[FlaxBert] Add ForCausalLM (#16995)

* [FlaxBert] Add ForCausalLM

* make style

* fix output attentions

* Add RobertaForCausalLM

* remove comment

* fix fx-to-pt model loading

* remove comment

* add modeling tests

* add enc-dec model tests

* add big_bird

* add electra

* make style

* make repo-consitency

* add to docs

* remove roberta test

* quality

* amend cookiecutter

* fix attention_mask bug in flax bert model tester

* tighten pt-fx thresholds to 1e-5

* add 'copied from' statements

* amend 'copied from' statements

* amend 'copied from' statements

* quality

docs/source/en/model_doc/bert.mdx

@@ -166,6 +166,11 @@ This model was contributed by [thomwolf](https://huggingface.co/thomwolf). The o
 [[autodoc]] FlaxBertForPreTraining
     - __call__
 
+## FlaxBertForCausalLM
+
+[[autodoc]] FlaxBertForCausalLM
+    - __call__
+
 ## FlaxBertForMaskedLM
 
 [[autodoc]] FlaxBertForMaskedLM

docs/source/en/model_doc/big_bird.mdx

@@ -120,6 +120,11 @@ This model was contributed by [vasudevgupta](https://huggingface.co/vasudevgupta
 [[autodoc]] FlaxBigBirdForPreTraining
     - __call__
 
+## FlaxBigBirdForCausalLM
+
+[[autodoc]] FlaxBigBirdForCausalLM
+    - __call__
+
 ## FlaxBigBirdForMaskedLM
 
 [[autodoc]] FlaxBigBirdForMaskedLM

docs/source/en/model_doc/electra.mdx

@@ -158,6 +158,11 @@ This model was contributed by [lysandre](https://huggingface.co/lysandre). The o
 [[autodoc]] FlaxElectraForPreTraining
     - __call__
 
+## FlaxElectraForCausalLM
+
+[[autodoc]] FlaxElectraForCausalLM
+    - __call__
+
 ## FlaxElectraForMaskedLM
 
 [[autodoc]] FlaxElectraForMaskedLM

docs/source/en/model_doc/roberta.mdx

@@ -136,6 +136,11 @@ This model was contributed by [julien-c](https://huggingface.co/julien-c). The o
 [[autodoc]] FlaxRobertaModel
     - __call__
 
+## FlaxRobertaForCausalLM
+
+[[autodoc]] FlaxRobertaForCausalLM
+    - __call__
+
 ## FlaxRobertaForMaskedLM
 
 [[autodoc]] FlaxRobertaForMaskedLM

src/transformers/__init__.py

@@ -2314,6 +2314,7 @@ if is_flax_available():
     )
     _import_structure["models.bert"].extend(
         [
+            "FlaxBertForCausalLM",
             "FlaxBertForMaskedLM",
             "FlaxBertForMultipleChoice",
             "FlaxBertForNextSentencePrediction",
@@ -2327,6 +2328,7 @@ if is_flax_available():
     )
     _import_structure["models.big_bird"].extend(
         [
+            "FlaxBigBirdForCausalLM",
             "FlaxBigBirdForMaskedLM",
             "FlaxBigBirdForMultipleChoice",
             "FlaxBigBirdForPreTraining",
@@ -2370,6 +2372,7 @@ if is_flax_available():
     )
     _import_structure["models.electra"].extend(
         [
+            "FlaxElectraForCausalLM",
             "FlaxElectraForMaskedLM",
             "FlaxElectraForMultipleChoice",
             "FlaxElectraForPreTraining",
@@ -2412,6 +2415,7 @@ if is_flax_available():
     )
     _import_structure["models.roberta"].extend(
         [
+            "FlaxRobertaForCausalLM",
             "FlaxRobertaForMaskedLM",
             "FlaxRobertaForMultipleChoice",
             "FlaxRobertaForQuestionAnswering",
@@ -4363,6 +4367,7 @@ if TYPE_CHECKING:
             FlaxBeitPreTrainedModel,
         )
         from .models.bert import (
+            FlaxBertForCausalLM,
             FlaxBertForMaskedLM,
             FlaxBertForMultipleChoice,
             FlaxBertForNextSentencePrediction,
@@ -4374,6 +4379,7 @@ if TYPE_CHECKING:
             FlaxBertPreTrainedModel,
         )
         from .models.big_bird import (
+            FlaxBigBirdForCausalLM,
             FlaxBigBirdForMaskedLM,
             FlaxBigBirdForMultipleChoice,
             FlaxBigBirdForPreTraining,
@@ -4411,6 +4417,7 @@ if TYPE_CHECKING:
             FlaxDistilBertPreTrainedModel,
         )
         from .models.electra import (
+            FlaxElectraForCausalLM,
             FlaxElectraForMaskedLM,
             FlaxElectraForMultipleChoice,
             FlaxElectraForPreTraining,
@@ -4435,6 +4442,7 @@ if TYPE_CHECKING:
         from .models.mt5 import FlaxMT5ForConditionalGeneration, FlaxMT5Model
         from .models.pegasus import FlaxPegasusForConditionalGeneration, FlaxPegasusModel, FlaxPegasusPreTrainedModel
         from .models.roberta import (
+            FlaxRobertaForCausalLM,
             FlaxRobertaForMaskedLM,
             FlaxRobertaForMultipleChoice,
             FlaxRobertaForQuestionAnswering,

src/transformers/modeling_flax_outputs.py

@@ -106,6 +106,55 @@ class FlaxBaseModelOutputWithPooling(ModelOutput):
     attentions: Optional[Tuple[jnp.ndarray]] = None
 
 
+@flax.struct.dataclass
+class FlaxBaseModelOutputWithPoolingAndCrossAttentions(ModelOutput):
+    """
+    Base class for model's outputs that also contains a pooling of the last hidden states.
+
+    Args:
+        last_hidden_state (`jnp.ndarray` of shape `(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (`jnp.ndarray` of shape `(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token) after further processing
+            through the layers used for the auxiliary pretraining task. E.g. for BERT-family of models, this returns
+            the classification token after processing through a linear layer and a tanh activation function. The linear
+            layer weights are trained from the next sentence prediction (classification) objective during pretraining.
+        hidden_states (`tuple(jnp.ndarray)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `jnp.ndarray` (one for the output of the embeddings, if the model has an embedding layer, + one
+            for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+        attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `jnp.ndarray` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        cross_attentions (`tuple(jnp.ndarray)`, *optional*, returned when `output_attentions=True` and `config.add_cross_attention=True` is passed or when `config.output_attentions=True`):
+            Tuple of `jnp.ndarray` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the
+            weighted average in the cross-attention heads.
+        past_key_values (`tuple(tuple(jnp.ndarray))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(jnp.ndarray)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and optionally if
+            `config.is_encoder_decoder=True` 2 additional tensors of shape `(batch_size, num_heads,
+            encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and optionally if
+            `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values`
+            input) to speed up sequential decoding.
+    """
+
+    last_hidden_state: jnp.ndarray = None
+    pooler_output: jnp.ndarray = None
+    hidden_states: Optional[Tuple[jnp.ndarray]] = None
+    past_key_values: Optional[Tuple[Tuple[jnp.ndarray]]] = None
+    attentions: Optional[Tuple[jnp.ndarray]] = None
+    cross_attentions: Optional[Tuple[jnp.ndarray]] = None
+
+
 @flax.struct.dataclass
 class FlaxBaseModelOutputWithPastAndCrossAttentions(ModelOutput):
     """

src/transformers/models/auto/modeling_flax_auto.py

@@ -127,6 +127,10 @@ FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gptj", "FlaxGPTJForCausalLM"),
         ("xglm", "FlaxXGLMForCausalLM"),
         ("bart", "FlaxBartForCausalLM"),
+        ("bert", "FlaxBertForCausalLM"),
+        ("roberta", "FlaxRobertaForCausalLM"),
+        ("big_bird", "FlaxBigBirdForCausalLM"),
+        ("electra", "FlaxElectraForCausalLM"),
     ]
 )
 

src/transformers/models/bert/__init__.py

@@ -65,6 +65,7 @@ if is_tf_available():
 
 if is_flax_available():
     _import_structure["modeling_flax_bert"] = [
+        "FlaxBertForCausalLM",
         "FlaxBertForMaskedLM",
         "FlaxBertForMultipleChoice",
         "FlaxBertForNextSentencePrediction",
@@ -119,6 +120,7 @@ if TYPE_CHECKING:
 
     if is_flax_available():
         from .modeling_flax_bert import (
+            FlaxBertForCausalLM,
             FlaxBertForMaskedLM,
             FlaxBertForMultipleChoice,
             FlaxBertForNextSentencePrediction,

src/transformers/models/bert/modeling_flax_bert.py

@@ -22,13 +22,16 @@ import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
 from flax.linen.attention import dot_product_attention_weights
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax import lax
 
 from ...modeling_flax_outputs import (
-    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
     FlaxBaseModelOutputWithPooling,
+    FlaxBaseModelOutputWithPoolingAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
     FlaxMaskedLMOutput,
     FlaxMultipleChoiceModelOutput,
     FlaxNextSentencePredictorOutput,
@@ -212,9 +215,11 @@ class FlaxBertEmbeddings(nn.Module):
 
 class FlaxBertSelfAttention(nn.Module):
     config: BertConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
+        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
         if self.config.hidden_size % self.config.num_attention_heads != 0:
             raise ValueError(
                 "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`\
@@ -237,30 +242,113 @@ class FlaxBertSelfAttention(nn.Module):
             kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
         )
 
