try load model from hub

model.py

@@ -9,11 +9,13 @@ from transformers import GPT2Tokenizer
 current_path = os.path.dirname(os.path.abspath(__file__))
 sys.path.append(current_path)
 
-# Main model -  ViTGPT2LM
-# from vit_gpt2.modeling_flax_vit_gpt2_lm import FlaxViTGPT2LMForConditionalGeneration
+Main model -  ViTGPT2LM
+from vit_gpt2.modeling_flax_vit_gpt2_lm import FlaxViTGPT2LMForConditionalGeneration
 
-def predict(image):
+model_name_or_path = 'flax-community/vit-gpt2/checkpoints/ckpt_5/'
+flax_vit_gpt2_lm = FlaxViTGPT2LMForConditionalGeneration.from_pretrained(model_name_or_path)
 
+def predict(image):
     return 'dummy caption!', ['dummy', 'caption', '!'], [1, 2, 3]
 
 

vit_gpt2/configuration_vit_gpt2.py

@@ -0,0 +1,45 @@
+import copy
+
+from transformers import GPT2Config, ViTConfig
+from transformers.configuration_utils import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class ViTGPT2Config(PretrainedConfig):
+
+    model_type = "vit-gpt2"
+    is_composition = True
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+
+        if "vit_config" not in kwargs:
+            raise ValueError("`vit_config` can not be `None`.")
+
+        if "gpt2_config" not in kwargs:
+            raise ValueError("`gpt2_config` can not be `None`.")
+
+        vit_config = kwargs.pop("vit_config")
+        gpt2_config = kwargs.pop("gpt2_config")
+
+        self.vit_config = ViTConfig(**vit_config)
+        self.gpt2_config = GPT2Config(**gpt2_config)
+
+    @classmethod
+    def from_vit_gpt2_configs(
+        cls, vit_config: PretrainedConfig, gpt2_config: PretrainedConfig, **kwargs
+    ):
+        return cls(
+            vit_config=vit_config.to_dict(),
+            gpt2_config=gpt2_config.to_dict(),
+            **kwargs
+        )
+
+    def to_dict(self):
+        output = copy.deepcopy(self.__dict__)
+        output["vit_config"] = self.vit_config.to_dict()
+        output["gpt2_config"] = self.gpt2_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

vit_gpt2/modeling_flax_gpt2.py

@@ -0,0 +1,752 @@
+# coding=utf-8
+# Copyright 2021 The Google Flax Team Authors and The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Any, Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from flax.linen import combine_masks, make_causal_mask
+from flax.linen.attention import dot_product_attention_weights
+from jax import lax
+
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_model_forward
+from transformers.modeling_flax_outputs import FlaxBaseModelOutput, FlaxBaseModelOutputWithPast, FlaxCausalLMOutput, FlaxBaseModelOutputWithPastAndCrossAttentions, FlaxSeq2SeqLMOutput
+from transformers.modeling_flax_utils import ACT2FN, FlaxPreTrainedModel, append_call_sample_docstring
+from transformers.utils import logging
+from transformers.models.gpt2.configuration_gpt2 import GPT2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "gpt2"
+_CONFIG_FOR_DOC = "GPT2Config"
+_TOKENIZER_FOR_DOC = "GPT2Tokenizer"
+
+
+GPT2_START_DOCSTRING = r"""
+
+    This model inherits from :class:`~transformers.FlaxPreTrainedModel`. Check the superclass documentation for the
+    generic methods the library implements for all its model (such as downloading or saving, resizing the input
+    embeddings, pruning heads etc.)
+
+    This model is also a Flax Linen `flax.nn.Module
+    <https://flax.readthedocs.io/en/latest/_autosummary/flax.nn.module.html>`__ subclass. Use it as a regular Flax
+    Module and refer to the Flax documentation for all matter related to general usage and behavior.
+
+    Finally, this model supports inherent JAX features such as:
+
+    - `Just-In-Time (JIT) compilation <https://jax.readthedocs.io/en/latest/jax.html#just-in-time-compilation-jit>`__
+    - `Automatic Differentiation <https://jax.readthedocs.io/en/latest/jax.html#automatic-differentiation>`__
+    - `Vectorization <https://jax.readthedocs.io/en/latest/jax.html#vectorization-vmap>`__
+    - `Parallelization <https://jax.readthedocs.io/en/latest/jax.html#parallelization-pmap>`__
+
+    Parameters:
+        config (:class:`~transformers.GPT2Config`): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the :meth:`~transformers.FlaxPreTrainedModel.from_pretrained` method to load the
+            model weights.
+"""
+
+GPT2_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`numpy.ndarray` of shape :obj:`(batch_size, input_ids_length)`):
+            :obj:`input_ids_length` = ``sequence_length``. Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`~transformers.GPT2Tokenizer`. See
+            :meth:`transformers.PreTrainedTokenizer.encode` and :meth:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`numpy.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on padding token indices. Mask values selected in ``[0, 1]``:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            `What are attention masks? <../glossary.html#attention-mask>`__
+        position_ids (:obj:`numpy.ndarray` of shape :obj:`(batch_size, sequence_length)`, `optional`):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range ``[0,
+            config.max_position_embeddings - 1]``.
+        past_key_values (:obj:`Dict[str, np.ndarray]`, `optional`, returned by ``init_cache`` or when passing previous ``past_key_values``):
+            Dictionary of pre-computed hidden-states (key and values in the attention blocks) that can be used for fast
+            auto-regressive decoding. Pre-computed key and value hidden-states are of shape `[batch_size, max_length]`.
