modeling_mt5.py

import copy
import math
import os
import warnings
from typing import List, Optional, Tuple, Union

import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss

from ...activations import ACT2FN
from ...modeling_outputs import (
    BaseModelOutput,
    BaseModelOutputWithPastAndCrossAttentions,
    Seq2SeqLMOutput,
    Seq2SeqModelOutput,
    Seq2SeqQuestionAnsweringModelOutput,
    Seq2SeqSequenceClassifierOutput,
    TokenClassifierOutput,
)
from ...modeling_utils import PreTrainedModel
from ...pytorch_utils import find_pruneable_heads_and_indices, prune_linear_layer
from ...utils import (
    DUMMY_INPUTS,
    DUMMY_MASK,
    add_start_docstrings,
    add_start_docstrings_to_model_forward,
    is_torch_fx_proxy,
    logging,
    replace_return_docstrings,
)
from ...utils.model_parallel_utils import assert_device_map, get_device_map
from .configuration_mt5 import MT5Config


logger = logging.get_logger(__name__)

_CONFIG_FOR_DOC = "MT5Config"
_CHECKPOINT_FOR_DOC = "mt5-small"


####################################################
# This dict contains ids and associated url
# for the pretrained weights provided with the models
####################################################


# Copied from transformers.models.t5.modeling_t5.T5LayerNorm with T5->MT5
class MT5LayerNorm(nn.Module):
    def __init__(self, hidden_size, eps=1e-6):
        """
        Construct a layernorm module in the MT5 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):
        # MT5 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


# Copied from transformers.models.t5.modeling_t5.T5DenseActDense with T5->MT5
class MT5DenseActDense(nn.Module):
    def __init__(self, config: MT5Config):
        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)
        if (
            isinstance(self.wo.weight, torch.Tensor)
            and hidden_states.dtype != self.wo.weight.dtype
            and self.wo.weight.dtype != torch.int8
        ):
            hidden_states = hidden_states.to(self.wo.weight.dtype)
        hidden_states = self.wo(hidden_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5DenseGatedActDense with T5->MT5
class MT5DenseGatedActDense(nn.Module):
    def __init__(self, config: MT5Config):
        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)

        # To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
        # See https://github.com/huggingface/transformers/issues/20287
        # we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
        if (
            isinstance(self.wo.weight, torch.Tensor)
            and hidden_states.dtype != self.wo.weight.dtype
            and self.wo.weight.dtype != torch.int8
        ):
            hidden_states = hidden_states.to(self.wo.weight.dtype)

        hidden_states = self.wo(hidden_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5LayerFF with T5->MT5
class MT5LayerFF(nn.Module):
    def __init__(self, config: MT5Config):
        super().__init__()
        if config.is_gated_act:
            self.DenseReluDense = MT5DenseGatedActDense(config)
        else:
            self.DenseReluDense = MT5DenseActDense(config)

        self.layer_norm = MT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    def forward(self, hidden_states):
        forwarded_states = self.layer_norm(hidden_states)
        forwarded_states = self.DenseReluDense(forwarded_states)
        hidden_states = hidden_states + self.dropout(forwarded_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5Attention with T5->MT5
class MT5Attention(nn.Module):
    def __init__(self, config: MT5Config, has_relative_attention_bias=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)
        self.k = nn.Linear(self.d_model, self.inner_dim, bias=False)
        self.v = nn.Linear(self.d_model, 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:
            if len(past_key_value) != 2:
                raise ValueError(
                    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)
                elif past_key_value.shape[2] != key_value_states.shape[1]:
                    # checking that the `sequence_length` of the `past_key_value` is the same as
                    # the provided `key_value_states` to support prefix tuning
                    # cross-attn
                    # (batch_size, n_heads, seq_length, dim_per_head)
                    hidden_states = shape(proj_layer(key_value_states))
                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


# Copied from transformers.models.t5.modeling_t5.T5LayerSelfAttention with T5->MT5
class MT5LayerSelfAttention(nn.Module):
    def __init__(self, config, has_relative_attention_bias=False):
        super().__init__()
        self.SelfAttention = MT5Attention(config, has_relative_attention_bias=has_relative_attention_bias)
        self.layer_norm = MT5LayerNorm(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