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
+
+    @nn.compact
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states: Optional[jnp.array] = None,
+        init_cache: bool = False,
         deterministic=True,
         output_attentions: bool = False,
     ):
-        head_dim = self.config.hidden_size // self.config.num_attention_heads
-
-        query_states = self.query(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
-        )
-        value_states = self.value(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.query(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.key(key_value_states)
+            value_states = self.value(key_value_states)
+        else:
+            # self_attention
+            key_states = self.key(hidden_states)
+            value_states = self.value(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                 )
-        key_states = self.key(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
             )
 
         # Convert the boolean attention mask to an attention bias.
         if attention_mask is not None:
             # attention mask in the form of attention bias
-            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
@@ -318,10 +406,11 @@ class FlaxBertSelfOutput(nn.Module):
 
 class FlaxBertAttention(nn.Module):
     config: BertConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
-        self.self = FlaxBertSelfAttention(self.config, dtype=self.dtype)
+        self.self = FlaxBertSelfAttention(self.config, causal=self.causal, dtype=self.dtype)
         self.output = FlaxBertSelfOutput(self.config, dtype=self.dtype)
 
     def __call__(
@@ -329,6 +418,8 @@ class FlaxBertAttention(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states=None,
+        init_cache=False,
         deterministic=True,
         output_attentions: bool = False,
     ):
@@ -339,6 +430,8 @@ class FlaxBertAttention(nn.Module):
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            key_value_states=key_value_states,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
@@ -396,27 +489,46 @@ class FlaxBertLayer(nn.Module):
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        self.attention = FlaxBertAttention(self.config, dtype=self.dtype)
+        self.attention = FlaxBertAttention(self.config, causal=self.config.is_decoder, dtype=self.dtype)
         self.intermediate = FlaxBertIntermediate(self.config, dtype=self.dtype)
         self.output = FlaxBertOutput(self.config, dtype=self.dtype)
+        if self.config.add_cross_attention:
+            self.crossattention = FlaxBertAttention(self.config, causal=False, dtype=self.dtype)
 
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
     ):
+        # Self Attention
         attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
         attention_output = attention_outputs[0]
 
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=layer_head_mask,
+                key_value_states=encoder_hidden_states,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+
         hidden_states = self.intermediate(attention_output)
         hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
 
@@ -424,6 +536,8 @@ class FlaxBertLayer(nn.Module):
 
         if output_attentions:
             outputs += (attention_outputs[1],)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attention_outputs[1],)
         return outputs
 
 
@@ -441,6 +555,9 @@ class FlaxBertLayerCollection(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -448,6 +565,7 @@ class FlaxBertLayerCollection(nn.Module):
     ):
         all_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
         # Check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
@@ -465,6 +583,9 @@ class FlaxBertLayerCollection(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=head_mask[i] if head_mask is not None else None,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -474,6 +595,9 @@ class FlaxBertLayerCollection(nn.Module):
             if output_attentions:
                 all_attentions += (layer_outputs[1],)
 
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
@@ -482,8 +606,11 @@ class FlaxBertLayerCollection(nn.Module):
         if not return_dict:
             return tuple(v for v in outputs if v is not None)
 
-        return FlaxBaseModelOutput(
-            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
         )
 
 
@@ -499,6 +626,9 @@ class FlaxBertEncoder(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -508,6 +638,9 @@ class FlaxBertEncoder(nn.Module):
             hidden_states,
             attention_mask,
             head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
@@ -639,9 +772,26 @@ class FlaxBertPreTrainedModel(FlaxPreTrainedModel):
         params_rng, dropout_rng = jax.random.split(rng)
         rngs = {"params": params_rng, "dropout": dropout_rng}
 
-        random_params = self.module.init(
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(
+                rngs,
+                input_ids,
+                attention_mask,
+                token_type_ids,
+                position_ids,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                return_dict=False,
+            )
+        else:
+            module_init_outputs = self.module.init(
                 rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
-        )["params"]
+            )
+
+        random_params = module_init_outputs["params"]
 
         if params is not None:
             random_params = flatten_dict(unfreeze(random_params))
@@ -653,6 +803,26 @@ class FlaxBertPreTrainedModel(FlaxPreTrainedModel):
         else:
             return random_params
 
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderPreTrainedModel.init_cache
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
     @add_start_docstrings_to_model_forward(BERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     def __call__(
         self,
@@ -661,12 +831,15 @@ class FlaxBertPreTrainedModel(FlaxPreTrainedModel):
         token_type_ids=None,
         position_ids=None,
         head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
         params: dict = None,
         dropout_rng: jax.random.PRNGKey = None,
         train: bool = False,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        past_key_values: dict = None,
     ):
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -692,20 +865,61 @@ class FlaxBertPreTrainedModel(FlaxPreTrainedModel):
         if dropout_rng is not None:
             rngs["dropout"] = dropout_rng
 
-        return self.module.apply(
-            {"params": params or self.params},
+        inputs = {"params": params or self.params}
+
+        if self.config.add_cross_attention:
+            # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+            # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+            # changed by FlaxBertAttention module
+            if past_key_values:
+                inputs["cache"] = past_key_values
+                mutable = ["cache"]
+            else:
+                mutable = False
+
+            outputs = self.module.apply(
+                inputs,
+                jnp.array(input_ids, dtype="i4"),
+                jnp.array(attention_mask, dtype="i4"),
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                rngs=rngs,
+                mutable=mutable,
+            )
+
+            # add updated cache to model output
+            if past_key_values is not None and return_dict:
+                outputs, past_key_values = outputs
+                outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+                return outputs
+            elif past_key_values is not None and not return_dict:
+                outputs, past_key_values = outputs
+                outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        else:
+            outputs = self.module.apply(
+                inputs,
                 jnp.array(input_ids, dtype="i4"),
                 jnp.array(attention_mask, dtype="i4"),
-            jnp.array(token_type_ids, dtype="i4"),
-            jnp.array(position_ids, dtype="i4"),
-            jnp.array(head_mask, dtype="i4"),
-            not train,
-            output_attentions,
-            output_hidden_states,
-            return_dict,
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
                 rngs=rngs,
             )
 
+        return outputs
+
 
 class FlaxBertModule(nn.Module):
     config: BertConfig
@@ -721,9 +935,12 @@ class FlaxBertModule(nn.Module):
         self,
         input_ids,
         attention_mask,
-        token_type_ids: Optional[np.ndarray] = None,
-        position_ids: Optional[np.ndarray] = None,
-        head_mask: Optional[np.ndarray] = None,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -745,6 +962,9 @@ class FlaxBertModule(nn.Module):
             attention_mask,
             head_mask=head_mask,
             deterministic=deterministic,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -758,11 +978,12 @@ class FlaxBertModule(nn.Module):
                 return (hidden_states,) + outputs[1:]
             return (hidden_states, pooled) + outputs[1:]
 
-        return FlaxBaseModelOutputWithPooling(
+        return FlaxBaseModelOutputWithPoolingAndCrossAttentions(
             last_hidden_state=hidden_states,
             pooler_output=pooled,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
         )
 
 
@@ -1313,3 +1534,108 @@ append_call_sample_docstring(
     FlaxQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+class FlaxBertForCausalLMModule(nn.Module):
+    config: BertConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.bert = FlaxBertModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.cls = FlaxBertOnlyMLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.bert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.bert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.cls(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Bert Model with a language modeling head on top (a linear layer on top of the hidden-states output) e.g for
+    autoregressive tasks.
+    """,
+    BERT_START_DOCSTRING,
+)
+class FlaxBertForCausalLM(FlaxBertPreTrainedModel):
+    module_class = FlaxBertForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxBertForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)

src/transformers/models/big_bird/__init__.py

@@ -55,6 +55,7 @@ if is_torch_available():
 
 if is_flax_available():
     _import_structure["modeling_flax_big_bird"] = [
+        "FlaxBigBirdForCausalLM",
         "FlaxBigBirdForMaskedLM",
         "FlaxBigBirdForMultipleChoice",
         "FlaxBigBirdForPreTraining",
@@ -92,6 +93,7 @@ if TYPE_CHECKING:
 
     if is_flax_available():
         from .modeling_flax_big_bird import (
+            FlaxBigBirdForCausalLM,
             FlaxBigBirdForMaskedLM,
             FlaxBigBirdForMultipleChoice,
             FlaxBigBirdForPreTraining,

src/transformers/models/big_bird/modeling_flax_big_bird.py

@@ -22,13 +22,16 @@ import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
 from flax.linen.attention import dot_product_attention_weights
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax import lax
 
 from ...modeling_flax_outputs import (
-    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
     FlaxBaseModelOutputWithPooling,
+    FlaxBaseModelOutputWithPoolingAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
     FlaxMaskedLMOutput,
     FlaxMultipleChoiceModelOutput,
     FlaxSequenceClassifierOutput,
@@ -234,9 +237,11 @@ class FlaxBigBirdEmbeddings(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfAttention with Bert->BigBird
 class FlaxBigBirdSelfAttention(nn.Module):
     config: BigBirdConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
+        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
         if self.config.hidden_size % self.config.num_attention_heads != 0:
             raise ValueError(
                 "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`\
@@ -259,30 +264,113 @@ class FlaxBigBirdSelfAttention(nn.Module):
             kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
         )
 