+        output_attentions (:obj:`bool`, `optional`):
+            Whether or not to return the attentions tensors of all attention layers. See ``attentions`` under returned
+            tensors for more detail.
+        output_hidden_states (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+
+class FlaxConv1D(nn.Module):
+    features: int
+    use_bias: bool = True
+    dtype: Any = jnp.float32
+    precision: Any = None
+
+    @nn.compact
+    def __call__(self, inputs):
+        inputs = jnp.asarray(inputs, self.dtype)
+        kernel = self.param("kernel", jax.nn.initializers.normal(stddev=0.02), (self.features, inputs.shape[-1]))
+        kernel = jnp.asarray(kernel.transpose(), self.dtype)
+        y = lax.dot_general(inputs, kernel, (((inputs.ndim - 1,), (0,)), ((), ())), precision=self.precision)
+        if self.use_bias:
+            bias = self.param("bias", jax.nn.initializers.zeros, (self.features,))
+            bias = jnp.asarray(bias, self.dtype)
+            y = y + bias
+        return y
+
+
+class FlaxGPT2Attention(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+    causal: bool = True
+
+    def setup(self):
+        config = self.config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+
+        self.c_attn = FlaxConv1D(features=3 * self.embed_dim, dtype=self.dtype)
+        self.c_proj = FlaxConv1D(self.embed_dim, dtype=self.dtype)
+
+        self.c_attn_for_k_v = FlaxConv1D(features=2 * self.embed_dim, dtype=self.dtype)
+
+        self.resid_dropout = nn.Dropout(rate=config.resid_pdrop)
+
+        if self.causal:
+            self.causal_mask = make_causal_mask(jnp.ones((1, config.max_position_embeddings), dtype="bool"), dtype="bool")
+
+    def _split_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.num_heads, self.head_dim))
+
+    def _merge_heads(self, hidden_states):
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
+
+    @nn.compact
+    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,
+        key_value_states: Optional[jnp.ndarray] = None,
+        attention_mask=None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+
+        # 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
+
+        qkv_out = self.c_attn(hidden_states)
+        query, key, value = jnp.split(qkv_out, 3, axis=2)
+
+        if is_cross_attention:
+            _qkv_out = self.c_attn_for_k_v(key_value_states)
+            key, value = jnp.split(_qkv_out, 2, axis=2)
+
+        query = self._split_heads(query)
+        key = self._split_heads(key)
+        value = self._split_heads(value)
+
+        query_length, key_length = query.shape[1], key.shape[1]
+
+        if self.causal:
+            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)
+                )
+            else:
+                causal_mask = self.causal_mask[:, :, :query_length, :key_length]
+
+            batch_size = hidden_states.shape[0]
+            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))
+
+        dropout_rng = None
+        if not deterministic and self.config.attn_pdrop > 0.0:
+            dropout_rng = self.make_rng("dropout")
+
+        # 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, value, attention_mask = self._concatenate_to_cache(key, value, query, attention_mask)
+
+        # transform boolean mask into float mask
+        if attention_mask is not None:
+            attention_bias = lax.select(
+                attention_mask > 0,
+                jnp.full(attention_mask.shape, 0.0).astype(self.dtype),
+                jnp.full(attention_mask.shape, -1e4).astype(self.dtype),
+            )
+        else:
+            attention_bias = None
+
+        # usual dot product attention
+        attn_weights = dot_product_attention_weights(
+            query,
+            key,
+            bias=attention_bias,
+            dropout_rng=dropout_rng,
+            dropout_rate=self.config.attn_pdrop,
+            deterministic=deterministic,
+            dtype=self.dtype,
+            precision=None,
+        )
+
+        attn_output = jnp.einsum("...hqk,...khd->...qhd", attn_weights, value)
+        attn_output = self._merge_heads(attn_output)
+        attn_output = self.c_proj(attn_output)
+        attn_output = self.resid_dropout(attn_output, deterministic=deterministic)
+
+        outputs = (attn_output, attn_weights) if output_attentions else (attn_output,)
+        return outputs
+
+
+class FlaxGPT2MLP(nn.Module):
+    config: GPT2Config
+    intermediate_size: int
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        embed_dim = self.config.hidden_size
+        self.c_fc = FlaxConv1D(self.intermediate_size, dtype=self.dtype)
+        self.c_proj = FlaxConv1D(embed_dim, dtype=self.dtype)
+        self.act = ACT2FN[self.config.activation_function]
+        self.dropout = nn.Dropout(rate=self.config.resid_pdrop)
+
+    def __call__(self, hidden_states, deterministic: bool = True):
+        hidden_states = self.c_fc(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.c_proj(hidden_states)
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+        return hidden_states
+
+
+class FlaxGPT2Block(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        hidden_size = self.config.hidden_size
+        inner_dim = self.config.n_inner if self.config.n_inner is not None else 4 * hidden_size
+
+        self.ln_1 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+        self.attn = FlaxGPT2Attention(self.config, dtype=self.dtype)
+        self.ln_3 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+        self.encoder_attn = FlaxGPT2Attention(config=self.config, dtype=self.dtype)
+        self.ln_2 = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+        self.mlp = FlaxGPT2MLP(self.config, inner_dim, dtype=self.dtype)
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+    ):
+        residual = hidden_states
+        hidden_states = self.ln_1(hidden_states)
+        outputs = self.attn(
+            hidden_states,
+            attention_mask=attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+        )
+        # residual connection
+        attn_output = outputs[0]
+        hidden_states = attn_output + residual
+
+        # Cross-Attention Block
+        if encoder_hidden_states is not None:
+
+            residual = hidden_states
