diff --git "a/m4/models/vt5/modeling_vt5.py" "b/m4/models/vt5/modeling_vt5.py"
new file mode 100644--- /dev/null
+++ "b/m4/models/vt5/modeling_vt5.py"
@@ -0,0 +1,2188 @@
+# coding=utf-8
+# Copyright 2018 Mesh TensorFlow authors, T5 Authors and 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.
+""" PyTorch T5 model."""
+
+
+import copy
+import math
+import os
+import warnings
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss
+from torch.utils.checkpoint import checkpoint
+from transformers.activations import ACT2FN
+from transformers.modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    Seq2SeqLMOutput,
+    Seq2SeqModelOutput,
+)
+from transformers.models.t5.configuration_t5 import T5Config
+from transformers.pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
+from transformers.utils import (
+    DUMMY_INPUTS,
+    DUMMY_MASK,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    is_torch_fx_proxy,
+    replace_return_docstrings,
+)
+from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
+
+from m4.models import DecoupledEmbedding, DecoupledLinear
+from m4.models.custom_modules import VLOOMPreTrainedModelBase
+from m4.models.vt5.configuration_vt5 import VT5Config
+from m4.training.packing import random_spans_helper
+from m4.training.utils import compute_tflops_per_batch_per_gpu, freeze_model
+from m4.utils import logging
+
+
+ALL_LAYERNORM_LAYERS = [nn.LayerNorm]
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "T5Config"
+_TOKENIZER_FOR_DOC = "T5Tokenizer"
+_CHECKPOINT_FOR_DOC = "t5-small"
+
+####################################################
+# This dict contains ids and associated url
+# for the pretrained weights provided with the models
+####################################################
+T5_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "t5-small",
+    "t5-base",
+    "t5-large",
+    "t5-3b",
+    "t5-11b",
+    # See all T5 models at https://huggingface.co/models?filter=t5
+]
+
+
+####################################################
+# This is a conversion method from TF 1.0 to PyTorch
+# More details: https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28
+####################################################
+def load_tf_weights_in_t5(model, config, tf_checkpoint_path):
+    """Load tf checkpoints in a pytorch model."""
+    try:
+        import re
+
+        import numpy as np
+        import tensorflow as tf
+    except ImportError:
+        logger.error(
+            "Loading a TensorFlow model in PyTorch, requires TensorFlow to be installed. Please see "
+            "https://www.tensorflow.org/install/ for installation instructions."
+        )
+        raise
+    tf_path = os.path.abspath(tf_checkpoint_path)
+    logger.info(f"Converting TensorFlow checkpoint from {tf_path}")
+    # Load weights from TF model
+    init_vars = tf.train.list_variables(tf_path)
+    names = []
+    tf_weights = {}
+    for name, shape in init_vars:
+        logger.info(f"Loading TF weight {name} with shape {shape}")
+        array = tf.train.load_variable(tf_path, name)
+        names.append(name)
+        tf_weights[name] = array
+
+    for txt_name in names:
+        name = txt_name.split("/")
+        # adam_v and adam_m are variables used in AdamWeightDecayOptimizer to calculated m and v
+        # which are not required for using pretrained model
+        if any(
+            n in ["adam_v", "adam_m", "AdamWeightDecayOptimizer", "AdamWeightDecayOptimizer_1", "global_step"]
+            for n in name
+        ):
+            logger.info(f"Skipping {'/'.join(name)}")
+            tf_weights.pop(txt_name, None)
+            continue
+        if "_slot_" in name[-1]:
+            logger.info(f"Skipping {'/'.join(name)}")
+            tf_weights.pop(txt_name, None)
+            continue
+        pointer = model
+        array = tf_weights[txt_name]
+
+        for m_name in name:
+            if re.fullmatch(r"[A-Za-z]+_\d+", m_name):
+                scope_names = re.split(r"_(\d+)", m_name)
+            else:
+                scope_names = [m_name]
+            if scope_names[0] in ["kernel", "scale", "embedding"]:
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "self_attention":
+                pointer = getattr(pointer, "layer")
+                pointer = pointer[0]
+            elif scope_names[0] == "enc_dec_attention":
+                pointer = getattr(pointer, "layer")
+                pointer = pointer[1]
+            elif scope_names[0] == "dense_relu_dense":
+                pointer = getattr(pointer, "layer")
+                pointer = pointer[2]
+            elif scope_names[0] == "rms_norm":
+                if hasattr(pointer, "layer_norm"):
+                    pointer = getattr(pointer, "layer_norm")
+                elif hasattr(pointer, "final_layer_norm"):
+                    pointer = getattr(pointer, "final_layer_norm")
+            elif scope_names[0] == "scale":
+                pointer = getattr(pointer, "weight")
+            elif scope_names[0] == "output_bias" or scope_names[0] == "beta":
+                pointer = getattr(pointer, "bias")
+            elif scope_names[0] == "squad":
+                pointer = getattr(pointer, "classifier")
+            elif scope_names[0] == "decoder" and name[1] == "logits":
+                continue
+            elif scope_names[0] == "logits":
+                pointer = getattr(pointer, "lm_head")
+            elif scope_names[0] == "wi" and len(scope_names) > 1 and scope_names[1].isdigit():
+                pointer = getattr(pointer, f"wi_{scope_names[1]}")
+                continue
+            else:
+                try:
+                    pointer = getattr(pointer, scope_names[0])
+                except AttributeError:
+                    logger.info(f"Skipping {'/'.join(name)}")
+                    continue
+            if len(scope_names) >= 2:
+                num = int(scope_names[1])
+                pointer = pointer[num]
+        if scope_names[0] not in ["kernel", "scale", "embedding"]:
+            pointer = getattr(pointer, "weight")
+        if scope_names[0] != "embedding":
+            logger.info(f"Transposing numpy weight of shape {array.shape} for {name}")
+            array = np.transpose(array)
+        try:
+            assert (
+                pointer.shape == array.shape
+            ), f"Pointer shape {pointer.shape} and array shape {array.shape} mismatched"
+        except AssertionError as e:
+            e.args += (pointer.shape, array.shape)
+            raise
+        logger.info(f"Initialize PyTorch weight {name}")
+        pointer.data = torch.from_numpy(array.astype(np.float32))
+        tf_weights.pop(txt_name, None)
+
+    logger.info(f"Weights not copied to PyTorch model: {', '.join(tf_weights.keys())}.")
+    return model
+
+
+####################################################
+# PyTorch Models are constructed by sub-classing
+# - torch.nn.Module for the layers and
+# - PreTrainedModel for the models (it-self a sub-class of nn.Module)
+####################################################
+PARALLELIZE_DOCSTRING = r"""
+    This is an experimental feature and is a subject to change at a moment's notice.
+
+    Uses a device map to distribute attention modules of the model across several devices. If no device map is given,
+    it will evenly distribute blocks across all devices.
+
+    Args:
+        device_map (`Dict[int, list]`, optional, defaults to None):
+            A dictionary that maps attention modules to devices. Note that the embedding module and LMHead are always
+            automatically mapped to the first device (for esoteric reasons). That means that the first device should
+            have fewer attention modules mapped to it than other devices. For reference, the t5 models have the
+            following number of attention modules:
+
+                - t5-small: 6
+                - t5-base: 12
+                - t5-large: 24
+                - t5-3b: 24
+                - t5-11b: 24
+
+    Example:
+
+    ```python
+    # Here is an example of a device map on a machine with 4 GPUs using t5-3b, which has a total of 24 attention modules:
+    model = T5ForConditionalGeneration.from_pretrained("t5-3b")
+    device_map = {
+        0: [0, 1, 2],
+        1: [3, 4, 5, 6, 7, 8, 9],
+        2: [10, 11, 12, 13, 14, 15, 16],
+        3: [17, 18, 19, 20, 21, 22, 23],
+    }
+    model.parallelize(device_map)
+    ```
+"""
+DEPARALLELIZE_DOCSTRING = r"""
+    Moves the model to cpu from a model parallel state.
+
+    Example:
+
+    ```python
+    # On a 4 GPU machine with t5-3b:
+    model = T5ForConditionalGeneration.from_pretrained("t5-3b")
+    device_map = {
+        0: [0, 1, 2],
+        1: [3, 4, 5, 6, 7, 8, 9],
+        2: [10, 11, 12, 13, 14, 15, 16],
+        3: [17, 18, 19, 20, 21, 22, 23],
+    }
+    model.parallelize(device_map)  # Splits the model across several devices
+    model.deparallelize()  # Put the model back on cpu and cleans memory by calling torch.cuda.empty_cache()
+    ```
+"""
+
+
+class T5LayerNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Construct a layernorm module in the T5 style. No bias and no subtraction of mean.