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
+
+    @nn.compact
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states: Optional[jnp.array] = None,
+        init_cache: bool = False,
         deterministic=True,
         output_attentions: bool = False,
     ):
-        head_dim = self.config.hidden_size // self.config.num_attention_heads
-
-        query_states = self.query(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
-        )
-        value_states = self.value(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.query(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.key(key_value_states)
+            value_states = self.value(key_value_states)
+        else:
+            # self_attention
+            key_states = self.key(hidden_states)
+            value_states = self.value(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                 )
-        key_states = self.key(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
             )
 
         # Convert the boolean attention mask to an attention bias.
         if attention_mask is not None:
             # attention mask in the form of attention bias
-            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
@@ -1118,11 +1206,12 @@ class FlaxBigBirdSelfOutput(nn.Module):
 class FlaxBigBirdAttention(nn.Module):
     config: BigBirdConfig
     layer_id: int = None
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
         if self.config.attention_type == "original_full":
-            self.self = FlaxBigBirdSelfAttention(self.config, dtype=self.dtype)
+            self.self = FlaxBigBirdSelfAttention(self.config, causal=self.causal, dtype=self.dtype)
         elif self.config.attention_type == "block_sparse":
             self.self = FlaxBigBirdBlockSparseAttention(self.config, block_sparse_seed=self.layer_id, dtype=self.dtype)
         else:
@@ -1137,6 +1226,8 @@ class FlaxBigBirdAttention(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states=None,
+        init_cache=False,
         deterministic=True,
         output_attentions: bool = False,
     ):
@@ -1148,6 +1239,8 @@ class FlaxBigBirdAttention(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=layer_head_mask,
+                key_value_states=key_value_states,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -1215,9 +1308,13 @@ class FlaxBigBirdLayer(nn.Module):
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        self.attention = FlaxBigBirdAttention(self.config, layer_id=self.layer_id, dtype=self.dtype)
+        self.attention = FlaxBigBirdAttention(
+            self.config, layer_id=self.layer_id, causal=self.config.is_decoder, dtype=self.dtype
+        )
         self.intermediate = FlaxBigBirdIntermediate(self.config, dtype=self.dtype)
         self.output = FlaxBigBirdOutput(self.config, dtype=self.dtype)
+        if self.config.add_cross_attention:
+            self.crossattention = FlaxBigBirdAttention(self.config, causal=False, dtype=self.dtype)
 
     # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertLayer.__call__ with Bert->BigBird
     def __call__(
@@ -1225,18 +1322,35 @@ class FlaxBigBirdLayer(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
     ):
+        # Self Attention
         attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
         attention_output = attention_outputs[0]
 
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=layer_head_mask,
+                key_value_states=encoder_hidden_states,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+
         hidden_states = self.intermediate(attention_output)
         hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
 
@@ -1244,6 +1358,8 @@ class FlaxBigBirdLayer(nn.Module):
 
         if output_attentions:
             outputs += (attention_outputs[1],)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attention_outputs[1],)
         return outputs
 
 
@@ -1263,6 +1379,9 @@ class FlaxBigBirdLayerCollection(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -1270,6 +1389,7 @@ class FlaxBigBirdLayerCollection(nn.Module):
     ):
         all_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
         # Check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
@@ -1287,6 +1407,9 @@ class FlaxBigBirdLayerCollection(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=head_mask[i] if head_mask is not None else None,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -1296,6 +1419,9 @@ class FlaxBigBirdLayerCollection(nn.Module):
             if output_attentions:
                 all_attentions += (layer_outputs[1],)
 
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
@@ -1304,8 +1430,11 @@ class FlaxBigBirdLayerCollection(nn.Module):
         if not return_dict:
             return tuple(v for v in outputs if v is not None)
 
-        return FlaxBaseModelOutput(
-            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
         )
 
 
@@ -1322,6 +1451,9 @@ class FlaxBigBirdEncoder(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -1331,6 +1463,9 @@ class FlaxBigBirdEncoder(nn.Module):
             hidden_states,
             attention_mask,
             head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
@@ -1432,6 +1567,7 @@ class FlaxBigBirdPreTrainedModel(FlaxPreTrainedModel):
 
         super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
 
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.init_weights
     def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
         # init input tensors
         input_ids = jnp.zeros(input_shape, dtype="i4")
@@ -1443,9 +1579,26 @@ class FlaxBigBirdPreTrainedModel(FlaxPreTrainedModel):
         params_rng, dropout_rng = jax.random.split(rng)
         rngs = {"params": params_rng, "dropout": dropout_rng}
 
-        random_params = self.module.init(
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(
+                rngs,
+                input_ids,
+                attention_mask,
+                token_type_ids,
+                position_ids,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                return_dict=False,
+            )
+        else:
+            module_init_outputs = self.module.init(
                 rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
-        )["params"]
+            )
+
+        random_params = module_init_outputs["params"]
 
         if params is not None:
             random_params = flatten_dict(unfreeze(random_params))
@@ -1457,7 +1610,28 @@ class FlaxBigBirdPreTrainedModel(FlaxPreTrainedModel):
         else:
             return random_params
 
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderPreTrainedModel.init_cache
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
     @add_start_docstrings_to_model_forward(BIG_BIRD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.__call__ with Bert->BigBird
     def __call__(
         self,
         input_ids,
@@ -1465,12 +1639,15 @@ class FlaxBigBirdPreTrainedModel(FlaxPreTrainedModel):
         token_type_ids=None,
         position_ids=None,
         head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
         params: dict = None,
         dropout_rng: jax.random.PRNGKey = None,
         train: bool = False,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        past_key_values: dict = None,
     ):
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -1496,20 +1673,61 @@ class FlaxBigBirdPreTrainedModel(FlaxPreTrainedModel):
         if dropout_rng is not None:
             rngs["dropout"] = dropout_rng
 
-        return self.module.apply(
-            {"params": params or self.params},
+        inputs = {"params": params or self.params}
+
+        if self.config.add_cross_attention:
+            # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+            # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+            # changed by FlaxBigBirdAttention module
+            if past_key_values:
+                inputs["cache"] = past_key_values
+                mutable = ["cache"]
+            else:
+                mutable = False
+
+            outputs = self.module.apply(
+                inputs,
                 jnp.array(input_ids, dtype="i4"),
                 jnp.array(attention_mask, dtype="i4"),
-            jnp.array(token_type_ids, dtype="i4"),
-            jnp.array(position_ids, dtype="i4"),
-            jnp.array(head_mask, dtype="i4"),
-            not train,
-            output_attentions,
-            output_hidden_states,
-            return_dict,
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                rngs=rngs,
+                mutable=mutable,
+            )
+
+            # add updated cache to model output
+            if past_key_values is not None and return_dict:
+                outputs, past_key_values = outputs
+                outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+                return outputs
+            elif past_key_values is not None and not return_dict:
+                outputs, past_key_values = outputs
+                outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        else:
+            outputs = self.module.apply(
+                inputs,
+                jnp.array(input_ids, dtype="i4"),
+                jnp.array(attention_mask, dtype="i4"),
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
                 rngs=rngs,
             )
 
+        return outputs
+
 
 class FlaxBigBirdModule(nn.Module):
     config: BigBirdConfig
@@ -1532,6 +1750,9 @@ class FlaxBigBirdModule(nn.Module):
         token_type_ids,
         position_ids,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -1545,6 +1766,9 @@ class FlaxBigBirdModule(nn.Module):
             attention_mask,
             head_mask=head_mask,
             deterministic=deterministic,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -1559,11 +1783,12 @@ class FlaxBigBirdModule(nn.Module):
                 return (hidden_states,) + outputs[1:]
             return (hidden_states, pooled) + outputs[1:]
 
-        return FlaxBaseModelOutputWithPooling(
+        return FlaxBaseModelOutputWithPoolingAndCrossAttentions(
             last_hidden_state=hidden_states,
             pooler_output=pooled,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
         )
 
