passing most tests

model/config.py

@@ -10,7 +10,7 @@ from model.utils import assertEqual, assertIn
 logger = logging.get_logger(__name__)
 
 
-class T5_VAE_Config(PretrainedConfig):
+class T5VaeConfig(PretrainedConfig):
     r"""
     This is the configuration class to store the configuration of :class:`FlaxT5VAE`.
     It is used to instantiate a T5-VAE model according to the specified arguments, defining the model architecture.
@@ -22,8 +22,8 @@ class T5_VAE_Config(PretrainedConfig):
     outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
 
     Arguments:
-        latent_size (:obj:`int`, `optional`, defaults to 1,000):
-            Number of dimensions to use for the sequences latent code.
+        latent_token_size (:obj:`int`, `optional`, defaults to 1,000):
+            Number of dimensions to use for each latent token.
         t5_name (:obj:`str`, `optional`, defaults to t5-base):
             Name of the Transformer model to use as a decoder.
         block_size (:obj:`int`, `optional`, defaults to 60):
@@ -37,8 +37,8 @@ class T5_VAE_Config(PretrainedConfig):
     def __init__(
         self,
         t5_model_name_or_path="t5-base",
-        n_latent_tokens=5,  # set to -1 for full sequence
-        latent_size=1_000,
+        n_latent_tokens=6,  # set to -1 for full sequence
+        latent_token_size=768,
         vae_encoder_model='',
         vae_decoder_model='',
         block_size=60,
@@ -51,7 +51,6 @@ class T5_VAE_Config(PretrainedConfig):
         num_layers=0,
         num_heads=0,
         tie_word_embeddings=True,
-        skip_upsample=False,
         **kwargs,
     ):
         assertIn(vae_encoder_model, VAE_ENCODER_MODELS.keys(), "Unexpected VAE encoder.")
@@ -65,28 +64,35 @@ class T5_VAE_Config(PretrainedConfig):
         self.vae_encoder_model = vae_encoder_model
         self.vae_decoder_model = vae_decoder_model
 
+        self.latent_token_size = latent_token_size
         assert(n_latent_tokens <= self.set_seq_size, 'Cannot use more latent tokens than input tokens.')
-        self.latent_size = latent_size
         self.n_latent_tokens = n_latent_tokens
-        self.skip_upsample = skip_upsample
 
         # T5
         if 't5' not in kwargs:
             self.t5 = AutoConfig.from_pretrained(t5_model_name_or_path, cache_dir=cache_dir)
+            assertEqual(self.t5.model_type, "t5", "Need t5 model type for transformer_decoder.")
             if num_layers:
                 self.t5.num_layers = num_layers
             if num_heads:
                 self.t5.num_heads = num_heads
             self.t5.decoder_start_token_id = decoder_start_token_id
             self.t5.n_positions = self.set_seq_size
-            assertEqual(self.t5.model_type, "t5", "Need t5 model type for transformer_decoder.")
         else:
-            self.t5 = T5Config(**kwargs.pop('t5'))
+            self.t5 = T5Config(
+                num_layers=num_layers, num_heads=num_heads,
+                decoder_start_token_id=decoder_start_token_id,
+                n_positions=self.set_seq_size, **kwargs
+            )
+
+        if self.t5.d_model < self.latent_token_size:
+            raise Exception('Using larger latent token dimension then T5 hidden dimension.')
 
-        # misc
+        # Add t5 config options
         self.tie_word_embeddings = tie_word_embeddings
         self.t5.tie_word_embeddings = self.tie_word_embeddings
-        self.use_cache = getattr(self.t5, "use_cache", False)
+        for attr in 'vocab_size hidden_size num_attention_heads num_hidden_layers use_cache'.split():
+            setattr(self, attr, getattr(self.t5, attr))
 
     def to_dict(self):
         """

model/encoders.py

@@ -9,12 +9,12 @@ class Encoder(nn.Module):
     '''
         Converts N hidden tokens into N seperate latent codes.
     '''
-    latent_size: int
+    latent_token_size: int
     n_latent_tokens: int
 
     @nn.compact
     def __call__(self, encoding):
-        latent_tokens = nn.Dense(self.latent_size)(encoding)
+        latent_tokens = nn.Dense(self.latent_token_size)(encoding)
         raw_latent_code = latent_tokens[:, : self.n_latent_tokens, :]
         # TODO does this just apply tanh to each latent token? Or across the whole batch
         latent_code = jnp.tanh(raw_latent_code)

model/outputs.py

@@ -12,7 +12,7 @@ class TransformerVAE_Output(ModelOutput):
     Base class for a Transformer-VAE's outputs.
 
     Args:
-        latent_codes (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, latent_size)`):
+        latent_codes (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, n_latent_tokens, latent_token_size)`):
             Latent codes representing encoded sequences.
         remade_encoder_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, n_tokens, model_dim)`):
             Reconstructed encoder hidden states representing sequences.

model/t5_vae.py

@@ -13,17 +13,23 @@ from transformers.models.t5.modeling_flax_t5 import FlaxT5ForConditionalGenerati
 
 from model.vae import VAE
 from model.outputs import TransformerVAE_Output
-from model.config import T5_VAE_Config
+from model.config import T5VaeConfig
 
 
 @add_start_docstrings("""T5 Model with a `language modeling` head on top converted into a VAE.""")
-class FlaxT5_VAE_ForAutoencodingModule(nn.Module):
-    config: T5_VAE_Config
+class FlaxT5VaeForAutoencodingModule(nn.Module):
+    config: T5VaeConfig
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def _get_encoder_module(self):
         return self.t5.encoder
 
+    def _get_vae_encoder_module(self):
+        return self.vae.encoder
+
+    def _get_vae_decoder_module(self):
+        return self.vae.decoder
+
     def _get_decoder_module(self):
         return self.t5.decoder
 
@@ -42,7 +48,6 @@ class FlaxT5_VAE_ForAutoencodingModule(nn.Module):
         return_dict=None,
         deterministic: bool = True,
     ):
-        # TODO should I use None args when everything has to be computed anyway?
         """
             Adapted from `FlaxT5ForConditionalGenerationModule`
         """
@@ -104,19 +109,19 @@ class FlaxT5_VAE_ForAutoencodingModule(nn.Module):
         )
 
 
-class FlaxT5_VAE_PreTrainedModel(FlaxPreTrainedModel):
+class FlaxT5VaePreTrainedModel(FlaxPreTrainedModel):
     """
     An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
     models.
     """
 
-    config_class = T5_VAE_Config
+    config_class = T5VaeConfig
     base_model_prefix = "transformer"
     module_class: nn.Module = None
 
     def __init__(
         self,
-        config: T5_VAE_Config,
+        config: T5VaeConfig,
         input_shape: Tuple[int] = (1, 1),
         seed: int = 0,
         dtype: jnp.dtype = jnp.float32,
@@ -208,19 +213,21 @@ class FlaxT5_VAE_PreTrainedModel(FlaxPreTrainedModel):
         decoder_input_ids = jnp.ones((batch_size, max_length), dtype="i4")
         decoder_attention_mask = jnp.ones_like(decoder_input_ids)
 
-        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, **kwargs):
+        def _decoder_forward(module, decoder_input_ids, latent_codes, decoder_attention_mask, **kwargs):
+            vae_decoder_module = module._get_vae_decoder_module()
             decoder_module = module._get_decoder_module()
             return decoder_module(
                 decoder_input_ids,
                 decoder_attention_mask,
+                encoder_hidden_states=vae_decoder_module(latent_codes),
                 **kwargs,
             )
 
         init_variables = self.module.init(
             jax.random.PRNGKey(0),
             decoder_input_ids=decoder_input_ids,
-            decoder_attention_mask=decoder_attention_mask,
             latent_codes=latent_codes,
+            decoder_attention_mask=decoder_attention_mask,
             init_cache=True,
             method=_decoder_forward,  # we only need to call the decoder to init the cache
         )
@@ -256,8 +263,8 @@ class FlaxT5_VAE_PreTrainedModel(FlaxPreTrainedModel):
         raise NotImplementedError()
 
 
-class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
-    module_class = FlaxT5_VAE_ForAutoencodingModule
+class FlaxT5VaeForAutoencoding(FlaxT5VaePreTrainedModel):
+    module_class = FlaxT5VaeForAutoencodingModule
 
     def __call__(
         self,
@@ -308,18 +315,6 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
         params: dict = None,
         dropout_rng: PRNGKey = None,
     ):
-        r"""
-        Returns:
-
-        Example::
-
-            >>> model = FlaxT5_VAE_ForAutoencoding.from_pretrained('t5-small')
-            >>> tokenizer = T5Tokenizer.from_pretrained('t5-small')
-
-            >>> text = "My friends are cool but they eat too many carbs."
-            >>> inputs = tokenizer(text, max_length=512, return_tensors='jax')
-            >>> latent_codes = model.encode(**inputs)
-        """
         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
@@ -335,20 +330,9 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
             rngs["dropout"] = dropout_rng
 
         def _encoder_forward(module, input_ids, attention_mask, **kwargs):
-            # Encode
-            encoder_outputs = self.t5.encoder(
-                input_ids=input_ids,
-                attention_mask=attention_mask,
-                output_attentions=output_attentions,
-                output_hidden_states=output_hidden_states,
-                return_dict=return_dict,
-                deterministic=not train,
-            )
-
-            hidden_states = encoder_outputs[0]
-
-            # Autoencode
-            return self.vae(hidden_states, kwargs.get('latent_codes'))
+            encode_module = module._get_encoder_module()
+            vae_encoder_module = module._get_vae_encoder_module()
+            return vae_encoder_module(encode_module(input_ids, attention_mask, **kwargs)[0])
 
         return self.module.apply(
             {"params": params or self.params},
@@ -381,7 +365,7 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
 