+            hidden_states = self.ln_3(hidden_states)
+
+            cross_attn_outputs = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                deterministic=deterministic,
+                output_attentions=output_attentions,
+            )
+
+            # residual connection
+            cross_attn_output = cross_attn_outputs[0]
+            hidden_states = cross_attn_output + residual
+
+        residual = hidden_states
+        hidden_states = self.ln_2(hidden_states)
+        feed_forward_hidden_states = self.mlp(hidden_states, deterministic=deterministic)
+        # residual connection
+        hidden_states = residual + feed_forward_hidden_states
+
+        output = (hidden_states,) + outputs[1:]
+        if encoder_hidden_states is not None:
+            output = output + cross_attn_outputs[1:]
+
+        return output
+
+
+class FlaxGPT2PreTrainedModel(FlaxPreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = GPT2Config
+    base_model_prefix = "transformer"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: GPT2Config,
+        input_shape: Tuple = (1, 1),
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs,
+    ):
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(config, module, input_shape=input_shape, seed=seed, dtype=dtype)
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        input_ids = jnp.zeros(input_shape, dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(jnp.arange(jnp.atleast_2d(input_ids).shape[-1]), input_shape)
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        if self.config.add_cross_attention:
+            encoder_hidden_states = jnp.zeros(input_shape + (self.config.n_embd,))
+            encoder_attention_mask = attention_mask
+            module_init_outputs = self.module.init(rngs, input_ids, attention_mask, position_ids, encoder_hidden_states, encoder_attention_mask, return_dict=False)
+        else:
+            module_init_outputs = self.module.init(rngs, input_ids, attention_mask, position_ids, return_dict=False)
+
+        return module_init_outputs["params"]
+
+    @classmethod
+    def _from_config(cls, config, **kwargs):
+        return super()._from_config(config, **kwargs)
+
+    def init_cache(self, batch_size, max_length):
+        r"""
+        Args:
+            batch_size (:obj:`int`):
+                batch_size used for fast auto-regressive decoding. Defines the batch size of the initialized cache.
+            max_length (:obj:`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 init_variables["cache"]
+
+    @add_start_docstrings_to_model_forward(GPT2_INPUTS_DOCSTRING)
+    def __call__(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        params: dict = None,
+        past_key_values: 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,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.return_dict
+
+        if encoder_hidden_states is not None and encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = input_ids.shape
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError("Make sure to provide `position_ids` when passing `past_key_values`.")
+
+            position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that it can be changed by FlaxGPT2Attention 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(position_ids, dtype="i4"),
+            encoder_hidden_states,
+            encoder_attention_mask,
+            not train,
+            False,
+            output_attentions,
+            output_hidden_states,
+            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:]
+
+        return outputs
+
+
+class FlaxGPT2BlockCollection(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.blocks = [
+            FlaxGPT2Block(self.config, name=str(i), dtype=self.dtype) for i in range(self.config.num_hidden_layers)
+        ]
+
+    def __call__(
+        self,
+        hidden_states,
+        attention_mask=None,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        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
+
+        for block in self.blocks:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            layer_outputs = block(
+                hidden_states,
+                attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_attention_mask=encoder_attention_mask,
+                deterministic=deterministic,
+                init_cache=init_cache,
+                output_attentions=output_attentions,
+            )
+            hidden_states = layer_outputs[0]
+
+            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,)
+
+        outputs = [hidden_states, all_hidden_states, all_attentions, all_cross_attentions]
+
+        if not return_dict:
+            return tuple(v for v in outputs if v is not None)
+
+        if encoder_hidden_states is None:
+            return FlaxBaseModelOutputWithPast(
+                last_hidden_state=hidden_states,
+                past_key_values=None,
+                hidden_states=all_hidden_states,
+                attentions=all_attentions,
+            )
+        else:
+            return FlaxBaseModelOutputWithPastAndCrossAttentions(
+                last_hidden_state=hidden_states,
+                past_key_values=None,
+                hidden_states=all_hidden_states,
+                attentions=all_attentions,
+                cross_attentions=all_cross_attentions,
+            )
+
+class FlaxGPT2Module(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.embed_dim = self.config.hidden_size
+
+        self.wte = nn.Embed(
+            self.config.vocab_size,
+            self.embed_dim,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.wpe = nn.Embed(
+            self.config.max_position_embeddings,
+            self.embed_dim,
+            embedding_init=jax.nn.initializers.normal(stddev=self.config.initializer_range),
+            dtype=self.dtype,
+        )
+        self.dropout = nn.Dropout(rate=self.config.embd_pdrop)
+        self.h = FlaxGPT2BlockCollection(self.config, dtype=self.dtype)
+        self.ln_f = nn.LayerNorm(epsilon=self.config.layer_norm_epsilon, dtype=self.dtype)
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic=True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        input_embeds = self.wte(input_ids.astype("i4"))
+        position_embeds = self.wpe(position_ids.astype("i4"))
+