 
@@ -2181,3 +2406,110 @@ append_call_sample_docstring(
     FlaxBigBirdForQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForCausalLMModule with Bert->BigBird
+class FlaxBigBirdForCausalLMModule(nn.Module):
+    config: BigBirdConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.bert = FlaxBigBirdModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.cls = FlaxBigBirdOnlyMLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.bert(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.bert.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.cls(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    BigBird Model with a language modeling head on top (a linear layer on top of the hidden-states output) e.g for
+    autoregressive tasks.
+    """,
+    BIG_BIRD_START_DOCSTRING,
+)
+# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForCausalLM with Bert->BigBird
+class FlaxBigBirdForCausalLM(FlaxBigBirdPreTrainedModel):
+    module_class = FlaxBigBirdForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxBigBirdForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)

src/transformers/models/electra/__init__.py

@@ -59,6 +59,7 @@ if is_tf_available():
 
 if is_flax_available():
     _import_structure["modeling_flax_electra"] = [
+        "FlaxElectraForCausalLM",
         "FlaxElectraForMaskedLM",
         "FlaxElectraForMultipleChoice",
         "FlaxElectraForPreTraining",
@@ -107,6 +108,7 @@ if TYPE_CHECKING:
 
     if is_flax_available():
         from .modeling_flax_electra import (
+            FlaxElectraForCausalLM,
             FlaxElectraForMaskedLM,
             FlaxElectraForMultipleChoice,
             FlaxElectraForPreTraining,

src/transformers/models/electra/modeling_flax_electra.py

@@ -22,13 +22,15 @@ import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
 from flax.linen.attention import dot_product_attention_weights
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax import lax
-from jax.random import PRNGKey
 
 from ...modeling_flax_outputs import (
     FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
     FlaxMaskedLMOutput,
     FlaxMultipleChoiceModelOutput,
     FlaxQuestionAnsweringModelOutput,
@@ -184,9 +186,11 @@ class FlaxElectraEmbeddings(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfAttention with Bert->Electra
 class FlaxElectraSelfAttention(nn.Module):
     config: ElectraConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
+        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
         if self.config.hidden_size % self.config.num_attention_heads != 0:
             raise ValueError(
                 "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`\
@@ -209,30 +213,113 @@ class FlaxElectraSelfAttention(nn.Module):
             kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
         )
 
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
+
+    @nn.compact
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states: Optional[jnp.array] = None,
+        init_cache: bool = False,
         deterministic=True,
         output_attentions: bool = False,
     ):
-        head_dim = self.config.hidden_size // self.config.num_attention_heads
-
-        query_states = self.query(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
-        )
-        value_states = self.value(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.query(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.key(key_value_states)
+            value_states = self.value(key_value_states)
+        else:
+            # self_attention
+            key_states = self.key(hidden_states)
+            value_states = self.value(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                 )
-        key_states = self.key(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
             )
 
         # Convert the boolean attention mask to an attention bias.
         if attention_mask is not None:
             # attention mask in the form of attention bias
-            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
@@ -292,10 +379,11 @@ class FlaxElectraSelfOutput(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertAttention with Bert->Electra
 class FlaxElectraAttention(nn.Module):
     config: ElectraConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
-        self.self = FlaxElectraSelfAttention(self.config, dtype=self.dtype)
+        self.self = FlaxElectraSelfAttention(self.config, causal=self.causal, dtype=self.dtype)
         self.output = FlaxElectraSelfOutput(self.config, dtype=self.dtype)
 
     def __call__(
@@ -303,6 +391,8 @@ class FlaxElectraAttention(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states=None,
+        init_cache=False,
         deterministic=True,
         output_attentions: bool = False,
     ):
@@ -313,6 +403,8 @@ class FlaxElectraAttention(nn.Module):
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            key_value_states=key_value_states,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
@@ -373,27 +465,46 @@ class FlaxElectraLayer(nn.Module):
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        self.attention = FlaxElectraAttention(self.config, dtype=self.dtype)
+        self.attention = FlaxElectraAttention(self.config, causal=self.config.is_decoder, dtype=self.dtype)
         self.intermediate = FlaxElectraIntermediate(self.config, dtype=self.dtype)
         self.output = FlaxElectraOutput(self.config, dtype=self.dtype)
+        if self.config.add_cross_attention:
+            self.crossattention = FlaxElectraAttention(self.config, causal=False, dtype=self.dtype)
 
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
     ):
+        # Self Attention
         attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
         attention_output = attention_outputs[0]
 
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=layer_head_mask,
+                key_value_states=encoder_hidden_states,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+
         hidden_states = self.intermediate(attention_output)
         hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
 
@@ -401,6 +512,8 @@ class FlaxElectraLayer(nn.Module):
 
         if output_attentions:
             outputs += (attention_outputs[1],)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attention_outputs[1],)
         return outputs
 
 
@@ -419,6 +532,9 @@ class FlaxElectraLayerCollection(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -426,6 +542,7 @@ class FlaxElectraLayerCollection(nn.Module):
     ):
         all_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
         # Check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
@@ -443,6 +560,9 @@ class FlaxElectraLayerCollection(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=head_mask[i] if head_mask is not None else None,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -452,6 +572,9 @@ class FlaxElectraLayerCollection(nn.Module):
             if output_attentions:
                 all_attentions += (layer_outputs[1],)
 
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
@@ -460,8 +583,11 @@ class FlaxElectraLayerCollection(nn.Module):
         if not return_dict:
             return tuple(v for v in outputs if v is not None)
 
-        return FlaxBaseModelOutput(
-            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
         )
 
 
@@ -478,6 +604,9 @@ class FlaxElectraEncoder(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -487,6 +616,9 @@ class FlaxElectraEncoder(nn.Module):
             hidden_states,
             attention_mask,
             head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
@@ -548,6 +680,7 @@ class FlaxElectraPreTrainedModel(FlaxPreTrainedModel):
         module = self.module_class(config=config, dtype=dtype, **kwargs)
         super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
 
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.init_weights
     def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
         # init input tensors
         input_ids = jnp.zeros(input_shape, dtype="i4")
@@ -559,9 +692,26 @@ class FlaxElectraPreTrainedModel(FlaxPreTrainedModel):
         params_rng, dropout_rng = jax.random.split(rng)
         rngs = {"params": params_rng, "dropout": dropout_rng}
 
-        random_params = self.module.init(
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(
+                rngs,
+                input_ids,
+                attention_mask,
+                token_type_ids,
+                position_ids,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                return_dict=False,
+            )
+        else:
+            module_init_outputs = self.module.init(
                 rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
-        )["params"]
+            )
+
+        random_params = module_init_outputs["params"]
 
         if params is not None:
             random_params = flatten_dict(unfreeze(random_params))
@@ -573,6 +723,26 @@ class FlaxElectraPreTrainedModel(FlaxPreTrainedModel):
         else:
             return random_params
 
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderPreTrainedModel.init_cache
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
     @add_start_docstrings_to_model_forward(ELECTRA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     def __call__(
         self,
@@ -581,12 +751,15 @@ class FlaxElectraPreTrainedModel(FlaxPreTrainedModel):
         token_type_ids=None,
         position_ids=None,
         head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
         params: dict = None,
-        dropout_rng: PRNGKey = None,
+        dropout_rng: jax.random.PRNGKey = None,
         train: bool = False,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        past_key_values: dict = None,
     ):
 
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -613,20 +786,61 @@ class FlaxElectraPreTrainedModel(FlaxPreTrainedModel):
         if dropout_rng is not None:
             rngs["dropout"] = dropout_rng
 
-        return self.module.apply(
-            {"params": params or self.params},
+        inputs = {"params": params or self.params}
+
+        if self.config.add_cross_attention:
+            # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+            # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+            # changed by FlaxElectraAttention module
+            if past_key_values:
+                inputs["cache"] = past_key_values
+                mutable = ["cache"]
+            else:
+                mutable = False
+
+            outputs = self.module.apply(
+                inputs,
+                jnp.array(input_ids, dtype="i4"),
+                jnp.array(attention_mask, dtype="i4"),
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                rngs=rngs,
+                mutable=mutable,
+            )
+
+            # add updated cache to model output
+            if past_key_values is not None and return_dict:
+                outputs, past_key_values = outputs
+                outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+                return outputs
+            elif past_key_values is not None and not return_dict:
+                outputs, past_key_values = outputs
+                outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        else:
+            outputs = self.module.apply(
+                inputs,
                 jnp.array(input_ids, dtype="i4"),
                 jnp.array(attention_mask, dtype="i4"),
-            jnp.array(token_type_ids, dtype="i4"),
-            jnp.array(position_ids, dtype="i4"),
-            jnp.array(head_mask, dtype="i4"),
-            not train,
-            output_attentions,
-            output_hidden_states,
-            return_dict,
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
                 rngs=rngs,
             )
 