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        # T5 uses a layer_norm which only scales and doesn't shift, which is also known as Root Mean
+        # Square Layer Normalization https://arxiv.org/abs/1910.07467 thus varience is calculated
+        # w/o mean and there is no bias. Additionally we want to make sure that the accumulation for
+        # half-precision inputs is done in fp32
+
+        variance = hidden_states.to(torch.float32).pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        # convert into half-precision if necessary
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            hidden_states = hidden_states.to(self.weight.dtype)
+
+        return self.weight * hidden_states
+
+
+try:
+    from apex.normalization import FusedRMSNorm
+
+    T5LayerNorm = FusedRMSNorm  # noqa
+
+    logger.info("Discovered apex.normalization.FusedRMSNorm - will use it instead of T5LayerNorm")
+except ImportError:
+    # using the normal T5LayerNorm
+    pass
+except Exception:
+    logger.warning("discovered apex but it failed to load, falling back to T5LayerNorm")
+    pass
+
+ALL_LAYERNORM_LAYERS = [nn.LayerNorm, T5LayerNorm]
+
+
+class T5DenseActDense(nn.Module):
+    def __init__(self, config: T5Config):
+        super().__init__()
+        self.wi = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+        self.act = ACT2FN[config.dense_act_fn]
+
+    def forward(self, hidden_states):
+        hidden_states = self.wi(hidden_states)
+        hidden_states = self.act(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.wo(hidden_states)
+        return hidden_states
+
+
+class T5DenseGatedActDense(nn.Module):
+    def __init__(self, config: T5Config):
+        super().__init__()
+        self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
+        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
+        self.dropout = nn.Dropout(config.dropout_rate)
+        self.act = ACT2FN[config.dense_act_fn]
+
+    def forward(self, hidden_states):
+        hidden_gelu = self.act(self.wi_0(hidden_states))
+        hidden_linear = self.wi_1(hidden_states)
+        hidden_states = hidden_gelu * hidden_linear
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.wo(hidden_states)
+        return hidden_states
+
+
+class T5LayerFF(nn.Module):
+    def __init__(self, config: T5Config, is_vision_cross_attention=False):
+        super().__init__()
+        if config.is_gated_act:
+            self.DenseReluDense = T5DenseGatedActDense(config)
+        else:
+            self.DenseReluDense = T5DenseActDense(config)
+
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        self.is_vision_cross_attention = is_vision_cross_attention
+        if is_vision_cross_attention:
+            self.act = nn.Tanh()
+            if config.alpha_initializer == "zeros":
+                if config.alpha_type == "vector":
+                    self.alpha_dense = nn.Parameter(torch.zeros(1, 1, config.d_model))
+                elif config.alpha_type == "float":
+                    self.alpha_dense = nn.Parameter(torch.zeros(1))
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            elif config.alpha_initializer == "ones":
+                if config.alpha_type == "vector":
+                    self.alpha_dense = nn.Parameter(torch.ones(1, 1, config.d_model))
+                elif config.alpha_type == "float":
+                    self.alpha_dense = nn.Parameter(torch.ones(1))
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            elif config.alpha_initializer in {"normal", "gaussian", "random"}:
+                if config.alpha_type == "vector":
+                    self.alpha_dense = nn.Parameter(
+                        torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, config.d_model))
+                    )
+                elif config.alpha_type == "float":
+                    self.alpha_dense = nn.Parameter(
+                        torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1))
+                    )
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            else:
+                raise NotImplementedError(
+                    f"Alpha initialization scheme {config.alpha_initializer} not yet implemented!"
+                )
+
+    def forward(self, hidden_states):
+        forwarded_states = self.layer_norm(hidden_states)
+        forwarded_states = self.DenseReluDense(forwarded_states)
+        if not self.is_vision_cross_attention:
+            hidden_states = hidden_states + self.dropout(forwarded_states)
+        else:
+            hidden_states = hidden_states + self.dropout(self.act(self.alpha_dense) * forwarded_states)
+        return hidden_states
+
+
+class T5Attention(nn.Module):
+    def __init__(self, config: T5Config, has_relative_attention_bias=False, is_vision_cross_attention=False):
+        super().__init__()
+        self.is_decoder = config.is_decoder
+        self.has_relative_attention_bias = has_relative_attention_bias
+        self.relative_attention_num_buckets = config.relative_attention_num_buckets
+        self.relative_attention_max_distance = config.relative_attention_max_distance
+        self.d_model = config.d_model
+        self.key_value_proj_dim = config.d_kv
+        self.n_heads = config.num_heads
+        self.dropout = config.dropout_rate
+        self.inner_dim = self.n_heads * self.key_value_proj_dim
+
+        # Mesh TensorFlow initialization to avoid scaling before softmax
+        self.q = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        if not is_vision_cross_attention:
+            self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
+            self.v = nn.Linear(self.d_model, self.inner_dim, bias=False)
+        else:
+            vision_embed_dim = self.embed_dim if not hasattr(config, "vision_embed_dim") else config.vision_embed_dim
+            self.k = nn.Linear(vision_embed_dim, self.inner_dim, bias=False)
+            self.v = nn.Linear(vision_embed_dim, self.inner_dim, bias=False)
+        self.o = nn.Linear(self.inner_dim, self.d_model, bias=False)
+
+        if self.has_relative_attention_bias:
+            self.relative_attention_bias = nn.Embedding(self.relative_attention_num_buckets, self.n_heads)
+        self.pruned_heads = set()
+        self.gradient_checkpointing = False
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.n_heads, self.key_value_proj_dim, self.pruned_heads
+        )
+        # Prune linear layers
+        self.q = prune_linear_layer(self.q, index)
+        self.k = prune_linear_layer(self.k, index)
+        self.v = prune_linear_layer(self.v, index)
+        self.o = prune_linear_layer(self.o, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.inner_dim = self.key_value_proj_dim * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    @staticmethod
+    def _relative_position_bucket(relative_position, bidirectional=True, num_buckets=32, max_distance=128):
+        """
+        Adapted from Mesh Tensorflow:
+        https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593
+
+        Translate relative position to a bucket number for relative attention. The relative position is defined as
+        memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
+        position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
+        small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
+        positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
+        This should allow for more graceful generalization to longer sequences than the model has been trained on
+
+        Args:
+            relative_position: an int32 Tensor
+            bidirectional: a boolean - whether the attention is bidirectional
+            num_buckets: an integer
+            max_distance: an integer
+
+        Returns:
+            a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
+        """
+        relative_buckets = 0
+        if bidirectional:
+            num_buckets //= 2
+            relative_buckets += (relative_position > 0).to(torch.long) * num_buckets
+            relative_position = torch.abs(relative_position)
+        else:
+            relative_position = -torch.min(relative_position, torch.zeros_like(relative_position))
+        # now relative_position is in the range [0, inf)
+
+        # half of the buckets are for exact increments in positions
+        max_exact = num_buckets // 2
+        is_small = relative_position < max_exact
+
+        # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
+        relative_position_if_large = max_exact + (
+            torch.log(relative_position.float() / max_exact)
+            / math.log(max_distance / max_exact)
+            * (num_buckets - max_exact)
+        ).to(torch.long)
+        relative_position_if_large = torch.min(
+            relative_position_if_large, torch.full_like(relative_position_if_large, num_buckets - 1)
+        )
+
+        relative_buckets += torch.where(is_small, relative_position, relative_position_if_large)
+        return relative_buckets
+
+    def compute_bias(self, query_length, key_length, device=None):
+        """Compute binned relative position bias"""
+        if device is None:
+            device = self.relative_attention_bias.weight.device
+        context_position = torch.arange(query_length, dtype=torch.long, device=device)[:, None]
+        memory_position = torch.arange(key_length, dtype=torch.long, device=device)[None, :]
+        relative_position = memory_position - context_position  # shape (query_length, key_length)
+        relative_position_bucket = self._relative_position_bucket(
+            relative_position,  # shape (query_length, key_length)
+            bidirectional=(not self.is_decoder),
+            num_buckets=self.relative_attention_num_buckets,
+            max_distance=self.relative_attention_max_distance,
+        )
+        values = self.relative_attention_bias(relative_position_bucket)  # shape (query_length, key_length, num_heads)
+        values = values.permute([2, 0, 1]).unsqueeze(0)  # shape (1, num_heads, query_length, key_length)
+        return values
+
+    def forward(
+        self,
+        hidden_states,
+        mask=None,
+        key_value_states=None,
+        position_bias=None,
+        past_key_value=None,
+        layer_head_mask=None,
+        query_length=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        """
+        Self-attention (if key_value_states is None) or attention over source sentence (provided by key_value_states).
