Removed XLMModel inheritance from FlaubertModel(torch+tf) (#19432)

* FlaubertModel inheritance from XLMModel removed

* Fix style and add FlaubertPreTrainedModel to __init__

* Fix formatting issue

* Fix Typo and repo-consistency

* Fix style

* add FlaubertPreTrainedModel to TYPE_HINT

* fix repo consistency

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_tf_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_tf_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/flaubert/modeling_flaubert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* removed redundant Copied from comments

* added missing copied from comments
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

src/transformers/__init__.py

@@ -1273,6 +1273,7 @@ else:
             "FlaubertForTokenClassification",
             "FlaubertModel",
             "FlaubertWithLMHeadModel",
+            "FlaubertPreTrainedModel",
         ]
     )
     _import_structure["models.flava"].extend(
@@ -4141,6 +4142,7 @@ if TYPE_CHECKING:
             FlaubertForSequenceClassification,
             FlaubertForTokenClassification,
             FlaubertModel,
+            FlaubertPreTrainedModel,
             FlaubertWithLMHeadModel,
         )
         from .models.flava import (

src/transformers/models/flaubert/__init__.py

@@ -41,6 +41,7 @@ else:
         "FlaubertForTokenClassification",
         "FlaubertModel",
         "FlaubertWithLMHeadModel",
+        "FlaubertPreTrainedModel",
     ]
 
 try:
@@ -79,6 +80,7 @@ if TYPE_CHECKING:
             FlaubertForSequenceClassification,
             FlaubertForTokenClassification,
             FlaubertModel,
+            FlaubertPreTrainedModel,
             FlaubertWithLMHeadModel,
         )
 

src/transformers/models/flaubert/modeling_flaubert.py

@@ -14,24 +14,35 @@
 # limitations under the License.
 """ PyTorch Flaubert model, based on XLM."""
 
-
+import itertools
+import math
 import random
+from dataclasses import dataclass
 from typing import Dict, Optional, Tuple, Union
 
+import numpy as np
 import torch
 from torch import nn
-
-from ...modeling_outputs import BaseModelOutput
-from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging
-from ..xlm.modeling_xlm import (
-    XLMForMultipleChoice,
-    XLMForQuestionAnswering,
-    XLMForQuestionAnsweringSimple,
-    XLMForSequenceClassification,
-    XLMForTokenClassification,
-    XLMModel,
-    XLMWithLMHeadModel,
-    get_masks,
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import gelu
+from ...modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel, SequenceSummary, SQuADHead
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_code_sample_docstrings,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
 )
 from .configuration_flaubert import FlaubertConfig
 
@@ -51,6 +62,161 @@ FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST = [
 ]
 