+        hidden_states = input_embeds + position_embeds
+        hidden_states = self.dropout(hidden_states, deterministic=deterministic)
+
+        outputs = self.h(
+            hidden_states,
+            attention_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+        hidden_states = self.ln_f(hidden_states)
+
+        if not return_dict:
+            return (hidden_states,) + outputs[1:]
+
+        if encoder_hidden_states is None:
+            return FlaxBaseModelOutput(
+                last_hidden_state=hidden_states,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )
+        else:
+            return FlaxBaseModelOutputWithPastAndCrossAttentions(
+                last_hidden_state=hidden_states,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+                cross_attentions=outputs.cross_attentions,
+            )
+
+@add_start_docstrings(
+    "The bare GPT2 Model transformer outputting raw hidden-states without any specific head on top.",
+    GPT2_START_DOCSTRING,
+)
+class FlaxGPT2Model(FlaxGPT2PreTrainedModel):
+    module_class = FlaxGPT2Module
+
+
+append_call_sample_docstring(
+    FlaxGPT2Model, _TOKENIZER_FOR_DOC, _CHECKPOINT_FOR_DOC, FlaxBaseModelOutput, _CONFIG_FOR_DOC
+)
+
+
+class FlaxGPT2LMHeadModule(nn.Module):
+    config: GPT2Config
+    dtype: jnp.dtype = jnp.float32
+
+    def setup(self):
+        self.transformer = FlaxGPT2Module(self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.config.vocab_size,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(stddev=self.config.initializer_range, dtype=self.dtype),
+        )
+
+    def __call__(
+        self,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_hidden_states: Optional[jnp.ndarray] = None,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        deterministic: bool = True,
+        init_cache: bool = False,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+    ):
+        outputs = self.transformer(
+            input_ids,
+            attention_mask,
+            position_ids,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            deterministic=deterministic,
+            init_cache=init_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        hidden_states = outputs[0]
+
+        if self.config.tie_word_embeddings:
+            shared_kernel = self.transformer.variables["params"]["wte"]["embedding"].T
+            lm_logits = self.lm_head.apply({"params": {"kernel": shared_kernel}}, hidden_states)
+        else:
+            lm_logits = self.lm_head(hidden_states)
+
+        if not return_dict:
+            return (lm_logits,) + outputs[1:]
+
+        if encoder_hidden_states is None:
+            return FlaxCausalLMOutput(logits=lm_logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions)
+        else:
+            return FlaxSeq2SeqLMOutput(
+                logits=lm_logits,
+                decoder_hidden_states=outputs.hidden_states,
+                decoder_attentions=outputs.attentions,
+                cross_attentions=outputs.cross_attentions,
+                encoder_last_hidden_state=encoder_hidden_states,
+                encoder_hidden_states=None,
+                encoder_attentions=None,
+            )
+
+@add_start_docstrings(
+    """
+    The GPT2 Model transformer with a language modeling head on top (linear layer with weights tied to the input
+    embeddings).
+    """,
+    GPT2_START_DOCSTRING,
+)
+class FlaxGPT2LMHeadModel(FlaxGPT2PreTrainedModel):
+    module_class = FlaxGPT2LMHeadModule
+
+    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 GPT2 uses a causal mask, those positions are masked anyways.
+        # 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(
+    FlaxGPT2LMHeadModel, _TOKENIZER_FOR_DOC, _CHECKPOINT_FOR_DOC, FlaxCausalLMOutput, _CONFIG_FOR_DOC
+)

vit_gpt2/modeling_flax_vit_gpt2.py

@@ -0,0 +1,704 @@
+from typing import Callable, Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from jax import lax
+from jax.random import PRNGKey
+from transformers import GPT2Config, FlaxViTModel, ViTConfig
+from transformers.modeling_flax_outputs import (
+    FlaxCausalLMOutputWithCrossAttentions,
+    FlaxSeq2SeqLMOutput,
+    FlaxSeq2SeqModelOutput,
+)
+from transformers.models.bart.modeling_flax_bart import (
+    shift_tokens_right,
+)
+from .modeling_flax_gpt2 import (
+    FlaxGPT2Module,
+    FlaxGPT2Model,
+    FlaxPreTrainedModel
+)
+from transformers.models.vit.modeling_flax_vit import FlaxViTModule
+
+from .configuration_vit_gpt2 import ViTGPT2Config
+
+
+class FlaxViTGPT2Module(nn.Module):
+    config: ViTGPT2Config
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+
+        self.encoder = FlaxViTModule(self.config.vit_config, dtype=self.dtype)
+        self.decoder = FlaxGPT2Module(self.config.gpt2_config, dtype=self.dtype)
+
+    def _get_encoder_module(self):
+        return self.encoder
+
+    def _get_decoder_module(self):
+        return self.decoder
+
+    def __call__(
+            self,
+            pixel_values,
+            input_ids,
+            attention_mask,
+            position_ids,
+            encoder_attention_mask: Optional[jnp.ndarray] = None,
+            output_attentions: bool = False,
+            output_hidden_states: bool = False,
+            return_dict: bool = True,
+            deterministic: bool = True,
+    ):
+        encoder_outputs = self.encoder(
+            pixel_values=pixel_values,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        decoder_outputs = self.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=encoder_attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        return FlaxSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+class FlaxViTGPT2ForConditionalGenerationModule(nn.Module):
+    config: ViTGPT2Config
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., jnp.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.model = FlaxViTGPT2Module(config=self.config, dtype=self.dtype)
+        self.lm_head = nn.Dense(
+            self.model.decoder.embed_dim,
+            use_bias=False,
+            dtype=self.dtype,
+            kernel_init=jax.nn.initializers.normal(
+                self.config.gpt2_config.initializer_range, self.dtype
+            ),
+        )
+        self.final_logits_bias = self.param(
+            "final_logits_bias", self.bias_init, (1, self.model.decoder.embed_dim)