         Example::
 
-            >>> model = FlaxT5_VAE_ForAutoencoding.from_pretrained('t5-small')
+            >>> model = FlaxT5VaeForAutoencoding.from_pretrained('t5-small')
             >>> tokenizer = T5Tokenizer.from_pretrained('t5-small')
 
             >>> text = "My friends are cool but they eat too many carbs."
@@ -400,10 +384,8 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
         )
         return_dict = return_dict if return_dict is not None else self.config.return_dict
 
-        # TODO match latent_codes to encoder hidden states size
-        encoder_hidden_states = latent_codes
         if encoder_attention_mask is None:
-            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            batch_size, sequence_length = latent_codes.shape[:2]
             encoder_attention_mask = jnp.ones((batch_size, sequence_length))
 
         batch_size, sequence_length = decoder_input_ids.shape
@@ -426,14 +408,15 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
         else:
             mutable = False
 
-        def _decoder_forward(module, decoder_input_ids, decoder_attention_mask, **kwargs):
+        def _decoder_forward(module, decoder_input_ids, latent_codes, decoder_attention_mask, **kwargs):
+            vae_decoder_module = module._get_vae_decoder_module()
             decoder_module = module._get_decoder_module()
             decoder_outputs = decoder_module(
                 decoder_input_ids,
                 decoder_attention_mask,
+                encoder_hidden_states=vae_decoder_module(latent_codes),
                 **kwargs,
             )
-
             sequence_output = decoder_outputs[0]
 
             if self.config.tie_word_embeddings:
@@ -442,18 +425,18 @@ class FlaxT5_VAE_ForAutoencoding(FlaxT5_VAE_PreTrainedModel):
                 sequence_output = sequence_output * (self.config.d_model ** -0.5)
 
             if self.config.tie_word_embeddings:
-                shared_embedding = module.shared.variables["params"]["embedding"]
-                lm_logits = module.lm_head.apply({"params": {"kernel": shared_embedding.T}}, sequence_output)
+                shared_embedding = module.t5.shared.variables["params"]["embedding"]
+                lm_logits = module.t5.lm_head.apply({"params": {"kernel": shared_embedding.T}}, sequence_output)
             else:
-                lm_logits = module.lm_head(sequence_output)
+                lm_logits = module.t5.lm_head(sequence_output)
 
             return lm_logits, decoder_outputs
 
         outputs = self.module.apply(
             inputs,
             decoder_input_ids=jnp.array(decoder_input_ids, dtype="i4"),
+            latent_codes=latent_codes,
             decoder_attention_mask=jnp.array(decoder_attention_mask, dtype="i4"),
-            encoder_hidden_states=encoder_hidden_states,
             encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,

model/vae.py

@@ -3,7 +3,7 @@ import flax.linen as nn
 
 from model.encoders import VAE_ENCODER_MODELS
 from model.decoders import VAE_DECODER_MODELS
-from model.config import T5_VAE_Config
+from model.config import T5VaeConfig
 
 
 class VAE(nn.Module):
@@ -12,11 +12,11 @@ class VAE(nn.Module):
         An MMD-VAE used with encoder-decoder models.
         Encodes all token encodings into a single latent & spits them back out.
     """
-    config: T5_VAE_Config
+    config: T5VaeConfig
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        self.encoder = VAE_ENCODER_MODELS[self.config.vae_encoder_model](self.config.latent_size, self.config.n_latent_tokens)
+        self.encoder = VAE_ENCODER_MODELS[self.config.vae_encoder_model](self.config.latent_token_size, self.config.n_latent_tokens)
         self.decoder = VAE_DECODER_MODELS[self.config.vae_decoder_model](self.config.t5.d_model,  self.config.n_latent_tokens)
 
     def __call__(self, encoding=None, latent_codes=None):

run_clm_flax.py

@@ -405,7 +405,7 @@ def main():
             k: [t[i : i + block_size] for i in range(0, total_length, block_size)]
             for k, t in concatenated_examples.items()
         }
-        result["label"] = result["input_ids"].copy()
+        result["labels"] = result["input_ids"].copy()
         return result
 
     # Note that with `batched=True`, this map processes 1,000 texts together, so group_texts throws away a remainder
@@ -421,8 +421,6 @@ def main():
         num_proc=data_args.preprocessing_num_workers,
         load_from_cache_file=not data_args.overwrite_cache,
     )
-    import pdb
-    pdb.set_trace()
 
     if training_args.do_train:
         if "train" not in tokenized_datasets:
@@ -624,7 +622,6 @@ def main():
 
                 # Save metrics
                 if has_tensorboard and jax.process_index() == 0:
-                    cur_step = epoch * (len(train_dataset) // train_batch_size)
                     write_eval_metric(summary_writer, eval_metrics, cur_step)
 
             if cur_step % training_args.save_steps == 0 and cur_step > 0:

tests/test_configuration_common.py

@@ -0,0 +1,157 @@
+# coding=utf-8
+# Copyright 2019 HuggingFace Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 json
+import os
+import tempfile
+import unittest
+
+from huggingface_hub import HfApi
+from requests.exceptions import HTTPError
+from transformers import BertConfig, GPT2Config
+from transformers.testing_utils import ENDPOINT_STAGING, PASS, USER, is_staging_test
+
+
+class ConfigTester(object):
+    def __init__(self, parent, config_class=None, has_text_modality=True, **kwargs):
+        self.parent = parent
+        self.config_class = config_class
+        self.has_text_modality = has_text_modality
+        self.inputs_dict = kwargs
+
+    def create_and_test_config_common_properties(self):
+        config = self.config_class(**self.inputs_dict)
+        if self.has_text_modality:
+            self.parent.assertTrue(hasattr(config, "vocab_size"))
+        self.parent.assertTrue(hasattr(config, "hidden_size"))
+        self.parent.assertTrue(hasattr(config, "num_attention_heads"))
+        self.parent.assertTrue(hasattr(config, "num_hidden_layers"))
+
+    def create_and_test_config_to_json_string(self):
+        config = self.config_class(**self.inputs_dict)
+        obj = json.loads(config.to_json_string())
+        for key, value in self.inputs_dict.items():
+            self.parent.assertEqual(obj[key], value)
+
+    def create_and_test_config_to_json_file(self):
+        config_first = self.config_class(**self.inputs_dict)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            json_file_path = os.path.join(tmpdirname, "config.json")
+            config_first.to_json_file(json_file_path)
+            config_second = self.config_class.from_json_file(json_file_path)
+
+        self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
+
+    def create_and_test_config_from_and_save_pretrained(self):
+        config_first = self.config_class(**self.inputs_dict)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            config_first.save_pretrained(tmpdirname)
+            config_second = self.config_class.from_pretrained(tmpdirname)
+
+        self.parent.assertEqual(config_second.to_dict(), config_first.to_dict())
+
+    def create_and_test_config_with_num_labels(self):
+        config = self.config_class(**self.inputs_dict, num_labels=5)
+        self.parent.assertEqual(len(config.id2label), 5)
+        self.parent.assertEqual(len(config.label2id), 5)
+
+        config.num_labels = 3
+        self.parent.assertEqual(len(config.id2label), 3)
+        self.parent.assertEqual(len(config.label2id), 3)
+
+    def check_config_can_be_init_without_params(self):
+        if self.config_class.is_composition:
+            return
+        config = self.config_class()
+        self.parent.assertIsNotNone(config)
+
+    def run_common_tests(self):
+        self.create_and_test_config_common_properties()
+        self.create_and_test_config_to_json_string()
+        self.create_and_test_config_to_json_file()
+        self.create_and_test_config_from_and_save_pretrained()
+        self.create_and_test_config_with_num_labels()
+        self.check_config_can_be_init_without_params()
+
+
+@is_staging_test
+class ConfigPushToHubTester(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls._api = HfApi(endpoint=ENDPOINT_STAGING)
+        cls._token = cls._api.login(username=USER, password=PASS)
+
+    @classmethod
+    def tearDownClass(cls):
+        try:
+            cls._api.delete_repo(token=cls._token, name="test-config")
+        except HTTPError:
+            pass
+
+        try:
+            cls._api.delete_repo(token=cls._token, name="test-config-org", organization="valid_org")
+        except HTTPError:
+            pass
+
+    def test_push_to_hub(self):
+        config = BertConfig(
+            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
+        )
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            config.save_pretrained(os.path.join(tmp_dir, "test-config"), push_to_hub=True, use_auth_token=self._token)
+
+            new_config = BertConfig.from_pretrained(f"{USER}/test-config")
+            for k, v in config.__dict__.items():
+                if k != "transformers_version":
+                    self.assertEqual(v, getattr(new_config, k))
+
+    def test_push_to_hub_in_organization(self):
+        config = BertConfig(
+            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
+        )
+
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            config.save_pretrained(
+                os.path.join(tmp_dir, "test-config-org"),
+                push_to_hub=True,
+                use_auth_token=self._token,
+                organization="valid_org",
+            )
+
+            new_config = BertConfig.from_pretrained("valid_org/test-config-org")
+            for k, v in config.__dict__.items():
+                if k != "transformers_version":
+                    self.assertEqual(v, getattr(new_config, k))
+
+
+class ConfigTestUtils(unittest.TestCase):
+    def test_config_from_string(self):
+        c = GPT2Config()
+
+        # attempt to modify each of int/float/bool/str config records and verify they were updated
+        n_embd = c.n_embd + 1  # int
+        resid_pdrop = c.resid_pdrop + 1.0  # float
+        scale_attn_weights = not c.scale_attn_weights  # bool
+        summary_type = c.summary_type + "foo"  # str
+        c.update_from_string(
+            f"n_embd={n_embd},resid_pdrop={resid_pdrop},scale_attn_weights={scale_attn_weights},summary_type={summary_type}"
+        )
+        self.assertEqual(n_embd, c.n_embd, "mismatch for key: n_embd")
+        self.assertEqual(resid_pdrop, c.resid_pdrop, "mismatch for key: resid_pdrop")
+        self.assertEqual(scale_attn_weights, c.scale_attn_weights, "mismatch for key: scale_attn_weights")
+        self.assertEqual(summary_type, c.summary_type, "mismatch for key: summary_type")

