Add X-MOD (#20939)

* Add X-MOD to Readme

* Add documentation for X-MOD

* Implement X-MOD

* Fix formatting of X-MOD docs

* Change signature of X-MOD forward methods to use lang_ids

* Minor changes

* Rebase with main and run make fix-copies

* Make suggested changes to docstrings

* Improve code readability
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

* Fix code style

* Conversion script: Remove asserts and type annotations

* Remove _TOKENIZER_FOR_DOC

* XMOD -> Xmod

* Update copyright note

* Fix doctests

* Fix docstring

* Add integration test for FillMaskPipeline

* Revert "Add integration test for FillMaskPipeline"

This reverts commit 4381eb3b1d0f5d85785f89caba83928e6efa6d1f.

* Add end-to-end integration test for mask fill

* make style

* Rebase with main and make fix-copies

---------
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>

src/transformers/models/auto/modeling_auto.py

@@ -184,6 +184,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaModel"),
         ("xlm-roberta-xl", "XLMRobertaXLModel"),
         ("xlnet", "XLNetModel"),
+        ("xmod", "XmodModel"),
         ("yolos", "YolosModel"),
         ("yoso", "YosoModel"),
     ]
@@ -244,6 +245,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForMaskedLM"),
         ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
         ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForMaskedLM"),
     ]
 )
 
@@ -317,6 +319,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForMaskedLM"),
         ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
         ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForMaskedLM"),
         ("yoso", "YosoForMaskedLM"),
     ]
 )
@@ -371,6 +374,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForCausalLM"),
         ("xlm-roberta-xl", "XLMRobertaXLForCausalLM"),
         ("xlnet", "XLNetLMHeadModel"),
+        ("xmod", "XmodForCausalLM"),
     ]
 )
 
@@ -531,6 +535,7 @@ MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
         ("xlm", "XLMWithLMHeadModel"),
         ("xlm-roberta", "XLMRobertaForMaskedLM"),
         ("xlm-roberta-xl", "XLMRobertaXLForMaskedLM"),
+        ("xmod", "XmodForMaskedLM"),
         ("yoso", "YosoForMaskedLM"),
     ]
 )
@@ -658,6 +663,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForSequenceClassification"),
         ("xlm-roberta-xl", "XLMRobertaXLForSequenceClassification"),
         ("xlnet", "XLNetForSequenceClassification"),
+        ("xmod", "XmodForSequenceClassification"),
         ("yoso", "YosoForSequenceClassification"),
     ]
 )
@@ -714,6 +720,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForQuestionAnswering"),
         ("xlm-roberta-xl", "XLMRobertaXLForQuestionAnswering"),
         ("xlnet", "XLNetForQuestionAnsweringSimple"),
+        ("xmod", "XmodForQuestionAnswering"),
         ("yoso", "YosoForQuestionAnswering"),
     ]
 )
@@ -785,6 +792,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForTokenClassification"),
         ("xlm-roberta-xl", "XLMRobertaXLForTokenClassification"),
         ("xlnet", "XLNetForTokenClassification"),
+        ("xmod", "XmodForTokenClassification"),
         ("yoso", "YosoForTokenClassification"),
     ]
 )
@@ -825,6 +833,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
         ("xlm-roberta", "XLMRobertaForMultipleChoice"),
         ("xlm-roberta-xl", "XLMRobertaXLForMultipleChoice"),
         ("xlnet", "XLNetForMultipleChoice"),
+        ("xmod", "XmodForMultipleChoice"),
         ("yoso", "YosoForMultipleChoice"),
     ]
 )

src/transformers/models/auto/tokenization_auto.py

@@ -325,6 +325,13 @@ else:
                     "XLNetTokenizerFast" if is_tokenizers_available() else None,
                 ),
             ),
+            (
+                "xmod",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             (
                 "yoso",
                 (

src/transformers/models/xmod/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import TYPE_CHECKING
+
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_xmod": [
+        "XMOD_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "XmodConfig",
+        "XmodOnnxConfig",
+    ],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_xmod"] = [
+        "XMOD_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "XmodForCausalLM",
+        "XmodForMaskedLM",
+        "XmodForMultipleChoice",
+        "XmodForQuestionAnswering",
+        "XmodForSequenceClassification",
+        "XmodForTokenClassification",
+        "XmodModel",
+        "XmodPreTrainedModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_xmod import XMOD_PRETRAINED_CONFIG_ARCHIVE_MAP, XmodConfig, XmodOnnxConfig
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_xmod import (
+            XMOD_PRETRAINED_MODEL_ARCHIVE_LIST,
+            XmodForCausalLM,
+            XmodForMaskedLM,
+            XmodForMultipleChoice,
+            XmodForQuestionAnswering,
+            XmodForSequenceClassification,
+            XmodForTokenClassification,
+            XmodModel,
+            XmodPreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/xmod/configuration_xmod.py