+        )
+
+    def _get_encoder_module(self):
+        return self.model.encoder
+
+    def _get_decoder_module(self):
+        return self.model.decoder
+
+    def __call__(
+        self,
+        pixel_values,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        outputs = self.model(
+            pixel_values=pixel_values,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        hidden_states = outputs[0]
+        lm_logits = self.lm_head(hidden_states)
+        lm_logits += self.final_logits_bias
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return output
+
+        return FlaxSeq2SeqLMOutput(
+            logits=lm_logits,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+class FlaxViTGPT2PreTrainedModel(FlaxPreTrainedModel):
+    config_class = ViTGPT2Config
+    base_model_prefix: str = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: ViTGPT2Config,
+        input_shape: Tuple = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs,
+    ):
+        if input_shape is None:
+            input_shape = (
+                (1, config.vit_config.image_size, config.vit_config.image_size, 3),
+                (1, 1),
+            )
+
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(
+            config, module, input_shape=input_shape, seed=seed, dtype=dtype
+        )
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        pixel_values = jax.random.normal(rng, input_shape[0])
+        # # make sure initialization pass will work for FlaxBartForSequenceClassificationModule
+        # input_ids = jax.ops.index_update(input_ids, (..., -1), self.config.eos_token_id)
+
+        input_ids = jnp.zeros(input_shape[1], dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+
+        batch_size, sequence_length = input_ids.shape
+        position_ids = jnp.broadcast_to(
+            jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+        )
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        return self.module.init(
+            rngs,
+            pixel_values,
+            input_ids,
+            attention_mask,
+            position_ids,
+        )["params"]
+
+    def init_cache(self, batch_size, max_length, encoder_outputs):
+
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(
+            jnp.arange(jnp.atleast_2d(input_ids).shape[-1]),
+            input_ids.shape,
+        )
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    def encode(
+        self,
+        pixel_values: jnp.ndarray,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, pixel_values, **kwargs):
+            encode_module = module._get_encoder_module()
+            return encode_module(pixel_values, **kwargs)
+
+        return self.module.apply(
+            {"params": params or self.params},
+            pixel_values=jnp.array(pixel_values, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    def decode(
+        self,
+        input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = input_ids.shape
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError(
+                    "Make sure to provide `position_ids` when passing `past_key_values`."
+                )
+
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+
+        outputs = self.module.apply(
+            inputs,
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past = outputs
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past = outputs
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def __call__(
+        self,
+        pixel_values: jnp.ndarray,
+        input_ids: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # # prepare encoder inputs
+        # if encoder_attention_mask is None:
+        #     encoder_attention_mask = jnp.ones_like(input_ids)
+
+        # if position_ids is None:
+        #     batch_size, sequence_length = input_ids.shape
+        #     position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        # prepare decoder inputs
+        # if decoder_input_ids is None:
+        #     decoder_input_ids = shift_tokens_right(
+        #         input_ids, self.config.pad_token_id, decoder_start_token_id=self.config.decoder_start_token_id
+        #     ) # TODO: Check how to use this
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+        if position_ids is None:
+            batch_size, sequence_length = input_ids.shape
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            pixel_values=jnp.array(pixel_values, dtype=jnp.float32),
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+
+class FlaxViTGPT2ForConditionalGeneration(FlaxViTGPT2PreTrainedModel):
+    module_class = FlaxViTGPT2ForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    def decode(
+        self,
+        input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        deterministic: bool = True,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = input_ids.shape
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError(
+                    "Make sure to provide `position_ids` when passing `past_key_values`."
+                )
+
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            outputs = decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+            hidden_states = outputs[0]
+
+            if self.config.tie_word_embeddings:
+                shared_embedding = module.model.variables["params"]["shared"][
+                    "embedding"
+                ]
+                lm_logits = module.lm_head.apply(
+                    {"params": {"kernel": shared_embedding.T}}, hidden_states
+                )
+            else:
+                lm_logits = module.lm_head(hidden_states)
+
+            lm_logits += module.final_logits_bias
+            return lm_logits, outputs
+
+        outputs = self.module.apply(
+            inputs,
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, outputs = outputs
+        else:
+            (lm_logits, outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+                cross_attentions=outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        max_length,
+        encoder_attention_mask: Optional[jnp.DeviceArray] = None,
+        attention_mask: Optional[jnp.DeviceArray] = None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
+        # 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 anyways.