# Copied from transformers.models.t5.modeling_t5.T5LayerCrossAttention with T5->MT5
class MT5LayerCrossAttention(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.EncDecAttention = MT5Attention(config, has_relative_attention_bias=False)
        self.layer_norm = MT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

    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,
        )
        layer_output = hidden_states + self.dropout(attention_output[0])
        outputs = (layer_output,) + attention_output[1:]  # add attentions if we output them
        return outputs


# Copied from transformers.models.t5.modeling_t5.T5Block with T5->MT5
class MT5Block(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(MT5LayerSelfAttention(config, has_relative_attention_bias=has_relative_attention_bias))
        if self.is_decoder:
            self.layer.append(MT5LayerCrossAttention(config))

        self.layer.append(MT5LayerFF(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 (key / value) 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:
            clamp_value = torch.where(
                torch.isinf(hidden_states).any(),
                torch.finfo(hidden_states.dtype).max - 1000,
                torch.finfo(hidden_states.dtype).max,
            )
            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:
                clamp_value = torch.where(
                    torch.isinf(hidden_states).any(),
                    torch.finfo(hidden_states.dtype).max - 1000,
                    torch.finfo(hidden_states.dtype).max,
                )
                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:
            clamp_value = torch.where(
                torch.isinf(hidden_states).any(),
                torch.finfo(hidden_states.dtype).max - 1000,
                torch.finfo(hidden_states.dtype).max,
            )
            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)


def load_tf_weights_in_mt5(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


# Copied from transformers.models.t5.modeling_t5.T5ClassificationHead with T5->MT5
class MT5ClassificationHead(nn.Module):
    """Head for sentence-level classification tasks."""

    def __init__(self, config: MT5Config):
        super().__init__()
        self.dense = nn.Linear(config.d_model, config.d_model)
        self.dropout = nn.Dropout(p=config.classifier_dropout)
        self.out_proj = nn.Linear(config.d_model, config.num_labels)

    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.dense(hidden_states)
        hidden_states = torch.tanh(hidden_states)
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.out_proj(hidden_states)
        return hidden_states


# Copied from transformers.models.t5.modeling_t5.T5PreTrainedModel with T5->MT5, t5->mt5
class MT5PreTrainedModel(PreTrainedModel):
    """
    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
    models.
    """

    config_class = MT5Config
    load_tf_weights = load_tf_weights_in_mt5
    base_model_prefix = "transformer"
    is_parallelizable = True
    supports_gradient_checkpointing = True
    _no_split_modules = ["MT5Block"]
    _keep_in_fp32_modules = ["wo"]

    @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, MT5LayerNorm):
            module.weight.data.fill_(factor * 1.0)
        elif isinstance(
            module,
            (MT5Model, MT5ForConditionalGeneration, MT5EncoderModel, MT5ForQuestionAnswering),
        ):
            # 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)
            if hasattr(module, "qa_outputs"):
                module.qa_outputs.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
                module.qa_outputs.bias.data.zero_()
        elif isinstance(module, MT5ForTokenClassification):
            if hasattr(module, "classifier"):
                module.classifier.weight.data.normal_(mean=0.0, std=factor * 1.0)
                module.classifier.bias.data.zero_()
        elif isinstance(module, MT5ClassificationHead):
            module.dense.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
            if hasattr(module.dense, "bias") and module.dense.bias is not None:
                module.dense.bias.data.zero_()
            module.out_proj.weight.data.normal_(mean=0.0, std=factor * ((self.config.d_model) ** -0.5))
            if hasattr(module.out_proj, "bias") and module.out_proj.bias is not None:
                module.out_proj.bias.data.zero_()
        elif isinstance(module, MT5DenseActDense):
            # 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, MT5DenseGatedActDense):
            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, MT5Attention):
            # 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 _shift_right(self, input_ids):
        decoder_start_token_id = self.config.decoder_start_token_id
        pad_token_id = self.config.pad_token_id

        if decoder_start_token_id is None:
            raise ValueError(
                "self.model.config.decoder_start_token_id has to be defined. In MT5 it is usually set to the pad_token_id. "
                "See MT5 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

        if pad_token_id is None:
            raise ValueError("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