+# coding=utf-8
+# Copyright 2023 The Meta AI Team Authors and The HuggingFace Inc. team.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" X-MOD configuration"""
+from collections import OrderedDict
+from typing import Mapping
+
+from ...configuration_utils import PretrainedConfig
+from ...onnx import OnnxConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+XMOD_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "jvamvas/xmod-base": "https://huggingface.co/jvamvas/xmod-base/resolve/main/config.json",
+    "jvamvas/xmod-large-prenorm": "https://huggingface.co/jvamvas/xmod-large-prenorm/resolve/main/config.json",
+    "jvamvas/xmod-base-13-125k": "https://huggingface.co/jvamvas/xmod-base-13-125k/resolve/main/config.json",
+    "jvamvas/xmod-base-30-125k": "https://huggingface.co/jvamvas/xmod-base-30-125k/resolve/main/config.json",
+    "jvamvas/xmod-base-30-195k": "https://huggingface.co/jvamvas/xmod-base-30-195k/resolve/main/config.json",
+    "jvamvas/xmod-base-60-125k": "https://huggingface.co/jvamvas/xmod-base-60-125k/resolve/main/config.json",
+    "jvamvas/xmod-base-60-265k": "https://huggingface.co/jvamvas/xmod-base-60-265k/resolve/main/config.json",
+    "jvamvas/xmod-base-75-125k": "https://huggingface.co/jvamvas/xmod-base-75-125k/resolve/main/config.json",
+    "jvamvas/xmod-base-75-269k": "https://huggingface.co/jvamvas/xmod-base-75-269k/resolve/main/config.json",
+}
+
+
+class XmodConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`XmodModel`]. It is used to instantiate an X-MOD
+    model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the [xmod-base](https://huggingface.co/jvamvas/xmod-base)
+    architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the X-MOD model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`XmodModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`XmodModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        is_decoder (`bool`, *optional*, defaults to `False`):
+            Whether the model is used as a decoder or not. If `False`, the model is used as an encoder.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+        pre_norm (`bool`, *optional*, defaults to `False`):
+            Whether to apply layer normalization before each block.
+        adapter_reduction_factor (`int` or `float`, *optional*, defaults to 2):
+            The factor by which the dimensionality of the adapter is reduced relative to `hidden_size`.
+        adapter_layer_norm (`bool`, *optional*, defaults to `False`):
+            Whether to apply a new layer normalization before the adapter modules (shared across all adapters).
+        adapter_reuse_layer_norm (`bool`, *optional*, defaults to `True`):
+            Whether to reuse the second layer normalization and apply it before the adapter modules as well.
+        ln_before_adapter (`bool`, *optional*, defaults to `True`):
+            Whether to apply the layer normalization before the residual connection around the adapter module.
+        languages (`Iterable[str]`, *optional*, defaults to `["en_XX"]`):
+            An iterable of language codes for which adapter modules should be initialized.
+        default_language (`str`, *optional*):
+            Language code of a default language. It will be assumed that the input is in this language if no language
+            codes are explicitly passed to the forward method.
+
+    Examples:
+
+    ```python
+    >>> from transformers import XmodConfig, XmodModel
+
+    >>> # Initializing an X-MOD jvamvas/xmod-base style configuration
+    >>> configuration = XmodConfig()
+
+    >>> # Initializing a model (with random weights) from the jvamvas/xmod-base style configuration
+    >>> model = XmodModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "xmod"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        position_embedding_type="absolute",
+        use_cache=True,
+        classifier_dropout=None,
+        pre_norm=False,
+        adapter_reduction_factor=2,
+        adapter_layer_norm=False,
+        adapter_reuse_layer_norm=True,
+        ln_before_adapter=True,
+        languages=("en_XX",),
+        default_language=None,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.classifier_dropout = classifier_dropout
+        self.pre_norm = pre_norm
+        self.adapter_reduction_factor = adapter_reduction_factor
+        self.adapter_layer_norm = adapter_layer_norm
+        self.adapter_reuse_layer_norm = adapter_reuse_layer_norm
+        self.ln_before_adapter = ln_before_adapter
+        self.languages = list(languages)
+        self.default_language = default_language
+
+
+# Copied from transformers.models.roberta.configuration_roberta.RobertaOnnxConfig with Roberta->Xmod
+class XmodOnnxConfig(OnnxConfig):
+    @property
+    def inputs(self) -> Mapping[str, Mapping[int, str]]:
+        if self.task == "multiple-choice":
+            dynamic_axis = {0: "batch", 1: "choice", 2: "sequence"}
+        else:
+            dynamic_axis = {0: "batch", 1: "sequence"}
+        return OrderedDict(
+            [
+                ("input_ids", dynamic_axis),
+                ("attention_mask", dynamic_axis),
+            ]
+        )

src/transformers/models/xmod/convert_xmod_original_pytorch_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert X-MOD checkpoint."""
+
+import argparse
+from pathlib import Path
+
+import fairseq
+import torch
+from fairseq.models.xmod import XMODModel as FairseqXmodModel
+from packaging import version
+
+from transformers import XmodConfig, XmodForMaskedLM, XmodForSequenceClassification
+from transformers.utils import logging
+
+
+if version.parse(fairseq.__version__) < version.parse("0.12.2"):
+    raise Exception("requires fairseq >= 0.12.2")
+if version.parse(fairseq.__version__) > version.parse("2"):
+    raise Exception("requires fairseq < v2")
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+SAMPLE_TEXT = "Hello, World!"
+SAMPLE_LANGUAGE = "en_XX"
+
+
+def convert_xmod_checkpoint_to_pytorch(
+    xmod_checkpoint_path: str, pytorch_dump_folder_path: str, classification_head: bool
+):
+    data_dir = Path("data_bin")
+    xmod = FairseqXmodModel.from_pretrained(
+        model_name_or_path=str(Path(xmod_checkpoint_path).parent),
+        checkpoint_file=Path(xmod_checkpoint_path).name,
+        _name="xmod_base",
+        arch="xmod_base",
+        task="multilingual_masked_lm",
+        data_name_or_path=str(data_dir),
+        bpe="sentencepiece",
+        sentencepiece_model=str(Path(xmod_checkpoint_path).parent / "sentencepiece.bpe.model"),
+        src_dict=str(data_dir / "dict.txt"),
+    )
+    xmod.eval()  # disable dropout
+    print(xmod)
+
+    xmod_sent_encoder = xmod.model.encoder.sentence_encoder
+    config = XmodConfig(
+        vocab_size=xmod_sent_encoder.embed_tokens.num_embeddings,
+        hidden_size=xmod.cfg.model.encoder_embed_dim,
+        num_hidden_layers=xmod.cfg.model.encoder_layers,
+        num_attention_heads=xmod.cfg.model.encoder_attention_heads,
+        intermediate_size=xmod.cfg.model.encoder_ffn_embed_dim,
+        max_position_embeddings=514,
+        type_vocab_size=1,
+        layer_norm_eps=1e-5,  # PyTorch default used in fairseq
+        pre_norm=xmod.cfg.model.encoder_normalize_before,
+        adapter_reduction_factor=getattr(xmod.cfg.model, "bottleneck", 2),
+        adapter_layer_norm=xmod.cfg.model.adapter_layer_norm,
+        adapter_reuse_layer_norm=xmod.cfg.model.adapter_reuse_layer_norm,
+        ln_before_adapter=xmod.cfg.model.ln_before_adapter,
+        languages=xmod.cfg.model.languages,
+    )
+    if classification_head:
+        config.num_labels = xmod.model.classification_heads["mnli"].out_proj.weight.shape[0]
+
+    print("Our X-MOD config:", config)
+
+    model = XmodForSequenceClassification(config) if classification_head else XmodForMaskedLM(config)
+    model.eval()
+
+    # Now let's copy all the weights.
+    # Embeddings
+    model.roberta.embeddings.word_embeddings.weight = xmod_sent_encoder.embed_tokens.weight
+    model.roberta.embeddings.position_embeddings.weight = xmod_sent_encoder.embed_positions.weight
+    model.roberta.embeddings.token_type_embeddings.weight.data = torch.zeros_like(
+        model.roberta.embeddings.token_type_embeddings.weight
+    )  # just zero them out b/c xmod doesn't use them.
+
+    model.roberta.embeddings.LayerNorm.weight = xmod_sent_encoder.layernorm_embedding.weight
+    model.roberta.embeddings.LayerNorm.bias = xmod_sent_encoder.layernorm_embedding.bias
+
+    for i in range(config.num_hidden_layers):
+        # Encoder: start of layer
+        layer = model.roberta.encoder.layer[i]
+        xmod_layer = xmod_sent_encoder.layers[i]
+
+        # self attention
+        self_attn = layer.attention.self
+        if not (
+            xmod_layer.self_attn.k_proj.weight.data.shape
+            == xmod_layer.self_attn.q_proj.weight.data.shape
+            == xmod_layer.self_attn.v_proj.weight.data.shape
+            == torch.Size((config.hidden_size, config.hidden_size))
+        ):
+            raise AssertionError("Dimensions of self-attention weights do not match.")
+
+        self_attn.query.weight.data = xmod_layer.self_attn.q_proj.weight
+        self_attn.query.bias.data = xmod_layer.self_attn.q_proj.bias
+        self_attn.key.weight.data = xmod_layer.self_attn.k_proj.weight
+        self_attn.key.bias.data = xmod_layer.self_attn.k_proj.bias
+        self_attn.value.weight.data = xmod_layer.self_attn.v_proj.weight
+        self_attn.value.bias.data = xmod_layer.self_attn.v_proj.bias
+
+        # self-attention output
+        self_output = layer.attention.output
+        if self_output.dense.weight.shape != xmod_layer.self_attn.out_proj.weight.shape:
+            raise AssertionError("Dimensions of self-attention output weights do not match.")
+        self_output.dense.weight = xmod_layer.self_attn.out_proj.weight
+        self_output.dense.bias = xmod_layer.self_attn.out_proj.bias
+        self_output.LayerNorm.weight = xmod_layer.self_attn_layer_norm.weight
+        self_output.LayerNorm.bias = xmod_layer.self_attn_layer_norm.bias
+
+        # intermediate
+        intermediate = layer.intermediate
+        if intermediate.dense.weight.shape != xmod_layer.fc1.weight.shape:
+            raise AssertionError("Dimensions of intermediate weights do not match.")