+        return outputs
+
 
 class FlaxElectraModule(nn.Module):
     config: ElectraConfig
@@ -645,6 +859,9 @@ class FlaxElectraModule(nn.Module):
         token_type_ids,
         position_ids,
         head_mask: Optional[np.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -661,6 +878,9 @@ class FlaxElectraModule(nn.Module):
             attention_mask,
             head_mask=head_mask,
             deterministic=deterministic,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -1232,3 +1452,111 @@ append_call_sample_docstring(
     FlaxSequenceClassifierOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+class FlaxElectraForCausalLMModule(nn.Module):
+    config: ElectraConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.electra = FlaxElectraModule(config=self.config, dtype=self.dtype)
+        self.generator_predictions = FlaxElectraGeneratorPredictions(config=self.config, dtype=self.dtype)
+        if self.config.tie_word_embeddings:
+            self.generator_lm_head = FlaxElectraTiedDense(self.config.vocab_size, dtype=self.dtype)
+        else:
+            self.generator_lm_head = nn.Dense(self.config.vocab_size, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask: Optional[jnp.ndarray] = None,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        outputs = self.electra(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        hidden_states = outputs[0]
+        prediction_scores = self.generator_predictions(hidden_states)
+
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.electra.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+            prediction_scores = self.generator_lm_head(prediction_scores, shared_embedding.T)
+        else:
+            prediction_scores = self.generator_lm_head(prediction_scores)
+
+        if not return_dict:
+            return (prediction_scores,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Electra Model with a language modeling head on top (a linear layer on top of the hidden-states output) e.g for
+    autoregressive tasks.
+    """,
+    ELECTRA_START_DOCSTRING,
+)
+# Copied from transformers.models.bert.modeling_flax_bert.FlaxBertForCausalLM with Bert->Electra
+class FlaxElectraForCausalLM(FlaxElectraPreTrainedModel):
+    module_class = FlaxElectraForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxElectraForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)

src/transformers/models/roberta/__init__.py

@@ -58,6 +58,7 @@ if is_tf_available():
 
 if is_flax_available():
     _import_structure["modeling_flax_roberta"] = [
+        "FlaxRobertaForCausalLM",
         "FlaxRobertaForMaskedLM",
         "FlaxRobertaForMultipleChoice",
         "FlaxRobertaForQuestionAnswering",
@@ -103,7 +104,8 @@ if TYPE_CHECKING:
         )
 
     if is_flax_available():
-        from .modeling_tf_roberta import (
+        from .modeling_flax_roberta import (
+            FlaxRobertaForCausalLM,
             FlaxRobertaForMaskedLM,
             FlaxRobertaForMultipleChoice,
             FlaxRobertaForQuestionAnswering,

src/transformers/models/roberta/modeling_flax_roberta.py

@@ -20,14 +20,16 @@ import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, freeze, unfreeze
+from flax.linen import combine_masks, make_causal_mask
 from flax.linen.attention import dot_product_attention_weights
 from flax.traverse_util import flatten_dict, unflatten_dict
 from jax import lax
-from jax.random import PRNGKey
 
 from ...modeling_flax_outputs import (
-    FlaxBaseModelOutput,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
     FlaxBaseModelOutputWithPooling,
+    FlaxBaseModelOutputWithPoolingAndCrossAttentions,
+    FlaxCausalLMOutputWithCrossAttentions,
     FlaxMaskedLMOutput,
     FlaxMultipleChoiceModelOutput,
     FlaxQuestionAnsweringModelOutput,
@@ -174,9 +176,11 @@ class FlaxRobertaEmbeddings(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfAttention with Bert->Roberta
 class FlaxRobertaSelfAttention(nn.Module):
     config: RobertaConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
+        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
         if self.config.hidden_size % self.config.num_attention_heads != 0:
             raise ValueError(
                 "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`\
@@ -199,30 +203,113 @@ class FlaxRobertaSelfAttention(nn.Module):
             kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
         )
 
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
+
+    @nn.compact
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states: Optional[jnp.array] = None,
+        init_cache: bool = False,
         deterministic=True,
         output_attentions: bool = False,
     ):
-        head_dim = self.config.hidden_size // self.config.num_attention_heads
-
-        query_states = self.query(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
-        )
-        value_states = self.value(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
+
+        # get query proj
+        query_states = self.query(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.key(key_value_states)
+            value_states = self.value(key_value_states)
+        else:
+            # self_attention
+            key_states = self.key(hidden_states)
+            value_states = self.value(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                 )
-        key_states = self.key(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
             )
 
         # Convert the boolean attention mask to an attention bias.
         if attention_mask is not None:
             # attention mask in the form of attention bias
-            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
@@ -282,10 +369,11 @@ class FlaxRobertaSelfOutput(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertAttention with Bert->Roberta
 class FlaxRobertaAttention(nn.Module):
     config: RobertaConfig
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
-        self.self = FlaxRobertaSelfAttention(self.config, dtype=self.dtype)
+        self.self = FlaxRobertaSelfAttention(self.config, causal=self.causal, dtype=self.dtype)
         self.output = FlaxRobertaSelfOutput(self.config, dtype=self.dtype)
 
     def __call__(
@@ -293,6 +381,8 @@ class FlaxRobertaAttention(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states=None,
+        init_cache=False,
         deterministic=True,
         output_attentions: bool = False,
     ):
@@ -303,6 +393,8 @@ class FlaxRobertaAttention(nn.Module):
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            key_value_states=key_value_states,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
@@ -363,27 +455,46 @@ class FlaxRobertaLayer(nn.Module):
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
-        self.attention = FlaxRobertaAttention(self.config, dtype=self.dtype)
+        self.attention = FlaxRobertaAttention(self.config, causal=self.config.is_decoder, dtype=self.dtype)
         self.intermediate = FlaxRobertaIntermediate(self.config, dtype=self.dtype)
         self.output = FlaxRobertaOutput(self.config, dtype=self.dtype)
+        if self.config.add_cross_attention:
+            self.crossattention = FlaxRobertaAttention(self.config, causal=False, dtype=self.dtype)
 
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
     ):
+        # Self Attention
         attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
         attention_output = attention_outputs[0]
 
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=layer_head_mask,
+                key_value_states=encoder_hidden_states,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+
         hidden_states = self.intermediate(attention_output)
         hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
 
@@ -391,6 +502,8 @@ class FlaxRobertaLayer(nn.Module):
 
         if output_attentions:
             outputs += (attention_outputs[1],)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attention_outputs[1],)
         return outputs
 
 
@@ -409,6 +522,9 @@ class FlaxRobertaLayerCollection(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -416,6 +532,7 @@ class FlaxRobertaLayerCollection(nn.Module):
     ):
         all_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
         # Check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
@@ -433,6 +550,9 @@ class FlaxRobertaLayerCollection(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=head_mask[i] if head_mask is not None else None,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -442,6 +562,9 @@ class FlaxRobertaLayerCollection(nn.Module):
             if output_attentions:
                 all_attentions += (layer_outputs[1],)
 
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
@@ -450,8 +573,11 @@ class FlaxRobertaLayerCollection(nn.Module):
         if not return_dict:
             return tuple(v for v in outputs if v is not None)
 
-        return FlaxBaseModelOutput(
-            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
         )
 
 
@@ -468,6 +594,9 @@ class FlaxRobertaEncoder(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -477,6 +606,9 @@ class FlaxRobertaEncoder(nn.Module):
             hidden_states,
             attention_mask,
             head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
@@ -603,9 +735,26 @@ class FlaxRobertaPreTrainedModel(FlaxPreTrainedModel):
         params_rng, dropout_rng = jax.random.split(rng)
         rngs = {"params": params_rng, "dropout": dropout_rng}
 
-        random_params = self.module.init(
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(
+                rngs,
+                input_ids,
+                attention_mask,
+                token_type_ids,
+                position_ids,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                return_dict=False,
+            )
+        else:
+            module_init_outputs = self.module.init(
                 rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
-        )["params"]
+            )
+
+        random_params = module_init_outputs["params"]
 
         if params is not None:
             random_params = flatten_dict(unfreeze(random_params))
@@ -617,6 +766,26 @@ class FlaxRobertaPreTrainedModel(FlaxPreTrainedModel):
         else:
             return random_params
 
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartDecoderPreTrainedModel.init_cache
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids, dtype="i4")
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
     @add_start_docstrings_to_model_forward(ROBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     def __call__(
         self,
@@ -625,12 +794,15 @@ class FlaxRobertaPreTrainedModel(FlaxPreTrainedModel):
         token_type_ids=None,
         position_ids=None,
         head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
         params: dict = None,
-        dropout_rng: PRNGKey = None,
+        dropout_rng: jax.random.PRNGKey = None,
         train: bool = False,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        past_key_values: dict = None,
     ):
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -656,20 +828,61 @@ class FlaxRobertaPreTrainedModel(FlaxPreTrainedModel):
         if dropout_rng is not None:
             rngs["dropout"] = dropout_rng
 