+        """
+        # Input is (batch_size, seq_length, dim)
+        # Mask is (batch_size, key_length) (non-causal) or (batch_size, key_length, key_length)
+        # past_key_value[0] is (batch_size, n_heads, q_len - 1, dim_per_head)
+        batch_size, seq_length = hidden_states.shape[:2]
+
+        real_seq_length = seq_length
+
+        if past_key_value is not None:
+            assert (
+                len(past_key_value) == 2
+            ), f"past_key_value should have 2 past states: keys and values. Got { len(past_key_value)} past states"
+            real_seq_length += past_key_value[0].shape[2] if query_length is None else query_length
+
+        key_length = real_seq_length if key_value_states is None else key_value_states.shape[1]
+
+        def shape(states):
+            """projection"""
+            return states.view(batch_size, -1, self.n_heads, self.key_value_proj_dim).transpose(1, 2)
+
+        def unshape(states):
+            """reshape"""
+            return states.transpose(1, 2).contiguous().view(batch_size, -1, self.inner_dim)
+
+        def project(hidden_states, proj_layer, key_value_states, past_key_value):
+            """projects hidden states correctly to key/query states"""
+            if key_value_states is None:
+                # self-attn
+                # (batch_size, n_heads, seq_length, dim_per_head)
+                hidden_states = shape(proj_layer(hidden_states))
+            elif past_key_value is None:
+                # cross-attn
+                # (batch_size, n_heads, seq_length, dim_per_head)
+                hidden_states = shape(proj_layer(key_value_states))
+
+            if past_key_value is not None:
+                if key_value_states is None:
+                    # self-attn
+                    # (batch_size, n_heads, key_length, dim_per_head)
+                    hidden_states = torch.cat([past_key_value, hidden_states], dim=2)
+                else:
+                    # cross-attn
+                    hidden_states = past_key_value
+            return hidden_states
+
+        # get query states
+        query_states = shape(self.q(hidden_states))  # (batch_size, n_heads, seq_length, dim_per_head)
+
+        # get key/value states
+        key_states = project(
+            hidden_states, self.k, key_value_states, past_key_value[0] if past_key_value is not None else None
+        )
+        value_states = project(
+            hidden_states, self.v, key_value_states, past_key_value[1] if past_key_value is not None else None
+        )
+
+        # compute scores
+        scores = torch.matmul(
+            query_states, key_states.transpose(3, 2)
+        )  # equivalent of torch.einsum("bnqd,bnkd->bnqk", query_states, key_states), compatible with onnx op>9
+
+        if position_bias is None:
+            if not self.has_relative_attention_bias:
+                position_bias = torch.zeros(
+                    (1, self.n_heads, real_seq_length, key_length), device=scores.device, dtype=scores.dtype
+                )
+                if self.gradient_checkpointing and self.training:
+                    position_bias.requires_grad = True
+            else:
+                position_bias = self.compute_bias(real_seq_length, key_length, device=scores.device)
+
+            # if key and values are already calculated
+            # we want only the last query position bias
+            if past_key_value is not None:
+                position_bias = position_bias[:, :, -hidden_states.size(1) :, :]
+
+            if mask is not None:
+                position_bias = position_bias + mask  # (batch_size, n_heads, seq_length, key_length)
+
+        if self.pruned_heads:
+            mask = torch.ones(position_bias.shape[1])
+            mask[list(self.pruned_heads)] = 0
+            position_bias_masked = position_bias[:, mask.bool()]
+        else:
+            position_bias_masked = position_bias
+
+        scores += position_bias_masked
+        attn_weights = nn.functional.softmax(scores.float(), dim=-1).type_as(
+            scores
+        )  # (batch_size, n_heads, seq_length, key_length)
+        attn_weights = nn.functional.dropout(
+            attn_weights, p=self.dropout, training=self.training
+        )  # (batch_size, n_heads, seq_length, key_length)
+
+        # Mask heads if we want to
+        if layer_head_mask is not None:
+            attn_weights = attn_weights * layer_head_mask
+
+        attn_output = unshape(torch.matmul(attn_weights, value_states))  # (batch_size, seq_length, dim)
+        attn_output = self.o(attn_output)
+
+        present_key_value_state = (key_states, value_states) if (self.is_decoder and use_cache) else None
+        outputs = (attn_output,) + (present_key_value_state,) + (position_bias,)
+
+        if output_attentions:
+            outputs = outputs + (attn_weights,)
+        return outputs
+
+
+class T5LayerSelfAttention(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super().__init__()
+        self.SelfAttention = T5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_bias=None,
+        layer_head_mask=None,
+        past_key_value=None,
+        use_cache=False,
+        output_attentions=False,
+    ):
+        normed_hidden_states = self.layer_norm(hidden_states)
+        attention_output = self.SelfAttention(
+            normed_hidden_states,
+            mask=attention_mask,
+            position_bias=position_bias,
+            layer_head_mask=layer_head_mask,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        hidden_states = hidden_states + self.dropout(attention_output[0])
+        outputs = (hidden_states,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5LayerCrossAttention(nn.Module):
+    def __init__(self, config, is_vision_cross_attention=False):
+        super().__init__()
+        self.EncDecAttention = T5Attention(
+            config, has_relative_attention_bias=False, is_vision_cross_attention=is_vision_cross_attention
+        )
+        self.layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        self.is_vision_cross_attention = is_vision_cross_attention
+        if is_vision_cross_attention:
+            self.act = nn.Tanh()
+            if config.alpha_initializer == "zeros":
+                if config.alpha_type == "vector":
+                    self.alpha_cross_attn = nn.Parameter(torch.zeros(1, 1, config.d_model))
+                elif config.alpha_type == "float":
+                    self.alpha_cross_attn = nn.Parameter(torch.zeros(1))
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            elif config.alpha_initializer == "ones":
+                if config.alpha_type == "vector":
+                    self.alpha_cross_attn = nn.Parameter(torch.ones(1, 1, config.d_model))
+                elif config.alpha_type == "float":
+                    self.alpha_cross_attn = nn.Parameter(torch.ones(1))
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            elif config.alpha_initializer in {"normal", "gaussian", "random"}:
+                if config.alpha_type == "vector":
+                    self.alpha_cross_attn = nn.Parameter(
+                        torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1, 1, config.d_model))
+                    )
+                elif config.alpha_type == "float":
+                    self.alpha_cross_attn = nn.Parameter(
+                        torch.normal(mean=0.0, std=config.alphas_initializer_range, size=(1))
+                    )
+                else:
+                    raise ValueError(f"Unknown value for `alpha_type` ({config.alpha_type})")
+
+            else:
+                raise NotImplementedError(
+                    f"Alpha initialization scheme {config.alpha_initializer} not yet implemented!"
+                )
+
+    def forward(
+        self,
+        hidden_states,
+        key_value_states,
+        attention_mask=None,
+        position_bias=None,
+        layer_head_mask=None,
+        past_key_value=None,
+        use_cache=False,
+        query_length=None,
+        output_attentions=False,
+    ):
+        normed_hidden_states = self.layer_norm(hidden_states)
+        attention_output = self.EncDecAttention(
+            normed_hidden_states,
+            mask=attention_mask,
+            key_value_states=key_value_states,
+            position_bias=position_bias,
+            layer_head_mask=layer_head_mask,
+            past_key_value=past_key_value,
+            use_cache=use_cache,
+            query_length=query_length,
+            output_attentions=output_attentions,
+        )
+        if not self.is_vision_cross_attention:
+            layer_output = hidden_states + self.dropout(attention_output[0])
+        else:
+            layer_output = hidden_states + self.dropout(self.act(self.alpha_cross_attn) * attention_output[0])
+        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
+        return outputs
+
+
+class T5Block(nn.Module):
+    def __init__(self, config, has_relative_attention_bias=False):
+        super().__init__()
+        self.is_decoder = config.is_decoder
+        self.layer = nn.ModuleList()
+        self.layer.append(T5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
+        if self.is_decoder:
+            self.layer.append(T5LayerCrossAttention(config))
+
+        self.layer.append(T5LayerFF(config))
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        position_bias=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        encoder_decoder_position_bias=None,
+        layer_head_mask=None,
+        cross_attn_layer_head_mask=None,
+        past_key_value=None,
+        use_cache=False,
+        output_attentions=False,
+        return_dict=True,
+    ):
+        if past_key_value is not None:
+            if not self.is_decoder:
+                logger.warning("`past_key_values` is passed to the encoder. Please make sure this is intended.")