+        # 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,
+            "encoder_outputs": encoder_outputs,
+            "encoder_attention_mask": encoder_attention_mask,
+            "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
+
+    @classmethod
+    def from_vit_gpt2_pretrained(
+        cls,
+        vit_model_name_or_path: str = None,
+        gpt2_model_name_or_path: str = None,
+        *model_args,
+        **kwargs,
+    ) -> FlaxViTGPT2PreTrainedModel:
+
+        kwargs_gpt2 = {
+            argument[len("gpt2_") :]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("gpt2_")
+        }
+
+        kwargs_vit = {
+            argument[len("vit_") :]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("vit_")
+        }
+
+        # remove gpt2, vit kwargs from kwargs
+        for key in kwargs_gpt2.keys():
+            del kwargs["gpt2_" + key]
+        for key in kwargs_vit.keys():
+            del kwargs["vit_" + key]
+
+        # Load and initialize the gpt2 and vit model
+        gpt2_model = kwargs_gpt2.pop("model", None)
+        if gpt2_model is None:
+            assert (
+                gpt2_model_name_or_path is not None
+            ), "If `model` is not defined as an argument, a `gpt2_model_name_or_path` has to be defined"
+
+            if "config" not in kwargs_gpt2:
+                gpt2_config = GPT2Config.from_pretrained(gpt2_model_name_or_path)
+                kwargs_gpt2["config"] = gpt2_config
+
+            kwargs_gpt2["config"].add_cross_attention = True
+            gpt2_model = FlaxGPT2Model.from_pretrained(
+                gpt2_model_name_or_path, *model_args, **kwargs_gpt2
+            )
+
+        vit_model = kwargs_vit.pop("model", None)
+        if vit_model is None:
+            assert (
+                vit_model_name_or_path is not None
+            ), "If `model` is not defined as an argument, a `vit_model_name_or_path` has to be defined"
+
+            if "config" not in kwargs_vit:
+                vit_config = ViTConfig.from_pretrained(vit_model_name_or_path)
+                kwargs_vit["config"] = vit_config
+
+            vit_model = FlaxViTModel.from_pretrained(
+                vit_model_name_or_path, *model_args, **kwargs_vit
+            )
+
+        # instantiate config with corresponding kwargs
+        dtype = kwargs.pop("dtype", jnp.float32)
+        config = ViTGPT2Config.from_vit_gpt2_configs(
+            vit_model.config, gpt2_model.config, **kwargs
+        )
+
+        # init model
+        model = cls(config, *model_args, dtype=dtype, **kwargs)
+        model.params["model"]["encoder"] = vit_model.params
+        model.params["model"]["decoder"] = gpt2_model.params
+
+        return model
+

vit_gpt2/modeling_flax_vit_gpt2_lm.py

@@ -0,0 +1,684 @@
+from typing import Callable, Optional, Tuple
+
+import flax.linen as nn
+import jax
+import jax.numpy as jnp
+from flax.core.frozen_dict import FrozenDict, unfreeze
+from jax import lax
+from jax.random import PRNGKey
+from transformers import GPT2Config, FlaxViTModel, ViTConfig
+from transformers.modeling_flax_outputs import (
+    FlaxCausalLMOutputWithCrossAttentions,
+    FlaxSeq2SeqLMOutput,
+    FlaxSeq2SeqModelOutput,
+)
+from transformers.models.bart.modeling_flax_bart import (
+    shift_tokens_right,
+)
+from .modeling_flax_gpt2 import (
+    FlaxGPT2Module,
+    FlaxGPT2Model,
+    FlaxGPT2LMHeadModule,
+    FlaxGPT2LMHeadModel,
+    FlaxPreTrainedModel
+)
+from transformers.models.vit.modeling_flax_vit import FlaxViTModule
+
+from .configuration_vit_gpt2 import ViTGPT2Config
+
+
+def shift_tokens_right(input_ids: jnp.ndarray, pad_token_id: int, decoder_start_token_id: int) -> jnp.ndarray:
+    """
+    Shift input ids one token to the right.
+    """
+    shifted_input_ids = jnp.roll(input_ids, 1, axis=-1)
+    shifted_input_ids = jax.ops.index_update(shifted_input_ids, (..., 0), decoder_start_token_id)
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids = jnp.where(shifted_input_ids == -100, pad_token_id, shifted_input_ids)
+
+    return shifted_input_ids
+
+class FlaxViTGPT2LMModule(nn.Module):
+    config: ViTGPT2Config
+    dtype: jnp.dtype = jnp.float32  # the dtype of the computation
+
+    def setup(self):
+
+        self.encoder = FlaxViTModule(self.config.vit_config, dtype=self.dtype)
+        self.decoder = FlaxGPT2LMHeadModule(self.config.gpt2_config, dtype=self.dtype)
+
+    def _get_encoder_module(self):
+        return self.encoder
+
+    def _get_decoder_module(self):
+        return self.decoder
+
+    def __call__(
+            self,
+            pixel_values,
+            input_ids,
+            attention_mask,
+            position_ids,
+            encoder_attention_mask: Optional[jnp.ndarray] = None,
+            output_attentions: bool = False,
+            output_hidden_states: bool = False,
+            return_dict: bool = True,
+            deterministic: bool = True,
+    ):
+        encoder_outputs = self.encoder(
+            pixel_values=pixel_values,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        decoder_outputs = self.decoder(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            encoder_attention_mask=encoder_attention_mask,
+            deterministic=deterministic,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return FlaxSeq2SeqLMOutput(
+            logits=decoder_outputs.logits,
+            decoder_hidden_states=decoder_outputs.decoder_hidden_states,
+            decoder_attentions=decoder_outputs.decoder_attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+class FlaxViTGPT2LMForConditionalGenerationModule(nn.Module):
+    config: ViTGPT2Config
+    dtype: jnp.dtype = jnp.float32
+    bias_init: Callable[..., jnp.ndarray] = jax.nn.initializers.zeros
+
+    def setup(self):
+        self.model = FlaxViTGPT2LMModule(config=self.config, dtype=self.dtype)
+
+    def _get_encoder_module(self):
+        return self.model.encoder
+
+    def _get_decoder_module(self):
+        return self.model.decoder
+
+    def __call__(
+        self,
+        pixel_values,
+        input_ids,
+        attention_mask,
+        position_ids,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: bool = False,
+        output_hidden_states: bool = False,
+        return_dict: bool = True,
+        deterministic: bool = True,
+    ):
+        outputs = self.model(
+            pixel_values=pixel_values,
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+        )
+
+        return outputs
+
+