 
+# Copied from transformers.models.xlm.modeling_xlm.create_sinusoidal_embeddings
+def create_sinusoidal_embeddings(n_pos, dim, out):
+    position_enc = np.array([[pos / np.power(10000, 2 * (j // 2) / dim) for j in range(dim)] for pos in range(n_pos)])
+    out[:, 0::2] = torch.FloatTensor(np.sin(position_enc[:, 0::2]))
+    out[:, 1::2] = torch.FloatTensor(np.cos(position_enc[:, 1::2]))
+    out.detach_()
+    out.requires_grad = False
+
+
+# Copied from transformers.models.xlm.modeling_xlm.get_masks
+def get_masks(slen, lengths, causal, padding_mask=None):
+    """
+    Generate hidden states mask, and optionally an attention mask.
+    """
+    alen = torch.arange(slen, dtype=torch.long, device=lengths.device)
+    if padding_mask is not None:
+        mask = padding_mask
+    else:
+        assert lengths.max().item() <= slen
+        mask = alen < lengths[:, None]
+
+    # attention mask is the same as mask, or triangular inferior attention (causal)
+    bs = lengths.size(0)
+    if causal:
+        attn_mask = alen[None, None, :].repeat(bs, slen, 1) <= alen[None, :, None]
+    else:
+        attn_mask = mask
+
+    # sanity check
+    assert mask.size() == (bs, slen)
+    assert causal is False or attn_mask.size() == (bs, slen, slen)
+
+    return mask, attn_mask
+
+
+# Copied from transformers.models.xlm.modeling_xlm.MultiHeadAttention
+class MultiHeadAttention(nn.Module):
+
+    NEW_ID = itertools.count()
+
+    def __init__(self, n_heads, dim, config):
+        super().__init__()
+        self.layer_id = next(MultiHeadAttention.NEW_ID)
+        self.dim = dim
+        self.n_heads = n_heads
+        self.dropout = config.attention_dropout
+        assert self.dim % self.n_heads == 0
+
+        self.q_lin = nn.Linear(dim, dim)
+        self.k_lin = nn.Linear(dim, dim)
+        self.v_lin = nn.Linear(dim, dim)
+        self.out_lin = nn.Linear(dim, dim)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        attention_head_size = self.dim // self.n_heads
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(heads, self.n_heads, attention_head_size, self.pruned_heads)
+        # Prune linear layers
+        self.q_lin = prune_linear_layer(self.q_lin, index)
+        self.k_lin = prune_linear_layer(self.k_lin, index)
+        self.v_lin = prune_linear_layer(self.v_lin, index)
+        self.out_lin = prune_linear_layer(self.out_lin, index, dim=1)
+        # Update hyper params
+        self.n_heads = self.n_heads - len(heads)
+        self.dim = attention_head_size * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(self, input, mask, kv=None, cache=None, head_mask=None, output_attentions=False):
+        """
+        Self-attention (if kv is None) or attention over source sentence (provided by kv).
+        """
+        # Input is (bs, qlen, dim)
+        # Mask is (bs, klen) (non-causal) or (bs, klen, klen)
+        bs, qlen, dim = input.size()
+        if kv is None:
+            klen = qlen if cache is None else cache["slen"] + qlen
+        else:
+            klen = kv.size(1)
+        # assert dim == self.dim, f'Dimensions do not match: {dim} input vs {self.dim} configured'
+        n_heads = self.n_heads
+        dim_per_head = self.dim // n_heads
+        mask_reshape = (bs, 1, qlen, klen) if mask.dim() == 3 else (bs, 1, 1, klen)
+
+        def shape(x):
+            """projection"""
+            return x.view(bs, -1, self.n_heads, dim_per_head).transpose(1, 2)
+
+        def unshape(x):
+            """compute context"""
+            return x.transpose(1, 2).contiguous().view(bs, -1, self.n_heads * dim_per_head)
+
+        q = shape(self.q_lin(input))  # (bs, n_heads, qlen, dim_per_head)
+        if kv is None:
+            k = shape(self.k_lin(input))  # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v_lin(input))  # (bs, n_heads, qlen, dim_per_head)
+        elif cache is None or self.layer_id not in cache:
+            k = v = kv
+            k = shape(self.k_lin(k))  # (bs, n_heads, qlen, dim_per_head)
+            v = shape(self.v_lin(v))  # (bs, n_heads, qlen, dim_per_head)
+
+        if cache is not None:
+            if self.layer_id in cache:
+                if kv is None:
+                    k_, v_ = cache[self.layer_id]
+                    k = torch.cat([k_, k], dim=2)  # (bs, n_heads, klen, dim_per_head)
+                    v = torch.cat([v_, v], dim=2)  # (bs, n_heads, klen, dim_per_head)
+                else:
+                    k, v = cache[self.layer_id]
+            cache[self.layer_id] = (k, v)
+
+        q = q / math.sqrt(dim_per_head)  # (bs, n_heads, qlen, dim_per_head)
+        scores = torch.matmul(q, k.transpose(2, 3))  # (bs, n_heads, qlen, klen)
+        mask = (mask == 0).view(mask_reshape).expand_as(scores)  # (bs, n_heads, qlen, klen)
+        scores.masked_fill_(mask, torch.finfo(scores.dtype).min)  # (bs, n_heads, qlen, klen)
+
+        weights = nn.functional.softmax(scores.float(), dim=-1).type_as(scores)  # (bs, n_heads, qlen, klen)
+        weights = nn.functional.dropout(weights, p=self.dropout, training=self.training)  # (bs, n_heads, qlen, klen)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            weights = weights * head_mask
+
+        context = torch.matmul(weights, v)  # (bs, n_heads, qlen, dim_per_head)
+        context = unshape(context)  # (bs, qlen, dim)
+
+        outputs = (self.out_lin(context),)
+        if output_attentions:
+            outputs = outputs + (weights,)
+        return outputs
+
+
+# Copied from transformers.models.xlm.modeling_xlm.TransformerFFN
+class TransformerFFN(nn.Module):
+    def __init__(self, in_dim, dim_hidden, out_dim, config):
+        super().__init__()
+        self.dropout = config.dropout
+        self.lin1 = nn.Linear(in_dim, dim_hidden)
+        self.lin2 = nn.Linear(dim_hidden, out_dim)
+        self.act = gelu if config.gelu_activation else nn.functional.relu
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+
+    def forward(self, input):
+        return apply_chunking_to_forward(self.ff_chunk, self.chunk_size_feed_forward, self.seq_len_dim, input)
+
+    def ff_chunk(self, input):
+        x = self.lin1(input)
+        x = self.act(x)
+        x = self.lin2(x)
+        x = nn.functional.dropout(x, p=self.dropout, training=self.training)
+        return x
+
+
 FLAUBERT_START_DOCSTRING = r"""
 