+        intermediate.dense.weight = xmod_layer.fc1.weight
+        intermediate.dense.bias = xmod_layer.fc1.bias
+
+        # output
+        bert_output = layer.output
+        if bert_output.dense.weight.shape != xmod_layer.fc2.weight.shape:
+            raise AssertionError("Dimensions of feed-forward weights do not match.")
+        bert_output.dense.weight = xmod_layer.fc2.weight
+        bert_output.dense.bias = xmod_layer.fc2.bias
+        bert_output.LayerNorm.weight = xmod_layer.final_layer_norm.weight
+        bert_output.LayerNorm.bias = xmod_layer.final_layer_norm.bias
+        if bert_output.adapter_layer_norm is not None:
+            bert_output.adapter_layer_norm.weight = xmod_layer.adapter_layer_norm.weight
+            bert_output.adapter_layer_norm.bias = xmod_layer.adapter_layer_norm.bias
+
+        if list(sorted(bert_output.adapter_modules.keys())) != list(sorted(xmod_layer.adapter_modules.keys())):
+            raise AssertionError("Lists of language adapters do not match.")
+        for lang_code, adapter in xmod_layer.adapter_modules.items():
+            to_adapter = bert_output.adapter_modules[lang_code]
+            from_adapter = xmod_layer.adapter_modules[lang_code]
+            to_adapter.dense1.weight = from_adapter.fc1.weight
+            to_adapter.dense1.bias = from_adapter.fc1.bias
+            to_adapter.dense2.weight = from_adapter.fc2.weight
+            to_adapter.dense2.bias = from_adapter.fc2.bias
+
+        # end of layer
+
+    if xmod_sent_encoder.layer_norm is not None:
+        model.roberta.encoder.LayerNorm.weight = xmod_sent_encoder.layer_norm.weight
+        model.roberta.encoder.LayerNorm.bias = xmod_sent_encoder.layer_norm.bias
+
+    if classification_head:
+        model.classifier.dense.weight = xmod.model.classification_heads["mnli"].dense.weight
+        model.classifier.dense.bias = xmod.model.classification_heads["mnli"].dense.bias
+        model.classifier.out_proj.weight = xmod.model.classification_heads["mnli"].out_proj.weight
+        model.classifier.out_proj.bias = xmod.model.classification_heads["mnli"].out_proj.bias
+    else:
+        # LM Head
+        model.lm_head.dense.weight = xmod.model.encoder.lm_head.dense.weight
+        model.lm_head.dense.bias = xmod.model.encoder.lm_head.dense.bias
+        model.lm_head.layer_norm.weight = xmod.model.encoder.lm_head.layer_norm.weight
+        model.lm_head.layer_norm.bias = xmod.model.encoder.lm_head.layer_norm.bias
+        model.lm_head.decoder.weight = xmod.model.encoder.lm_head.weight
+        model.lm_head.decoder.bias = xmod.model.encoder.lm_head.bias
+
+    # Let's check that we get the same results.
+    input_ids = xmod.encode(SAMPLE_TEXT).unsqueeze(0)  # batch of size 1
+    model.roberta.set_default_language(SAMPLE_LANGUAGE)
+
+    our_output = model(input_ids)[0]
+    if classification_head:
+        their_output = xmod.model.classification_heads["mnli"](xmod.extract_features(input_ids))
+    else:
+        their_output = xmod.model(input_ids, lang_id=[SAMPLE_LANGUAGE])[0]
+    print(our_output.shape, their_output.shape)
+    max_absolute_diff = torch.max(torch.abs(our_output - their_output)).item()
+    print(f"max_absolute_diff = {max_absolute_diff}")  # ~ 1e-7
+    success = torch.allclose(our_output, their_output, atol=1e-3)
+    print("Do both models output the same tensors?", "🔥" if success else "💩")
+    if not success:
+        raise Exception("Something went wRoNg")
+
+    Path(pytorch_dump_folder_path).mkdir(parents=True, exist_ok=True)
+    print(f"Saving model to {pytorch_dump_folder_path}")
+    model.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--xmod_checkpoint_path", default=None, type=str, required=True, help="Path the official PyTorch dump."
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    parser.add_argument(
+        "--classification_head", action="store_true", help="Whether to convert a final classification head."
+    )
+    args = parser.parse_args()
+    convert_xmod_checkpoint_to_pytorch(
+        args.xmod_checkpoint_path, args.pytorch_dump_folder_path, args.classification_head
+    )

src/transformers/models/xmod/modeling_xmod.py

+# coding=utf-8
+# Copyright 2023 Meta AI Team and the HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch X-MOD model."""
+
+import math
+from typing import List, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
+
+from ...activations import ACT2FN, gelu
+from ...modeling_outputs import (
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+    CausalLMOutputWithCrossAttentions,
+    MaskedLMOutput,
+    MultipleChoiceModelOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging
+from .configuration_xmod import XmodConfig
+
+
+logger = logging.get_logger(__name__)
+
+XMOD_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "jvamvas/xmod-base",
+    "jvamvas/xmod-large-prenorm",
+    "jvamvas/xmod-base-13-125k",
+    "jvamvas/xmod-base-30-125k",
+    "jvamvas/xmod-base-30-195k",
+    "jvamvas/xmod-base-60-125k",
+    "jvamvas/xmod-base-60-265k",
+    "jvamvas/xmod-base-75-125k",
+    "jvamvas/xmod-base-75-269k",
+    # See all X-MOD models at https://huggingface.co/models?filter=xmod
+]
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaEmbeddings with Roberta->Xmod
+class XmodEmbeddings(nn.Module):
+    """
+    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
+    """
+
+    # Copied from transformers.models.bert.modeling_bert.BertEmbeddings.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size, padding_idx=config.pad_token_id)
+        self.position_embeddings = nn.Embedding(config.max_position_embeddings, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(config.type_vocab_size, config.hidden_size)
+
+        # self.LayerNorm is not snake-cased to stick with TensorFlow model variable name and be able to load
+        # any TensorFlow checkpoint file
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+        self.register_buffer(
+            "token_type_ids", torch.zeros(self.position_ids.size(), dtype=torch.long), persistent=False
+        )
+
+        # End copy
+        self.padding_idx = config.pad_token_id
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size, padding_idx=self.padding_idx
+        )
+
+    def forward(
+        self, input_ids=None, token_type_ids=None, position_ids=None, inputs_embeds=None, past_key_values_length=0
+    ):
+        if position_ids is None:
+            if input_ids is not None:
+                # Create the position ids from the input token ids. Any padded tokens remain padded.
+                position_ids = create_position_ids_from_input_ids(input_ids, self.padding_idx, past_key_values_length)
+            else:
+                position_ids = self.create_position_ids_from_inputs_embeds(inputs_embeds)
+
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        # Setting the token_type_ids to the registered buffer in constructor where it is all zeros, which usually occurs
+        # when its auto-generated, registered buffer helps users when tracing the model without passing token_type_ids, solves
+        # issue #5664
+        if token_type_ids is None:
+            if hasattr(self, "token_type_ids"):
+                buffered_token_type_ids = self.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(input_shape[0], seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=self.position_ids.device)
+
+        if inputs_embeds is None:
+            inputs_embeds = self.word_embeddings(input_ids)
+        token_type_embeddings = self.token_type_embeddings(token_type_ids)
+
+        embeddings = inputs_embeds + token_type_embeddings
+        if self.position_embedding_type == "absolute":
+            position_embeddings = self.position_embeddings(position_ids)
+            embeddings += position_embeddings
+        embeddings = self.LayerNorm(embeddings)
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+    def create_position_ids_from_inputs_embeds(self, inputs_embeds):
+        """
+        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.
+
+        Args:
+            inputs_embeds: torch.Tensor
+
+        Returns: torch.Tensor
+        """
+        input_shape = inputs_embeds.size()[:-1]
+        sequence_length = input_shape[1]
+
+        position_ids = torch.arange(
+            self.padding_idx + 1, sequence_length + self.padding_idx + 1, dtype=torch.long, device=inputs_embeds.device
+        )
+        return position_ids.unsqueeze(0).expand(input_shape)
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaSelfAttention with Roberta->Xmod
+class XmodSelfAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                f"The hidden size ({config.hidden_size}) is not a multiple of the number of attention "
+                f"heads ({config.num_attention_heads})"
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        self.key = nn.Linear(config.hidden_size, self.all_head_size)
+        self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = position_embedding_type or getattr(
+            config, "position_embedding_type", "absolute"
+        )
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+
+        self.is_decoder = config.is_decoder
+
+    def transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        mixed_query_layer = self.query(hidden_states)
+
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_layer = past_key_value[0]
+            value_layer = past_key_value[1]
+            attention_mask = encoder_attention_mask
+        elif is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        use_cache = past_key_value is not None
+        if self.is_decoder:
+            # if cross_attention save Tuple(torch.Tensor, torch.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(torch.Tensor, torch.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            query_length, key_length = query_layer.shape[2], key_layer.shape[2]
+            if use_cache:
+                position_ids_l = torch.tensor(key_length - 1, dtype=torch.long, device=hidden_states.device).view(
+                    -1, 1
+                )
+            else:
+                position_ids_l = torch.arange(query_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(key_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in XmodModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        if self.is_decoder:
+            outputs = outputs + (past_key_value,)
+        return outputs
+
+
+class XmodSelfOutput(nn.Module):
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaSelfOutput.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = hidden_states + input_tensor
+        return hidden_states
+
+
+class XmodAttention(nn.Module):
+    def __init__(self, config, position_embedding_type=None):
+        super().__init__()
+        self.self = XmodSelfAttention(config, position_embedding_type=position_embedding_type)
+        self.output = XmodSelfOutput(config)
+        self.pruned_heads = set()
+        self.pre_norm = config.pre_norm
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaAttention.prune_heads
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.self.num_attention_heads, self.self.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.self.query = prune_linear_layer(self.self.query, index)
+        self.self.key = prune_linear_layer(self.self.key, index)
+        self.self.value = prune_linear_layer(self.self.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
+        self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        residual = hidden_states
+        if self.pre_norm:
+            hidden_states = self.output.LayerNorm(hidden_states)
+        self_outputs = self.self(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], residual)
+        if not self.pre_norm:
+            attention_output = self.output.LayerNorm(attention_output)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaIntermediate
+class XmodIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+class XmodAdapter(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.bottleneck_size = config.hidden_size // config.adapter_reduction_factor
+        self.dense1 = nn.Linear(config.hidden_size, self.bottleneck_size)
+        self.dense2 = nn.Linear(self.bottleneck_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.adapter_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.adapter_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense1(hidden_states)
+        hidden_states = self.adapter_act_fn(hidden_states)
+        hidden_states = self.dense2(hidden_states)
+        return hidden_states
+
+
+class XmodOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.ln_before_adapter = config.ln_before_adapter
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        if config.adapter_layer_norm:
+            self.adapter_layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        else:
+            self.adapter_layer_norm = None
+        self.adapter_reuse_layer_norm = config.adapter_reuse_layer_norm
+        self.adapter_modules = nn.ModuleDict(dict())
+        for language in config.languages:
+            self.adapter_modules[str(language)] = XmodAdapter(config)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor, lang_ids: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = hidden_states + input_tensor
+        hidden_states = self.lang_adapter(lang_ids, hidden_states)
+        return hidden_states
+
+    def lang_adapter(self, lang_ids: torch.Tensor, hidden_states: torch.Tensor):
+        # Process subsequent samples with the same lang_id in parallel
+        lang_ids, lang_lengths = torch.unique_consecutive(lang_ids, return_counts=True)
+
+        if not self.ln_before_adapter:
+            residual = hidden_states
+
+        if self.adapter_layer_norm is not None:
+            hidden_states = self.adapter_layer_norm(hidden_states)
+        elif self.adapter_reuse_layer_norm:
+            hidden_states = self.LayerNorm(hidden_states)
+
+        if self.ln_before_adapter:
+            residual = hidden_states
+
+        split_hidden_states = torch.split(hidden_states, lang_lengths.tolist(), 0)
+        lang_wise_outputs = []
+        for i, (lang_id, split_hidden_state) in enumerate(zip(lang_ids, split_hidden_states)):
+            lang = list(self.adapter_modules.keys())[int(lang_id.item())]
+            lang_wise_outputs.append(self.adapter_modules[lang](split_hidden_state))
+        hidden_states = torch.cat(lang_wise_outputs, 0)
+
+        hidden_states = self.dropout(hidden_states)
+        hidden_states += residual
+        return hidden_states
+
+
+class XmodLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = XmodAttention(config)
+        self.is_decoder = config.is_decoder
+        self.add_cross_attention = config.add_cross_attention
+        if self.add_cross_attention:
+            if not self.is_decoder:
+                raise ValueError(f"{self} should be used as a decoder model if cross attention is added")
+            self.crossattention = XmodAttention(config, position_embedding_type="absolute")
+        self.intermediate = XmodIntermediate(config)
+        self.output = XmodOutput(config)
+        self.pre_norm = config.pre_norm
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        lang_ids: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+
+        # if decoder, the last output is tuple of self-attn cache
+        if self.is_decoder:
+            outputs = self_attention_outputs[1:-1]
+            present_key_value = self_attention_outputs[-1]
+        else:
+            outputs = self_attention_outputs[1:]  # add self attentions if we output attention weights
+
+        cross_attn_present_key_value = None
+        if self.is_decoder and encoder_hidden_states is not None:
+            if not hasattr(self, "crossattention"):
+                raise ValueError(
+                    f"If `encoder_hidden_states` are passed, {self} has to be instantiated with cross-attention layers"
+                    " by setting `config.add_cross_attention=True`"
+                )
+
+            # cross_attn cached key/values tuple is at positions 3,4 of past_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            cross_attention_outputs = self.crossattention(
+                attention_output,
+                attention_mask,
+                head_mask,
+                encoder_hidden_states,
+                encoder_attention_mask,
+                cross_attn_past_key_value,
+                output_attentions,
+            )
+            attention_output = cross_attention_outputs[0]
+            outputs = outputs + cross_attention_outputs[1:-1]  # add cross attentions if we output attention weights
+
+            # add cross-attn cache to positions 3,4 of present_key_value tuple
+            cross_attn_present_key_value = cross_attention_outputs[-1]
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        residual = attention_output
+        if self.pre_norm:
+            attention_output = self.output.LayerNorm(attention_output)
+        intermediate_output = apply_chunking_to_forward(
+            self.feed_forward_chunk,
+            self.chunk_size_feed_forward,
+            self.seq_len_dim,
+            attention_output,
+        )
+        layer_output = self.output(intermediate_output, residual, lang_ids)
+        if not self.pre_norm:
+            layer_output = self.output.LayerNorm(layer_output)
+        outputs = (layer_output,) + outputs
+
+        # if decoder, return the attn key/values as the last output
+        if self.is_decoder:
+            outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        return self.intermediate(attention_output)
+
+
+class XmodEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList([XmodLayer(config) for _ in range(config.num_hidden_layers)])
+        self.is_pre_norm = config.pre_norm
+        if self.is_pre_norm:
+            self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        lang_ids: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = False,
+        output_hidden_states: Optional[bool] = False,
+        return_dict: Optional[bool] = True,
+    ) -> Union[Tuple[torch.Tensor], BaseModelOutputWithPastAndCrossAttentions]:
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions and self.config.add_cross_attention else None
+
+        next_decoder_cache = () if use_cache else None
+        for i, layer_module in enumerate(self.layer):
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if self.gradient_checkpointing and self.training:
+                if use_cache:
+                    logger.warning(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    lang_ids,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    lang_ids,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if self.config.add_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if self.is_pre_norm:
+            hidden_states = self.LayerNorm(hidden_states)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaPooler
+class XmodPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.activation = nn.Tanh()
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+        first_token_tensor = hidden_states[:, 0]
+        pooled_output = self.dense(first_token_tensor)
+        pooled_output = self.activation(pooled_output)
+        return pooled_output
+
+
+class XmodPreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = XmodConfig
+    base_model_prefix = "roberta"
+    supports_gradient_checkpointing = True
+
+    # Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Linear):
+            # Slightly different from the TF version which uses truncated_normal for initialization
+            # cf https://github.com/pytorch/pytorch/pull/5617
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.bias is not None:
+                module.bias.data.zero_()
+        elif isinstance(module, nn.Embedding):
+            module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+            if module.padding_idx is not None:
+                module.weight.data[module.padding_idx].zero_()
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaPreTrainedModel._set_gradient_checkpointing with Roberta->Xmod
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, XmodEncoder):
+            module.gradient_checkpointing = value
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaPreTrainedModel.update_keys_to_ignore
+    def update_keys_to_ignore(self, config, del_keys_to_ignore):
+        """Remove some keys from ignore list"""
+        if not config.tie_word_embeddings:
+            # must make a new list, or the class variable gets modified!
+            self._keys_to_ignore_on_save = [k for k in self._keys_to_ignore_on_save if k not in del_keys_to_ignore]
+            self._keys_to_ignore_on_load_missing = [
+                k for k in self._keys_to_ignore_on_load_missing if k not in del_keys_to_ignore
+            ]
+
+    def set_default_language(self, language: str):
+        """
+        Set the default language code for the model. This is used when the language is not specified in the input.
+
+        Args:
+            language (`str`): The language code, such as `"en_XX"` or `"de_DE"`.
+        """
+        if language not in self.config.languages:
+            raise ValueError(
+                f"{self} does not have an adapter for {language}. Supported languages: {list(self.config.languages)}"
+            )
+        self.config.default_language = language
+
+    def freeze_embeddings_and_language_adapters(self):
+        """
+        Freeze the embeddings and language adapters of the model. Usually, this is applied before the model is
+        fine-tuned on a downstream task.
+        """
+        logger.info("Freezing embeddings")
+        for parameter in self.roberta.embeddings.parameters():
+            parameter.requires_grad = False
+        logger.info("Freezing adapters")
+        for layer in self.roberta.encoder.layer:
+            if layer.output.adapter_layer_norm is not None:
+                for parameter in layer.output.adapter_layer_norm.parameters():
+                    parameter.requires_grad = False
+            for parameter in layer.output.adapter_modules.parameters():
+                parameter.requires_grad = False
+
+
+XMOD_START_DOCSTRING = r"""
+
+    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 ([`XmodConfig`]): 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.
+"""
+
+XMOD_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (`torch.LongTensor` of shape `({0})`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        lang_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
+            that corresponds to `self.config.default_language`.
+        attention_mask (`torch.FloatTensor` of shape `({0})`, *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)
+        token_type_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
+            1]`:
+
+            - 0 corresponds to a *sentence A* token,
+            - 1 corresponds to a *sentence B* token.
+
+            [What are token type IDs?](../glossary#token-type-ids)
+        position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
+            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+            config.max_position_embeddings - 1]`.
+
+            [What are position IDs?](../glossary#position-ids)
+        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 `({0}, 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.
+"""
+
+
+@add_start_docstrings(
+    "The bare X-MOD Model transformer outputting raw hidden-states without any specific head on top.",
+    XMOD_START_DOCSTRING,
+)
+class XmodModel(XmodPreTrainedModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
+    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
+    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
+    Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
+    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
+    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.
+
+    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
+
+    """
+
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.bert.modeling_bert.BertModel.__init__ with Bert->Xmod
+    def __init__(self, config, add_pooling_layer=True):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = XmodEmbeddings(config)
+        self.encoder = XmodEncoder(config)
+
+        self.pooler = XmodPooler(config) if add_pooling_layer else None
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaModel.get_input_embeddings
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaModel.set_input_embeddings
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaModel._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(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.Tensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        token_type_ids: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        head_mask: Optional[torch.Tensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        past_key_values: Optional[List[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.Tensor], BaseModelOutputWithPoolingAndCrossAttentions]:
+        r"""
+        encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        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)`.
+        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 = 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 self.config.is_decoder:
+            use_cache = use_cache if use_cache is not None else self.config.use_cache
+        else:
+            use_cache = False
+
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        batch_size, seq_length = input_shape
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if lang_ids is None:
+            if self.config.default_language is None:
+                raise ValueError("Input language unknown. Please call `XmodPreTrainedModel.set_default_language()`")
+            adapter_languages = list(self.encoder.layer[0].output.adapter_modules.keys())
+            default_lang_id = adapter_languages.index(self.config.default_language)
+            lang_ids = default_lang_id * torch.ones(batch_size, device=device)
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        if token_type_ids is None:
+            if hasattr(self.embeddings, "token_type_ids"):
+                buffered_token_type_ids = self.embeddings.token_type_ids[:, :seq_length]
+                buffered_token_type_ids_expanded = buffered_token_type_ids.expand(batch_size, seq_length)
+                token_type_ids = buffered_token_type_ids_expanded
+            else:
+                token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=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: torch.Tensor = 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.config.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=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            position_ids=position_ids,
+            token_type_ids=token_type_ids,
+            inputs_embeds=inputs_embeds,
+            past_key_values_length=past_key_values_length,
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            lang_ids=lang_ids,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output) if self.pooler is not None else None
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "X-MOD Model with a `language modeling` head on top for CLM fine-tuning.",
+    XMOD_START_DOCSTRING,
+)
+class XmodForCausalLM(XmodPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+
+        if not config.is_decoder:
+            logger.warning("If you want to use `XmodLMHeadModel` as a standalone, add `is_decoder=True.`")
+
+        self.roberta = XmodModel(config, add_pooling_layer=False)
+        self.lm_head = XmodLMHead(config)
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        past_key_values: Tuple[Tuple[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.Tensor], CausalLMOutputWithCrossAttentions]:
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the left-to-right language modeling loss (next word prediction). Indices should be in
+            `[-100, 0, ..., config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are
+            ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        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)`.
+        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`).
+
+        Returns: `transformers.modeling_outputs.CausalLMOutputWithCrossAttentions` or `tuple(torch.FloatTensor)`
+
+        Example:
+
+        ```python
+        >>> from transformers import AutoTokenizer, XmodForCausalLM, AutoConfig
+        >>> import torch
+
+        >>> tokenizer = AutoTokenizer.from_pretrained("xlm-roberta-base")
+        >>> config = AutoConfig.from_pretrained("jvamvas/xmod-base")
+        >>> config.is_decoder = True
+        >>> model = XmodForCausalLM.from_pretrained("jvamvas/xmod-base", config=config)
+        >>> model.set_default_language("en_XX")
+
+        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
+        >>> outputs = model(**inputs)
+
+        >>> prediction_logits = outputs.logits
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+        if labels is not None:
+            use_cache = False
+
+        outputs = self.roberta(
+            input_ids,
+            lang_ids=lang_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        lm_loss = None
+        if labels is not None:
+            # we are doing next-token prediction; shift prediction scores and input ids by one
+            shifted_prediction_scores = prediction_scores[:, :-1, :].contiguous()
+            labels = labels[:, 1:].contiguous()
+            loss_fct = CrossEntropyLoss()
+            lm_loss = loss_fct(shifted_prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((lm_loss,) + output) if lm_loss is not None else output
+
+        return CausalLMOutputWithCrossAttentions(
+            loss=lm_loss,
+            logits=prediction_scores,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            cross_attentions=outputs.cross_attentions,
+        )
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM.prepare_inputs_for_generation
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, attention_mask=None, **model_kwargs):
+        input_shape = input_ids.shape
+        # if model is used as a decoder in encoder-decoder model, the decoder attention mask is created on the fly
+        if attention_mask is None:
+            attention_mask = input_ids.new_ones(input_shape)
+
+        # cut decoder_input_ids if past is used
+        if past_key_values is not None:
+            input_ids = input_ids[:, -1:]
+
+        return {"input_ids": input_ids, "attention_mask": attention_mask, "past_key_values": past_key_values}
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForCausalLM._reorder_cache