# Copied from transformers.models.t5.modeling_t5.T5Stack with T5->MT5
class MT5Stack(MT5PreTrainedModel):
    def __init__(self, config, embed_tokens=None):
        super().__init__(config)

        self.embed_tokens = embed_tokens
        self.is_decoder = config.is_decoder

        self.block = nn.ModuleList(
            [MT5Block(config, has_relative_attention_bias=bool(i == 0)) for i in range(config.num_layers)]
        )
        self.final_layer_norm = MT5LayerNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dropout = nn.Dropout(config.dropout_rate)

        # Initialize weights and apply final processing
        self.post_init()
        # Model parallel
        self.model_parallel = False
        self.device_map = None
        self.gradient_checkpointing = False

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    def parallelize(self, device_map=None):
        warnings.warn(
            "`MT5Stack.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your model"
            " with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
            " `device_map` but it needs to be a dictionary module_name to device, so for instance {'block.0': 0,"
            " 'block.1': 1, ...}",
            FutureWarning,
        )
        # 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(DEPARALLELIZE_DOCSTRING)
    def deparallelize(self):
        warnings.warn(
            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
            FutureWarning,
        )
        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,
        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:
            if self.embed_tokens is None:
                raise ValueError("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:
            if not self.is_decoder:
                raise ValueError(f"`use_cache` can only be set to `True` if {self} is used as a decoder")

        # initialize past_key_values with `None` if past does not exist
        if past_key_values is None:
            past_key_values = [None] * len(self.block)

        if attention_mask is None:
            attention_mask = torch.ones(batch_size, mask_seq_length, device=inputs_embeds.device)

        # 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, dtype=torch.long
                )
            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
        else:
            encoder_extended_attention_mask = None

        if self.gradient_checkpointing and self.training:
            if use_cache:
                logger.warning_once(
                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
                )
                use_cache = False

        # 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)

        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,)

            if self.gradient_checkpointing and self.training:
                layer_outputs = self._gradient_checkpointing_func(
                    layer_module.forward,
                    hidden_states,
                    extended_attention_mask,
                    position_bias,
                    encoder_hidden_states,
                    encoder_extended_attention_mask,
                    encoder_decoder_position_bias,
                    layer_head_mask,
                    cross_attn_layer_head_mask,
                    None,  # past_key_value is always None with gradient checkpointing
                    use_cache,
                    output_attentions,
                )
            else:
                layer_outputs = layer_module(
                    hidden_states,
                    attention_mask=extended_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,
                )

            # 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,
        )


MT5_START_DOCSTRING = r"""

    The MT5 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 ([`MT5Config`]): 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.
"""

MT5_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. MT5 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 [`AutoTokenizer`]. 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 [MT5 Training](./mt5#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 [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are decoder input IDs?](../glossary#decoder-input-ids)

            MT5 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 [MT5
            Training](./mt5#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.
"""

MT5_ENCODER_INPUTS_DOCSTRING = r"""
    Args:
        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
            Indices of input sequence tokens in the vocabulary. MT5 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 [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for detail.

            To know more on how to prepare `input_ids` for pretraining take a look a [MT5 Training](./mt5#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 MT5 Model transformer outputting raw hidden-states without any specific head on top.",
    MT5_START_DOCSTRING,
)
class MT5Model(MT5PreTrainedModel):
    r"""
    Examples:

    ```python
    >>> from transformers import MT5Model, AutoTokenizer

    >>> model = MT5Model.from_pretrained("google/mt5-small")
    >>> tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
    >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien."
    >>> summary = "Weiter Verhandlung in Syrien."
    >>> inputs = tokenizer(article, return_tensors="pt")
    >>> labels = tokenizer(text_target=summary, return_tensors="pt")

    >>> outputs = model(input_ids=inputs["input_ids"], decoder_input_ids=labels["input_ids"])
    >>> hidden_states = outputs.last_hidden_state
    ```"""

    model_type = "mt5"
    config_class = MT5Config
    _keys_to_ignore_on_load_unexpected = ["decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight"]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5Model.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = MT5Stack(encoder_config, self.shared)

        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 = MT5Stack(decoder_config, self.shared)