tests/test_generation_flax_utils.py

@@ -0,0 +1,247 @@
+# Copyright 2021 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 random
+
+import numpy as np
+
+from transformers import is_flax_available
+from transformers.testing_utils import require_flax
+
+
+if is_flax_available():
+    import os
+
+    import jax
+    import jax.numpy as jnp
+    from jax import jit
+
+    os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "0.12"  # assumed parallelism: 8
+
+
+def ids_tensor(shape, vocab_size, rng=None):
+    """Creates a random int32 tensor of the shape within the vocab size."""
+    if rng is None:
+        rng = random.Random()
+
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.randint(0, vocab_size - 1))
+
+    output = np.array(values, dtype=jnp.int32).reshape(shape)
+
+    return output
+
+
+def random_attention_mask(shape, rng=None):
+    attn_mask = ids_tensor(shape, vocab_size=2, rng=rng)
+    # make sure that at least one token is attended to for each batch
+    attn_mask[:, -1] = 1
+    return attn_mask
+
+
+@require_flax
+class FlaxGenerationTesterMixin:
+    model_tester = None
+    all_generative_model_classes = ()
+
+    def _get_input_ids_and_config(self):
+        config, inputs = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # cut to half length & take max batch_size 3
+        max_batch_size = 2
+        sequence_length = inputs["input_ids"].shape[-1] // 2
+        input_ids = inputs["input_ids"][:max_batch_size, :sequence_length]
+
+        attention_mask = jnp.ones_like(input_ids)
+        attention_mask = attention_mask[:max_batch_size, :sequence_length]
+
+        # generate max 5 tokens
+        max_length = input_ids.shape[-1] + 5
+        if config.eos_token_id is not None and config.pad_token_id is None:
+            # hack to allow generate for models such as GPT2 as is done in `generate()`
+            config.pad_token_id = config.eos_token_id
+        return config, input_ids, attention_mask, max_length
+
+    def test_greedy_generate(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.do_sample = False
+        config.max_length = max_length
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_sample_generate(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.do_sample = True
+        config.max_length = max_length
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_beam_search_generate(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.do_sample = False
+        config.max_length = max_length
+        config.num_beams = 2
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_sample_generate_logits_warper(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.do_sample = True
+        config.max_length = max_length
+        config.temperature = 0.8
+        config.top_k = 10
+        config.top_p = 0.3
+        config.min_length = 1
+        config.forced_bos_token_id = 8
+        config.forced_eos_token_id = 9
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_greedy_generate_logits_warper(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.max_length = max_length
+        config.min_length = 1
+        config.forced_bos_token_id = 8
+        config.forced_eos_token_id = 9
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_beam_search_generate_logits_warper(self):
+        config, input_ids, _, max_length = self._get_input_ids_and_config()
+        config.max_length = max_length
+        config.num_beams = 2
+        config.min_length = 1
+        config.forced_bos_token_id = 8
+        config.forced_eos_token_id = 9
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_greedy_generate_attn_mask(self):
+        config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
+
+        # pad attention mask on the left
+        attention_mask = jax.ops.index_update(attention_mask, (0, 0), 0)
+
+        config.do_sample = False
+        config.max_length = max_length
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids, attention_mask=attention_mask).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids, attention_mask=attention_mask).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_sample_generate_attn_mask(self):
+        config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
+
+        # pad attention mask on the left
+        attention_mask = jax.ops.index_update(attention_mask, (0, 0), 0)
+
+        config.do_sample = True
+        config.max_length = max_length
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids, attention_mask=attention_mask).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids, attention_mask=attention_mask).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())
+
+    def test_beam_search_generate_attn_mask(self):
+        config, input_ids, attention_mask, max_length = self._get_input_ids_and_config()
+
+        # pad attention mask on the left
+        attention_mask = jax.ops.index_update(attention_mask, (0, 0), 0)
+
+        config.num_beams = 2
+        config.max_length = max_length
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            generation_outputs = model.generate(input_ids, attention_mask=attention_mask).sequences
+            self.assertEqual(generation_outputs.shape[-1], max_length)
+
+            jit_generate = jit(model.generate)
+            jit_generation_outputs = jit_generate(input_ids, attention_mask=attention_mask).sequences
+
+            self.assertListEqual(generation_outputs.tolist(), jit_generation_outputs.tolist())