+class FlaxViTGPT2LMPreTrainedModel(FlaxPreTrainedModel):
+    config_class = ViTGPT2Config
+    base_model_prefix: str = "model"
+    module_class: nn.Module = None
+
+    def __init__(
+        self,
+        config: ViTGPT2Config,
+        input_shape: Tuple = None,
+        seed: int = 0,
+        dtype: jnp.dtype = jnp.float32,
+        **kwargs,
+    ):
+        if input_shape is None:
+            input_shape = (
+                (1, config.vit_config.image_size, config.vit_config.image_size, 3),
+                (1, 1),
+            )
+
+        module = self.module_class(config=config, dtype=dtype, **kwargs)
+        super().__init__(
+            config, module, input_shape=input_shape, seed=seed, dtype=dtype
+        )
+
+    def init_weights(self, rng: jax.random.PRNGKey, input_shape: Tuple) -> FrozenDict:
+        # init input tensors
+        pixel_values = jax.random.normal(rng, input_shape[0])
+        # # make sure initialization pass will work for FlaxBartForSequenceClassificationModule
+        # input_ids = jax.ops.index_update(input_ids, (..., -1), self.config.eos_token_id)
+
+        input_ids = jnp.zeros(input_shape[1], dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+
+        batch_size, sequence_length = input_ids.shape
+        position_ids = jnp.broadcast_to(
+            jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+        )
+
+        params_rng, dropout_rng = jax.random.split(rng)
+        rngs = {"params": params_rng, "dropout": dropout_rng}
+
+        return self.module.init(
+            rngs,
+            pixel_values,
+            input_ids,
+            attention_mask,
+            position_ids,
+        )["params"]
+
+    def init_cache(self, batch_size, max_length, encoder_outputs):
+
+        input_ids = jnp.ones((batch_size, max_length), dtype="i4")
+        attention_mask = jnp.ones_like(input_ids)
+        position_ids = jnp.broadcast_to(
+            jnp.arange(jnp.atleast_2d(input_ids).shape[-1]),
+            input_ids.shape,
+        )
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+
+        init_variables = self.module.init(
+            jax.random.PRNGKey(0),
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            init_cache=True,
+            method=_decoder_forward,  # we only need to call the decoder to init the cache
+        )
+        return unfreeze(init_variables["cache"])
+
+    def encode(
+        self,
+        pixel_values: jnp.ndarray,
+        attention_mask: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        def _encoder_forward(module, pixel_values, **kwargs):
+            encode_module = module._get_encoder_module()
+            return encode_module(pixel_values, **kwargs)
+
+        return self.module.apply(
+            {"params": params or self.params},
+            pixel_values=jnp.array(pixel_values, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            method=_encoder_forward,
+        )
+
+    def decode(
+        self,
+        input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = input_ids.shape
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError(
+                    "Make sure to provide `position_ids` when passing `past_key_values`."
+                )
+
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            return decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+
+        outputs = self.module.apply(
+            inputs,
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs, past = outputs
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs, past = outputs
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def __call__(
+        self,
+        pixel_values: jnp.ndarray,
+        input_ids: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        train: bool = False,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        pixel_values = jnp.transpose(pixel_values, (0, 2, 3, 1))
+
+        # # prepare encoder inputs
+        # if encoder_attention_mask is None:
+        #     encoder_attention_mask = jnp.ones_like(input_ids)
+
+        # if position_ids is None:
+        #     batch_size, sequence_length = input_ids.shape
+        #     position_ids = jnp.broadcast_to(jnp.arange(sequence_length)[None, :], (batch_size, sequence_length))
+
+        # prepare decoder inputs
+        # if decoder_input_ids is None:
+        #     decoder_input_ids = shift_tokens_right(
+        #         input_ids, self.config.pad_token_id, decoder_start_token_id=self.config.decoder_start_token_id
+        #     ) # TODO: Check how to use this
+
+        if attention_mask is None:
+            attention_mask = jnp.ones_like(input_ids)
+        if position_ids is None:
+            batch_size, sequence_length = input_ids.shape
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {"dropout": dropout_rng} if dropout_rng is not None else {}
+
+        return self.module.apply(
+            {"params": params or self.params},
+            pixel_values=jnp.array(pixel_values, dtype=jnp.float32),
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=not train,
+            rngs=rngs,
+        )
+
+
+class FlaxViTGPT2LMForConditionalGeneration(FlaxViTGPT2LMPreTrainedModel):
+    module_class = FlaxViTGPT2LMForConditionalGenerationModule
+    dtype: jnp.dtype = jnp.float32
+
+    def decode(
+        self,
+        input_ids,
+        encoder_outputs,
+        encoder_attention_mask: Optional[jnp.ndarray] = None,
+        attention_mask: Optional[jnp.ndarray] = None,
+        position_ids: Optional[jnp.ndarray] = None,
+        past_key_values: dict = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        deterministic: bool = True,
+        params: dict = None,
+        dropout_rng: PRNGKey = None,
+    ):
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+        return_dict = (
+            return_dict if return_dict is not None else self.config.return_dict
+        )
+
+        encoder_hidden_states = encoder_outputs[0]
+        if encoder_attention_mask is None:
+            batch_size, sequence_length = encoder_hidden_states.shape[:2]
+            encoder_attention_mask = jnp.ones((batch_size, sequence_length))
+
+        batch_size, sequence_length = input_ids.shape
+        if attention_mask is None:
+            attention_mask = jnp.ones((batch_size, sequence_length))
+
+        if position_ids is None:
+            if past_key_values is not None:
+                raise ValueError(
+                    "Make sure to provide `position_ids` when passing `past_key_values`."
