.py

@@ -11,17 +11,16 @@
 # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 # See the License for the specific language governing permissions and
 # limitations under the License.
-
 import unittest
 
 import numpy as np
 import timeout_decorator  # noqa
 
-from transformers import BartConfig, is_flax_available
+from transformers import BartConfig, BartTokenizer, is_flax_available
 from transformers.testing_utils import require_flax, slow
 
 from ..generation.test_generation_flax_utils import FlaxGenerationTesterMixin
-from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
+from ..test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
 
 
 if is_flax_available():
@@ -34,7 +33,6 @@ if is_flax_available():
 
     import jax
     import jax.numpy as jnp
-    from transformers import BartTokenizer
     from transformers.models.bart.modeling_flax_bart import (
         FlaxBartForConditionalGeneration,
         FlaxBartForQuestionAnswering,
@@ -475,3 +473,95 @@ class FlaxBartModelTest(FlaxModelTesterMixin, unittest.TestCase, FlaxGenerationT
             hypotheses_batch.tolist(), clean_up_tokenization_spaces=True, skip_special_tokens=True
         )
         assert generated_summaries == EXPECTED
+
+
+class FlaxBartStandaloneDecoderModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_attention_mask=True,
+        use_labels=False,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        intermediate_size=4,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=32,
+        eos_token_id=2,
+        pad_token_id=1,
+        bos_token_id=0,
+        initializer_range=0.02,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_attention_mask = use_attention_mask
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.initializer_range = initializer_range
+
+    def prepare_config_and_inputs(self):
+        input_ids = jnp.clip(ids_tensor([self.batch_size, self.seq_length], self.vocab_size), 3, self.vocab_size)
+
+        attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        config = BartConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            encoder_layers=self.num_hidden_layers,
+            decoder_layers=self.num_hidden_layers,
+            encoder_attention_heads=self.num_attention_heads,
+            decoder_attention_heads=self.num_attention_heads,
+            encoder_ffn_dim=self.intermediate_size,
+            decoder_ffn_dim=self.intermediate_size,
+            dropout=self.hidden_dropout_prob,
+            attention_dropout=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
+            initializer_range=self.initializer_range,
+            use_cache=False,
+        )
+
+        return config, input_ids, attention_mask
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, attention_mask = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "attention_mask": attention_mask}
+        return config, inputs_dict
+
+    def prepare_config_and_inputs_for_decoder(self):
+        config, input_ids, attention_mask = self.prepare_config_and_inputs()
+
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )

tests/encoder_decoder/test_modeling_flax_encoder_decoder.py

@@ -22,6 +22,7 @@ import numpy as np
 from transformers import is_flax_available, is_torch_available
 from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow, torch_device
 
+from ..bart.test_modeling_flax_bart import FlaxBartStandaloneDecoderModelTester
 from ..bert.test_modeling_flax_bert import FlaxBertModelTester
 from ..gpt2.test_modeling_flax_gpt2 import FlaxGPT2ModelTester
 from ..test_modeling_flax_common import ids_tensor
@@ -31,6 +32,7 @@ if is_flax_available():
     from transformers import (
         AutoTokenizer,
         EncoderDecoderConfig,
+        FlaxBartForCausalLM,
         FlaxBertModel,
         FlaxEncoderDecoderModel,
         FlaxGPT2LMHeadModel,
@@ -360,6 +362,7 @@ class FlaxEncoderDecoderMixin:
         self.assertTrue(decoder_config.cross_attention_hidden_size is None)
 
         # check without `enc_to_dec_proj` projection
+        decoder_config.hidden_size = config.hidden_size
         self.assertTrue(config.hidden_size == decoder_config.hidden_size)
         self.check_equivalence_pt_to_flax(config, decoder_config, inputs_dict)
         self.check_equivalence_flax_to_pt(config, decoder_config, inputs_dict)
@@ -456,6 +459,43 @@ class FlaxGPT2EncoderDecoderModelTest(FlaxEncoderDecoderMixin, unittest.TestCase
         self.assertEqual(summary, [EXPECTED_SUMMARY_STUDENTS])
 