tests/test_modeling_flax_common.py

@@ -0,0 +1,579 @@
+# Copyright 2020 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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 copy
+import inspect
+import random
+import tempfile
+import unittest
+from typing import List, Tuple
+
+import numpy as np
+
+import transformers
+from huggingface_hub import HfApi
+from requests.exceptions import HTTPError
+from transformers import BertConfig, FlaxBertModel, is_flax_available, is_torch_available
+from transformers.models.auto import get_values
+from transformers.testing_utils import (
+    ENDPOINT_STAGING,
+    PASS,
+    USER,
+    is_pt_flax_cross_test,
+    is_staging_test,
+    require_flax,
+    slow,
+)
+
+
+if is_flax_available():
+    import os
+
+    import jax
+    import jax.numpy as jnp
+    import jaxlib.xla_extension as jax_xla
+    from flax.core.frozen_dict import unfreeze
+    from flax.traverse_util import flatten_dict
+    from transformers import FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING, FLAX_MODEL_MAPPING
+    from transformers.modeling_flax_pytorch_utils import (
+        convert_pytorch_state_dict_to_flax,
+        load_flax_weights_in_pytorch_model,
+    )
+
+    os.environ["XLA_PYTHON_CLIENT_MEM_FRACTION"] = "0.12"  # assumed parallelism: 8
+
+if is_torch_available():
+    import torch
+
+
+def _config_zero_init(config):
+    configs_no_init = copy.deepcopy(config)
+    for key in configs_no_init.__dict__.keys():
+        if "_range" in key or "_std" in key or "initializer_factor" in key:
+            setattr(configs_no_init, key, 1e-10)
+    return configs_no_init
+
+
+def ids_tensor(shape, vocab_size, rng=None):
+    """Creates a random int32 tensor of the shape within the vocab size."""
+    if rng is None:
+        rng = random.Random()
+
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.randint(0, vocab_size - 1))
+
+    output = np.array(values, dtype=jnp.int32).reshape(shape)
+
+    return output
+
+
+def floats_tensor(shape, scale=1.0, rng=None, name=None):
+    """Creates a random float32 tensor"""
+    if rng is None:
+        rng = random.Random()
+
+    total_dims = 1
+    for dim in shape:
+        total_dims *= dim
+
+    values = []
+    for _ in range(total_dims):
+        values.append(rng.random() * scale)
+
+    return np.array(values, dtype=jnp.float32).reshape(shape)
+
+
+def random_attention_mask(shape, rng=None):
+    attn_mask = ids_tensor(shape, vocab_size=2, rng=rng)
+    # make sure that at least one token is attended to for each batch
+    attn_mask[:, -1] = 1
+    return attn_mask
+
+
+@require_flax
+class FlaxModelTesterMixin:
+    model_tester = None
+    all_model_classes = ()
+    is_encoder_decoder = False
+
+    def _prepare_for_class(self, inputs_dict, model_class):
+        inputs_dict = copy.deepcopy(inputs_dict)
+
+        # hack for now until we have AutoModel classes
+        if "ForMultipleChoice" in model_class.__name__:
+            inputs_dict = {
+                k: jnp.broadcast_to(v[:, None], (v.shape[0], self.model_tester.num_choices, v.shape[-1]))
+                if isinstance(v, (jax_xla.DeviceArray, np.ndarray))
+                else v
+                for k, v in inputs_dict.items()
+            }
+
+        return inputs_dict
+
+    def assert_almost_equals(self, a: np.ndarray, b: np.ndarray, tol: float):
+        diff = np.abs((a - b)).max()
+        self.assertLessEqual(diff, tol, f"Difference between torch and flax is {diff} (>= {tol}).")
+
+    def test_model_outputs_equivalence(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        def set_nan_tensor_to_zero(t):
+            t[t != t] = 0
+            return t
+
+        def check_equivalence(model, tuple_inputs, dict_inputs, additional_kwargs={}):
+            tuple_output = model(**tuple_inputs, return_dict=False, **additional_kwargs)
+            dict_output = model(**dict_inputs, return_dict=True, **additional_kwargs).to_tuple()
+
+            def recursive_check(tuple_object, dict_object):
+                if isinstance(tuple_object, (List, Tuple)):
+                    for tuple_iterable_value, dict_iterable_value in zip(tuple_object, dict_object):
+                        recursive_check(tuple_iterable_value, dict_iterable_value)
+                elif tuple_object is None:
+                    return
+                else:
+                    self.assert_almost_equals(
+                        set_nan_tensor_to_zero(tuple_object), set_nan_tensor_to_zero(dict_object), 1e-5
+                    )
+
+                recursive_check(tuple_output, dict_output)
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+
+            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
+            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
+            check_equivalence(model, tuple_inputs, dict_inputs)
+
+            tuple_inputs = self._prepare_for_class(inputs_dict, model_class)
+            dict_inputs = self._prepare_for_class(inputs_dict, model_class)
+            check_equivalence(model, tuple_inputs, dict_inputs, {"output_hidden_states": True})
+
+    @is_pt_flax_cross_test
+    def test_equivalence_pt_to_flax(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                # prepare inputs
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
+
+                # load corresponding PyTorch class
+                pt_model_class_name = model_class.__name__[4:]  # Skip the "Flax" at the beginning
+                pt_model_class = getattr(transformers, pt_model_class_name)
+
+                pt_model = pt_model_class(config).eval()
+                # Flax models don't use the `use_cache` option and cache is not returned as a default.
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+                fx_model = model_class(config, dtype=jnp.float32)
+
+                fx_state = convert_pytorch_state_dict_to_flax(pt_model.state_dict(), fx_model)
+                fx_model.params = fx_state
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+                for fx_output, pt_output in zip(fx_outputs, pt_outputs):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    pt_model.save_pretrained(tmpdirname)
+                    fx_model_loaded = model_class.from_pretrained(tmpdirname, from_pt=True)
+
+                fx_outputs_loaded = fx_model_loaded(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(
+                    len(fx_outputs_loaded), len(pt_outputs), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output_loaded, pt_output in zip(fx_outputs_loaded, pt_outputs):
+                    self.assert_almost_equals(fx_output_loaded, pt_output.numpy(), 4e-2)
+
+    @is_pt_flax_cross_test
+    def test_equivalence_flax_to_pt(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                # prepare inputs
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                pt_inputs = {k: torch.tensor(v.tolist()) for k, v in prepared_inputs_dict.items()}
+
+                # load corresponding PyTorch class
+                pt_model_class_name = model_class.__name__[4:]  # Skip the "Flax" at the beginning
+                pt_model_class = getattr(transformers, pt_model_class_name)
+
+                pt_model = pt_model_class(config).eval()
+                # Flax models don't use the `use_cache` option and cache is not returned as a default.
+                # So we disable `use_cache` here for PyTorch model.
+                pt_model.config.use_cache = False
+                fx_model = model_class(config, dtype=jnp.float32)
+
+                pt_model = load_flax_weights_in_pytorch_model(pt_model, fx_model.params)
+
+                # make sure weights are tied in PyTorch
+                pt_model.tie_weights()
+
+                with torch.no_grad():
+                    pt_outputs = pt_model(**pt_inputs).to_tuple()
+
+                fx_outputs = fx_model(**prepared_inputs_dict).to_tuple()
+                self.assertEqual(len(fx_outputs), len(pt_outputs), "Output lengths differ between Flax and PyTorch")
+
+                for fx_output, pt_output in zip(fx_outputs, pt_outputs):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    fx_model.save_pretrained(tmpdirname)
+                    pt_model_loaded = pt_model_class.from_pretrained(tmpdirname, from_flax=True)
+
+                with torch.no_grad():
+                    pt_outputs_loaded = pt_model_loaded(**pt_inputs).to_tuple()
+
+                self.assertEqual(
+                    len(fx_outputs), len(pt_outputs_loaded), "Output lengths differ between Flax and PyTorch"
+                )
+                for fx_output, pt_output in zip(fx_outputs, pt_outputs_loaded):
+                    self.assert_almost_equals(fx_output, pt_output.numpy(), 4e-2)
+
+    def test_from_pretrained_save_pretrained(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                model = model_class(config)
+
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                outputs = model(**prepared_inputs_dict).to_tuple()
+
+                # verify that normal save_pretrained works as expected
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname)
+                    model_loaded = model_class.from_pretrained(tmpdirname)
+
+                outputs_loaded = model_loaded(**prepared_inputs_dict).to_tuple()
+                for output_loaded, output in zip(outputs_loaded, outputs):
+                    self.assert_almost_equals(output_loaded, output, 1e-3)
+
+                # verify that save_pretrained for distributed training
+                # with `params=params` works as expected
+                with tempfile.TemporaryDirectory() as tmpdirname:
+                    model.save_pretrained(tmpdirname, params=model.params)
+                    model_loaded = model_class.from_pretrained(tmpdirname)
+
+                outputs_loaded = model_loaded(**prepared_inputs_dict).to_tuple()
+                for output_loaded, output in zip(outputs_loaded, outputs):
+                    self.assert_almost_equals(output_loaded, output, 1e-3)
+
+    def test_save_load_from_base(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        base_class = FLAX_MODEL_MAPPING[config.__class__]
+
+        for model_class in self.all_model_classes:
+            if model_class == base_class:
+                continue
+
+            model = base_class(config)
+            base_params = flatten_dict(unfreeze(model.params))
+
+            # check that all base model weights are loaded correctly
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                head_model = model_class.from_pretrained(tmpdirname)
+
+                base_param_from_head = flatten_dict(unfreeze(head_model.params[head_model.base_model_prefix]))
+
+                for key in base_param_from_head.keys():
+                    max_diff = (base_params[key] - base_param_from_head[key]).sum().item()
+                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
+
+    def test_save_load_to_base(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        base_class = FLAX_MODEL_MAPPING[config.__class__]
+
+        for model_class in self.all_model_classes:
+            if model_class == base_class:
+                continue
+
+            model = model_class(config)
+            base_params_from_head = flatten_dict(unfreeze(model.params[model.base_model_prefix]))
+
+            # check that all base model weights are loaded correctly
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                base_model = base_class.from_pretrained(tmpdirname)
+
+                base_params = flatten_dict(unfreeze(base_model.params))
+
+                for key in base_params_from_head.keys():
+                    max_diff = (base_params[key] - base_params_from_head[key]).sum().item()
+                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
+
+    @slow
+    def test_jit_compilation(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                model = model_class(config)
+
+                @jax.jit
+                def model_jitted(input_ids, attention_mask=None, **kwargs):
+                    return model(input_ids=input_ids, attention_mask=attention_mask, **kwargs)
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = model_jitted(**prepared_inputs_dict).to_tuple()
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = model_jitted(**prepared_inputs_dict).to_tuple()
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs, outputs):
+
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.__call__)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            if model.config.is_encoder_decoder:
+                expected_arg_names = [
+                    "input_ids",
+                    "attention_mask",
+                    "decoder_input_ids",
+                    "decoder_attention_mask",
+                ]
+                self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+            else:
+                expected_arg_names = ["input_ids", "attention_mask"]
+                self.assertListEqual(arg_names[:2], expected_arg_names)
+
+    def test_naming_convention(self):
+        for model_class in self.all_model_classes:
+            model_class_name = model_class.__name__
+            module_class_name = (
+                model_class_name[:-5] + "Module" if model_class_name[-5:] == "Model" else model_class_name + "Module"
+            )
+            bert_modeling_flax_module = __import__(model_class.__module__, fromlist=[module_class_name])
+            module_cls = getattr(bert_modeling_flax_module, module_class_name)
+
+            self.assertIsNotNone(module_cls)
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            if hasattr(self.model_tester, "encoder_seq_length"):
+                seq_length = self.model_tester.encoder_seq_length
+            else:
+                seq_length = self.model_tester.seq_length
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [seq_length, self.model_tester.hidden_size],
+            )
+
+            if config.is_encoder_decoder:
+                hidden_states = outputs.decoder_hidden_states
+
+                self.assertIsInstance(hidden_states, (list, tuple))
+                self.assertEqual(len(hidden_states), expected_num_layers)
+                seq_len = getattr(self.model_tester, "seq_length", None)
+                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
+
+                self.assertListEqual(
+                    list(hidden_states[0].shape[-2:]),
+                    [decoder_seq_length, self.model_tester.hidden_size],
+                )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        seq_length = getattr(self.model_tester, "seq_length", None)
+        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_length)
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_length)
+        decoder_key_length = getattr(self.model_tester, "decoder_key_length", decoder_seq_length)
+        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+            )
+            out_len = len(outputs)
+
+            if self.is_encoder_decoder:
+                correct_outlen = 5
+
+                # Question Answering model returns start_logits and end_logits
+                if model_class in get_values(FLAX_MODEL_FOR_QUESTION_ANSWERING_MAPPING):
+                    correct_outlen += 1  # start_logits and end_logits instead of only 1 output
+
+                self.assertEqual(out_len, correct_outlen)
+
+                # decoder attentions
+                decoder_attentions = outputs.decoder_attentions
+                self.assertIsInstance(decoder_attentions, (list, tuple))
+                self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+                self.assertListEqual(
+                    list(decoder_attentions[0].shape[-3:]),
+                    [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
+                )
+
+                # cross attentions
+                cross_attentions = outputs.cross_attentions
+                self.assertIsInstance(cross_attentions, (list, tuple))
+                self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
+                self.assertListEqual(
+                    list(cross_attentions[0].shape[-3:]),
+                    [
+                        self.model_tester.num_attention_heads,
+                        decoder_seq_length,
+                        encoder_key_length,
+                    ],
+                )
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            if hasattr(self.model_tester, "num_hidden_states_types"):
+                added_hidden_states = self.model_tester.num_hidden_states_types
+            elif self.is_encoder_decoder:
+                added_hidden_states = 2
+            else:
+                added_hidden_states = 1
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
+            )
+
+
+@require_flax
+@is_staging_test
+class FlaxModelPushToHubTester(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls._api = HfApi(endpoint=ENDPOINT_STAGING)
+        cls._token = cls._api.login(username=USER, password=PASS)
+
+    @classmethod
+    def tearDownClass(cls):
+        try:
+            cls._api.delete_repo(token=cls._token, name="test-model-flax")
+        except HTTPError:
+            pass
+
+        try:
+            cls._api.delete_repo(token=cls._token, name="test-model-flax-org", organization="valid_org")
+        except HTTPError:
+            pass
+
+    def test_push_to_hub(self):
+        config = BertConfig(
+            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
+        )
+        model = FlaxBertModel(config)
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(
+                os.path.join(tmp_dir, "test-model-flax"), push_to_hub=True, use_auth_token=self._token
+            )
+
+            new_model = FlaxBertModel.from_pretrained(f"{USER}/test-model-flax")
+
+            base_params = flatten_dict(unfreeze(model.params))
+            new_params = flatten_dict(unfreeze(new_model.params))
+
+            for key in base_params.keys():
+                max_diff = (base_params[key] - new_params[key]).sum().item()
+                self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
+
+    def test_push_to_hub_in_organization(self):
+        config = BertConfig(
+            vocab_size=99, hidden_size=32, num_hidden_layers=5, num_attention_heads=4, intermediate_size=37
+        )
+        model = FlaxBertModel(config)
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(
+                os.path.join(tmp_dir, "test-model-flax-org"),
+                push_to_hub=True,
+                use_auth_token=self._token,
+                organization="valid_org",
+            )
+
+            new_model = FlaxBertModel.from_pretrained("valid_org/test-model-flax-org")
+
+            base_params = flatten_dict(unfreeze(model.params))
+            new_params = flatten_dict(unfreeze(new_model.params))
+
+            for key in base_params.keys():
+                max_diff = (base_params[key] - new_params[key]).sum().item()
+                self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")