-        return self.module.apply(
-            {"params": params or self.params},
+        inputs = {"params": params or self.params}
+
+        if self.config.add_cross_attention:
+            # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+            # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+            # changed by FlaxRobertaAttention module
+            if past_key_values:
+                inputs["cache"] = past_key_values
+                mutable = ["cache"]
+            else:
+                mutable = False
+
+            outputs = self.module.apply(
+                inputs,
+                jnp.array(input_ids, dtype="i4"),
+                jnp.array(attention_mask, dtype="i4"),
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                rngs=rngs,
+                mutable=mutable,
+            )
+
+            # add updated cache to model output
+            if past_key_values is not None and return_dict:
+                outputs, past_key_values = outputs
+                outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+                return outputs
+            elif past_key_values is not None and not return_dict:
+                outputs, past_key_values = outputs
+                outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        else:
+            outputs = self.module.apply(
+                inputs,
                 jnp.array(input_ids, dtype="i4"),
                 jnp.array(attention_mask, dtype="i4"),
-            jnp.array(token_type_ids, dtype="i4"),
-            jnp.array(position_ids, dtype="i4"),
-            jnp.array(head_mask, dtype="i4"),
-            not train,
-            output_attentions,
-            output_hidden_states,
-            return_dict,
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
                 rngs=rngs,
             )
 
+        return outputs
+
 
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertModule with Bert->Roberta
 class FlaxRobertaModule(nn.Module):
@@ -686,9 +899,12 @@ class FlaxRobertaModule(nn.Module):
         self,
         input_ids,
         attention_mask,
-        token_type_ids: Optional[np.ndarray] = None,
-        position_ids: Optional[np.ndarray] = None,
-        head_mask: Optional[np.ndarray] = None,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -710,6 +926,9 @@ class FlaxRobertaModule(nn.Module):
             attention_mask,
             head_mask=head_mask,
             deterministic=deterministic,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -723,11 +942,12 @@ class FlaxRobertaModule(nn.Module):
                 return (hidden_states,) + outputs[1:]
             return (hidden_states, pooled) + outputs[1:]
 
-        return FlaxBaseModelOutputWithPooling(
+        return FlaxBaseModelOutputWithPoolingAndCrossAttentions(
             last_hidden_state=hidden_states,
             pooler_output=pooled,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
         )
 
 
@@ -1101,3 +1321,108 @@ append_call_sample_docstring(
     FlaxQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+class FlaxRobertaForCausalLMModule(nn.Module):
+    config: RobertaConfig
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.roberta = FlaxRobertaModule(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.lm_head = FlaxRobertaLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.roberta(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.roberta.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.lm_head(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    Roberta Model with a language modeling head on top (a linear layer on top of the hidden-states output) e.g for
+    autoregressive tasks.
+    """,
+    ROBERTA_START_DOCSTRING,
+)
+class FlaxRobertaForCausalLM(FlaxRobertaPreTrainedModel):
+    module_class = FlaxRobertaForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    FlaxRobertaForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)

src/transformers/utils/dummy_flax_objects.py

@@ -326,6 +326,13 @@ class FlaxBeitPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxBertForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxBertForMaskedLM(metaclass=DummyObject):
     _backends = ["flax"]
 
@@ -389,6 +396,13 @@ class FlaxBertPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxBigBirdForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxBigBirdForMaskedLM(metaclass=DummyObject):
     _backends = ["flax"]
 
@@ -578,6 +592,13 @@ class FlaxDistilBertPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxElectraForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxElectraForMaskedLM(metaclass=DummyObject):
     _backends = ["flax"]
 
@@ -795,6 +816,13 @@ class FlaxPegasusPreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["flax"])
 
 
+class FlaxRobertaForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
 class FlaxRobertaForMaskedLM(metaclass=DummyObject):
     _backends = ["flax"]
 

templates/adding_a_new_model/cookiecutter-template-{{cookiecutter.modelname}}/modeling_flax_{{cookiecutter.lowercase_modelname}}.py

@@ -24,15 +24,17 @@ import flax.linen as nn
 import jax
 import jax.numpy as jnp
 from flax.core.frozen_dict import FrozenDict, unfreeze, freeze
+from flax.linen import combine_masks, make_causal_mask
 from flax.traverse_util import flatten_dict, unflatten_dict
 from flax.linen.attention import dot_product_attention_weights
 from jax import lax
 
 from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward
 from ...modeling_flax_outputs import (
-    FlaxBaseModelOutput,
-    FlaxBaseModelOutputWithPooling,
+    FlaxBaseModelOutputWithPastAndCrossAttentions,
+    FlaxBaseModelOutputWithPoolingAndCrossAttentions,
     FlaxCausalLMOutput,
+    FlaxCausalLMOutputWithCrossAttentions,
     FlaxMaskedLMOutput,
     FlaxMultipleChoiceModelOutput,
     FlaxQuestionAnsweringModelOutput,
@@ -170,9 +172,11 @@ class Flax{{cookiecutter.camelcase_modelname}}Embeddings(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertSelfAttention with Bert->{{cookiecutter.camelcase_modelname}}
 class Flax{{cookiecutter.camelcase_modelname}}SelfAttention(nn.Module):
     config: {{cookiecutter.camelcase_modelname}}Config
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32  # the dtype of the computation
 
     def setup(self):
+        self.head_dim = self.config.hidden_size // self.config.num_attention_heads
         if self.config.hidden_size % self.config.num_attention_heads != 0:
             raise ValueError(
                 "`config.hidden_size`: {self.config.hidden_size} has to be a multiple of `config.num_attention_heads`\
@@ -195,30 +199,113 @@ class Flax{{cookiecutter.camelcase_modelname}}SelfAttention(nn.Module):
             kernel_init=jax.nn.initializers.normal(self.config.initializer_range),
         )
 
+        if self.causal:
+            self.causal_mask = make_causal_mask(
+                jnp.ones((1, self.config.max_position_embeddings), dtype="bool"), dtype="bool"
+            )
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.num_attention_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.config.hidden_size,))
+
+    @nn.compact
+    # Copied from transformers.models.bart.modeling_flax_bart.FlaxBartAttention._concatenate_to_cache
+    def _concatenate_to_cache(self, key, value, query, attention_mask):
+        """
+        This function takes projected key, value states from a single input token and concatenates the states to cached
+        states from previous steps. This function is slighly adapted from the official Flax repository:
+        https://github.com/google/flax/blob/491ce18759622506588784b4fca0e4bf05f8c8cd/flax/linen/attention.py#L252
+        """
+        # detect if we're initializing by absence of existing cache data.
+        is_initialized = self.has_variable("cache", "cached_key")
+        cached_key = self.variable("cache", "cached_key", jnp.zeros, key.shape, key.dtype)
+        cached_value = self.variable("cache", "cached_value", jnp.zeros, value.shape, value.dtype)
+        cache_index = self.variable("cache", "cache_index", lambda: jnp.array(0, dtype=jnp.int32))
+
+        if is_initialized:
+            *batch_dims, max_length, num_heads, depth_per_head = cached_key.value.shape
+            # update key, value caches with our new 1d spatial slices
+            cur_index = cache_index.value
+            indices = (0,) * len(batch_dims) + (cur_index, 0, 0)
+            key = lax.dynamic_update_slice(cached_key.value, key, indices)
+            value = lax.dynamic_update_slice(cached_value.value, value, indices)
+            cached_key.value = key
+            cached_value.value = value
+            num_updated_cache_vectors = query.shape[1]
+            cache_index.value = cache_index.value + num_updated_cache_vectors
+            # causal mask for cached decoder self-attention: our single query position should only attend to those key positions that have already been generated and cached, not the remaining zero elements.
+            pad_mask = jnp.broadcast_to(
+                jnp.arange(max_length) < cur_index + num_updated_cache_vectors,
+                tuple(batch_dims) + (1, num_updated_cache_vectors, max_length),
+            )
+            attention_mask = combine_masks(pad_mask, attention_mask)
+        return key, value, attention_mask
+
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states: Optional[jnp.array] = None,
+        init_cache: bool = False,
         deterministic=True,
-        output_attentions: bool = False
+        output_attentions: bool = False,
     ):
-        head_dim = self.config.hidden_size // self.config.num_attention_heads
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        batch_size = hidden_states.shape[0]
 