+            expected_num_past_key_values = 2 if encoder_hidden_states is None else 4
+
+            if len(past_key_value) != expected_num_past_key_values:
+                raise ValueError(
+                    f"There should be {expected_num_past_key_values} past states. "
+                    f"{'2 (past / key) for cross attention. ' if expected_num_past_key_values == 4 else ''}"
+                    f"Got {len(past_key_value)} past key / value states"
+                )
+
+            self_attn_past_key_value = past_key_value[:2]
+            cross_attn_past_key_value = past_key_value[2:]
+        else:
+            self_attn_past_key_value, cross_attn_past_key_value = None, None
+
+        self_attention_outputs = self.layer[0](
+            hidden_states,
+            attention_mask=attention_mask,
+            position_bias=position_bias,
+            layer_head_mask=layer_head_mask,
+            past_key_value=self_attn_past_key_value,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        hidden_states, present_key_value_state = self_attention_outputs[:2]
+        attention_outputs = self_attention_outputs[2:]  # Keep self-attention outputs and relative position weights
+
+        # clamp inf values to enable fp16 training
+        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        do_cross_attention = self.is_decoder and encoder_hidden_states is not None
+        if do_cross_attention:
+            # the actual query length is unknown for cross attention
+            # if using past key value states. Need to inject it here
+            if present_key_value_state is not None:
+                query_length = present_key_value_state[0].shape[2]
+            else:
+                query_length = None
+
+            cross_attention_outputs = self.layer[1](
+                hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                position_bias=encoder_decoder_position_bias,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                query_length=query_length,
+                use_cache=use_cache,
+                output_attentions=output_attentions,
+            )
+            hidden_states = cross_attention_outputs[0]
+
+            # clamp inf values to enable fp16 training
+            if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+                clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+                hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+            # Combine self attn and cross attn key value states
+            if present_key_value_state is not None:
+                present_key_value_state = present_key_value_state + cross_attention_outputs[1]
+
+            # Keep cross-attention outputs and relative position weights
+            attention_outputs = attention_outputs + cross_attention_outputs[2:]
+
+        # Apply Feed Forward layer
+        hidden_states = self.layer[-1](hidden_states)
+
+        # clamp inf values to enable fp16 training
+        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if use_cache:
+            outputs = outputs + (present_key_value_state,) + attention_outputs
+        else:
+            outputs = outputs + attention_outputs
+
+        return outputs  # hidden-states, present_key_value_states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
+
+
+class VT5GatedCrossAttentionBlock(nn.Module):
+    """Implementing the gated cross attention from the text LM to the vision encoder output."""
+
+    def __init__(self, config, has_relative_attention_bias=False):
+        super().__init__()
+        self.is_decoder = config.is_decoder
+        self.layer = nn.ModuleList()
+        self.layer.append(T5LayerCrossAttention(config, is_vision_cross_attention=True))
+        self.layer.append(T5LayerFF(config, is_vision_cross_attention=True))
+
+    def forward(
+        self,
+        hidden_states,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        encoder_decoder_position_bias=None,
+        cross_attn_layer_head_mask=None,
+        use_cache=False,
+        output_attentions=False,
+        return_dict=True,
+    ):
+        cross_attention_outputs = self.layer[0](
+            hidden_states,
+            key_value_states=encoder_hidden_states,
+            attention_mask=encoder_attention_mask,
+            position_bias=encoder_decoder_position_bias,
+            layer_head_mask=cross_attn_layer_head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+        )
+        hidden_states, _ = cross_attention_outputs[
+            :2
+        ]  # In the standard case, `_` would be `present_key_value_state`. But in the case of vision cross attention for the lm encoder, `present_key_value_state` is always None. So I am directly simplifying the logic by simply removing it and setting it to None where it should be None.
+
+        # clamp inf values to enable fp16 training
+        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        # Apply Feed Forward layer
+        hidden_states = self.layer[-1](hidden_states)
+
+        # clamp inf values to enable fp16 training
+        if hidden_states.dtype == torch.float16 and torch.isinf(hidden_states).any():
+            clamp_value = torch.finfo(hidden_states.dtype).max - 1000
+            hidden_states = torch.clamp(hidden_states, min=-clamp_value, max=clamp_value)
+
+        outputs = (hidden_states,)
+
+        if use_cache:
+            outputs = outputs + (None,) + cross_attention_outputs[2:]
+        else:
+            outputs = outputs + cross_attention_outputs[2:]
+
+        return outputs
+
+
+class VT5PreTrainedModel(VLOOMPreTrainedModelBase):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = VT5Config
+    load_tf_weights = load_tf_weights_in_t5
+    base_model_prefix = "transformer"
+    is_parallelizable = True
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["T5Block"]
+
+    @property
+    def dummy_inputs(self):
+        input_ids = torch.tensor(DUMMY_INPUTS)
+        input_mask = torch.tensor(DUMMY_MASK)
+        dummy_inputs = {
+            "decoder_input_ids": input_ids,
+            "input_ids": input_ids,
+            "decoder_attention_mask": input_mask,
+        }
+        return dummy_inputs
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_factor  # Used for testing weights initialization
+        if isinstance(module, T5LayerNorm):
+            module.weight.data.fill_(factor * 1.0)
+        elif isinstance(module, (VT5Model, VT5ForConditionalGeneration)):
+            # Mesh TensorFlow embeddings initialization
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L1624
+            module.shared.weight.data.normal_(mean=0.0, std=factor * 1.0)
+            if hasattr(module, "lm_head") and not self.config.tie_word_embeddings:
+                module.lm_head.weight.data.normal_(mean=0.0, std=factor * 1.0)
+        elif isinstance(module, T5DenseActDense):
+            # Mesh TensorFlow FF initialization
+            # See https://github.com/tensorflow/mesh/blob/master/mesh_tensorflow/transformer/transformer_layers.py#L56
+            # and https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L89
+            module.wi.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi, "bias") and module.wi.bias is not None:
+                module.wi.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, "bias") and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5DenseGatedActDense):
+            module.wi_0.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi_0, "bias") and module.wi_0.bias is not None:
+                module.wi_0.bias.data.zero_()
+            module.wi_1.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
+            if hasattr(module.wi_1, "bias") and module.wi_1.bias is not None:
+                module.wi_1.bias.data.zero_()
+            module.wo.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_ff) ** -0.5))
+            if hasattr(module.wo, "bias") and module.wo.bias is not None:
+                module.wo.bias.data.zero_()
+        elif isinstance(module, T5Attention):
+            # Mesh TensorFlow attention initialization to avoid scaling before softmax
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/attention.py#L136
+            d_model = self.config.d_model
+            key_value_proj_dim = self.config.d_kv
+            n_heads = self.config.num_heads
+            module.q.weight.data.normal_(mean=0.0, std=factor * ((d_model * key_value_proj_dim) ** -0.5))
+            module.k.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+            module.v.weight.data.normal_(mean=0.0, std=factor * (d_model**-0.5))
+            module.o.weight.data.normal_(mean=0.0, std=factor * ((n_heads * key_value_proj_dim) ** -0.5))
+            if module.has_relative_attention_bias:
+                module.relative_attention_bias.weight.data.normal_(mean=0.0, std=factor * ((d_model) ** -0.5))
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, (T5Attention, VT5Stack)):
+            module.gradient_checkpointing = value
+
+    @classmethod
+    def override_vision_model_wrapper(cls, model, config, vision_model_name, vision_model_params, torch_dtype):
+        # this can be called via from_pretrained from a class w/ head or w/o head so we extract the beheaded model version
+        beheaded_model = model.encoder
+        cls.override_vision_model(beheaded_model, vision_model_name, vision_model_params, torch_dtype)
+        beheaded_model.freeze_relevant_params(config)
+
+    def _shift_right(self, input_ids):
+        decoder_start_token_id = self.config.decoder_start_token_id
+        pad_token_id = self.config.pad_token_id
+
+        assert decoder_start_token_id is not None, (
+            "self.model.config.decoder_start_token_id has to be defined. In T5 it is usually set to the pad_token_id."