tests/test_t5_vae.py

@@ -0,0 +1,491 @@
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers import is_flax_available
+from transformers.models.t5.modeling_flax_t5 import shift_tokens_right
+from transformers.testing_utils import require_flax
+
+from tests.test_configuration_common import ConfigTester
+from tests.test_generation_flax_utils import FlaxGenerationTesterMixin
+from tests.test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor
+
+
+if is_flax_available():
+    import os
+
+    # The slow tests are often failing with OOM error on GPU
+    # This makes JAX allocate exactly what is needed on demand, and deallocate memory that is no longer needed
+    # but will be slower as stated here https://jax.readthedocs.io/en/latest/gpu_memory_allocation.html
+    os.environ["XLA_PYTHON_CLIENT_ALLOCATOR"] = "platform"
+
+    import jax
+    import jax.numpy as jnp
+    from flax.core.frozen_dict import unfreeze
+    from flax.traverse_util import flatten_dict
+    from transformers import FLAX_MODEL_MAPPING
+    from model.t5_vae import FlaxT5VaeForAutoencoding, T5VaeConfig
+
+
+class FlaxVaeModelTester:
+    def __init__(
+        self,
+        parent,
+        vocab_size=99,
+        batch_size=13,
+        seq_length=7,
+        latent_token_size=10,
+        n_latent_tokens=3,
+        # For common tests
+        is_training=True,
+        use_attention_mask=True,
+        use_labels=True,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        d_ff=37,
+        relative_attention_num_buckets=8,
+        dropout_rate=0.1,
+        initializer_factor=0.002,
+        eos_token_id=1,
+        pad_token_id=0,
+        decoder_start_token_id=0,
+        scope=None,
+        decoder_layers=None,
+    ):
+
+        self.parent = parent
+        self.batch_size = batch_size
+        self.latent_token_size = latent_token_size
+        self.n_latent_tokens = n_latent_tokens
+        # For common tests
+        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.d_ff = d_ff
+        self.relative_attention_num_buckets = relative_attention_num_buckets
+        self.dropout_rate = dropout_rate
+        self.initializer_factor = initializer_factor
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.scope = None
+        self.decoder_layers = decoder_layers
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+        decoder_input_ids = shift_tokens_right(input_ids, self.pad_token_id, self.pad_token_id)
+
+        attention_mask = None
+        decoder_attention_mask = None
+        if self.use_attention_mask:
+            attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+            decoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        config = T5VaeConfig(
+            latent_token_size=self.latent_token_size,
+            n_latent_tokens=self.n_latent_tokens,
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            block_size=self.seq_length,
+            d_ff=self.d_ff,
+            d_kv=self.hidden_size // self.num_attention_heads,
+            num_layers=self.num_hidden_layers,
+            num_decoder_layers=self.decoder_layers,
+            num_heads=self.num_attention_heads,
+            relative_attention_num_buckets=self.relative_attention_num_buckets,
+            dropout_rate=self.dropout_rate,
+            initializer_factor=self.initializer_factor,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.pad_token_id,
+            pad_token_id=self.pad_token_id,
+            decoder_start_token_id=self.decoder_start_token_id,
+        )
+
+        return (
+            config,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+    ):
+        model = FlaxT5VaeForAutoencoding(config=config)
+        result = model(
+            input_ids=input_ids,
+            decoder_input_ids=decoder_input_ids,
+            attention_mask=attention_mask,
+            decoder_attention_mask=decoder_attention_mask,
+        )
+        result = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+        decoder_output = result.last_hidden_state
+        encoder_output = result.encoder_last_hidden_state
+
+        self.parent.assertEqual(encoder_output.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(decoder_output.shape, (self.batch_size, self.decoder_seq_length+1, self.hidden_size))
+
+    def check_use_cache_forward_with_attn_mask(
+        self,
+        model_class_name,
+        config,
+        input_ids,
+        decoder_input_ids,
+        attention_mask,
+        decoder_attention_mask,
+    ):
+        max_decoder_length = 20
+        model = model_class_name(config)
+
+        latent_codes = model.encode(input_ids)
+
+        # prevent fully zero'd out attention mask
+        decoder_attention_mask = jnp.ones_like(decoder_attention_mask)
+
+        decoder_attention_mask_cache = jnp.concatenate(
+            [
+                decoder_attention_mask,
+                jnp.zeros((decoder_attention_mask.shape[0], max_decoder_length - decoder_attention_mask.shape[1])),
+            ],
+            axis=-1,
+        )
+
+        past_key_values = model.init_cache(decoder_input_ids.shape[0], max_decoder_length, encoder_outputs)
+
+        outputs_cache = model.decode(
+            decoder_input_ids[:, :-1],
+            latent_codes,
+            decoder_attention_mask=decoder_attention_mask_cache,
+            past_key_values=past_key_values,
+        )
+        outputs_cache_next = model.decode(
+            decoder_input_ids[:, -1:],
+            latent_codes,
+            past_key_values=outputs_cache.past_key_values,
+            decoder_attention_mask=decoder_attention_mask_cache,
+        )
+
+        outputs = model.decode(decoder_input_ids, latent_codes, decoder_attention_mask=decoder_attention_mask)
+
+        diff = np.max(np.abs((outputs_cache_next[0][:, -1, :5] - outputs[0][:, -1, :5])))
+        self.parent.assertTrue(diff < 1e-3, msg=f"Max diff is {diff}")
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            decoder_input_ids,
+            attention_mask,
+            decoder_attention_mask,
+        ) = config_and_inputs
+
+        inputs_dict = {
+            "input_ids": input_ids,
+            "attention_mask": attention_mask,
+            "decoder_input_ids": decoder_input_ids,
+            "decoder_attention_mask": decoder_attention_mask,
+        }
+        return config, inputs_dict
+
+
+@require_flax
+class FlaxT5VaeModelTest(FlaxModelTesterMixin, FlaxGenerationTesterMixin, unittest.TestCase):
+
+    all_model_classes = (FlaxT5VaeForAutoencoding,) if is_flax_available() else ()
+    is_encoder_decoder = True
+
+    def setUp(self):
+        self.model_tester = FlaxVaeModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=T5VaeConfig, d_model=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_v1_1(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        # check that gated gelu feed forward and different word embeddings work
+        config = config_and_inputs[0]
+        config.tie_word_embeddings = False
+        config.feed_forward_proj = "gated-gelu"
+        self.model_tester.create_and_check_model(config, *config_and_inputs[1:])
+
+    def test_use_cache_forward_with_attn_mask(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for model_class in self.all_model_classes:
+            self.model_tester.check_use_cache_forward_with_attn_mask(model_class, *config_and_inputs)
+
+    def test_encode(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                prepared_inputs_dict = self._prepare_for_class(inputs_dict, model_class)
+                model = model_class(config)
+
+                @jax.jit
+                def encode_jitted(input_ids, attention_mask=None, **kwargs):
+                    return model.encode(input_ids=input_ids, attention_mask=attention_mask)
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = encode_jitted(**prepared_inputs_dict)
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = encode_jitted(**prepared_inputs_dict)
+
+                self.assertEqual(outputs.shape, (inputs_dict['input_ids'].shape[0], config.n_latent_tokens, config.latent_token_size))
+                self.assertEqual(jitted_outputs.shape, (inputs_dict['input_ids'].shape[0], config.n_latent_tokens, config.latent_token_size))
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs, outputs):
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+    def test_decode(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            with self.subTest(model_class.__name__):
+                model = model_class(config)
+                latent_codes = model.encode(inputs_dict["input_ids"], inputs_dict["attention_mask"])
+
+                prepared_inputs_dict = {
+                    "decoder_input_ids": inputs_dict["decoder_input_ids"],
+                    "decoder_attention_mask": inputs_dict["decoder_attention_mask"],
+                    "latent_codes": latent_codes,
+                }
+
+                @jax.jit
+                def decode_jitted(decoder_input_ids, decoder_attention_mask, latent_codes):
+                    return model.decode(
+                        decoder_input_ids=decoder_input_ids,
+                        latent_codes=latent_codes,
+                        decoder_attention_mask=decoder_attention_mask,
+                    )
+
+                with self.subTest("JIT Enabled"):
+                    jitted_outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
+
+                with self.subTest("JIT Disabled"):
+                    with jax.disable_jit():
+                        outputs = decode_jitted(**prepared_inputs_dict).to_tuple()
+
+                self.assertEqual(len(outputs), len(jitted_outputs))
+                for jitted_output, output in zip(jitted_outputs, outputs):
+                    self.assertEqual(jitted_output.shape, output.shape)
+
+    # overwrite since special base model prefix is used
+    def test_save_load_from_base(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        base_class = FLAX_MODEL_MAPPING[config.__class__]
+
+        for model_class in self.all_model_classes:
+            if model_class == base_class:
+                continue
+
+            model = base_class(config)
+            base_params = flatten_dict(unfreeze(model.params))
+
+            # check that all base model weights are loaded correctly
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                head_model = model_class.from_pretrained(tmpdirname)
+
+                base_param_from_head = flatten_dict(unfreeze(head_model.params))
+
+                for key in base_param_from_head.keys():
+                    max_diff = (base_params[key] - base_param_from_head[key]).sum().item()
+                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
+
+    # overwrite since special base model prefix is used
+    def test_save_load_to_base(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+        base_class = FLAX_MODEL_MAPPING[config.__class__]
+
+        for model_class in self.all_model_classes:
+            if model_class == base_class:
+                continue
+
+            model = model_class(config)
+            base_params_from_head = flatten_dict(unfreeze(model.params))
+
+            # check that all base model weights are loaded correctly
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname)
+                base_model = base_class.from_pretrained(tmpdirname)
+
+                base_params = flatten_dict(unfreeze(base_model.params))
+
+                for key in base_params_from_head.keys():
+                    max_diff = (base_params[key] - base_params_from_head[key]).sum().item()
+                    self.assertLessEqual(max_diff, 1e-3, msg=f"{key} not identical")
+
+
+'''
+# Not using for now.
+
+@require_sentencepiece
+@require_tokenizers
+@require_flax
+class FlaxT5ModelIntegrationTests(unittest.TestCase):
+    @slow
+    def test_small_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.7.1
+        >>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
+
+        >>> path_to_mtf_small_t5_checkpoint = '<fill_in>'