        # Initialize weights and apply final processing
        self.post_init()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    # Copied from transformers.models.t5.modeling_t5.T5Model.parallelize
    def parallelize(self, device_map=None):
        warnings.warn(
            "`T5Model.parallelize` is deprecated and will be removed in v5 of Transformers, you should load your model"
            " with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
            " `device_map` but it needs to be a dictionary module_name to device, so for instance {'encoder.block.0':"
            " 0, 'encoder.block.1': 1, ...}",
            FutureWarning,
        )
        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)
    # Copied from transformers.models.t5.modeling_t5.T5Model.deparallelize
    def deparallelize(self):
        warnings.warn(
            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
            FutureWarning,
        )
        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()

    # Copied from transformers.models.t5.modeling_t5.T5Model.get_input_embeddings
    def get_input_embeddings(self):
        return self.shared

    # Copied from transformers.models.t5.modeling_t5.T5Model.set_input_embeddings
    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)

    # Copied from transformers.models.t5.modeling_t5.T5Model.get_encoder
    def get_encoder(self):
        return self.encoder

    # Copied from transformers.models.t5.modeling_t5.T5Model.get_decoder
    def get_decoder(self):
        return self.decoder

    # Copied from transformers.models.t5.modeling_t5.T5Model._prune_heads
    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(MT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqModelOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5Model.forward with T5->MT5, t5->mt5
    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,
        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 AutoTokenizer, MT5Model

        >>> tokenizer = AutoTokenizer.from_pretrained("google-mt5/mt5-small")
        >>> model = MT5Model.from_pretrained("google-mt5/mt5-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 MT5Model.
        >>> # This is not needed for torch's MT5ForConditionalGeneration 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,
                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("""MT5 Model with a `language modeling` head on top.""", MT5_START_DOCSTRING)
class MT5ForConditionalGeneration(MT5PreTrainedModel):
    r"""
    Examples:

    ```python
    >>> from transformers import MT5ForConditionalGeneration, AutoTokenizer

    >>> model = MT5ForConditionalGeneration.from_pretrained("google/mt5-small")
    >>> tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
    >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien."
    >>> summary = "Weiter Verhandlung in Syrien."
    >>> inputs = tokenizer(article, text_target=summary, return_tensors="pt")

    >>> outputs = model(**inputs)
    >>> loss = outputs.loss
    ```"""

    model_type = "mt5"
    config_class = MT5Config
    _keys_to_ignore_on_load_unexpected = ["decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight"]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight", "lm_head.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.model_dim = config.d_model

        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = MT5Stack(encoder_config, self.shared)

        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 = MT5Stack(decoder_config, self.shared)

        self.lm_head = nn.Linear(config.d_model, config.vocab_size, bias=False)

        # Initialize weights and apply final processing
        self.post_init()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.parallelize
    def parallelize(self, device_map=None):
        warnings.warn(
            "`T5ForConditionalGeneration.parallelize` is deprecated and will be removed in v5 of Transformers, you"
            " should load your model with `device_map='balanced'` in the call to `from_pretrained`. You can also"
            " provide your own `device_map` but it needs to be a dictionary module_name to device, so for instance"
            " {'encoder.block.0': 0, 'encoder.block.1': 1, ...}",
            FutureWarning,
        )
        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)
    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.deparallelize
    def deparallelize(self):
        warnings.warn(
            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
            FutureWarning,
        )
        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()

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.get_input_embeddings
    def get_input_embeddings(self):
        return self.shared

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.set_input_embeddings
    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)

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.set_output_embeddings
    def set_output_embeddings(self, new_embeddings):
        self.lm_head = new_embeddings

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.get_output_embeddings
    def get_output_embeddings(self):
        return self.lm_head

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.get_encoder
    def get_encoder(self):
        return self.encoder

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.get_decoder
    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(MT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.forward with T5->MT5, t5->mt5
    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,
        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 AutoTokenizer, MT5ForConditionalGeneration

        >>> tokenizer = AutoTokenizer.from_pretrained("google-mt5/mt5-small")
        >>> model = MT5ForConditionalGeneration.from_pretrained("google-mt5/mt5-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,
                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)
            # move labels to correct device to enable PP
            labels = labels.to(lm_logits.device)
            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,
        )