     This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
@@ -130,18 +296,186 @@ FLAUBERT_INPUTS_DOCSTRING = r"""
     "The bare Flaubert Model transformer outputting raw hidden-states without any specific head on top.",
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertModel(XLMModel):
+# Copied from transformers.models.xlm.modeling_xlm.XLMPredLayer with XLM->Flaubert
+class FlaubertPredLayer(nn.Module):
+    """
+    Prediction layer (cross_entropy or adaptive_softmax).
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.asm = config.asm
+        self.n_words = config.n_words
+        self.pad_index = config.pad_index
+        dim = config.emb_dim
+
+        if config.asm is False:
+            self.proj = nn.Linear(dim, config.n_words, bias=True)
+        else:
+            self.proj = nn.AdaptiveLogSoftmaxWithLoss(
+                in_features=dim,
+                n_classes=config.n_words,
+                cutoffs=config.asm_cutoffs,
+                div_value=config.asm_div_value,
+                head_bias=True,  # default is False
+            )
+
+    def forward(self, x, y=None):
+        """Compute the loss, and optionally the scores."""
+        outputs = ()
+        if self.asm is False:
+            scores = self.proj(x)
+            outputs = (scores,) + outputs
+            if y is not None:
+                loss = nn.functional.cross_entropy(scores.view(-1, self.n_words), y.view(-1), reduction="mean")
+                outputs = (loss,) + outputs
+        else:
+            scores = self.proj.log_prob(x)
+            outputs = (scores,) + outputs
+            if y is not None:
+                _, loss = self.proj(x, y)
+                outputs = (loss,) + outputs
+
+        return outputs
+
+
+# Copied from transformers.models.xlm.modeling_xlm.XLMPreTrainedModel with XLM->Flaubert
+class FlaubertPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
 
     config_class = FlaubertConfig
+    load_tf_weights = None
+    base_model_prefix = "transformer"
+
+    def __init__(self, *inputs, **kwargs):
+        super().__init__(*inputs, **kwargs)
+
+    @property
+    def dummy_inputs(self):
+        inputs_list = torch.tensor([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
+        attns_list = torch.tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        if self.config.use_lang_emb and self.config.n_langs > 1:
+            langs_list = torch.tensor([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
+        else:
+            langs_list = None
+        return {"input_ids": inputs_list, "attention_mask": attns_list, "langs": langs_list}
+
+    def _init_weights(self, module):
+        """Initialize the weights."""
+        if isinstance(module, nn.Embedding):
+            if self.config is not None and self.config.embed_init_std is not None:
+                nn.init.normal_(module.weight, mean=0, std=self.config.embed_init_std)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        if isinstance(module, nn.Linear):
+            if self.config is not None and self.config.init_std is not None:
+                nn.init.normal_(module.weight, mean=0, std=self.config.init_std)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0.0)
+        if isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+
+class FlaubertModel(FlaubertPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
 
     def __init__(self, config):  # , dico, is_encoder, with_output):
         super().__init__(config)
+
+        # encoder / decoder, output layer
+        self.is_encoder = config.is_encoder
+        self.is_decoder = not config.is_encoder
+        if self.is_decoder:
+            raise NotImplementedError("Currently Flaubert can only be used as an encoder")
+        # self.with_output = with_output
+        self.causal = config.causal
+
+        # dictionary / languages
+        self.n_langs = config.n_langs
+        self.use_lang_emb = config.use_lang_emb
+        self.n_words = config.n_words
+        self.eos_index = config.eos_index
+        self.pad_index = config.pad_index
+        # self.dico = dico
+        # self.id2lang = config.id2lang
+        # self.lang2id = config.lang2id
+        # assert len(self.dico) == self.n_words
+        # assert len(self.id2lang) == len(self.lang2id) == self.n_langs
+
+        # model parameters
+        self.dim = config.emb_dim  # 512 by default
+        self.hidden_dim = self.dim * 4  # 2048 by default
+        self.n_heads = config.n_heads  # 8 by default
+        self.n_layers = config.n_layers
+        self.dropout = config.dropout
+        self.attention_dropout = config.attention_dropout
+        assert self.dim % self.n_heads == 0, "transformer dim must be a multiple of n_heads"
+
+        # embeddings
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.dim)
+        if config.sinusoidal_embeddings:
+            create_sinusoidal_embeddings(config.max_position_embeddings, self.dim, out=self.position_embeddings.weight)
+        if config.n_langs > 1 and config.use_lang_emb:
+            self.lang_embeddings = nn.Embedding(self.n_langs, self.dim)
+        self.embeddings = nn.Embedding(self.n_words, self.dim, padding_idx=self.pad_index)
+        self.layer_norm_emb = nn.LayerNorm(self.dim, eps=config.layer_norm_eps)
+
+        # transformer layers
+        self.attentions = nn.ModuleList()
+        self.layer_norm1 = nn.ModuleList()
+        self.ffns = nn.ModuleList()
+        self.layer_norm2 = nn.ModuleList()
+        # if self.is_decoder:
+        #     self.layer_norm15 = nn.ModuleList()
+        #     self.encoder_attn = nn.ModuleList()
+
+        for _ in range(self.n_layers):
+            self.attentions.append(MultiHeadAttention(self.n_heads, self.dim, config=config))
+            self.layer_norm1.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
+            # if self.is_decoder:
+            #     self.layer_norm15.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
+            #     self.encoder_attn.append(MultiHeadAttention(self.n_heads, self.dim, dropout=self.attention_dropout))
+            self.ffns.append(TransformerFFN(self.dim, self.hidden_dim, self.dim, config=config))
+            self.layer_norm2.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
+
+        if hasattr(config, "pruned_heads"):
+            pruned_heads = config.pruned_heads.copy().items()
+            config.pruned_heads = {}
+            for layer, heads in pruned_heads:
+                if self.attentions[int(layer)].n_heads == config.n_heads:
+                    self.prune_heads({int(layer): list(map(int, heads))})
+
+        # Initialize weights and apply final processing
+        self.post_init()
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
         self.layerdrop = getattr(config, "layerdrop", 0.0)
         self.pre_norm = getattr(config, "pre_norm", False)
         self.register_buffer(
             "position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)), persistent=False
         )
 