+        >>> path_to_mtf_small_spm_model_path = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_t5_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = SentencePieceVocabulary(path_to_mtf_small_spm_model_path, extra_ids=100)
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = FlaxT5ForConditionalGeneration.from_pretrained("t5-small")
+        tokenizer = T5Tokenizer.from_pretrained("t5-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="np").input_ids
+        labels = tokenizer("Hi I am", return_tensors="np").input_ids
+
+        decoder_input_ids = shift_tokens_right(labels, model.config.pad_token_id, model.config.decoder_start_token_id)
+
+        logits = model(input_ids, decoder_input_ids=decoder_input_ids).logits
+
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])).mean()
+        mtf_score = -(labels.shape[-1] * loss.item())
+
+        EXPECTED_SCORE = -19.0845
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
+
+    @slow
+    def test_small_v1_1_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.7.1
+        >>> from t5.data.sentencepiece_vocabulary import SentencePieceVocabulary
+
+        >>> path_to_mtf_small_t5_v1_1_checkpoint = '<fill_in>'
+        >>> path_to_mtf_small_spm_model_path = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_mtf_small_t5_v1_1_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = SentencePieceVocabulary(path_to_mtf_small_spm_model_path, extra_ids=100)
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = FlaxT5ForConditionalGeneration.from_pretrained("google/t5-v1_1-small")
+        tokenizer = T5Tokenizer.from_pretrained("google/t5-v1_1-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="np").input_ids
+        labels = tokenizer("Hi I am", return_tensors="np").input_ids
+
+        decoder_input_ids = shift_tokens_right(labels, model.config.pad_token_id, model.config.decoder_start_token_id)
+
+        logits = model(input_ids, decoder_input_ids=decoder_input_ids).logits
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])).mean()
+
+        mtf_score = -(labels.shape[-1] * loss.item())
+
+        EXPECTED_SCORE = -59.0293
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
+
+    @slow
+    def test_small_byt5_integration_test(self):
+        """
+        For comparision run:
+        >>> import t5  # pip install t5==0.9.1
+
+        >>> path_to_byt5_small_checkpoint = '<fill_in>'
+        >>> t5_model = t5.models.MtfModel(model_dir=path_to_tf_checkpoint, batch_size=1, tpu=None)
+        >>> vocab = t5.data.ByteVocabulary()
+        >>> score = t5_model.score(inputs=["Hello there"], targets=["Hi I am"], vocabulary=vocab)
+        """
+
+        model = FlaxT5ForConditionalGeneration.from_pretrained("google/byt5-small")
+        tokenizer = ByT5Tokenizer.from_pretrained("google/byt5-small")
+
+        input_ids = tokenizer("Hello there", return_tensors="np").input_ids
+        labels = tokenizer("Hi I am", return_tensors="np").input_ids
+
+        decoder_input_ids = shift_tokens_right(labels, model.config.pad_token_id, model.config.decoder_start_token_id)
+
+        logits = model(input_ids, decoder_input_ids=decoder_input_ids).logits
+        loss = optax.softmax_cross_entropy(logits, onehot(labels, logits.shape[-1])).mean()
+
+        mtf_score = -(labels.shape[-1] * loss.item())
+
+        EXPECTED_SCORE = -60.7397
+        self.assertTrue(abs(mtf_score - EXPECTED_SCORE) < 1e-4)
+
+    @slow
+    def test_small_generation(self):
+        model = FlaxT5ForConditionalGeneration.from_pretrained("t5-small")
+        model.config.max_length = 8
+        model.config.num_beams = 1
+        model.config.do_sample = False
+        tokenizer = T5Tokenizer.from_pretrained("t5-small")
+
+        input_ids = tokenizer("summarize: Hello there", return_tensors="np").input_ids
+
+        sequences = model.generate(input_ids).sequences
+
+        output_str = tokenizer.batch_decode(sequences, skip_special_tokens=True)[0]
+        self.assertTrue(output_str == "Hello there!")
+
+    @slow
+    def test_summarization(self):
+        model = FlaxT5ForConditionalGeneration.from_pretrained("t5-base")
+        tok = T5Tokenizer.from_pretrained("t5-base")
+
+        FRANCE_ARTICLE = 'Marseille, France (CNN)The French prosecutor leading an investigation into the crash of Germanwings Flight 9525 insisted Wednesday that he was not aware of any video footage from on board the plane. Marseille prosecutor Brice Robin told CNN that "so far no videos were used in the crash investigation." He added, "A person who has such a video needs to immediately give it to the investigators." Robin\'s comments follow claims by two magazines, German daily Bild and French Paris Match, of a cell phone video showing the harrowing final seconds from on board Germanwings Flight 9525 as it crashed into the French Alps. All 150 on board were killed. Paris Match and Bild reported that the video was recovered from a phone at the wreckage site. The two publications described the supposed video, but did not post it on their websites. The publications said that they watched the video, which was found by a source close to the investigation. "One can hear cries of \'My God\' in several languages," Paris Match reported. "Metallic banging can also be heard more than three times, perhaps of the pilot trying to open the cockpit door with a heavy object.  Towards the end, after a heavy shake, stronger than the others, the screaming intensifies. Then nothing." "It is a very disturbing scene," said Julian Reichelt, editor-in-chief of Bild online. An official with France\'s accident investigation agency, the BEA, said the agency is not aware of any such video. Lt. Col. Jean-Marc Menichini, a French Gendarmerie spokesman in charge of communications on rescue efforts around the Germanwings crash site, told CNN that the reports were "completely wrong" and "unwarranted." Cell phones have been collected at the site, he said, but that they "hadn\'t been exploited yet." Menichini said he believed the cell phones would need to be sent to the Criminal Research Institute in Rosny sous-Bois, near Paris, in order to be analyzed by specialized technicians working hand-in-hand with investigators. But none of the cell phones found so far have been sent to the institute, Menichini said. Asked whether staff involved in the search could have leaked a memory card to the media, Menichini answered with a categorical "no." Reichelt told "Erin Burnett: Outfront" that he had watched the video and stood by the report, saying Bild and Paris Match are "very confident" that the clip is real. He noted that investigators only revealed they\'d recovered cell phones from the crash site after Bild and Paris Match published their reports. "That is something we did not know before. ... Overall we can say many things of the investigation weren\'t revealed by the investigation at the beginning," he said. What was mental state of Germanwings co-pilot? German airline Lufthansa confirmed Tuesday that co-pilot Andreas Lubitz had battled depression years before he took the controls of Germanwings Flight 9525, which he\'s accused of deliberately crashing last week in the French Alps. Lubitz told his Lufthansa flight training school in 2009 that he had a "previous episode of severe depression," the airline said Tuesday. Email correspondence between Lubitz and the school discovered in an internal investigation, Lufthansa said, included medical documents he submitted in connection with resuming his flight training. The announcement indicates that Lufthansa, the parent company of Germanwings, knew of Lubitz\'s battle with depression, allowed him to continue training and ultimately put him in the cockpit. Lufthansa, whose CEO Carsten Spohr previously said Lubitz was 100% fit to fly, described its statement Tuesday as a "swift and seamless clarification" and said it was sharing the information and documents -- including training and medical records -- with public prosecutors. Spohr traveled to the crash site Wednesday, where recovery teams have been working for the past week to recover human remains and plane debris scattered across a steep mountainside. He saw the crisis center set up in Seyne-les-Alpes, laid a wreath in the village of Le Vernet, closer to the crash site, where grieving families have left flowers at a simple stone memorial. Menichini told CNN late Tuesday that no visible human remains were left at the site but recovery teams would keep searching. French President Francois Hollande, speaking Tuesday, said that it should be possible to identify all the victims using DNA analysis by the end of the week, sooner than authorities had previously suggested. In the meantime, the recovery of the victims\' personal belongings will start Wednesday, Menichini said. Among those personal belongings could be more cell phones belonging to the 144 passengers and six crew on board. Check out the latest from our correspondents . The details about Lubitz\'s correspondence with the flight school during his training were among several developments as investigators continued to delve into what caused the crash and Lubitz\'s possible motive for downing the jet. A Lufthansa spokesperson told CNN on Tuesday that Lubitz had a valid medical certificate, had passed all his examinations and "held all the licenses required." Earlier, a spokesman for the prosecutor\'s office in Dusseldorf, Christoph Kumpa, said medical records reveal Lubitz suffered from suicidal tendencies at some point before his aviation career and underwent psychotherapy before he got his pilot\'s license. Kumpa emphasized there\'s no evidence suggesting Lubitz was suicidal or acting aggressively before the crash. Investigators are looking into whether Lubitz feared his medical condition would cause him to lose his pilot\'s license, a European government official briefed on the investigation told CNN on Tuesday. While flying was "a big part of his life," the source said, it\'s only one theory being considered. Another source, a law enforcement official briefed on the investigation, also told CNN that authorities believe the primary motive for Lubitz to bring down the plane was that he feared he would not be allowed to fly because of his medical problems. Lubitz\'s girlfriend told investigators he had seen an eye doctor and a neuropsychologist, both of whom deemed him unfit to work recently and concluded he had psychological issues, the European government official said. But no matter what details emerge about his previous mental health struggles, there\'s more to the story, said Brian Russell, a forensic psychologist. "Psychology can explain why somebody would turn rage inward on themselves about the fact that maybe they weren\'t going to keep doing their job and they\'re upset about that and so they\'re suicidal," he said. "But there is no mental illness that explains why somebody then feels entitled to also take that rage and turn it outward on 149 other people who had nothing to do with the person\'s problems." Germanwings crash compensation: What we know . Who was the captain of Germanwings Flight 9525? CNN\'s Margot Haddad reported from Marseille and Pamela Brown from Dusseldorf, while Laura Smith-Spark wrote from London. CNN\'s Frederik Pleitgen, Pamela Boykoff, Antonia Mortensen, Sandrine Amiel and Anna-Maja Rappard contributed to this report.'  # @noqa