+    def _reorder_cache(self, past_key_values, beam_idx):
+        reordered_past = ()
+        for layer_past in past_key_values:
+            reordered_past += (tuple(past_state.index_select(0, beam_idx) for past_state in layer_past),)
+        return reordered_past
+
+
+@add_start_docstrings(
+    """X-MOD Model with a `language modeling` head on top.""",
+    XMOD_START_DOCSTRING,
+)
+class XmodForMaskedLM(XmodPreTrainedModel):
+    _keys_to_ignore_on_save = [r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"lm_head.decoder.weight", r"lm_head.decoder.bias"]
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+
+        if config.is_decoder:
+            logger.warning(
+                "If you want to use `XmodForMaskedLM` make sure `config.is_decoder=False` for "
+                "bi-directional self-attention."
+            )
+
+        self.roberta = XmodModel(config, add_pooling_layer=False)
+        self.lm_head = XmodLMHead(config)
+
+        # The LM head weights require special treatment only when they are tied with the word embeddings
+        self.update_keys_to_ignore(config, ["lm_head.decoder.weight"])
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.get_output_embeddings
+    def get_output_embeddings(self):
+        return self.lm_head.decoder
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM.set_output_embeddings
+    def set_output_embeddings(self, new_embeddings):
+        self.lm_head.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], MaskedLMOutput]:
+        r"""
+        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the masked language modeling loss. Indices should be in `[-100, 0, ...,
+            config.vocab_size]` (see `input_ids` docstring) Tokens with indices set to `-100` are ignored (masked), the
+            loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`
+        kwargs (`Dict[str, any]`, optional, defaults to *{}*):
+            Used to hide legacy arguments that have been deprecated.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.roberta(
+            input_ids,
+            lang_ids=lang_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        sequence_output = outputs[0]
+        prediction_scores = self.lm_head(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[2:]
+            return ((masked_lm_loss,) + output) if masked_lm_loss is not None else output
+
+        return MaskedLMOutput(
+            loss=masked_lm_loss,
+            logits=prediction_scores,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaLMHead
+class XmodLMHead(nn.Module):
+    """Roberta Head for masked language modeling."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.layer_norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size)
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+        self.decoder.bias = self.bias
+
+    def forward(self, features, **kwargs):
+        x = self.dense(features)
+        x = gelu(x)
+        x = self.layer_norm(x)
+
+        # project back to size of vocabulary with bias
+        x = self.decoder(x)
+
+        return x
+
+    def _tie_weights(self):
+        # To tie those two weights if they get disconnected (on TPU or when the bias is resized)
+        # For accelerate compatibility and to not break backward compatibility
+        if self.decoder.bias.device.type == "meta":
+            self.decoder.bias = self.bias
+        else:
+            self.bias = self.decoder.bias
+
+
+@add_start_docstrings(
+    """
+    X-MOD Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
+    output) e.g. for GLUE tasks.
+    """,
+    XMOD_START_DOCSTRING,
+)
+class XmodForSequenceClassification(XmodPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForSequenceClassification.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.config = config
+
+        self.roberta = XmodModel(config, add_pooling_layer=False)
+        self.classifier = XmodClassificationHead(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], 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
+
+        outputs = self.roberta(
+            input_ids,
+            lang_ids=lang_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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]
+        logits = self.classifier(sequence_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,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return SequenceClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    X-MOD Model with a multiple choice classification head on top (a linear layer on top of the pooled output and a
+    softmax) e.g. for RocStories/SWAG tasks.
+    """,
+    XMOD_START_DOCSTRING,
+)
+class XmodForMultipleChoice(XmodPreTrainedModel):
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForMultipleChoice.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.roberta = XmodModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, 1)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(XMOD_INPUTS_DOCSTRING.format("batch_size, num_choices, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = 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.Tensor], 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]
+
+        flat_input_ids = input_ids.view(-1, input_ids.size(-1)) if input_ids is not None else None
+        flat_lang_ids = lang_ids.repeat(input_ids.size(0) * input_ids.size(1)) if lang_ids is not None else None
+        flat_position_ids = position_ids.view(-1, position_ids.size(-1)) if position_ids is not None else None
+        flat_token_type_ids = token_type_ids.view(-1, token_type_ids.size(-1)) if token_type_ids is not None else None
+        flat_attention_mask = attention_mask.view(-1, attention_mask.size(-1)) if attention_mask is not None else None
+        flat_inputs_embeds = (
+            inputs_embeds.view(-1, inputs_embeds.size(-2), inputs_embeds.size(-1))
+            if inputs_embeds is not None
+            else None
+        )
+
+        outputs = self.roberta(
+            flat_input_ids,
+            lang_ids=flat_lang_ids,
+            position_ids=flat_position_ids,
+            token_type_ids=flat_token_type_ids,
+            attention_mask=flat_attention_mask,
+            head_mask=head_mask,
+            inputs_embeds=flat_inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        pooled_output = outputs[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+        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,) + outputs[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return MultipleChoiceModelOutput(
+            loss=loss,
+            logits=reshaped_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    X-MOD 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.
+    """,
+    XMOD_START_DOCSTRING,
+)
+class XmodForTokenClassification(XmodPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForTokenClassification.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = XmodModel(config, add_pooling_layer=False)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(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(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = 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]`.
+        """
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.roberta(
+            input_ids,
+            lang_ids=lang_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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[2:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.RobertaClassificationHead
+class XmodClassificationHead(nn.Module):
+    """Head for sentence-level classification tasks."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        classifier_dropout = (
+            config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
+        )
+        self.dropout = nn.Dropout(classifier_dropout)
+        self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, features, **kwargs):
+        x = features[:, 0, :]  # take <s> token (equiv. to [CLS])
+        x = self.dropout(x)
+        x = self.dense(x)
+        x = torch.tanh(x)
+        x = self.dropout(x)
+        x = self.out_proj(x)
+        return x
+
+
+@add_start_docstrings(
+    """
+    X-MOD Model with a span classification head on top for extractive question-answering tasks like SQuAD (a linear
+    layers on top of the hidden-states output to compute `span start logits` and `span end logits`).
+    """,
+    XMOD_START_DOCSTRING,
+)
+class XmodForQuestionAnswering(XmodPreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids"]
+
+    # Copied from transformers.models.roberta.modeling_roberta.RobertaForQuestionAnswering.__init__ with Roberta->Xmod
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.roberta = XmodModel(config, add_pooling_layer=False)
+        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(XMOD_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        lang_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        token_type_ids: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        start_positions: Optional[torch.LongTensor] = None,
+        end_positions: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple[torch.Tensor], 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
+
+        outputs = self.roberta(
+            input_ids,
+            lang_ids=lang_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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]
+
+        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) + outputs[2:]
+            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=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+# Copied from transformers.models.roberta.modeling_roberta.create_position_ids_from_input_ids
+def create_position_ids_from_input_ids(input_ids, padding_idx, past_key_values_length=0):
+    """
+    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
+    are ignored. This is modified from fairseq's `utils.make_positions`.
+
+    Args:
+        x: torch.Tensor x:
+
+    Returns: torch.Tensor
+    """
+    # The series of casts and type-conversions here are carefully balanced to both work with ONNX export and XLA.
+    mask = input_ids.ne(padding_idx).int()
+    incremental_indices = (torch.cumsum(mask, dim=1).type_as(mask) + past_key_values_length) * mask
+    return incremental_indices.long() + padding_idx