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.prepare_inputs_for_generation
    def prepare_inputs_for_generation(
        self,
        input_ids,
        past_key_values=None,
        attention_mask=None,
        head_mask=None,
        decoder_head_mask=None,
        decoder_attention_mask=None,
        cross_attn_head_mask=None,
        use_cache=None,
        encoder_outputs=None,
        **kwargs,
    ):
        # cut decoder_input_ids if past_key_values is used
        if past_key_values is not None:
            past_length = past_key_values[0][0].shape[2]

            # Some generation methods already pass only the last input ID
            if input_ids.shape[1] > past_length:
                remove_prefix_length = past_length
            else:
                # Default to old behavior: keep only final ID
                remove_prefix_length = input_ids.shape[1] - 1

            input_ids = input_ids[:, remove_prefix_length:]

        return {
            "decoder_input_ids": input_ids,
            "past_key_values": past_key_values,
            "encoder_outputs": encoder_outputs,
            "attention_mask": attention_mask,
            "head_mask": head_mask,
            "decoder_head_mask": decoder_head_mask,
            "decoder_attention_mask": decoder_attention_mask,
            "cross_attn_head_mask": cross_attn_head_mask,
            "use_cache": use_cache,
        }

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration.prepare_decoder_input_ids_from_labels
    def prepare_decoder_input_ids_from_labels(self, labels: torch.Tensor):
        return self._shift_right(labels)

    # Copied from transformers.models.t5.modeling_t5.T5ForConditionalGeneration._reorder_cache
    def _reorder_cache(self, past_key_values, beam_idx):
        # if decoder past is not included in output
        # speedy decoding is disabled and no need to reorder
        if past_key_values is None:
            logger.warning("You might want to consider setting `use_cache=True` to speed up decoding")
            return past_key_values

        reordered_decoder_past = ()
        for layer_past_states in past_key_values:
            # 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)),
                )

            if reordered_layer_past_states[0].shape != layer_past_states[0].shape:
                raise ValueError(
                    f"reordered_layer_past_states[0] shape {reordered_layer_past_states[0].shape} and layer_past_states[0] shape {layer_past_states[0].shape} mismatched"
                )
            if len(reordered_layer_past_states) != len(layer_past_states):
                raise ValueError(
                    f"length of reordered_layer_past_states {len(reordered_layer_past_states)} and length of layer_past_states {len(layer_past_states)} mismatched"
                )

            reordered_decoder_past = reordered_decoder_past + (reordered_layer_past_states,)
        return reordered_decoder_past


@add_start_docstrings(
    "The bare MT5 Model transformer outputting encoder's raw hidden-states without any specific head on top.",
    MT5_START_DOCSTRING,
)
class MT5EncoderModel(MT5PreTrainedModel):
    r"""
    Examples:

    ```python
    >>> from transformers import MT5EncoderModel, AutoTokenizer

    >>> model = MT5EncoderModel.from_pretrained("google/mt5-small")
    >>> tokenizer = AutoTokenizer.from_pretrained("google/mt5-small")
    >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien."
    >>> input_ids = tokenizer(article, return_tensors="pt").input_ids
    >>> outputs = model(input_ids)
    >>> hidden_state = outputs.last_hidden_state
    ```"""

    model_type = "mt5"
    config_class = MT5Config
    _tied_weights_keys = ["encoder.embed_tokens.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = MT5Stack(encoder_config, self.shared)

        # Initialize weights and apply final processing
        self.post_init()

        # Model parallel
        self.model_parallel = False
        self.device_map = None

    @add_start_docstrings(PARALLELIZE_DOCSTRING)
    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.parallelize
    def parallelize(self, device_map=None):
        warnings.warn(
            "`T5EncoderModel.parallelize` is deprecated and will be removed in v5 of Transformers, you should load"
            " your model with `device_map='balanced'` in the call to `from_pretrained`. You can also provide your own"
            " `device_map` but it needs to be a dictionary module_name to device, so for instance {'block.0': 0,"
            " 'block.1': 1, ...}",
            FutureWarning,
        )
        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.model_parallel = True

    @add_start_docstrings(DEPARALLELIZE_DOCSTRING)
    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.deparallelize
    def deparallelize(self):
        warnings.warn(
            "Like `parallelize`, `deparallelize` is deprecated and will be removed in v5 of Transformers.",
            FutureWarning,
        )
        self.encoder.deparallelize()
        self.encoder = self.encoder.to("cpu")
        self.model_parallel = False
        self.device_map = None
        torch.cuda.empty_cache()