+        SHORTER_ARTICLE = '(CNN)The Palestinian Authority officially became the 123rd member of the International Criminal Court on Wednesday, a step that gives the court jurisdiction over alleged crimes in Palestinian territories. The formal accession was marked with a ceremony at The Hague, in the Netherlands, where the court is based. The Palestinians signed the ICC\'s founding Rome Statute in January, when they also accepted its jurisdiction over alleged crimes committed "in the occupied Palestinian territory, including East Jerusalem, since June 13, 2014." Later that month, the ICC opened a preliminary examination into the situation in Palestinian territories, paving the way for possible war crimes investigations against Israelis. As members of the court, Palestinians may be subject to counter-charges as well. Israel and the United States, neither of which is an ICC member, opposed the Palestinians\' efforts to join the body. But Palestinian Foreign Minister Riad al-Malki, speaking at Wednesday\'s ceremony, said it was a move toward greater justice. "As Palestine formally becomes a State Party to the Rome Statute today, the world is also a step closer to ending a long era of impunity and injustice," he said, according to an ICC news release. "Indeed, today brings us closer to our shared goals of justice and peace." Judge Kuniko Ozaki, a vice president of the ICC, said acceding to the treaty was just the first step for the Palestinians. "As the Rome Statute today enters into force for the State of Palestine, Palestine acquires all the rights as well as responsibilities that come with being a State Party to the Statute. These are substantive commitments, which cannot be taken lightly," she said. Rights group Human Rights Watch welcomed the development. "Governments seeking to penalize Palestine for joining the ICC should immediately end their pressure, and countries that support universal acceptance of the court\'s treaty should speak out to welcome its membership," said Balkees Jarrah, international justice counsel for the group. "What\'s objectionable is the attempts to undermine international justice, not Palestine\'s decision to join a treaty to which over 100 countries around the world are members." In January, when the preliminary ICC examination was opened, Israeli Prime Minister Benjamin Netanyahu described it as an outrage, saying the court was overstepping its boundaries. The United States also said it "strongly" disagreed with the court\'s decision. "As we have said repeatedly, we do not believe that Palestine is a state and therefore we do not believe that it is eligible to join the ICC," the State Department said in a statement. It urged the warring sides to resolve their differences through direct negotiations. "We will continue to oppose actions against Israel at the ICC as counterproductive to the cause of peace," it said. But the ICC begs to differ with the definition of a state for its purposes and refers to the territories as "Palestine." While a preliminary examination is not a formal investigation, it allows the court to review evidence and determine whether to investigate suspects on both sides. Prosecutor Fatou Bensouda said her office would "conduct its analysis in full independence and impartiality." The war between Israel and Hamas militants in Gaza last summer left more than 2,000 people dead. The inquiry will include alleged war crimes committed since June. The International Criminal Court was set up in 2002 to prosecute genocide, crimes against humanity and war crimes. CNN\'s Vasco Cotovio, Kareem Khadder and Faith Karimi contributed to this report.'
+        IRAN_ARTICLE = "(CNN)The United States and its negotiating partners reached a very strong framework agreement with Iran in Lausanne, Switzerland, on Thursday that limits Iran's nuclear program in such a way as to effectively block it from building a nuclear weapon. Expect pushback anyway, if the recent past is any harbinger. Just last month, in an attempt to head off such an agreement, House Speaker John Boehner invited Israeli Prime Minister Benjamin Netanyahu to preemptively blast it before Congress, and 47 senators sent a letter to the Iranian leadership warning them away from a deal. The debate that has already begun since the announcement of the new framework will likely result in more heat than light. It will not be helped by the gathering swirl of dubious assumptions and doubtful assertions. Let us address some of these: . The most misleading assertion, despite universal rejection by experts, is that the negotiations' objective at the outset was the total elimination of any nuclear program in Iran. That is the position of Netanyahu and his acolytes in the U.S. Congress. But that is not and never was the objective. If it had been, there would have been no Iranian team at the negotiating table. Rather, the objective has always been to structure an agreement or series of agreements so that Iran could not covertly develop a nuclear arsenal before the United States and its allies could respond. The new framework has exceeded expectations in achieving that goal. It would reduce Iran's low-enriched uranium stockpile, cut by two-thirds its number of installed centrifuges and implement a rigorous inspection regime. Another dubious assumption of opponents is that the Iranian nuclear program is a covert weapons program. Despite sharp accusations by some in the United States and its allies, Iran denies having such a program, and U.S. intelligence contends that Iran has not yet made the decision to build a nuclear weapon. Iran's continued cooperation with International Atomic Energy Agency inspections is further evidence on this point, and we'll know even more about Iran's program in the coming months and years because of the deal. In fact, the inspections provisions that are part of this agreement are designed to protect against any covert action by the Iranians. What's more, the rhetoric of some members of Congress has implied that the negotiations have been between only the United States and Iran (i.e., the 47 senators' letter warning that a deal might be killed by Congress or a future president). This of course is not the case. The talks were between Iran and the five permanent members of the U.N. Security Council (United States, United Kingdom, France, China and Russia) plus Germany, dubbed the P5+1. While the United States has played a leading role in the effort, it negotiated the terms alongside its partners. If the agreement reached by the P5+1 is rejected by Congress, it could result in an unraveling of the sanctions on Iran and threaten NATO cohesion in other areas. Another questionable assertion is that this agreement contains a sunset clause, after which Iran will be free to do as it pleases. Again, this is not the case. Some of the restrictions on Iran's nuclear activities, such as uranium enrichment, will be eased or eliminated over time, as long as 15 years. But most importantly, the framework agreement includes Iran's ratification of the Additional Protocol, which allows IAEA inspectors expanded access to nuclear sites both declared and nondeclared. This provision will be permanent. It does not sunset. Thus, going forward, if Iran decides to enrich uranium to weapons-grade levels, monitors will be able to detect such a move in a matter of days and alert the U.N. Security Council. Many in Congress have said that the agreement should be a formal treaty requiring the Senate to \"advise and consent.\" But the issue is not suited for a treaty. Treaties impose equivalent obligations on all signatories. For example, the New START treaty limits Russia and the United States to 1,550 deployed strategic warheads. But any agreement with Iran will not be so balanced.  The restrictions and obligations in the final framework agreement will be imposed almost exclusively on Iran. The P5+1 are obligated only to ease and eventually remove most but not all economic sanctions, which were imposed as leverage to gain this final deal. Finally some insist that any agreement must address Iranian missile programs, human rights violations or support for Hamas or Hezbollah.  As important as these issues are, and they must indeed be addressed, they are unrelated to the most important aim of a nuclear deal: preventing a nuclear Iran.  To include them in the negotiations would be a poison pill. This agreement should be judged on its merits and on how it affects the security of our negotiating partners and allies, including Israel. Those judgments should be fact-based, not based on questionable assertions or dubious assumptions."
+        ARTICLE_SUBWAY = 'New York (CNN)When Liana Barrientos was 23 years old, she got married in Westchester County, New York. A year later, she got married again in Westchester County, but to a different man and without divorcing her first husband.  Only 18 days after that marriage, she got hitched yet again. Then, Barrientos declared "I do" five more times, sometimes only within two weeks of each other. In 2010, she married once more, this time in the Bronx. In an application for a marriage license, she stated it was her "first and only" marriage. Barrientos, now 39, is facing two criminal counts of "offering a false instrument for filing in the first degree," referring to her false statements on the 2010 marriage license application, according to court documents. Prosecutors said the marriages were part of an immigration scam. On Friday, she pleaded not guilty at State Supreme Court in the Bronx, according to her attorney, Christopher Wright, who declined to comment further. After leaving court, Barrientos was arrested and charged with theft of service and criminal trespass for allegedly sneaking into the New York subway through an emergency exit, said Detective Annette Markowski, a police spokeswoman. In total, Barrientos has been married 10 times, with nine of her marriages occurring between 1999 and 2002.  All occurred either in Westchester County, Long Island, New Jersey or the Bronx. She is believed to still be married to four men, and at one time, she was married to eight men at once, prosecutors say. Prosecutors said the immigration scam involved some of her husbands, who filed for permanent residence status shortly after the marriages.  Any divorces happened only after such filings were approved. It was unclear whether any of the men will be prosecuted. The case was referred to the Bronx District Attorney\'s Office by Immigration and Customs Enforcement and the Department of Homeland Security\'s Investigation Division. Seven of the men are from so-called "red-flagged" countries, including Egypt, Turkey, Georgia, Pakistan and Mali. Her eighth husband, Rashid Rajput, was deported in 2006 to his native Pakistan after an investigation by the Joint Terrorism Task Force. If convicted, Barrientos faces up to four years in prison.  Her next court appearance is scheduled for May 18.'
+
+        expected_summaries = [
+            'prosecutor: "so far no videos were used in the crash investigation" two magazines claim to have found a cell phone video of the final seconds . "one can hear cries of \'My God\' in several languages," the magazine says . all 150 on board the germanwings flight were killed .',
+            "the Palestinians become the 123rd member of the international criminal court . the accession was marked by a ceremony at the Hague, where the court is based . as members of the court, Palestinians may be subject to counter-charges as well .",
+            "the u.s. and its negotiating partners reached a very strong framework agreement with Iran . aaron miller: the debate that has already begun since the announcement of the new framework will likely result in more heat than light . he says the new framework would reduce Iran's low-enriched uranium stockpile and cut centrifuges . miller: if it had been, there would have been no Iranian team at the table .",
+            'prosecutors say the marriages were part of an immigration scam . if convicted, barrientos faces two criminal counts of "offering a false instrument for filing in the first degree" she has been married 10 times, with nine of her marriages occurring between 1999 and 2002 .',
+        ]
+
+        dct = tok(
+            ["summarize: " + x for x in [FRANCE_ARTICLE, SHORTER_ARTICLE, IRAN_ARTICLE, ARTICLE_SUBWAY]],
+            padding="max_length",
+            truncation=True,
+            return_tensors="np",
+        )
+        self.assertEqual(512, dct["input_ids"].shape[1])
+
+        hypotheses_batch = model.generate(
+            **dct,
+            num_beams=4,
+            length_penalty=2.0,
+            max_length=142,
+            min_length=56,
+            do_sample=False,
+            early_stopping=True,
+        ).sequences
+
+        decoded = tok.batch_decode(hypotheses_batch, skip_special_tokens=True, clean_up_tokenization_spaces=False)
+        self.assertListEqual(
+            expected_summaries,
+            decoded,
+        )
+'''