+            " See T5 docs for more information"
+        )
+
+        # shift inputs to the right
+        if is_torch_fx_proxy(input_ids):
+            # Item assignment is not supported natively for proxies.
+            shifted_input_ids = torch.full(input_ids.shape[:-1] + (1,), decoder_start_token_id)
+            shifted_input_ids = torch.cat([shifted_input_ids, input_ids[..., :-1]], dim=-1)
+        else:
+            shifted_input_ids = input_ids.new_zeros(input_ids.shape)
+            shifted_input_ids[..., 1:] = input_ids[..., :-1].clone()
+            shifted_input_ids[..., 0] = decoder_start_token_id
+
+        assert pad_token_id is not None, "self.model.config.pad_token_id has to be defined."
+        # replace possible -100 values in labels by `pad_token_id`
+        shifted_input_ids.masked_fill_(shifted_input_ids == -100, pad_token_id)
+
+        return shifted_input_ids
+
+
+class VT5Stack(VT5PreTrainedModel):
+    def __init__(self, config, embed_tokens=None, vision_model=None):
+        super().__init__(config)
+
+        self.embed_tokens = embed_tokens
+        self.is_decoder = config.is_decoder
+
+        self.block = nn.ModuleList(
+            [T5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)]
+        )
+        self.final_layer_norm = T5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
+        self.dropout = nn.Dropout(config.dropout_rate)
+
+        if not self.is_decoder:
+            self.cross_layer_interval = config.cross_layer_interval
+            num_cross_layers = config.num_layers // self.cross_layer_interval
+            self.gated_cross_attn_layers = nn.ModuleList(
+                [VT5GatedCrossAttentionBlock(config) for i in range(num_cross_layers)]
+            )
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+        self.gradient_checkpointing = False
+
+        # Load an uninitialized model and later in from_pretrained will load the pre-trained model -
+        # this solves the losing of weights in `from_pretrained` on the main model
+        if not self.is_decoder:
+            self.vision_model = vision_model
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        self.freeze_relevant_params(config)
+
+    def freeze_relevant_params(self, config=None):
+        if config is None:
+            config = self.config
+
+        if config.freeze_text_layers:
+            self.freeze_text_layers()
+
+        if not self.is_decoder and config.freeze_vision_layers:
+            freeze_model(self.vision_model)
+
+    def freeze_text_layers(self):
+        for module in [self.block, self.final_layer_norm]:
+            freeze_model(module)
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        # Check validity of device_map
+        self.device_map = (
+            get_device_map(len(self.block), range(torch.cuda.device_count())) if device_map is None else device_map
+        )
+        assert_device_map(self.device_map, len(self.block))
+        self.model_parallel = True
+        self.first_device = "cpu" if "cpu" in self.device_map.keys() else "cuda:" + str(min(self.device_map.keys()))
+        self.last_device = "cuda:" + str(max(self.device_map.keys()))
+        # Load onto devices
+        for k, v in self.device_map.items():
+            for layer in v:
+                cuda_device = "cuda:" + str(k)
+                self.block[layer] = self.block[layer].to(cuda_device)
+
+        # Set embed_tokens to first layer
+        self.embed_tokens = self.embed_tokens.to(self.first_device)
+        # Set final layer norm to last device
+        self.final_layer_norm = self.final_layer_norm.to(self.last_device)
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.model_parallel = False
+        self.device_map = None
+        self.first_device = "cpu"
+        self.last_device = "cpu"
+        for i in range(len(self.block)):
+            self.block[i] = self.block[i].to("cpu")
+        self.embed_tokens = self.embed_tokens.to("cpu")
+        self.final_layer_norm = self.final_layer_norm.to("cpu")
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embed_tokens = new_embeddings
+
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        inputs_embeds=None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        image_attention_mask: Optional[torch.Tensor] = None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        # Model parallel
+        if self.model_parallel:
+            torch.cuda.set_device(self.first_device)
+            self.embed_tokens = self.embed_tokens.to(self.first_device)
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        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.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            err_msg_prefix = "decoder_" if self.is_decoder else ""
+            raise ValueError(
+                f"You cannot specify both {err_msg_prefix}input_ids and {err_msg_prefix}inputs_embeds at the same time"
+            )
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+            input_ids = input_ids.view(-1, input_shape[-1])
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            err_msg_prefix = "decoder_" if self.is_decoder else ""
+            raise ValueError(f"You have to specify either {err_msg_prefix}input_ids or {err_msg_prefix}inputs_embeds")
+
+        if inputs_embeds is None:
+            assert self.embed_tokens is not None, "You have to initialize the model with valid token embeddings"
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        batch_size, seq_length = input_shape
+
+        # required mask seq length can be calculated via length of past
+        mask_seq_length = past_key_values[0][0].shape[2] + seq_length if past_key_values is not None else seq_length
+
+        if use_cache is True:
+            assert self.is_decoder, f"`use_cache` can only be set to `True` if {self} is used as a decoder"
+
+        if attention_mask is None:
+            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)
+        if self.is_decoder and encoder_attention_mask is None and encoder_hidden_states is not None:
+            encoder_seq_length = encoder_hidden_states.shape[1]
+            encoder_attention_mask = torch.ones(
+                batch_size, encoder_seq_length, device=inputs_embeds.device, dtype=torch.long
+            )
+
+        # initialize past_key_values with `None` if past does not exist
+        if past_key_values is None:
+            past_key_values = [None] * len(self.block)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if self.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=inputs_embeds.device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        head_mask = self.get_head_mask(head_mask, self.config.num_layers)
+        cross_attn_head_mask = self.get_head_mask(cross_attn_head_mask, self.config.num_layers)
+        present_key_value_states = () if use_cache else None
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and self.is_decoder) else None
+        position_bias = None
+        encoder_decoder_position_bias = None
+
+        hidden_states = self.dropout(inputs_embeds)
+
+        if not self.is_decoder:
+            pixel_values = pixel_values.to(dtype=self.dtype)  # fp16 compatibility
+            batch_size, num_images = pixel_values.size(0), pixel_values.size(1)
+            pixel_values = pixel_values.contiguous().view(batch_size * num_images, *pixel_values.shape[2:])
+            # Get sequence from the vision encoder
+            image_hidden_states = self.vision_model(pixel_values=pixel_values).last_hidden_state
+            image_seq_len, image_hidden_size = image_hidden_states.size(1), image_hidden_states.size(2)
+            image_hidden_states = image_hidden_states.view(batch_size, num_images * image_seq_len, image_hidden_size)
+            # Make image_attention_mask compatible with hidden states
+            # TODO: these steps are overly complex -> refactor
+            text_seq_len = image_attention_mask.size(1)
+            image_attention_mask = image_attention_mask.unsqueeze(-1)
+            image_attention_mask = image_attention_mask.repeat(1, 1, 1, image_seq_len)
+            image_attention_mask = image_attention_mask.view(batch_size, text_seq_len, num_images * image_seq_len)
+
+            if image_hidden_states is not None:
+                image_batch_size, image_sequence_length, _ = image_hidden_states.size()
+                image_hidden_shape = (image_batch_size, image_sequence_length)
+                if image_attention_mask is None:
+                    image_attention_mask = torch.ones(image_hidden_shape, device=hidden_states.device)
+                extended_image_attention_mask = self.invert_attention_mask(image_attention_mask)
+            else:
+                extended_image_attention_mask = None
+
+        for i, (layer_module, past_key_value) in enumerate(zip(self.block, past_key_values)):
+            layer_head_mask = head_mask[i]
+            cross_attn_layer_head_mask = cross_attn_head_mask[i]
+            # Model parallel
+            if self.model_parallel:
+                torch.cuda.set_device(hidden_states.device)
+                # Ensure that attention_mask is always on the same device as hidden_states
+                if attention_mask is not None:
+                    attention_mask = attention_mask.to(hidden_states.device)
+                if position_bias is not None:
+                    position_bias = position_bias.to(hidden_states.device)
+                if encoder_hidden_states is not None:
+                    encoder_hidden_states = encoder_hidden_states.to(hidden_states.device)
+                if encoder_extended_attention_mask is not None:
+                    encoder_extended_attention_mask = encoder_extended_attention_mask.to(hidden_states.device)
+                if encoder_decoder_position_bias is not None:
+                    encoder_decoder_position_bias = encoder_decoder_position_bias.to(hidden_states.device)
+                if layer_head_mask is not None:
+                    layer_head_mask = layer_head_mask.to(hidden_states.device)
+                if cross_attn_layer_head_mask is not None:
+                    cross_attn_layer_head_mask = cross_attn_layer_head_mask.to(hidden_states.device)
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            def vblock(
+                main_block,
+                hidden_states,
+                attention_mask,
+                position_bias,
+                encoder_hidden_states,
+                encoder_extended_attention_mask,
+                encoder_decoder_position_bias,
+                layer_head_mask,
+                cross_attn_layer_head_mask,
+                past_key_value,
+                use_cache,
+                output_attentions,
+                image_hidden_states,
+                image_attention_mask,
+                layer_idx,
+                cross_layer_interval,
+                gated_cross_attn_layers,
+                is_decoder,
+            ):
+                if not is_decoder and (layer_idx % cross_layer_interval == 0):
+                    xblock = gated_cross_attn_layers[layer_idx // cross_layer_interval]
+                    outputs = xblock(
+                        hidden_states,
+                        encoder_hidden_states=image_hidden_states,
+                        encoder_attention_mask=image_attention_mask,
+                        encoder_decoder_position_bias=None,
+                        use_cache=use_cache,
+                        output_attentions=output_attentions,
+                    )
+                    hidden_states = outputs[0]
+
+                layer_outputs = main_block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    position_bias=position_bias,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_attention_mask=encoder_extended_attention_mask,
+                    encoder_decoder_position_bias=encoder_decoder_position_bias,
+                    layer_head_mask=layer_head_mask,
+                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
+                    past_key_value=past_key_value,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                )