-        query_states = self.query(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
-        )
-        value_states = self.value(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+        # get query proj
+        query_states = self.query(hidden_states)
+        # get key, value proj
+        if is_cross_attention:
+            # cross_attentions
+            key_states = self.key(key_value_states)
+            value_states = self.value(key_value_states)
+        else:
+            # self_attention
+            key_states = self.key(hidden_states)
+            value_states = self.value(hidden_states)
+
+        query_states = self._split_heads(query_states)
+        key_states = self._split_heads(key_states)
+        value_states = self._split_heads(value_states)
+
+        # handle cache prepare causal attention mask
+        if self.causal:
+            query_length, key_length = query_states.shape[1], key_states.shape[1]
+            if self.has_variable("cache", "cached_key"):
+                mask_shift = self.variables["cache"]["cache_index"]
+                max_decoder_length = self.variables["cache"]["cached_key"].shape[1]
+                causal_mask = lax.dynamic_slice(
+                    self.causal_mask, (0, 0, mask_shift, 0), (1, 1, query_length, max_decoder_length)
                 )
-        key_states = self.key(hidden_states).reshape(
-            hidden_states.shape[:2] + (self.config.num_attention_heads, head_dim)
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+            causal_mask = jnp.broadcast_to(causal_mask, (batch_size,) + causal_mask.shape[1:])
+
+        # combine masks if needed
+        if attention_mask is not None and self.causal:
+            attention_mask = jnp.broadcast_to(jnp.expand_dims(attention_mask, axis=(-3, -2)), causal_mask.shape)
+            attention_mask = combine_masks(attention_mask, causal_mask)
+        elif self.causal:
+            attention_mask = causal_mask
+        elif attention_mask is not None:
+            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
+
+        # During fast autoregressive decoding, we feed one position at a time,
+        # and cache the keys and values step by step.
+        if self.causal and (self.has_variable("cache", "cached_key") or init_cache):
+            key_states, value_states, attention_mask = self._concatenate_to_cache(
+                key_states, value_states, query_states, attention_mask
             )
 
         # Convert the boolean attention mask to an attention bias.
         if attention_mask is not None:
             # attention mask in the form of attention bias
-            attention_mask = jnp.expand_dims(attention_mask, axis=(-3, -2))
             attention_bias = lax.select(
                 attention_mask > 0,
                 jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
@@ -278,6 +365,7 @@ class Flax{{cookiecutter.camelcase_modelname}}SelfOutput(nn.Module):
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertAttention with Bert->{{cookiecutter.camelcase_modelname}}
 class Flax{{cookiecutter.camelcase_modelname}}Attention(nn.Module):
     config: {{cookiecutter.camelcase_modelname}}Config
+    causal: bool = False
     dtype: jnp.dtype = jnp.float32
 
     def setup(self):
@@ -289,6 +377,8 @@ class Flax{{cookiecutter.camelcase_modelname}}Attention(nn.Module):
         hidden_states,
         attention_mask,
         layer_head_mask,
+        key_value_states=None,
+        init_cache=False,
         deterministic=True,
         output_attentions: bool = False,
     ):
@@ -299,6 +389,8 @@ class Flax{{cookiecutter.camelcase_modelname}}Attention(nn.Module):
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            key_value_states=key_value_states,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
@@ -362,24 +454,43 @@ class Flax{{cookiecutter.camelcase_modelname}}Layer(nn.Module):
         self.attention = Flax{{cookiecutter.camelcase_modelname}}Attention(self.config, dtype=self.dtype)
         self.intermediate = Flax{{cookiecutter.camelcase_modelname}}Intermediate(self.config, dtype=self.dtype)
         self.output = Flax{{cookiecutter.camelcase_modelname}}Output(self.config, dtype=self.dtype)
+        if self.config.add_cross_attention:
+            self.crossattention = Flax{{cookiecutter.camelcase_modelname}}Attention(self.config, causal=False, dtype=self.dtype)
 
     def __call__(
         self,
         hidden_states,
         attention_mask,
         layer_head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
     ):
+        # Self Attention
         attention_outputs = self.attention(
             hidden_states,
             attention_mask,
             layer_head_mask=layer_head_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
         )
         attention_output = attention_outputs[0]
 
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=layer_head_mask,
+                key_value_states=encoder_hidden_states,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+
         hidden_states = self.intermediate(attention_output)
         hidden_states = self.output(hidden_states, attention_output, deterministic=deterministic)
 
@@ -387,6 +498,8 @@ class Flax{{cookiecutter.camelcase_modelname}}Layer(nn.Module):
 
         if output_attentions:
             outputs += (attention_outputs[1],)
+            if encoder_hidden_states is not None:
+                outputs += (cross_attention_outputs[1],)
         return outputs
 
 
@@ -405,6 +518,9 @@ class Flax{{cookiecutter.camelcase_modelname}}LayerCollection(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -412,6 +528,7 @@ class Flax{{cookiecutter.camelcase_modelname}}LayerCollection(nn.Module):
     ):
         all_attentions = () if output_attentions else None
         all_hidden_states = () if output_hidden_states else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
 
         # Check if head_mask has a correct number of layers specified if desired
         if head_mask is not None:
@@ -429,6 +546,9 @@ class Flax{{cookiecutter.camelcase_modelname}}LayerCollection(nn.Module):
                 hidden_states,
                 attention_mask,
                 layer_head_mask=head_mask[i] if head_mask is not None else None,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                init_cache=init_cache,
                 deterministic=deterministic,
                 output_attentions=output_attentions,
             )
@@ -438,6 +558,9 @@ class Flax{{cookiecutter.camelcase_modelname}}LayerCollection(nn.Module):
             if output_attentions:
                 all_attentions += (layer_outputs[1],)
 
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
         if output_hidden_states:
             all_hidden_states += (hidden_states,)
 
@@ -446,8 +569,11 @@ class Flax{{cookiecutter.camelcase_modelname}}LayerCollection(nn.Module):
         if not return_dict:
             return tuple(v for v in outputs if v is not None)
 
-        return FlaxBaseModelOutput(
-            last_hidden_state=hidden_states, hidden_states=all_hidden_states, attentions=all_attentions
+        return FlaxBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
         )
 
 
@@ -464,6 +590,9 @@ class Flax{{cookiecutter.camelcase_modelname}}Encoder(nn.Module):
         hidden_states,
         attention_mask,
         head_mask,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
@@ -473,6 +602,9 @@ class Flax{{cookiecutter.camelcase_modelname}}Encoder(nn.Module):
             hidden_states,
             attention_mask,
             head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             deterministic=deterministic,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
@@ -598,6 +730,7 @@ class Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel(FlaxPreTrainedMode
         module = self.module_class(config=config, dtype=dtype, **kwargs)
         super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype, _do_init=_do_init)
 
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.init_weights with Bert->{{cookiecutter.camelcase_modelname}}
     def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple, params: FrozenDict = None) -> FrozenDict:
         # init input tensors
         input_ids = jnp.zeros(input_shape, dtype="i4")
@@ -609,9 +742,26 @@ class Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel(FlaxPreTrainedMode
         params_rng, dropout_rng = jax.random.split(rng)
         rngs = {"params": params_rng, "dropout": dropout_rng}
 
-        random_params =  self.module.init(
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.hidden_size,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(
+                rngs,
+                input_ids,
+                attention_mask,
+                token_type_ids,
+                position_ids,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                return_dict=False,
+            )
+        else:
+            module_init_outputs = self.module.init(
                 rngs, input_ids, attention_mask, token_type_ids, position_ids, head_mask, return_dict=False
-        )["params"]
+            )
+
+        random_params = module_init_outputs["params"]
 
         if params is not None:
             random_params = flatten_dict(unfreeze(random_params))
@@ -623,7 +773,29 @@ class Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel(FlaxPreTrainedMode
         else:
             return random_params
 
+
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.init_cache with Bert->{{cookiecutter.camelcase_modelname}}
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (`int`):
+                maximum possible length for auto-regressive decoding. Defines the sequence length of the initialized
+                cache.
+        """
+        # init input variables to retrieve cache
+        input_ids = jnp.ones((batch_size, max_length))
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0), input_ids, attention_mask, position_ids, return_dict=False, init_cache=True
+        )
+        return unfreeze(init_variables["cache"])
+
     @add_start_docstrings_to_model_forward({{cookiecutter.uppercase_modelname}}_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertPreTrainedModel.__call__ with Bert->{{cookiecutter.camelcase_modelname}}
     def __call__(
         self,
         input_ids,
@@ -631,12 +803,15 @@ class Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel(FlaxPreTrainedMode
         token_type_ids=None,
         position_ids=None,
         head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
         params: dict = None,
         dropout_rng: jax.random.PRNGKey = None,
         train: bool = False,
         output_attentions: Optional[bool] = None,
         output_hidden_states: Optional[bool] = None,
         return_dict: Optional[bool] = None,
+        past_key_values: dict = None,
     ):
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
         output_hidden_states = (
@@ -662,20 +837,61 @@ class Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel(FlaxPreTrainedMode
         if dropout_rng is not None:
             rngs["dropout"] = dropout_rng
 