    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.get_input_embeddings
    def get_input_embeddings(self):
        return self.shared

    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.set_input_embeddings
    def set_input_embeddings(self, new_embeddings):
        self.shared = new_embeddings
        self.encoder.set_input_embeddings(new_embeddings)

    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.get_encoder
    def get_encoder(self):
        return self.encoder

    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel._prune_heads
    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.block[layer].layer[0].SelfAttention.prune_heads(heads)

    @add_start_docstrings_to_model_forward(MT5_ENCODER_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=BaseModelOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5EncoderModel.forward with T5->MT5, t5->mt5
    def forward(
        self,
        input_ids: Optional[torch.LongTensor] = None,
        attention_mask: Optional[torch.FloatTensor] = None,
        head_mask: Optional[torch.FloatTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.FloatTensor], BaseModelOutput]:
        r"""
        Returns:

        Example:

        ```python
        >>> from transformers import AutoTokenizer, MT5EncoderModel

        >>> tokenizer = AutoTokenizer.from_pretrained("google-mt5/mt5-small")
        >>> model = MT5EncoderModel.from_pretrained("google-mt5/mt5-small")
        >>> input_ids = tokenizer(
        ...     "Studies have been shown that owning a dog is good for you", return_tensors="pt"
        ... ).input_ids  # Batch size 1
        >>> outputs = model(input_ids=input_ids)
        >>> last_hidden_states = outputs.last_hidden_state
        ```"""
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        encoder_outputs = self.encoder(
            input_ids=input_ids,
            attention_mask=attention_mask,
            inputs_embeds=inputs_embeds,
            head_mask=head_mask,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        return encoder_outputs


@add_start_docstrings(
    """
    MT5 model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE
    tasks.
    """,
    MT5_START_DOCSTRING,
)
class MT5ForSequenceClassification(MT5PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = ["decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight"]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5ForSequenceClassification.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.transformer = MT5Model(config)
        self.classification_head = MT5ClassificationHead(config)

        # Initialize weights and apply final processing
        self.post_init()

        self.model_parallel = False

    @add_start_docstrings_to_model_forward(MT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqSequenceClassifierOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5ForSequenceClassification.forward
    def forward(
        self,
        input_ids: torch.LongTensor = None,
        attention_mask: Optional[torch.Tensor] = None,
        decoder_input_ids: Optional[torch.LongTensor] = None,
        decoder_attention_mask: Optional[torch.LongTensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        decoder_head_mask: Optional[torch.Tensor] = None,
        cross_attn_head_mask: Optional[torch.Tensor] = None,
        encoder_outputs: Optional[List[torch.FloatTensor]] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = 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, Seq2SeqSequenceClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        Returns:
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        if labels is not None:
            use_cache = False

        if input_ids is None and inputs_embeds is not None:
            raise NotImplementedError(
                f"Passing input embeddings is currently not supported for {self.__class__.__name__}"
            )

        # Copied from models.bart.modeling_bart.BartModel.forward different to other models, T5 automatically creates
        # decoder_input_ids from input_ids if no decoder_input_ids are provided
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            if input_ids is None:
                raise ValueError(
                    "If no `decoder_input_ids` or `decoder_inputs_embeds` are "
                    "passed, `input_ids` cannot be `None`. Please pass either "
                    "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`."
                )
            decoder_input_ids = self._shift_right(input_ids)

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            decoder_input_ids=decoder_input_ids,
            decoder_attention_mask=decoder_attention_mask,
            head_mask=head_mask,
            decoder_head_mask=decoder_head_mask,
            cross_attn_head_mask=cross_attn_head_mask,
            encoder_outputs=encoder_outputs,
            inputs_embeds=inputs_embeds,
            decoder_inputs_embeds=decoder_inputs_embeds,
            use_cache=use_cache,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )
        sequence_output = outputs[0]

        eos_mask = input_ids.eq(self.config.eos_token_id).to(sequence_output.device)