+
+                return layer_outputs
+
+            if self.is_decoder:
+                self.cross_layer_interval = None
+                self.gated_cross_attn_layers = None
+                image_hidden_states = None
+                extended_image_attention_mask = None
+
+            if self.gradient_checkpointing and self.training:
+                # past_key_value is always None with gradient checkpointing
+                past_key_value = None
+                if use_cache:
+                    logger.warning_once(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                layer_outputs = checkpoint(
+                    vblock,
+                    layer_module,
+                    hidden_states,
+                    extended_attention_mask,
+                    position_bias,
+                    encoder_hidden_states,
+                    encoder_extended_attention_mask,
+                    encoder_decoder_position_bias,
+                    head_mask[i],
+                    cross_attn_layer_head_mask,
+                    past_key_value,
+                    use_cache,
+                    output_attentions,
+                    image_hidden_states,
+                    extended_image_attention_mask,
+                    i,
+                    self.cross_layer_interval,
+                    self.gated_cross_attn_layers,
+                    self.is_decoder,
+                )
+            else:
+                layer_outputs = vblock(
+                    layer_module,
+                    hidden_states,
+                    attention_mask=extended_attention_mask,
+                    position_bias=position_bias,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_extended_attention_mask=encoder_extended_attention_mask,
+                    encoder_decoder_position_bias=encoder_decoder_position_bias,
+                    layer_head_mask=head_mask[i],
+                    cross_attn_layer_head_mask=cross_attn_layer_head_mask,
+                    past_key_value=past_key_value,
+                    use_cache=use_cache,
+                    output_attentions=output_attentions,
+                    image_hidden_states=image_hidden_states,
+                    image_attention_mask=extended_image_attention_mask,
+                    layer_idx=i,
+                    cross_layer_interval=self.cross_layer_interval,
+                    gated_cross_attn_layers=self.gated_cross_attn_layers,
+                    is_decoder=self.is_decoder,
+                )
+
+            # layer_outputs is a tuple with:
+            # hidden-states, key-value-states, (self-attention position bias), (self-attention weights), (cross-attention position bias), (cross-attention weights)
+            if use_cache is False:
+                layer_outputs = layer_outputs[:1] + (None,) + layer_outputs[1:]
+
+            hidden_states, present_key_value_state = layer_outputs[:2]
+
+            # We share the position biases between the layers - the first layer store them
+            # layer_outputs = hidden-states, key-value-states (self-attention position bias), (self-attention weights),
+            # (cross-attention position bias), (cross-attention weights)
+            position_bias = layer_outputs[2]
+            if self.is_decoder and encoder_hidden_states is not None:
+                encoder_decoder_position_bias = layer_outputs[4 if output_attentions else 3]
+            # append next layer key value states
+            if use_cache:
+                present_key_value_states = present_key_value_states + (present_key_value_state,)
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[3],)
+                if self.is_decoder:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[5],)
+
+            # Model Parallel: If it's the last layer for that device, put things on the next device
+            if self.model_parallel:
+                for k, v in self.device_map.items():
+                    if i == v[-1] and "cuda:" + str(k) != self.last_device:
+                        hidden_states = hidden_states.to("cuda:" + str(k + 1))
+
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+
+        # Add last layer
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    present_key_value_states,
+                    all_hidden_states,
+                    all_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=present_key_value_states,
+            hidden_states=all_hidden_states,
+            attentions=all_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+T5_START_DOCSTRING = r"""
+
+    The T5 model was proposed in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text
+    Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan
+    Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. It's an encoder decoder transformer pre-trained in a
+    text-to-text denoising generative setting.
+
+    This model inherits from [`PreTrainedModel`]. 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 PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`T5Config`]): 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 [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+T5_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. T5 is a model with relative position embeddings so you
+            should be able to pad the inputs on both the right and the left.
+
+            Indices can be obtained using [`T5Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for detail.
+
+            [What are input IDs?](../glossary#input-ids)
+
+            To know more on how to prepare `input_ids` for pretraining take a look a [T5 Training](./t5#training).
+        attention_mask (`torch.FloatTensor` of shape `(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#attention-mask)
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`T5Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            T5 uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values`
+            is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`).
+
+            To know more on how to prepare `decoder_input_ids` for pretraining take a look at [T5
+            Training](./t5#training).
+        decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules in the encoder. Mask values selected in `[0,
+            1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0,
+            1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`torch.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+                Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in
+                `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, `optional`: *hidden_states*, `optional`: *attentions*)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` is a sequence of hidden states at
+            the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`):
+            Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded
+            representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be
+            input (see `past_key_values`). This is useful if you want more control over how to convert
+            `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix.
+
+            If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value
+            of `inputs_embeds`.
+
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+
+        output_attentions (`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 (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+T5_ENCODER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+            Indices of input sequence tokens in the vocabulary. T5 is a model with relative position embeddings so you
+            should be able to pad the inputs on both the right and the left.
+
+            Indices can be obtained using [`T5Tokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for detail.
+
+            To know more on how to prepare `input_ids` for pretraining take a look a [T5 Training](./t5#training).
+        attention_mask (`torch.FloatTensor` of shape `(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#attention-mask)
+        head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*):
+            Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+            is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+            model's internal embedding lookup matrix.
+        output_attentions (`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 (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+# Warning message for FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+__HEAD_MASK_WARNING_MSG = """
+The input argument `head_mask` was split into two arguments `head_mask` and `decoder_head_mask`. Currently,
+`decoder_head_mask` is set to copy `head_mask`, but this feature is deprecated and will be removed in future versions.
+If you do not want to use any `decoder_head_mask` now, please set `decoder_head_mask = torch.ones(num_layers,
+num_heads)`.
+"""
+
+
+@add_start_docstrings(
+    "The bare T5 Model transformer outputting raw hidden-states without any specific head on top.",
+    T5_START_DOCSTRING,
+)
+class VT5Model(VT5PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.embed_tokens.weight",
+        r"decoder.embed_tokens.weight",
+    ]
+    _keys_to_ignore_on_load_unexpected = [
+        r"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+    ]
+
+    def __init__(self, config: T5Config, vision_model=None):
+        super().__init__(config)
+        self.shared = DecoupledEmbedding(
+            num_embeddings=config.vocab_size,
+            num_additional_embeddings=config.additional_vocab_size,
+            embedding_dim=config.d_model,
+            partially_freeze=config.freeze_text_layers,
+        )
+
+        encoder_config = copy.deepcopy(config)
+        encoder_config.is_decoder = False
+        encoder_config.use_cache = False
+        encoder_config.is_encoder_decoder = False
+        self.encoder = VT5Stack(encoder_config, self.shared, vision_model)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        decoder_config.is_encoder_decoder = False
+        decoder_config.num_layers = config.num_decoder_layers
+        self.decoder = VT5Stack(decoder_config, self.shared)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        self.freeze_relevant_params(config)
+
+    def freeze_relevant_params(self, config=None):
+        self.encoder.freeze_relevant_params(config)
+        self.decoder.freeze_relevant_params(config)
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.encoder.block))
+        self.encoder.parallelize(self.device_map)
+        self.decoder.parallelize(self.device_map)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.encoder.deparallelize()
+        self.decoder.deparallelize()
+        self.encoder = self.encoder.to("cpu")
+        self.decoder = self.decoder.to("cpu")
+        self.model_parallel = False
+        self.device_map = None
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+        self.encoder.set_input_embeddings(new_embeddings)
+        self.decoder.set_input_embeddings(new_embeddings)
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.BoolTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        decoder_head_mask: Optional[torch.FloatTensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        decoder_inputs_embeds: Optional[torch.Tensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        image_attention_mask: Optional[torch.Tensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqModelOutput]:
+        r"""
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import T5Tokenizer, T5Model
+
+        >>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
+        >>> model = T5Model.from_pretrained("t5-small")
+
+        >>> input_ids = tokenizer(
+        ...     "Studies have been shown that owning a dog is good for you", return_tensors="pt"
+        ... ).input_ids  # Batch size 1
+        >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids  # Batch size 1
+
+        >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for T5Model.