-        return self.module.apply(
-            {"params": params or self.params},
+        inputs = {"params": params or self.params}
+
+        if self.config.add_cross_attention:
+            # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed
+            # down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be
+            # changed by FlaxBertAttention module
+            if past_key_values:
+                inputs["cache"] = past_key_values
+                mutable = ["cache"]
+            else:
+                mutable = False
+
+            outputs = self.module.apply(
+                inputs,
                 jnp.array(input_ids, dtype="i4"),
                 jnp.array(attention_mask, dtype="i4"),
-            jnp.array(token_type_ids, dtype="i4"),
-            jnp.array(position_ids, dtype="i4"),
-            jnp.array(head_mask, dtype="i4"),
-            not train,
-            output_attentions,
-            output_hidden_states,
-            return_dict,
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                rngs=rngs,
+                mutable=mutable,
+            )
+
+            # add updated cache to model output
+            if past_key_values is not None and return_dict:
+                outputs, past_key_values = outputs
+                outputs["past_key_values"] = unfreeze(past_key_values["cache"])
+                return outputs
+            elif past_key_values is not None and not return_dict:
+                outputs, past_key_values = outputs
+                outputs = outputs[:1] + (unfreeze(past_key_values["cache"]),) + outputs[1:]
+
+        else:
+            outputs = self.module.apply(
+                inputs,
+                jnp.array(input_ids, dtype="i4"),
+                jnp.array(attention_mask, dtype="i4"),
+                token_type_ids=jnp.array(token_type_ids, dtype="i4"),
+                position_ids=jnp.array(position_ids, dtype="i4"),
+                head_mask=jnp.array(head_mask, dtype="i4"),
+                deterministic=not train,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
                 rngs=rngs,
             )
 
+        return outputs
+
 # Copied from transformers.models.bert.modeling_flax_bert.FlaxBertModule with Bert->{{cookiecutter.camelcase_modelname}}
 class Flax{{cookiecutter.camelcase_modelname}}Module(nn.Module):
     config: {{cookiecutter.camelcase_modelname}}Config
@@ -691,14 +907,25 @@ class Flax{{cookiecutter.camelcase_modelname}}Module(nn.Module):
         self,
         input_ids,
         attention_mask,
-        token_type_ids,
-        position_ids,
-        head_mask,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
         deterministic: bool = True,
         output_attentions: bool = False,
         output_hidden_states: bool = False,
         return_dict: bool = True,
     ):
+        # make sure `token_type_ids` is correctly initialized when not passed
+        if token_type_ids is None:
+            token_type_ids = jnp.zeros_like(input_ids)
+
+        # make sure `position_ids` is correctly initialized when not passed
+        if position_ids is None:
+            position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_ids.shape)
+
         hidden_states = self.embeddings(
             input_ids, token_type_ids, position_ids, attention_mask, deterministic=deterministic
         )
@@ -707,6 +934,9 @@ class Flax{{cookiecutter.camelcase_modelname}}Module(nn.Module):
             attention_mask,
             head_mask=head_mask,
             deterministic=deterministic,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
@@ -720,11 +950,12 @@ class Flax{{cookiecutter.camelcase_modelname}}Module(nn.Module):
                 return (hidden_states,) + outputs[1:]
             return (hidden_states, pooled) + outputs[1:]
 
-        return FlaxBaseModelOutputWithPooling(
+        return FlaxBaseModelOutputWithPoolingAndCrossAttentions(
             last_hidden_state=hidden_states,
             pooler_output=pooled,
             hidden_states=outputs.hidden_states,
             attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
         )
 
 add_start_docstrings(
@@ -1137,6 +1368,112 @@ append_call_sample_docstring(
     FlaxQuestionAnsweringModelOutput,
     _CONFIG_FOR_DOC,
 )
+
+
+class Flax{{cookiecutter.camelcase_modelname}}ForCausalLMModule(nn.Module):
+    config: {{cookiecutter.camelcase_modelname}}Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.{{cookiecutter.lowercase_modelname}} = Flax{{cookiecutter.camelcase_modelname}}Module(config=self.config, add_pooling_layer=False, dtype=self.dtype)
+        self.cls = Flax{{cookiecutter.camelcase_modelname}}OnlyMLMHead(config=self.config, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        token_type_ids: Optional[jnp.ndarray] = None,
+        head_mask: Optional[jnp.ndarray] = None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        init_cache: bool = False,
+        deterministic: bool = True,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        # Model
+        outputs = self.{{cookiecutter.lowercase_modelname}}(
+            input_ids,
+            attention_mask,
+            token_type_ids,
+            position_ids,
+            head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            init_cache=init_cache,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        if self.config.tie_word_embeddings:
+            shared_embedding = self.{{cookiecutter.lowercase_modelname}}.variables["params"]["embeddings"]["word_embeddings"]["embedding"]
+        else:
+            shared_embedding = None
+
+        # Compute the prediction scores
+        logits = self.cls(hidden_states, shared_embedding=shared_embedding)
+
+        if not return_dict:
+            return (logits,) + outputs[1:]
+
+        return FlaxCausalLMOutputWithCrossAttentions(
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    {{cookiecutter.camelcase_modelname}} Model with a language modeling head on top (a linear layer on top of the hidden-states output) e.g for
+    autoregressive tasks.
+    """,
+    {{cookiecutter.uppercase_modelname}}_START_DOCSTRING,
+)
+
+class Flax{{cookiecutter.camelcase_modelname}}ForCausalLM(Flax{{cookiecutter.camelcase_modelname}}PreTrainedModel):
+    module_class = Flax{{cookiecutter.camelcase_modelname}}ForCausalLMModule
+
+    def prepare_inputs_for_generation(self, input_ids, max_length, attention_mask: Optional[jnp.DeviceArray] = None):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length)
+        # Note that usually one would have to put 0's in the attention_mask for x > input_ids.shape[-1] and x < cache_length.
+        # But since the decoder uses a causal mask, those positions are masked anyway.
+        # Thus, we can create a single static attention_mask here, which is more efficient for compilation
+        extended_attention_mask = jnp.ones((batch_size, max_length), dtype="i4")
+        if attention_mask is not None:
+            position_ids = attention_mask.cumsum(axis=-1) - 1
+            extended_attention_mask = lax.dynamic_update_slice(extended_attention_mask, attention_mask, (0, 0))
+        else:
+            position_ids = jnp.broadcast_to(jnp.arange(seq_length, dtype="i4")[None, :], (batch_size, seq_length))
+
+        return {
+            "past_key_values": past_key_values,
+            "attention_mask": extended_attention_mask,
+            "position_ids": position_ids,
+        }
+
+    def update_inputs_for_generation(self, model_outputs, model_kwargs):
+        model_kwargs["past_key_values"] = model_outputs.past_key_values
+        model_kwargs["position_ids"] = model_kwargs["position_ids"][:, -1:] + 1
+        return model_kwargs
+
+
+append_call_sample_docstring(
+    Flax{{cookiecutter.camelcase_modelname}}ForCausalLM,
+    _TOKENIZER_FOR_DOC,
+    _CHECKPOINT_FOR_DOC,
+    FlaxCausalLMOutputWithCrossAttentions,
+    _CONFIG_FOR_DOC,
+)
 {# encoder_decoder #}
 {% else %}
 import math

tests/bert/test_modeling_flax_bert.py

@@ -19,7 +19,7 @@ import numpy as np
 from transformers import BertConfig, is_flax_available
 from transformers.testing_utils import require_flax, slow
 
-from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random_attention_mask
+from ..test_modeling_flax_common import FlaxModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
 
 
 if is_flax_available():
@@ -114,6 +114,22 @@ class FlaxBertModelTester(unittest.TestCase):
         inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": attention_mask}
         return config, inputs_dict
 
+    def prepare_config_and_inputs_for_decoder(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, token_type_ids, attention_mask = config_and_inputs
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
 
 @require_flax
 class FlaxBertModelTest(FlaxModelTesterMixin, unittest.TestCase):

tests/big_bird/test_modeling_flax_big_bird.py

@@ -25,6 +25,7 @@ from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random
 if is_flax_available():
     import jax
     from transformers.models.big_bird.modeling_flax_big_bird import (
+        FlaxBigBirdForCausalLM,
         FlaxBigBirdForMaskedLM,
         FlaxBigBirdForMultipleChoice,
         FlaxBigBirdForPreTraining,
@@ -136,6 +137,7 @@ class FlaxBigBirdModelTest(FlaxModelTesterMixin, unittest.TestCase):
 
     all_model_classes = (
         (
+            FlaxBigBirdForCausalLM,
             FlaxBigBirdModel,
             FlaxBigBirdForPreTraining,
             FlaxBigBirdForMaskedLM,

tests/electra/test_modeling_flax_electra.py

@@ -10,6 +10,7 @@ from ..test_modeling_flax_common import FlaxModelTesterMixin, ids_tensor, random
 
 if is_flax_available():
     from transformers.models.electra.modeling_flax_electra import (
+        FlaxElectraForCausalLM,
         FlaxElectraForMaskedLM,
         FlaxElectraForMultipleChoice,
         FlaxElectraForPreTraining,
@@ -110,6 +111,7 @@ class FlaxElectraModelTest(FlaxModelTesterMixin, unittest.TestCase):
     all_model_classes = (
         (
             FlaxElectraModel,
+            FlaxElectraForCausalLM,
             FlaxElectraForMaskedLM,
             FlaxElectraForPreTraining,
             FlaxElectraForTokenClassification,