src/transformers/utils/dummy_pt_objects.py

@@ -6937,6 +6937,65 @@ def load_tf_weights_in_xlnet(*args, **kwargs):
     requires_backends(load_tf_weights_in_xlnet, ["torch"])
 
 
+XMOD_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class XmodForCausalLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodForMaskedLM(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodForMultipleChoice(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodForQuestionAnswering(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodForSequenceClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodForTokenClassification(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class XmodPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 YOLOS_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/models/xmod/test_modeling_xmod.py

+# coding=utf-8
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import unittest
+
+from transformers import XLMRobertaTokenizer, is_torch_available
+from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
+
+from ...generation.test_utils import GenerationTesterMixin
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        XmodConfig,
+        XmodForCausalLM,
+        XmodForMaskedLM,
+        XmodForMultipleChoice,
+        XmodForQuestionAnswering,
+        XmodForSequenceClassification,
+        XmodForTokenClassification,
+        XmodModel,
+    )
+    from transformers.models.xmod.modeling_xmod import XmodEmbeddings, create_position_ids_from_input_ids
+
+
+class XmodModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        return XmodConfig(
+            vocab_size=self.vocab_size,
+            hidden_size=self.hidden_size,
+            num_hidden_layers=self.num_hidden_layers,
+            num_attention_heads=self.num_attention_heads,
+            intermediate_size=self.intermediate_size,
+            hidden_act=self.hidden_act,
+            hidden_dropout_prob=self.hidden_dropout_prob,
+            attention_probs_dropout_prob=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            type_vocab_size=self.type_vocab_size,
+            initializer_range=self.initializer_range,
+            default_language="en_XX",
+        )
+
+    def prepare_config_and_inputs_for_decoder(self):
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = self.prepare_config_and_inputs()
+
+        config.is_decoder = True
+        encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
+        encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+        return (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = XmodModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_model_as_decoder(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.add_cross_attention = True
+        model = XmodModel(config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+        )
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            encoder_hidden_states=encoder_hidden_states,
+        )
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_for_causal_lm(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        model = XmodForCausalLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_decoder_model_past_large_inputs(
+        self,
+        config,
+        input_ids,
+        token_type_ids,
+        input_mask,
+        sequence_labels,
+        token_labels,
+        choice_labels,
+        encoder_hidden_states,
+        encoder_attention_mask,
+    ):
+        config.is_decoder = True
+        config.add_cross_attention = True
+        model = XmodForCausalLM(config=config).to(torch_device).eval()
+
+        # make sure that ids don't start with pad token
+        mask = input_ids.ne(config.pad_token_id).long()
+        input_ids = input_ids * mask
+
+        # first forward pass
+        outputs = model(
+            input_ids,
+            attention_mask=input_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            use_cache=True,
+        )
+        past_key_values = outputs.past_key_values
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
+
+        # make sure that ids don't start with pad token
+        mask = next_tokens.ne(config.pad_token_id).long()
+        next_tokens = next_tokens * mask
+        next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
+
+        # append to next input_ids and
+        next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
+        next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
+
+        output_from_no_past = model(
+            next_input_ids,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+        output_from_past = model(
+            next_tokens,
+            attention_mask=next_attention_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_attention_mask,
+            past_key_values=past_key_values,
+            output_hidden_states=True,
+        )["hidden_states"][0]
+
+        # select random slice
+        random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
+        output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
+
+    def create_and_check_for_masked_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = XmodForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_token_classification(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_labels = self.num_labels
+        model = XmodForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = XmodForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            attention_mask=multiple_choice_input_mask,
+            token_type_ids=multiple_choice_token_type_ids,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def create_and_check_for_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = XmodForQuestionAnswering(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(
+            input_ids,
+            attention_mask=input_mask,
+            token_type_ids=token_type_ids,
+            start_positions=sequence_labels,
+            end_positions=sequence_labels,
+        )
+        self.parent.assertEqual(result.start_logits.shape, (self.batch_size, self.seq_length))
+        self.parent.assertEqual(result.end_logits.shape, (self.batch_size, self.seq_length))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class XmodModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
+    all_model_classes = (
+        (
+            XmodForCausalLM,
+            XmodForMaskedLM,
+            XmodModel,
+            XmodForSequenceClassification,
+            XmodForTokenClassification,
+            XmodForMultipleChoice,
+            XmodForQuestionAnswering,
+        )
+        if is_torch_available()
+        else ()
+    )
+    all_generative_model_classes = (XmodForCausalLM,) if is_torch_available() else ()
+
+    def setUp(self):
+        self.model_tester = XmodModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=XmodConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_as_decoder(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_model_as_decoder(*config_and_inputs)
+
+    def test_model_as_decoder_with_default_input_mask(self):
+        # This regression test was failing with PyTorch < 1.3
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        ) = self.model_tester.prepare_config_and_inputs_for_decoder()
+
+        input_mask = None
+
+        self.model_tester.create_and_check_model_as_decoder(
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+            encoder_hidden_states,
+            encoder_attention_mask,
+        )
+
+    def test_for_causal_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs_relative_pos_emb(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
+        config_and_inputs[0].position_embedding_type = "relative_key"
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_token_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_token_classification(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_create_position_ids_respects_padding_index(self):
+        """Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is XmodEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        model = XmodEmbeddings(config=config)
+
+        input_ids = torch.as_tensor([[12, 31, 13, model.padding_idx]])
+        expected_positions = torch.as_tensor(
+            [[0 + model.padding_idx + 1, 1 + model.padding_idx + 1, 2 + model.padding_idx + 1, model.padding_idx]]
+        )
+
+        position_ids = create_position_ids_from_input_ids(input_ids, model.padding_idx)
+        self.assertEqual(position_ids.shape, expected_positions.shape)
+        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
+
+    def test_create_position_ids_from_inputs_embeds(self):
+        """Ensure that the default position ids only assign a sequential . This is a regression
+        test for https://github.com/huggingface/transformers/issues/1761
+