+        >>> # This is not needed for torch's T5ForConditionalGeneration as it does this internally using labels arg.
+        >>> decoder_input_ids = model._shift_right(decoder_input_ids)
+
+        >>> # forward pass
+        >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids)
+        >>> last_hidden_states = outputs.last_hidden_state
+        ```"""
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+        if head_mask is not None and decoder_head_mask is None:
+            if self.config.num_layers == self.config.num_decoder_layers:
+                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
+                decoder_head_mask = head_mask
+
+        # Encode if needed (training, first prediction pass)
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                pixel_values=pixel_values,
+                image_attention_mask=image_attention_mask,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        hidden_states = encoder_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+            hidden_states = hidden_states.to(self.decoder.first_device)
+            if decoder_input_ids is not None:
+                decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(self.decoder.first_device)
+            if decoder_attention_mask is not None:
+                decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            past_key_values=past_key_values,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return Seq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            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,
+        )
+
+
+@add_start_docstrings("""T5 Model with a `language modeling` head on top.""", T5_START_DOCSTRING)
+class VT5ForConditionalGeneration(VT5PreTrainedModel):
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.embed_tokens.weight",
+        r"decoder.embed_tokens.weight",
+        r"lm_head.weight",
+    ]
+    _keys_to_ignore_on_load_unexpected = [
+        r"decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight",
+    ]
+
+    def __init__(self, config: T5Config, vision_model=None):
+        super().__init__(config)
+        self.model_dim = config.d_model
+
+        self.shared = DecoupledEmbedding(
+            num_embeddings=config.vocab_size,
+            num_additional_embeddings=config.additional_vocab_size,
+            embedding_dim=config.d_model,
+            partially_freeze=config.freeze_text_layers,
+        )
+
+        encoder_config = copy.deepcopy(config)
+        encoder_config.is_decoder = False
+        encoder_config.use_cache = False
+        encoder_config.is_encoder_decoder = False
+        self.encoder = VT5Stack(encoder_config, self.shared, vision_model=vision_model)
+
+        decoder_config = copy.deepcopy(config)
+        decoder_config.is_decoder = True
+        decoder_config.is_encoder_decoder = False
+        decoder_config.num_layers = config.num_decoder_layers
+        self.decoder = VT5Stack(decoder_config, self.shared)
+
+        self.lm_head = DecoupledLinear(
+            in_features=config.d_model,
+            out_features=config.vocab_size,
+            out_additional_features=config.additional_vocab_size,
+            bias=False,
+            partially_freeze=config.freeze_lm_head,
+        )
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+        # Model parallel
+        self.model_parallel = False
+        self.device_map = None
+
+        self.freeze_relevant_params(config)
+
+    def freeze_relevant_params(self, config=None):
+        self.encoder.freeze_relevant_params(config)
+        self.decoder.freeze_relevant_params(config)
+
+    @add_start_docstrings(PARALLELIZE_DOCSTRING)
+    def parallelize(self, device_map=None):
+        self.device_map = (
+            get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
+            if device_map is None
+            else device_map
+        )
+        assert_device_map(self.device_map, len(self.encoder.block))
+        self.encoder.parallelize(self.device_map)
+        self.decoder.parallelize(self.device_map)
+        self.lm_head = self.lm_head.to(self.decoder.first_device)
+        self.model_parallel = True
+
+    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
+    def deparallelize(self):
+        self.encoder.deparallelize()
+        self.decoder.deparallelize()
+        self.encoder = self.encoder.to("cpu")
+        self.decoder = self.decoder.to("cpu")
+        self.lm_head = self.lm_head.to("cpu")
+        self.model_parallel = False
+        self.device_map = None
+        torch.cuda.empty_cache()
+
+    def get_input_embeddings(self):
+        return self.shared
+
+    def set_input_embeddings(self, new_embeddings):
+        self.shared = new_embeddings
+        self.encoder.set_input_embeddings(new_embeddings)
+        self.decoder.set_input_embeddings(new_embeddings)
+
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head = new_embeddings
+
+    def get_output_embeddings(self):
+        return self.lm_head
+
+    def tie_weights(self):
+        """
+        Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of DecoupledLinear and DecoupledEmbedding.
+        """
+        output_embeddings = self.get_output_embeddings()
+        input_embeddings = self.get_input_embeddings()
+
+        if getattr(self.config, "tie_word_embeddings", True):
+            output_embeddings.weight = input_embeddings.weight
+            if input_embeddings.num_additional_embeddings > 0:
+                assert output_embeddings.out_additional_features == input_embeddings.num_additional_embeddings
+                output_embeddings.additional_fc.weight = input_embeddings.additional_embedding.weight
+
+        if hasattr(output_embeddings, "out_features") and hasattr(input_embeddings, "num_embeddings"):
+            output_embeddings.out_features = input_embeddings.num_embeddings
+            if hasattr(output_embeddings, "out_additional_features") and hasattr(
+                input_embeddings, "num_additional_embeddings"
+            ):
+                output_embeddings.out_additional_features = input_embeddings.num_additional_embeddings
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(T5_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.BoolTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        decoder_head_mask: Optional[torch.FloatTensor] = None,
+        cross_attn_head_mask: Optional[torch.Tensor] = None,
+        encoder_outputs: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        decoder_inputs_embeds: Optional[torch.FloatTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        image_attention_mask: Optional[torch.Tensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.FloatTensor], Seq2SeqLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for
+            labels in `[0, ..., config.vocab_size]`
+
+        Returns:
+
+        Examples:
+
+        ```python
+        >>> from transformers import T5Tokenizer, T5ForConditionalGeneration
+
+        >>> tokenizer = T5Tokenizer.from_pretrained("t5-small")
+        >>> model = T5ForConditionalGeneration.from_pretrained("t5-small")
+
+        >>> # training
+        >>> input_ids = tokenizer("The <extra_id_0> walks in <extra_id_1> park", return_tensors="pt").input_ids
+        >>> labels = tokenizer("<extra_id_0> cute dog <extra_id_1> the <extra_id_2>", return_tensors="pt").input_ids
+        >>> outputs = model(input_ids=input_ids, labels=labels)
+        >>> loss = outputs.loss
+        >>> logits = outputs.logits
+
+        >>> # inference
+        >>> input_ids = tokenizer(
+        ...     "summarize: studies have shown that owning a dog is good for you", return_tensors="pt"
+        ... ).input_ids  # Batch size 1
+        >>> outputs = model.generate(input_ids)
+        >>> print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+        >>> # studies have shown that owning a dog is good for you.