 
+@require_flax
+class FlaxBartEncoderDecoderModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
+    def get_encoder_decoder_model(self, config, decoder_config):
+        encoder_model = FlaxBertModel(config)
+        decoder_model = FlaxBartForCausalLM(decoder_config)
+        return encoder_model, decoder_model
+
+    def prepare_config_and_inputs(self):
+        model_tester_encoder = FlaxBertModelTester(self, batch_size=13)
+        model_tester_decoder = FlaxBartStandaloneDecoderModelTester(self, batch_size=13)
+        encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
+        decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
+        (config, input_ids, token_type_ids, attention_mask) = encoder_config_and_inputs
+        (
+            decoder_config,
+            decoder_input_ids,
+            decoder_attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        ) = decoder_config_and_inputs
+
+        # make sure that cross attention layers are added
+        decoder_config.add_cross_attention = True
+        return {
+            "config": config,
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "decoder_config": decoder_config,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "encoder_hidden_states": encoder_hidden_states,
+        }
+
+    def get_pretrained_model(self):
+        return FlaxEncoderDecoderModel.from_encoder_decoder_pretrained("bert-base-cased", "facebook/bart-base")
+
+
 @require_flax
 class FlaxEncoderDecoderModelTest(unittest.TestCase):
     def get_from_encoderdecoder_pretrained_model(self):

tests/speech_encoder_decoder/test_modeling_flax_speech_encoder_decoder.py

@@ -21,6 +21,7 @@ import numpy as np
 from transformers import is_flax_available, is_torch_available
 from transformers.testing_utils import is_pt_flax_cross_test, require_flax, slow, torch_device
 
+from ..bart.test_modeling_flax_bart import FlaxBartStandaloneDecoderModelTester
 from ..gpt2.test_modeling_flax_gpt2 import FlaxGPT2ModelTester
 from ..test_modeling_flax_common import floats_tensor, ids_tensor, random_attention_mask
 from ..wav2vec2.test_modeling_flax_wav2vec2 import FlaxWav2Vec2ModelTester
@@ -28,6 +29,7 @@ from ..wav2vec2.test_modeling_flax_wav2vec2 import FlaxWav2Vec2ModelTester
 
 if is_flax_available():
     from transformers import (
+        FlaxBartForCausalLM,
         FlaxGPT2LMHeadModel,
         FlaxSpeechEncoderDecoderModel,
         FlaxWav2Vec2Model,
@@ -553,3 +555,120 @@ class FlaxWav2Vec2GPT2ModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
 