train.py

@@ -47,8 +47,8 @@ from transformers import (
 )
 from transformers.testing_utils import CaptureLogger
 
-from model.t5_vae import FlaxT5_VAE_ForAutoencoding
-from model.config import T5_VAE_Config
+from model.t5_vae import FlaxT5VaeForAutoencoding
+from model.config import T5VaeConfig
 
 
 logger = logging.getLogger(__name__)
@@ -316,15 +316,15 @@ def main():
     # download model & vocab.
 
     if model_args.config_path:
-        config = T5_VAE_Config.from_pretrained(
+        config = T5VaeConfig.from_pretrained(
             model_args.config_path, cache_dir=model_args.cache_dir
         )
     elif model_args.model_name_or_path:
-        config = T5_VAE_Config.from_pretrained(
+        config = T5VaeConfig.from_pretrained(
             model_args.model_name_or_path, cache_dir=model_args.cache_dir
         )
     else:
-        config = T5_VAE_Config(**model_args.__dict__)
+        config = T5VaeConfig(**model_args.__dict__)
         logger.warning("You are instantiating a new config instance from scratch.")
 
     if model_args.tokenizer_name:
@@ -346,7 +346,7 @@ def main():
         )
 
     if model_args.model_name_or_path:
-        model = FlaxT5_VAE_ForAutoencoding.from_pretrained(
+        model = FlaxT5VaeForAutoencoding.from_pretrained(
             model_args.model_name_or_path, config=config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
         )
         # TODO assert token embedding size == len(tokenizer)
@@ -355,7 +355,7 @@ def main():
         config.t5.vocab_size = vocab_size
         config.vocab_size = vocab_size
         logger.info("Training new model from scratch.")
-        model = FlaxT5_VAE_ForAutoencoding(
+        model = FlaxT5VaeForAutoencoding(
             config, seed=training_args.seed, dtype=getattr(jnp, model_args.dtype)
         )
 
@@ -402,7 +402,7 @@ def main():
     # Limits each input sequence to size block_size.
     pad_token_id = tokenizer.pad_token_id
 
-    def limit_length(examples):
+    def clip_texts(examples):
         examples["labels"] = examples["input_ids"].copy()
 
         for i, input_ids in enumerate(examples["input_ids"]):
@@ -425,7 +425,7 @@ def main():
     # https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map
 
     lm_datasets = tokenized_datasets.map(
-        limit_length,
+        clip_texts,
         batched=True,
         num_proc=data_args.preprocessing_num_workers,
         load_from_cache_file=not data_args.overwrite_cache,
@@ -536,22 +536,23 @@ def main():
         true_samples = jax.random.normal(rng, latent_codes.shape())
         return compute_mmd(true_samples, latent_codes)
 
-    def loss_fn(logits, labels, latent_codes, rng: jax.random.PRNGKey):
+    def loss_fn(logits, labels, latent_codes, regulariser_rng: jax.random.PRNGKey):
         shift_logits = logits[..., :-1, :]
         shift_labels = labels[..., 1:]
         loss = optax.softmax_cross_entropy(shift_logits, onehot(shift_labels, shift_logits.shape[-1]))
 
-        reg_loss = regulariser_loss(latent_codes, rng)
+        reg_loss = regulariser_loss(latent_codes, regulariser_rng)
         return loss.mean() + reg_loss.mean()
 
     # Define gradient update step fn
     def train_step(state, batch):
         dropout_rng, new_dropout_rng = jax.random.split(state.dropout_rng)
+        new_dropout_rng, regulariser_rng = jax.random.split(new_dropout_rng)
 
         def compute_loss(params):
             labels = batch.pop("labels")
             outputs = state.apply_fn(**batch, params=params, dropout_rng=dropout_rng, train=True)
-            loss = loss_fn(outputs.logits, labels, outputs.latent_codes, state.dropout_rng)
+            loss = loss_fn(outputs.logits, labels, outputs.latent_codes, regulariser_rng)
             return loss
 
         grad_fn = jax.value_and_grad(compute_loss)