+                )
+
+            position_ids = jnp.broadcast_to(
+                jnp.arange(sequence_length)[None, :], (batch_size, sequence_length)
+            )
+
+        # Handle any PRNG if needed
+        rngs = {}
+        if dropout_rng is not None:
+            rngs["dropout"] = dropout_rng
+
+        inputs = {"params": params or self.params}
+
+        # if past_key_values are passed then cache is already initialized a private flag init_cache has to be
+        # passed down to ensure cache is used. It has to be made sure that cache is marked as mutable so that
+        # it can be changed by FlaxGPT2Attention module
+        if past_key_values:
+            inputs["cache"] = past_key_values
+            mutable = ["cache"]
+        else:
+            mutable = False
+
+        def _decoder_forward(
+            module,
+            input_ids,
+            attention_mask,
+            position_ids,
+            **kwargs,
+        ):
+            decoder_module = module._get_decoder_module()
+            outputs = decoder_module(
+                input_ids,
+                attention_mask,
+                position_ids,
+                **kwargs,
+            )
+            lm_logits = outputs[0]
+
+            return lm_logits, outputs
+
+        outputs = self.module.apply(
+            inputs,
+            input_ids=jnp.array(input_ids, dtype="i4"),
+            attention_mask=jnp.array(attention_mask, dtype="i4"),
+            position_ids=jnp.array(position_ids, dtype="i4"),
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=jnp.array(encoder_attention_mask, dtype="i4"),
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            deterministic=deterministic,
+            rngs=rngs,
+            mutable=mutable,
+            method=_decoder_forward,
+        )
+
+        if past_key_values is None:
+            lm_logits, outputs = outputs
+        else:
+            (lm_logits, outputs), past = outputs
+
+        if return_dict:
+            outputs = FlaxCausalLMOutputWithCrossAttentions(
+                logits=lm_logits,
+                hidden_states=outputs.decoder_hidden_states,
+                attentions=outputs.decoder_attentions,
+                cross_attentions=outputs.cross_attentions,
+            )
+        else:
+            outputs = (lm_logits,) + outputs[1:]
+
+        # add updated cache to model output
+        if past_key_values is not None and return_dict:
+            outputs["past_key_values"] = unfreeze(past["cache"])
+            return outputs
+        elif past_key_values is not None and not return_dict:
+            outputs = outputs[:1] + (unfreeze(past["cache"]),) + outputs[1:]
+
+        return outputs
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        max_length,
+        encoder_attention_mask: Optional[jnp.DeviceArray] = None,
+        attention_mask: Optional[jnp.DeviceArray] = None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # initializing the cache
+        batch_size, seq_length = input_ids.shape
+
+        past_key_values = self.init_cache(batch_size, max_length, encoder_outputs)
+        # 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 anyways.
+        # 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,
+            "encoder_outputs": encoder_outputs,
+            "encoder_attention_mask": encoder_attention_mask,
+            "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
+
+    @classmethod
+    def from_vit_gpt2_pretrained(
+        cls,
+        vit_model_name_or_path: str = None,
+        gpt2_model_name_or_path: str = None,
+        *model_args,
+        **kwargs,
+    ) -> FlaxViTGPT2LMPreTrainedModel:
+
+        kwargs_gpt2 = {
+            argument[len("gpt2_") :]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("gpt2_")
+        }
+
+        kwargs_vit = {
+            argument[len("vit_") :]: value
+            for argument, value in kwargs.items()
+            if argument.startswith("vit_")
+        }
+
+        # remove gpt2, vit kwargs from kwargs
+        for key in kwargs_gpt2.keys():
+            del kwargs["gpt2_" + key]
+        for key in kwargs_vit.keys():
+            del kwargs["vit_" + key]
+
+        # Load and initialize the gpt2 and vit model
+        gpt2_model = kwargs_gpt2.pop("model", None)
+        if gpt2_model is None:
+            assert (
+                gpt2_model_name_or_path is not None
+            ), "If `model` is not defined as an argument, a `gpt2_model_name_or_path` has to be defined"
+
+            if "config" not in kwargs_gpt2:
+                gpt2_config = GPT2Config.from_pretrained(gpt2_model_name_or_path)
+                kwargs_gpt2["config"] = gpt2_config
+
+            kwargs_gpt2["config"].add_cross_attention = True
+            gpt2_model = FlaxGPT2LMHeadModel.from_pretrained(
+                gpt2_model_name_or_path, *model_args, **kwargs_gpt2
+            )
+
+        vit_model = kwargs_vit.pop("model", None)
+        if vit_model is None:
+            assert (
+                vit_model_name_or_path is not None
+            ), "If `model` is not defined as an argument, a `vit_model_name_or_path` has to be defined"
+
+            if "config" not in kwargs_vit:
+                vit_config = ViTConfig.from_pretrained(vit_model_name_or_path)
+                kwargs_vit["config"] = vit_config
+
+            vit_model = FlaxViTModel.from_pretrained(
+                vit_model_name_or_path, *model_args, **kwargs_vit
+            )
+
+        # instantiate config with corresponding kwargs
+        dtype = kwargs.pop("dtype", jnp.float32)
+        config = ViTGPT2Config.from_vit_gpt2_configs(
+            vit_model.config, gpt2_model.config, **kwargs
+        )
+
+        # init model
+        model = cls(config, *model_args, dtype=dtype, **kwargs)
+        model.params["model"]["encoder"] = vit_model.params
+        model.params["model"]["decoder"] = gpt2_model.params
+
+        return model