+    # Copied from transformers.models.xlm.modeling_xlm.XLMModel.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.embeddings
+
+    # Copied from transformers.models.xlm.modeling_xlm.XLMModel.set_input_embeddings
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings = new_embeddings
+
+    # Copied from transformers.models.xlm.modeling_xlm.XLMModel._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.attentions[layer].prune_heads(heads)
+
     @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING)
     @add_code_sample_docstrings(
         processor_class=_TOKENIZER_FOR_DOC,
@@ -321,20 +655,95 @@ class FlaubertModel(XLMModel):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertWithLMHeadModel(XLMWithLMHeadModel):
-    """
-    This class overrides [`XLMWithLMHeadModel`]. Please check the superclass for the appropriate documentation
-    alongside usage examples.
-    """
-
-    config_class = FlaubertConfig
-
+# Copied transformers.models.xlm.modeling_xlm.XLMWithLMHeadModel with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertWithLMHeadModel(FlaubertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
         self.transformer = FlaubertModel(config)
+        self.pred_layer = FlaubertPredLayer(config)
+
         # Initialize weights and apply final processing
         self.post_init()
 
+    def get_output_embeddings(self):
+        return self.pred_layer.proj
+
+    def set_output_embeddings(self, new_embeddings):
+        self.pred_layer.proj = new_embeddings
+
+    def prepare_inputs_for_generation(self, input_ids, **kwargs):
+        mask_token_id = self.config.mask_token_id
+        lang_id = self.config.lang_id
+
+        effective_batch_size = input_ids.shape[0]
+        mask_token = torch.full((effective_batch_size, 1), mask_token_id, dtype=torch.long, device=input_ids.device)
+        input_ids = torch.cat([input_ids, mask_token], dim=1)
+        if lang_id is not None:
+            langs = torch.full_like(input_ids, lang_id)
+        else:
+            langs = None
+        return {"input_ids": input_ids, "langs": langs}
+
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+        mask="<special1>",
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, 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, MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set
+            `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100`
+            are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]`
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        output = transformer_outputs[0]
+        outputs = self.pred_layer(output, labels)  # (loss, logits) or (logits,) depending on if labels are provided.
+
+        if not return_dict:
+            return outputs + transformer_outputs[1:]
+
+        return MaskedLMOutput(
+            loss=outputs[0] if labels is not None else None,
+            logits=outputs[0] if labels is None else outputs[1],
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
 
 @add_start_docstrings(
     """
@@ -343,20 +752,102 @@ class FlaubertWithLMHeadModel(XLMWithLMHeadModel):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertForSequenceClassification(XLMForSequenceClassification):
-    """
-    This class overrides [`XLMForSequenceClassification`]. Please check the superclass for the appropriate
-    documentation alongside usage examples.
-    """
-
-    config_class = FlaubertConfig
-
+# Copied transformers.models.xlm.modeling_xlm.XLMForSequenceClassification with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertForSequenceClassification(FlaubertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
         self.transformer = FlaubertModel(config)
+        self.sequence_summary = SequenceSummary(config)
+
         # Initialize weights and apply final processing
         self.post_init()
 
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, 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, SequenceClassifierOutput]:
+        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 regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        output = transformer_outputs[0]
+        logits = self.sequence_summary(output)
+
+        loss = None
+        if labels is not None:
+            if self.config.problem_type is None:
+                if self.num_labels == 1:
+                    self.config.problem_type = "regression"
+                elif self.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.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.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,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
 
 @add_start_docstrings(
     """
@@ -365,20 +856,84 @@ class FlaubertForSequenceClassification(XLMForSequenceClassification):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertForTokenClassification(XLMForTokenClassification):
-    """
-    This class overrides [`XLMForTokenClassification`]. Please check the superclass for the appropriate documentation
-    alongside usage examples.
-    """
-
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_xlm.XLMForTokenClassification with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertForTokenClassification(FlaubertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+        self.num_labels = config.num_labels
+
         self.transformer = FlaubertModel(config)
+        self.dropout = nn.Dropout(config.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(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, 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, 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]`.
+        """
+        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,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+
+        sequence_output = self.dropout(sequence_output)
+        logits = self.classifier(sequence_output)
+
+        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[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(
     """
@@ -387,20 +942,104 @@ class FlaubertForTokenClassification(XLMForTokenClassification):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertForQuestionAnsweringSimple(XLMForQuestionAnsweringSimple):
-    """
-    This class overrides [`XLMForQuestionAnsweringSimple`]. Please check the superclass for the appropriate
-    documentation alongside usage examples.
-    """
-
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_xlm.XLMForQuestionAnsweringSimple with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertForQuestionAnsweringSimple(FlaubertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+
         self.transformer = FlaubertModel(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
         # Initialize weights and apply final processing
         self.post_init()
 