        if len(torch.unique_consecutive(eos_mask.sum(1))) > 1:
            raise ValueError("All examples must have the same number of <eos> tokens.")
        batch_size, _, hidden_size = sequence_output.shape
        sentence_representation = sequence_output[eos_mask, :].view(batch_size, -1, hidden_size)[:, -1, :]
        logits = self.classification_head(sentence_representation)

        loss = None
        if labels is not None:
            labels = labels.to(logits.device)
            if self.config.problem_type is None:
                if self.config.num_labels == 1:
                    self.config.problem_type = "regression"
                elif self.config.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
                    self.config.problem_type = "single_label_classification"
                else:
                    self.config.problem_type = "multi_label_classification"

            if self.config.problem_type == "regression":
                loss_fct = MSELoss()
                if self.config.num_labels == 1:
                    loss = loss_fct(logits.squeeze(), labels.squeeze())
                else:
                    loss = loss_fct(logits, labels)
            elif self.config.problem_type == "single_label_classification":
                loss_fct = CrossEntropyLoss()
                loss = loss_fct(logits.view(-1, self.config.num_labels), labels.view(-1))
            elif self.config.problem_type == "multi_label_classification":
                loss_fct = BCEWithLogitsLoss()
                loss = loss_fct(logits, labels)
        if not return_dict:
            output = (logits,) + outputs[1:]
            return ((loss,) + output) if loss is not None else output

        return Seq2SeqSequenceClassifierOutput(
            loss=loss,
            logits=logits,
            past_key_values=outputs.past_key_values,
            decoder_hidden_states=outputs.decoder_hidden_states,
            decoder_attentions=outputs.decoder_attentions,
            cross_attentions=outputs.cross_attentions,
            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
            encoder_hidden_states=outputs.encoder_hidden_states,
            encoder_attentions=outputs.encoder_attentions,
        )


@add_start_docstrings(
    """
    MT5 Encoder Model with a token classification head on top (a linear layer on top of the hidden-states output)
    e.g. for Named-Entity-Recognition (NER) tasks.
    """,
    MT5_START_DOCSTRING,
)
class MT5ForTokenClassification(MT5PreTrainedModel):
    _tied_weights_keys = ["transformer.encoder.embed_tokens.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5ForTokenClassification.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.num_labels = config.num_labels

        self.transformer = MT5EncoderModel(config)
        self.dropout = nn.Dropout(config.classifier_dropout)
        self.classifier = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

    @add_start_docstrings_to_model_forward(MT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=TokenClassifierOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5ForTokenClassification.forward with T5->MT5
    def forward(
        self,
        input_ids: Optional[torch.Tensor] = None,
        attention_mask: Optional[torch.Tensor] = None,
        head_mask: Optional[torch.Tensor] = None,
        inputs_embeds: Optional[torch.Tensor] = None,
        labels: Optional[torch.Tensor] = None,
        output_attentions: Optional[bool] = None,
        output_hidden_states: Optional[bool] = None,
        return_dict: Optional[bool] = None,
    ) -> Union[Tuple[torch.Tensor], TokenClassifierOutput]:
        r"""
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        Returns:
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict

        outputs = self.transformer(
            input_ids,
            attention_mask=attention_mask,
            head_mask=head_mask,
            inputs_embeds=inputs_embeds,
            output_attentions=output_attentions,
            output_hidden_states=output_hidden_states,
            return_dict=return_dict,
        )

        hidden_states = outputs[0]
        hidden_states = self.dropout(hidden_states)
        logits = self.classifier(hidden_states)

        loss = None
        if labels is not None:
            loss_fct = CrossEntropyLoss()
            loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))

        if not return_dict:
            output = (logits, outputs[2:-1])
            return ((loss,) + output) if loss is not None else output

        return TokenClassifierOutput(
            loss=loss,
            logits=logits,
            hidden_states=outputs.hidden_states,
            attentions=outputs.attentions,
        )


@add_start_docstrings(
    """
    MT5 Model with a span classification head on top for extractive question-answering tasks like SQuAD (linear layers
    on top of the hidden-states output to compute `span start logits` and `span end logits`).
    """,
    MT5_START_DOCSTRING,
)
class MT5ForQuestionAnswering(MT5PreTrainedModel):
    _keys_to_ignore_on_load_unexpected = ["decoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weight"]
    _tied_weights_keys = ["encoder.embed_tokens.weight", "decoder.embed_tokens.weight"]