+        ```"""
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # FutureWarning: head_mask was separated into two input args - head_mask, decoder_head_mask
+        if head_mask is not None and decoder_head_mask is None:
+            if self.config.num_layers == self.config.num_decoder_layers:
+                warnings.warn(__HEAD_MASK_WARNING_MSG, FutureWarning)
+                decoder_head_mask = head_mask
+
+        # Encode if needed (training, first prediction pass)
+        if encoder_outputs is None:
+            # Convert encoder inputs in embeddings if needed
+            encoder_outputs = self.encoder(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                inputs_embeds=inputs_embeds,
+                pixel_values=pixel_values,
+                image_attention_mask=image_attention_mask,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+        elif return_dict and not isinstance(encoder_outputs, BaseModelOutput):
+            encoder_outputs = BaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        hidden_states = encoder_outputs[0]
+
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+
+        if labels is not None and decoder_input_ids is None and decoder_inputs_embeds is None:
+            # get decoder inputs from shifting lm labels to the right
+            decoder_input_ids = self._shift_right(labels)
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.decoder.first_device)
+            hidden_states = hidden_states.to(self.decoder.first_device)
+            if decoder_input_ids is not None:
+                decoder_input_ids = decoder_input_ids.to(self.decoder.first_device)
+            if attention_mask is not None:
+                attention_mask = attention_mask.to(self.decoder.first_device)
+            if decoder_attention_mask is not None:
+                decoder_attention_mask = decoder_attention_mask.to(self.decoder.first_device)
+
+        # Decode
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            inputs_embeds=decoder_inputs_embeds,
+            past_key_values=past_key_values,
+            encoder_hidden_states=hidden_states,
+            encoder_attention_mask=attention_mask,
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = decoder_outputs[0]
+
+        # Set device for model parallelism
+        if self.model_parallel:
+            torch.cuda.set_device(self.encoder.first_device)
+            self.lm_head = self.lm_head.to(self.encoder.first_device)
+            sequence_output = sequence_output.to(self.lm_head.weight.device)
+
+        if self.config.tie_word_embeddings:
+            # Rescale output before projecting on vocab
+            # See https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/transformer/transformer.py#L586
+            sequence_output = sequence_output * (self.model_dim**-0.5)
+
+        lm_logits = self.lm_head(sequence_output)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss(ignore_index=-100)
+            loss = loss_fct(lm_logits.view(-1, lm_logits.size(-1)), labels.view(-1))
+            # TODO(thom): Add z_loss https://github.com/tensorflow/mesh/blob/fa19d69eafc9a482aff0b59ddd96b025c0cb207d/mesh_tensorflow/layers.py#L666
+
+        if not return_dict:
+            output = (lm_logits,) + decoder_outputs[1:] + encoder_outputs
+            return ((loss,) + output) if loss is not None else output
+
+        return Seq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=decoder_outputs.past_key_values,
+            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,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids,
+        past=None,
+        attention_mask=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        use_cache=None,
+        encoder_outputs=None,
+        **kwargs,
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {
+            "decoder_input_ids": input_ids,
+            "past_key_values": past,
+            "encoder_outputs": encoder_outputs,
+            "attention_mask": attention_mask,
+            "head_mask": head_mask,
+            "decoder_head_mask": decoder_head_mask,
+            "cross_attn_head_mask": cross_attn_head_mask,
+            "use_cache": use_cache,
+        }
+
+    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
+        return self._shift_right(labels)
+
+    def _reorder_cache(self, past, beam_idx):
+        # if decoder past is not included in output
+        # speedy decoding is disabled and no need to reorder
+        if past is None:
+            logger.warning("You might want to consider setting `use_cache=True` to speed up decoding")
+            return past
+
+        reordered_decoder_past = ()
+        for layer_past_states in past:
+            # get the correct batch idx from layer past batch dim
+            # batch dim of `past` is at 2nd position
+            reordered_layer_past_states = ()
+            for layer_past_state in layer_past_states:
+                # need to set correct `past` for each of the four key / value states
+                reordered_layer_past_states = reordered_layer_past_states + (
+                    layer_past_state.index_select(0, beam_idx.to(layer_past_state.device)),
+                )
+
+            assert reordered_layer_past_states[0].shape == layer_past_states[0].shape
+            assert len(reordered_layer_past_states) == len(layer_past_states)
+
+            reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,)
+        return reordered_decoder_past
+
+    def get_model_tflops_per_batch_per_gpu(self, hparams, data_param, tokenizer, max_num_images):
+        config_vl_model = self.config
+
+        vloom_embed_size = config_vl_model.d_model
+        vision_config = self.encoder.vision_model.config
+        num_language_layers = config_vl_model.num_layers
+        ffn_inner_size = config_vl_model.d_ff
+
+        # Get vision model blocks infos
+        vision_patch_size = vision_config.patch_size
+        vision_hidden_size = vision_config.hidden_size
+        num_vision_layers = vision_config.num_hidden_layers
+        # The +1 is for the CLS token
+        single_image_seq_len = (vision_config.image_size // vision_patch_size) ** 2 + 1
+        vision_exp_factor = vision_config.intermediate_size // vision_hidden_size
+
+        # Get language and cross-att blocks infos
+        num_cross_attn_layers = num_language_layers // config_vl_model.cross_layer_interval
+        language_seq_len = data_param.max_seq_len
+        language_exp_factor = (ffn_inner_size // vloom_embed_size) if ffn_inner_size is not None else 4
+        cross_att_exp_factor = (ffn_inner_size // vloom_embed_size) if ffn_inner_size is not None else 4
+
+        encoder_language_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
+            num_layers=num_language_layers,
+            batch_size=hparams.batch_size_per_gpu,
+            q_seq_len=language_seq_len,
+            k_seq_len=language_seq_len,
+            hidden_size=vloom_embed_size,
+            kv_in_dim=vloom_embed_size,
+            ff_exp_factor=language_exp_factor,
+            grad_acc_size=hparams.grad_acc_size,
+            vocab_size=None,
+            count_backward=True,  # Always True regardless of freezing, because gradients are computed for cross-attentions
+            use_grad_checkpointing=False,
+        )
+        image_gated_cross_attention_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
+            num_layers=num_cross_attn_layers,
+            batch_size=hparams.batch_size_per_gpu,
+            q_seq_len=language_seq_len,
+            k_seq_len=single_image_seq_len * max_num_images,
+            hidden_size=vloom_embed_size,
+            kv_in_dim=vision_hidden_size,
+            ff_exp_factor=cross_att_exp_factor,
+            grad_acc_size=hparams.grad_acc_size,
+            vocab_size=None,
+            count_backward=True,
+            use_grad_checkpointing=False,
+        )
+        vision_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
+            num_layers=num_vision_layers,
+            batch_size=hparams.batch_size_per_gpu * max_num_images,
+            q_seq_len=single_image_seq_len,
+            k_seq_len=single_image_seq_len,
+            hidden_size=vision_hidden_size,
+            kv_in_dim=vision_hidden_size,
+            ff_exp_factor=vision_exp_factor,
+            grad_acc_size=hparams.grad_acc_size,
+            vocab_size=None,
+            count_backward=not hparams.model_params["freeze_vision_layers"],
+            use_grad_checkpointing=False,
+        )
+
+        _, target_seq_len = random_spans_helper(
+            inputs_length=data_param.max_seq_len,
+            noise_density=data_param.t5_mlm_noise_density,
+            mean_noise_span_length=data_param.t5_mlm_mean_noise_span_length,
+            extra_tokens_per_span_inputs=1,
+            extra_tokens_per_span_targets=1,
+            verbose=False,
+        )
+
+        decoder_language_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
+            num_layers=config_vl_model.num_decoder_layers,
+            batch_size=hparams.batch_size_per_gpu,
+            q_seq_len=target_seq_len,
+            k_seq_len=target_seq_len,
+            hidden_size=vloom_embed_size,
+            kv_in_dim=vloom_embed_size,
+            ff_exp_factor=language_exp_factor,
+            grad_acc_size=hparams.grad_acc_size,
+            vocab_size=tokenizer.vocab_size,
+            count_backward=True,  # Always True regardless of freezing, because gradients are computed for cross-attentions
+            use_grad_checkpointing=False,
+        )
+
+        encoder_decoder_cross_attention_tflops_per_batch_per_gpu = compute_tflops_per_batch_per_gpu(
+            num_layers=config_vl_model.num_decoder_layers,
+            batch_size=hparams.batch_size_per_gpu,
+            q_seq_len=target_seq_len,
+            k_seq_len=language_seq_len,
+            hidden_size=vloom_embed_size,
+            kv_in_dim=vloom_embed_size,
+            ff_exp_factor=0,  # There is only one pair of expansion linear layers per pair of self attention and cross attention blocks
+            grad_acc_size=hparams.grad_acc_size,
+            vocab_size=None,
+            count_backward=True,
+            use_grad_checkpointing=False,
+        )
+        return (
+            encoder_language_tflops_per_batch_per_gpu
+            + image_gated_cross_attention_tflops_per_batch_per_gpu
+            + vision_tflops_per_batch_per_gpu
+            + decoder_language_tflops_per_batch_per_gpu
+            + encoder_decoder_cross_attention_tflops_per_batch_per_gpu
+        )