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, QuestionAnsweringModelOutput]:
+        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.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = transformer_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)
+            if len(end_positions.size()) > 1:
+                end_positions = end_positions.squeeze(-1)
+            # 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) + transformer_outputs[1:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
 
 @add_start_docstrings(
     """
@@ -409,20 +1048,164 @@ class FlaubertForQuestionAnsweringSimple(XLMForQuestionAnsweringSimple):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertForQuestionAnswering(XLMForQuestionAnswering):
+@dataclass
+# Copied from transformer.models.xlm.modeling_xlm.XLMForQuestionAnsweringOutput with XLM->Flaubert
+class FlaubertForQuestionAnsweringOutput(ModelOutput):
     """
-    This class overrides [`XLMForQuestionAnswering`]. Please check the superclass for the appropriate documentation
-    alongside usage examples.
+    Base class for outputs of question answering models using a `SquadHead`.
+
+    Args:
+        loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned if both `start_positions` and `end_positions` are provided):
+            Classification loss as the sum of start token, end token (and is_impossible if provided) classification
+            losses.
+        start_top_log_probs (`torch.FloatTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
+            Log probabilities for the top config.start_n_top start token possibilities (beam-search).
+        start_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
+            Indices for the top config.start_n_top start token possibilities (beam-search).
+        end_top_log_probs (`torch.FloatTensor` of shape `(batch_size, config.start_n_top * config.end_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
+            Log probabilities for the top `config.start_n_top * config.end_n_top` end token possibilities
+            (beam-search).
+        end_top_index (`torch.LongTensor` of shape `(batch_size, config.start_n_top * config.end_n_top)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
+            Indices for the top `config.start_n_top * config.end_n_top` end token possibilities (beam-search).
+        cls_logits (`torch.FloatTensor` of shape `(batch_size,)`, *optional*, returned if `start_positions` or `end_positions` is not provided):
+            Log probabilities for the `is_impossible` label of the answers.
+        hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+            Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of
+            shape `(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+            Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+            sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
     """
 
-    config_class = FlaubertConfig
+    loss: Optional[torch.FloatTensor] = None
+    start_top_log_probs: Optional[torch.FloatTensor] = None
+    start_top_index: Optional[torch.LongTensor] = None
+    end_top_log_probs: Optional[torch.FloatTensor] = None
+    end_top_index: Optional[torch.LongTensor] = None
+    cls_logits: Optional[torch.FloatTensor] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+
 
+# Copied from transformer.models.xlm.modeling_xlm.XLMForQuestionAnswering with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertForQuestionAnswering(FlaubertPreTrainedModel):
     def __init__(self, config):
         super().__init__(config)
+
         self.transformer = FlaubertModel(config)
+        self.qa_outputs = SQuADHead(config)
+
         # Initialize weights and apply final processing
         self.post_init()
 
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @replace_return_docstrings(output_type=FlaubertForQuestionAnsweringOutput, config_class=_CONFIG_FOR_DOC)
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, torch.Tensor]] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        start_positions: Optional[torch.Tensor] = None,
+        end_positions: Optional[torch.Tensor] = None,
+        is_impossible: Optional[torch.Tensor] = None,
+        cls_index: Optional[torch.Tensor] = None,
+        p_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, FlaubertForQuestionAnsweringOutput]:
+        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.
+        is_impossible (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels whether a question has an answer or no answer (SQuAD 2.0)
+        cls_index (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for position (index) of the classification token to use as input for computing plausibility of the
+            answer.
+        p_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Optional mask of tokens which can't be in answers (e.g. [CLS], [PAD], ...). 1.0 means token should be
+            masked. 0.0 mean token is not masked.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> from transformers import XLMTokenizer, XLMForQuestionAnswering
+        >>> import torch
+
+        >>> tokenizer = XLMTokenizer.from_pretrained("xlm-mlm-en-2048")
+        >>> model = XLMForQuestionAnswering.from_pretrained("xlm-mlm-en-2048")
+
+        >>> input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(
+        ...     0
+        ... )  # Batch size 1
+        >>> start_positions = torch.tensor([1])
+        >>> end_positions = torch.tensor([3])
+
+        >>> outputs = model(input_ids, start_positions=start_positions, end_positions=end_positions)
+        >>> loss = outputs.loss
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        transformer_outputs = self.transformer(
+            input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        output = transformer_outputs[0]
+
+        outputs = self.qa_outputs(
+            output,
+            start_positions=start_positions,
+            end_positions=end_positions,
+            cls_index=cls_index,
+            is_impossible=is_impossible,
+            p_mask=p_mask,
+            return_dict=return_dict,
+        )
+
+        if not return_dict:
+            return outputs + transformer_outputs[1:]
+
+        return FlaubertForQuestionAnsweringOutput(
+            loss=outputs.loss,
+            start_top_log_probs=outputs.start_top_log_probs,
+            start_top_index=outputs.start_top_index,
+            end_top_log_probs=outputs.end_top_log_probs,
+            end_top_index=outputs.end_top_index,
+            cls_logits=outputs.cls_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
 