    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.__init__ with T5->MT5
    def __init__(self, config: MT5Config):
        super().__init__(config)
        self.model_dim = config.d_model

        self.shared = nn.Embedding(config.vocab_size, config.d_model)

        encoder_config = copy.deepcopy(config)
        encoder_config.is_decoder = False
        encoder_config.use_cache = False
        encoder_config.is_encoder_decoder = False
        self.encoder = MT5Stack(encoder_config, self.shared)

        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 = MT5Stack(decoder_config, self.shared)

        self.num_labels = config.num_labels
        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)

        # Initialize weights and apply final processing
        self.post_init()

        self.model_parallel = False

    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.get_input_embeddings
    def get_input_embeddings(self):
        return self.shared

    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.set_input_embeddings
    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)

    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.get_encoder
    def get_encoder(self):
        return self.encoder

    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.get_decoder
    def get_decoder(self):
        return self.decoder

    @add_start_docstrings_to_model_forward(MT5_INPUTS_DOCSTRING)
    @replace_return_docstrings(output_type=Seq2SeqQuestionAnsweringModelOutput, config_class=_CONFIG_FOR_DOC)
    # Copied from transformers.models.t5.modeling_t5.T5ForQuestionAnswering.forward
    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,
        start_positions: Optional[torch.LongTensor] = None,
        end_positions: Optional[torch.LongTensor] = None,
        inputs_embeds: Optional[torch.FloatTensor] = None,
        decoder_inputs_embeds: Optional[torch.FloatTensor] = 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], Seq2SeqQuestionAnsweringModelOutput]:
        r"""
        start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the start of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence
            are not taken into account for computing the loss.
        end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for position (index) of the end of the labelled span for computing the token classification loss.
            Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence
            are not taken into account for computing the loss.
        Returns:
        """
        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
        use_cache = use_cache if use_cache is not None else self.config.use_cache
        if start_positions is not None and end_positions is not None:
            use_cache = False

        # Copied from models.bart.modeling_bart.BartModel.forward
        #   different to other models, T5 automatically creates decoder_input_ids from
        #   input_ids if no decoder_input_ids are provided
        if decoder_input_ids is None and decoder_inputs_embeds is None:
            if input_ids is None:
                raise ValueError(
                    "If no `decoder_input_ids` or `decoder_inputs_embeds` are "
                    "passed, `input_ids` cannot be `None`. Please pass either "
                    "`input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`."
                )
            decoder_input_ids = self._shift_right(input_ids)

        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,
                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]

        # Decode
        decoder_outputs = self.decoder(
            input_ids=decoder_input_ids,
            attention_mask=decoder_attention_mask,
            inputs_embeds=decoder_inputs_embeds,
            past_key_values=None,
            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]

        logits = self.qa_outputs(sequence_output)
        start_logits, end_logits = logits.split(1, dim=-1)
        start_logits = start_logits.squeeze(-1).contiguous()
        end_logits = end_logits.squeeze(-1).contiguous()

        total_loss = None
        if start_positions is not None and end_positions is not None:
            # If we are on multi-GPU, split add a dimension
            if len(start_positions.size()) > 1:
                start_positions = start_positions.squeeze(-1).to(start_logits.device)
            if len(end_positions.size()) > 1:
                end_positions = end_positions.squeeze(-1).to(end_logits.device)
            # sometimes the start/end positions are outside our model inputs, we ignore these terms
            ignored_index = start_logits.size(1)
            start_positions = start_positions.clamp(0, ignored_index)
            end_positions = end_positions.clamp(0, ignored_index)

            loss_fct = CrossEntropyLoss(ignore_index=ignored_index)
            start_loss = loss_fct(start_logits, start_positions)
            end_loss = loss_fct(end_logits, end_positions)
            total_loss = (start_loss + end_loss) / 2

        if not return_dict:
            output = (start_logits, end_logits) + decoder_outputs[1:] + encoder_outputs
            return ((total_loss,) + output) if total_loss is not None else output

        return Seq2SeqQuestionAnsweringModelOutput(
            loss=total_loss,
            start_logits=start_logits,
            end_logits=end_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,
        )