         self.assertEqual(len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch")
         self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)
+
+
+@require_flax
+class FlaxWav2Vec2BartModelTest(FlaxEncoderDecoderMixin, unittest.TestCase):
+    def get_pretrained_model_and_inputs(self):
+        model = FlaxSpeechEncoderDecoderModel.from_encoder_decoder_pretrained(
+            "facebook/wav2vec2-large-lv60", "bart-large"
+        )
+        batch_size = 13
+        input_values = floats_tensor([batch_size, 512], model.config.encoder.vocab_size)
+        attention_mask = random_attention_mask([batch_size, 512])
+        decoder_input_ids = ids_tensor([batch_size, 4], model.config.decoder.vocab_size)
+        decoder_attention_mask = random_attention_mask([batch_size, 4])
+        inputs = {
+            "inputs": input_values,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+        }
+
+        return model, inputs
+
+    def get_encoder_decoder_model(self, config, decoder_config):
+        encoder_model = FlaxWav2Vec2Model(config)
+        decoder_model = FlaxBartForCausalLM(decoder_config)
+        return encoder_model, decoder_model
+
+    def prepare_config_and_inputs(self):
+        model_tester_encoder = FlaxWav2Vec2ModelTester(self, batch_size=13)
+        model_tester_decoder = FlaxBartStandaloneDecoderModelTester(self, batch_size=13)
+        encoder_config_and_inputs = model_tester_encoder.prepare_config_and_inputs()
+        decoder_config_and_inputs = model_tester_decoder.prepare_config_and_inputs_for_decoder()
+        (config, inputs, attention_mask) = encoder_config_and_inputs
+        (
+            decoder_config,
+            decoder_input_ids,
+            decoder_attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        ) = decoder_config_and_inputs
+
+        # make sure that cross attention layers are added
+        decoder_config.add_cross_attention = True
+        return {
+            "config": config,
+            "inputs": inputs,
+            "attention_mask": attention_mask,
+            "decoder_config": decoder_config,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+            "encoder_hidden_states": encoder_hidden_states,
+        }
+
+    @slow
+    def test_flaxwav2vec2bart_pt_flax_equivalence(self):
+        pt_model = SpeechEncoderDecoderModel.from_pretrained("patrickvonplaten/wav2vec2-2-bart-large")
+        fx_model = FlaxSpeechEncoderDecoderModel.from_pretrained(
+            "patrickvonplaten/wav2vec2-2-bart-large", from_pt=True
+        )
+
+        pt_model.to(torch_device)
+        pt_model.eval()
+
+        # prepare inputs
+        batch_size = 13
+        input_values = floats_tensor([batch_size, 512], fx_model.config.encoder.vocab_size)
+        attention_mask = random_attention_mask([batch_size, 512])
+        decoder_input_ids = ids_tensor([batch_size, 4], fx_model.config.decoder.vocab_size)
+        decoder_attention_mask = random_attention_mask([batch_size, 4])
+        inputs_dict = {
+            "inputs": input_values,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+        }
+
+        flax_inputs = inputs_dict
+        pt_inputs = {k: torch.tensor(v.tolist()) for k, v in flax_inputs.items()}
+
+        with torch.no_grad():
+            pt_outputs = pt_model(**pt_inputs)
+        pt_logits = pt_outputs.logits
+        pt_outputs = pt_outputs.to_tuple()
+
+        fx_outputs = fx_model(**inputs_dict)
+        fx_logits = fx_outputs.logits
+        fx_outputs = fx_outputs.to_tuple()
+
+        self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+        self.assert_almost_equals(fx_logits, pt_logits.numpy(), 4e-2)
+
+        # PT -> Flax
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            pt_model.save_pretrained(tmpdirname)
+            fx_model_loaded = FlaxSpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_pt=True)
+
+        fx_outputs_loaded = fx_model_loaded(**inputs_dict)
+        fx_logits_loaded = fx_outputs_loaded.logits
+        fx_outputs_loaded = fx_outputs_loaded.to_tuple()
+        self.assertEqual(len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+        self.assert_almost_equals(fx_logits_loaded, pt_logits.numpy(), 4e-2)
+
+        # Flax -> PT
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            fx_model.save_pretrained(tmpdirname)
+            pt_model_loaded = SpeechEncoderDecoderModel.from_pretrained(tmpdirname, from_flax=True)
+
+        pt_model_loaded.to(torch_device)
+        pt_model_loaded.eval()
+
+        with torch.no_grad():
+            pt_outputs_loaded = pt_model_loaded(**pt_inputs)
+        pt_logits_loaded = pt_outputs_loaded.logits
+        pt_outputs_loaded = pt_outputs_loaded.to_tuple()
+
+        self.assertEqual(len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch")
+        self.assert_almost_equals(fx_logits, pt_logits_loaded.numpy(), 4e-2)

utils/check_repo.py

@@ -89,6 +89,7 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     "TFRobertaForMultipleChoice",  # TODO: fix
     "TrOCRDecoderWrapper",  # Building part of bigger (tested) model.
     "SeparableConv1D",  # Building part of bigger (tested) model.
+    "FlaxBartForCausalLM",  # Building part of bigger (tested) model.
 ]
 
 # Update this list with test files that don't have a tester with a `all_model_classes` variable and which don't