 @add_start_docstrings(
     """
@@ -431,16 +1214,101 @@ class FlaubertForQuestionAnswering(XLMForQuestionAnswering):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class FlaubertForMultipleChoice(XLMForMultipleChoice):
-    """
-    This class overrides [`XLMForMultipleChoice`]. Please check the superclass for the appropriate documentation
-    alongside usage examples.
-    """
+# Copied from transformer.models.xlm.modeling_xlm.XLMForMultipleChoice with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class FlaubertForMultipleChoice(FlaubertPreTrainedModel):
+    def __init__(self, config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
 
-    config_class = FlaubertConfig
-
-    def __init__(self, config):
-        super().__init__(config)
         self.transformer = FlaubertModel(config)
+        self.sequence_summary = SequenceSummary(config)
+        self.logits_proj = nn.Linear(config.num_labels, 1)
+
         # Initialize weights and apply final processing
         self.post_init()
+
+    @add_start_docstrings_to_model_forward(
+        FLAUBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        langs: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        lengths: Optional[torch.Tensor] = None,
+        cache: Optional[Dict[str, 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, MultipleChoiceModelOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
+            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
+            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
+            `input_ids` above)
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        num_choices = input_ids.shape[1] if input_ids is not None else inputs_embeds.shape[1]
+
+        input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        langs = langs.view(-1, langs.size(-1)) if langs is not None else None
+        inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        if lengths is not None:
+            logger.warning(
+                "The `lengths` parameter cannot be used with the Flaubert multiple choice models. Please use the "
+                "attention mask instead."
+            )
+            lengths = None
+
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        output = transformer_outputs[0]
+        logits = self.sequence_summary(output)
+        logits = self.logits_proj(logits)
+        reshaped_logits = logits.view(-1, num_choices)
+
+        loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            loss = loss_fct(reshaped_logits, labels)
+
+        if not return_dict:
+            output = (reshaped_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )

src/transformers/models/flaubert/modeling_tf_flaubert.py

@@ -26,28 +26,35 @@ import numpy as np
 import tensorflow as tf
 
 from ...activations_tf import get_tf_activation
-from ...modeling_tf_outputs import TFBaseModelOutput
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFMultipleChoiceModelOutput,
+    TFQuestionAnsweringModelOutput,
+    TFSequenceClassifierOutput,
+    TFTokenClassifierOutput,
+)
 from ...modeling_tf_utils import (
+    TFModelInputType,
+    TFMultipleChoiceLoss,
     TFPreTrainedModel,
+    TFQuestionAnsweringLoss,
+    TFSequenceClassificationLoss,
+    TFSequenceSummary,
     TFSharedEmbeddings,
+    TFTokenClassificationLoss,
     get_initializer,
     keras_serializable,
     unpack_inputs,
 )
 from ...tf_utils import shape_list, stable_softmax
 from ...utils import (
+    MULTIPLE_CHOICE_DUMMY_INPUTS,
     ModelOutput,
     add_code_sample_docstrings,
     add_start_docstrings,
     add_start_docstrings_to_model_forward,
     logging,
 )
-from ..xlm.modeling_tf_xlm import (
-    TFXLMForMultipleChoice,
-    TFXLMForQuestionAnsweringSimple,
-    TFXLMForSequenceClassification,
-    TFXLMForTokenClassification,
-)
 from .configuration_flaubert import FlaubertConfig
 
 
@@ -218,7 +225,7 @@ class TFFlaubertPreTrainedModel(TFPreTrainedModel):
 