+        The position ids should be masked with the embedding object's padding index. Therefore, the
+        first available non-padding position index is XmodEmbeddings.padding_idx + 1
+        """
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        embeddings = XmodEmbeddings(config=config)
+
+        inputs_embeds = torch.empty(2, 4, 30)
+        expected_single_positions = [
+            0 + embeddings.padding_idx + 1,
+            1 + embeddings.padding_idx + 1,
+            2 + embeddings.padding_idx + 1,
+            3 + embeddings.padding_idx + 1,
+        ]
+        expected_positions = torch.as_tensor([expected_single_positions, expected_single_positions])
+        position_ids = embeddings.create_position_ids_from_inputs_embeds(inputs_embeds)
+        self.assertEqual(position_ids.shape, expected_positions.shape)
+        self.assertTrue(torch.all(torch.eq(position_ids, expected_positions)))
+
+    def test_set_default_language(self):
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        model = XmodForMaskedLM(config=config)
+        model.set_default_language("en_XX")
+        self.assertEqual(model.config.default_language, "en_XX")
+        with self.assertRaises(ValueError):
+            model.set_default_language("xx_XX")
+
+    def test_freeze_embeddings_and_language_adapters(self):
+        config = self.model_tester.prepare_config_and_inputs()[0]
+        model = XmodForMaskedLM(config=config)
+        num_trainable_params_before = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        model.freeze_embeddings_and_language_adapters()
+        num_trainable_params_after = sum(p.numel() for p in model.parameters() if p.requires_grad)
+        self.assertLess(num_trainable_params_after, num_trainable_params_before)
+
+
+@require_sentencepiece
+@require_tokenizers
+@require_torch
+class XmodModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_xmod_base(self):
+        model = XmodModel.from_pretrained("jvamvas/xmod-base")
+
+        # language en_XX
+        model.set_default_language("en_XX")
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # The dog is cute and lives in the garden house
+        expected_output_shape = torch.Size((1, 12, 768))  # batch_size, sequence_length, embedding_vector_dim
+        expected_output_values_last_dim = torch.tensor(
+            [[-0.2394, -0.0036, 0.1252, -0.0087, 0.1325, 0.0580, -0.2049, -0.1978, -0.1223, 0.0648, -0.2599, -0.3724]]
+        )
+        output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+        # language de_DE
+        model.set_default_language("de_DE")
+        input_ids = torch.tensor([[0, 1310, 49083, 443, 269, 71, 5486, 165, 60429, 660, 23, 2315, 58761, 18391, 5, 2]])
+        # Der Hund ist niedlich und wohnt in einem Gartenhaus.
+        expected_output_shape = torch.Size((1, 16, 768))  # batch_size, sequence_length, embedding_vector_dim
+        # fmt: off
+        expected_output_values_last_dim = torch.tensor(
+            [[0.0162, 0.0075, -0.1882, 0.2335, -0.0952, -0.3994, -0.0317, -0.1174, 0.0177, 0.4280, -0.0240, -0.2138,
+              0.0785, -0.1045, -0.2811, -0.3220]]
+        )
+        # fmt: on
+        output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+    @slow
+    def test_xmod_large_prenorm(self):
+        model = XmodModel.from_pretrained("jvamvas/xmod-large-prenorm")
+
+        # language en_XX
+        model.set_default_language("en_XX")
+        input_ids = torch.tensor([[0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2]])
+        # The dog is cute and lives in the garden house
+        expected_output_shape = torch.Size((1, 12, 1024))  # batch_size, sequence_length, embedding_vector_dim
+        # fmt: off
+        expected_output_values_last_dim = torch.tensor(
+            [[-0.0121, -0.0194, -0.0240, -0.0160, -0.0205, -0.0159, -0.0243, -0.0206, -0.0161, -0.0335, -0.0196,
+              -0.0141]]
+        )
+        # fmt: on
+        output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+        # language de_DE
+        model.set_default_language("de_DE")
+        input_ids = torch.tensor([[0, 1310, 49083, 443, 269, 71, 5486, 165, 60429, 660, 23, 2315, 58761, 18391, 5, 2]])
+        # Der Hund ist niedlich und wohnt in einem Gartenhaus.
+        expected_output_shape = torch.Size((1, 16, 1024))  # batch_size, sequence_length, embedding_vector_dim
+        # fmt: off
+        expected_output_values_last_dim = torch.tensor(
+            [[-0.0120, -0.0262, -0.0253, -0.0112, -0.0128, -0.0164, -0.0080, -0.0081, -0.0192, -0.0117, -0.0170,
+              -0.0120, -0.0210, -0.0173, -0.0078, -0.0122]]
+        )
+        # fmt: on
+        output = model(input_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+    @slow
+    def test_multilingual_batch(self):
+        model = XmodModel.from_pretrained("jvamvas/xmod-base")
+        # fmt: off
+        input_ids = torch.tensor([
+            [0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2],
+            [0, 1310, 49083, 443, 269, 71, 5486, 165, 60429, 660, 23, 2],
+            [0, 1310, 49083, 443, 269, 71, 5486, 165, 60429, 660, 23, 2],
+            [0, 581, 10269, 83, 99942, 136, 60742, 23, 70, 80583, 18276, 2],
+        ])
+        # fmt: on
+        lang_ids = torch.LongTensor([0, 8, 8, 0])
+        expected_output_shape = torch.Size((4, 12, 768))  # batch_size, sequence_length, embedding_vector_dim
+        # fmt: off
+        expected_output_values_last_dim = torch.tensor([
+            [-0.2394, -0.0036, 0.1252, -0.0087, 0.1325, 0.0580, -0.2049, -0.1978, -0.1223, 0.0648, -0.2599, -0.3724],
+            [-0.2668, -0.0235, -0.1739, 0.2266, -0.0901, -0.3482, 0.0105, -0.1915, 0.0397, 0.3822, 0.1836, -0.3407],
+            [-0.2668, -0.0235, -0.1739, 0.2266, -0.0901, -0.3482, 0.0105, -0.1915, 0.0397, 0.3822, 0.1836, -0.3407],
+            [-0.2394, -0.0036, 0.1252, -0.0087, 0.1325, 0.0580, -0.2049, -0.1978, -0.1223, 0.0648, -0.2599, -0.3724],
+        ])
+        # fmt: on
+        output = model(input_ids, lang_ids=lang_ids)["last_hidden_state"].detach()
+        self.assertEqual(output.shape, expected_output_shape)
+        # compare the actual values for a slice of last dim
+        self.assertTrue(torch.allclose(output[:, :, -1], expected_output_values_last_dim, atol=1e-3))
+
+    @slow
+    def test_end_to_end_mask_fill(self):
+        tokenizer = XLMRobertaTokenizer.from_pretrained("xlm-roberta-base")
+        model = XmodForMaskedLM.from_pretrained("jvamvas/xmod-base", default_language="en_XX")
+        model.to(torch_device)
+
+        sentences = [
+            "Hello, my dog is a little <mask>.",
+            "Hi <mask>!",
+        ]
+
+        inputs = tokenizer(sentences, return_tensors="pt", padding=True)
+        input_ids = inputs["input_ids"].to(torch_device)
+
+        outputs = model(
+            input_ids=input_ids,
+            attention_mask=inputs["attention_mask"].to(torch_device),
+        )
+        probs = outputs.logits.softmax(dim=-1)
+        _, predictions = probs.topk(1)
+        predictions = predictions.squeeze(-1)
+
+        inputs_non_padded = tokenizer(sentences[0], return_tensors="pt").input_ids.to(torch_device)
+        output_non_padded = model(input_ids=inputs_non_padded)
+        probs_non_padded = output_non_padded.logits.softmax(dim=-1)
+        _, predictions_non_padded = probs_non_padded.topk(1)
+        predictions_non_padded = predictions_non_padded.squeeze(-1)
+
+        inputs_padded = tokenizer(sentences[1], return_tensors="pt").input_ids.to(torch_device)
+        output_padded = model(input_ids=inputs_padded)
+        probs_padded = output_padded.logits.softmax(dim=-1)
+        _, predictions_padded = probs_padded.topk(1)
+        predictions_padded = predictions_padded.squeeze(-1)
+
+        batch_out_sentence = tokenizer.batch_decode(predictions, skip_special_tokens=True)
+        non_padded_sentence = tokenizer.decode(predictions_non_padded[0], skip_special_tokens=True)
+        padded_sentence = tokenizer.decode(predictions_padded[0], skip_special_tokens=True)
+
+        expected_output_sentence = [
+            "Hello, my dog is a little girl.",
+            "Hi everyone!",
+        ]
+        self.assertListEqual(expected_output_sentence, batch_out_sentence)
+        self.assertListEqual(batch_out_sentence, [non_padded_sentence, padded_sentence])

utils/documentation_tests.txt

@@ -217,6 +217,8 @@ src/transformers/models/xlm/configuration_xlm.py
 src/transformers/models/xlm_roberta/configuration_xlm_roberta.py
 src/transformers/models/xlm_roberta_xl/configuration_xlm_roberta_xl.py
 src/transformers/models/xlnet/configuration_xlnet.py
+src/transformers/models/xmod/configuration_xmod.py
+src/transformers/models/xmod/modeling_xmod.py
 src/transformers/models/yolos/configuration_yolos.py
 src/transformers/models/yolos/modeling_yolos.py
 src/transformers/models/x_clip/modeling_x_clip.py