     @property
     def dummy_inputs(self):
-        # Sometimes XLM has language embeddings so don't forget to build them as well if needed
+        # Sometimes Flaubert has language embeddings so don't forget to build them as well if needed
         inputs_list = tf.constant([[7, 6, 0, 0, 1], [1, 2, 3, 0, 0], [0, 0, 0, 4, 5]])
         attns_list = tf.constant([[1, 1, 0, 0, 1], [1, 1, 1, 0, 0], [1, 0, 0, 1, 1]])
         if self.config.use_lang_emb and self.config.n_langs > 1:
@@ -862,12 +869,84 @@ class TFFlaubertWithLMHeadModel(TFFlaubertPreTrainedModel):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class TFFlaubertForSequenceClassification(TFXLMForSequenceClassification):
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMForSequenceClassification with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class TFFlaubertForSequenceClassification(TFFlaubertPreTrainedModel, TFSequenceClassificationLoss):
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
         self.transformer = TFFlaubertMainLayer(config, name="transformer")
+        self.sequence_summary = TFSequenceSummary(config, initializer_range=config.init_std, name="sequence_summary")
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFSequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        langs: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        lengths: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        cache: Optional[Dict[str, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: bool = False,
+    ):
+        r"""
+        labels (`tf.Tensor` 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 regression loss is computed (Mean-Square loss), If
+            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+        """
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        output = transformer_outputs[0]
+
+        logits = self.sequence_summary(output)
+
+        loss = None if labels is None else self.hf_compute_loss(labels, logits)
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForSequenceClassification.serving_output
+    def serving_output(self, output: TFSequenceClassifierOutput) -> TFSequenceClassifierOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFSequenceClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
 
 
 @add_start_docstrings(
@@ -877,12 +956,99 @@ class TFFlaubertForSequenceClassification(TFXLMForSequenceClassification):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class TFFlaubertForQuestionAnsweringSimple(TFXLMForQuestionAnsweringSimple):
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMForQuestionAnsweringSimple with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class TFFlaubertForQuestionAnsweringSimple(TFFlaubertPreTrainedModel, TFQuestionAnsweringLoss):
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
         self.transformer = TFFlaubertMainLayer(config, name="transformer")
+        self.qa_outputs = tf.keras.layers.Dense(
+            config.num_labels, kernel_initializer=get_initializer(config.init_std), name="qa_outputs"
+        )
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFQuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        langs: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        lengths: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        cache: Optional[Dict[str, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        start_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        end_positions: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: bool = False,
+    ):
+        r"""
+        start_positions (`tf.Tensor` 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 (`tf.Tensor` 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.
+        """
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = transformer_outputs[0]
+
+        logits = self.qa_outputs(sequence_output)
+        start_logits, end_logits = tf.split(logits, 2, axis=-1)
+        start_logits = tf.squeeze(start_logits, axis=-1)
+        end_logits = tf.squeeze(end_logits, axis=-1)
+
+        loss = None
+        if start_positions is not None and end_positions is not None:
+            labels = {"start_position": start_positions}
+            labels["end_position"] = end_positions
+            loss = self.hf_compute_loss(labels, (start_logits, end_logits))
+
+        if not return_dict:
+            output = (start_logits, end_logits) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFQuestionAnsweringModelOutput(
+            loss=loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForQuestionAnswering.serving_output
+    def serving_output(self, output: TFQuestionAnsweringModelOutput) -> TFQuestionAnsweringModelOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFQuestionAnsweringModelOutput(
+            start_logits=output.start_logits, end_logits=output.end_logits, hidden_states=hs, attentions=attns
+        )
 
 
 @add_start_docstrings(
@@ -892,12 +1058,86 @@ class TFFlaubertForQuestionAnsweringSimple(TFXLMForQuestionAnsweringSimple):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class TFFlaubertForTokenClassification(TFXLMForTokenClassification):
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMForTokenClassification with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class TFFlaubertForTokenClassification(TFFlaubertPreTrainedModel, TFTokenClassificationLoss):
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
+        self.num_labels = config.num_labels
+
         self.transformer = TFFlaubertMainLayer(config, name="transformer")
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.classifier = tf.keras.layers.Dense(
+            config.num_labels, kernel_initializer=get_initializer(config.init_std), name="classifier"
+        )
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(FLAUBERT_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFTokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        langs: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        lengths: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        cache: Optional[Dict[str, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: bool = False,
+    ):
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
+        """
+        transformer_outputs = self.transformer(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            langs=langs,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            lengths=lengths,
+            cache=cache,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        sequence_output = transformer_outputs[0]
+
+        sequence_output = self.dropout(sequence_output, training=training)
+        logits = self.classifier(sequence_output)
+
+        loss = None if labels is None else self.hf_compute_loss(labels, logits)
+
+        if not return_dict:
+            output = (logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFTokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForTokenClassification.serving_output
+    def serving_output(self, output: TFTokenClassifierOutput) -> TFTokenClassifierOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFTokenClassifierOutput(logits=output.logits, hidden_states=hs, attentions=attns)
 
 
 @add_start_docstrings(
@@ -907,9 +1147,139 @@ class TFFlaubertForTokenClassification(TFXLMForTokenClassification):
     """,
     FLAUBERT_START_DOCSTRING,
 )
-class TFFlaubertForMultipleChoice(TFXLMForMultipleChoice):
-    config_class = FlaubertConfig
-
+# Copied from transformers.models.xlm.modeling_tf_xlm.TFXLMForMultipleChoice with XLM_INPUTS->FLAUBERT_INPUTS,XLM->Flaubert
+class TFFlaubertForMultipleChoice(TFFlaubertPreTrainedModel, TFMultipleChoiceLoss):
     def __init__(self, config, *inputs, **kwargs):
         super().__init__(config, *inputs, **kwargs)
+
         self.transformer = TFFlaubertMainLayer(config, name="transformer")
+        self.sequence_summary = TFSequenceSummary(config, initializer_range=config.init_std, name="sequence_summary")
+        self.logits_proj = tf.keras.layers.Dense(
+            1, kernel_initializer=get_initializer(config.initializer_range), name="logits_proj"
+        )
+
+    @property
+    def dummy_inputs(self):
+        """
+        Dummy inputs to build the network.
+
+        Returns:
+            tf.Tensor with dummy inputs
+        """
+        # Sometimes Flaubert has language embeddings so don't forget to build them as well if needed
+        if self.config.use_lang_emb and self.config.n_langs > 1:
+            return {
+                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
+                "langs": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
+            }
+        else:
+            return {
+                "input_ids": tf.constant(MULTIPLE_CHOICE_DUMMY_INPUTS),
+            }
+
+    @unpack_inputs
+    @add_start_docstrings_to_model_forward(
+        FLAUBERT_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length")
+    )
+    @add_code_sample_docstrings(
+        processor_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TFMultipleChoiceModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def call(
+        self,
+        input_ids: Optional[TFModelInputType] = None,
+        attention_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        langs: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        token_type_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        position_ids: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        lengths: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        cache: Optional[Dict[str, tf.Tensor]] = None,
+        head_mask: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        inputs_embeds: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        labels: Optional[Union[np.ndarray, tf.Tensor]] = None,
+        training: bool = False,
+    ):
+        if input_ids is not None:
+            num_choices = shape_list(input_ids)[1]
+            seq_length = shape_list(input_ids)[2]
+        else:
+            num_choices = shape_list(inputs_embeds)[1]
+            seq_length = shape_list(inputs_embeds)[2]
+
+        flat_input_ids = tf.reshape(input_ids, (-1, seq_length)) if input_ids is not None else None
+        flat_attention_mask = tf.reshape(attention_mask, (-1, seq_length)) if attention_mask is not None else None
+        flat_token_type_ids = tf.reshape(token_type_ids, (-1, seq_length)) if token_type_ids is not None else None
+        flat_position_ids = tf.reshape(position_ids, (-1, seq_length)) if position_ids is not None else None
+        flat_langs = tf.reshape(langs, (-1, seq_length)) if langs is not None else None
+        flat_inputs_embeds = (
+            tf.reshape(inputs_embeds, (-1, seq_length, shape_list(inputs_embeds)[3]))
+            if inputs_embeds is not None
+            else None
+        )
+
+        if lengths is not None:
+            logger.warning(
+                "The `lengths` parameter cannot be used with the Flaubert multiple choice models. Please use the "
+                "attention mask instead.",
+            )
+            lengths = None
+
+        transformer_outputs = self.transformer(
+            flat_input_ids,
+            flat_attention_mask,
+            flat_langs,
+            flat_token_type_ids,
+            flat_position_ids,
+            lengths,
+            cache,
+            head_mask,
+            flat_inputs_embeds,
+            output_attentions,
+            output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        output = transformer_outputs[0]
+        logits = self.sequence_summary(output)
+        logits = self.logits_proj(logits)
+        reshaped_logits = tf.reshape(logits, (-1, num_choices))
+
+        loss = None if labels is None else self.hf_compute_loss(labels, reshaped_logits)
+
+        if not return_dict:
+            output = (reshaped_logits,) + transformer_outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFMultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=transformer_outputs.hidden_states,
+            attentions=transformer_outputs.attentions,
+        )
+
+    @tf.function(
+        input_signature=[
+            {
+                "input_ids": tf.TensorSpec((None, None, None), tf.int32, name="input_ids"),
+                "attention_mask": tf.TensorSpec((None, None, None), tf.int32, name="attention_mask"),
+                "token_type_ids": tf.TensorSpec((None, None, None), tf.int32, name="token_type_ids"),
+            }
+        ]
+    )
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving
+    def serving(self, inputs: Dict[str, tf.Tensor]):
+        output = self.call(input_ids=inputs)
+
+        return self.serving_output(output)
+
+    # Copied from transformers.models.bert.modeling_tf_bert.TFBertForMultipleChoice.serving_output
+    def serving_output(self, output: TFMultipleChoiceModelOutput) -> TFMultipleChoiceModelOutput:
+        hs = tf.convert_to_tensor(output.hidden_states) if self.config.output_hidden_states else None
+        attns = tf.convert_to_tensor(output.attentions) if self.config.output_attentions else None
+
+        return TFMultipleChoiceModelOutput(logits=output.logits, hidden_states=hs, attentions=attns)

src/transformers/utils/dummy_pt_objects.py

@@ -2065,6 +2065,13 @@ class FlaubertModel(metaclass=DummyObject):
         requires_backends(self, ["torch"])
 
 
+class FlaubertPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class FlaubertWithLMHeadModel(metaclass=DummyObject):
     _backends = ["torch"]