Add XGLM models (#14876)

* add xglm

* update vocab size

* fix model name

* style and tokenizer

* typo

* no mask token

* fix pos embed compute

* fix args

* fix tokenizer

* fix positions

* fix tokenization

* style and dic fixes

* fix imports

* add fast tokenizer

* update names

* add pt tests

* fix tokenizer

* fix typo

* fix tokenizer import

* fix fast tokenizer

* fix tokenizer

* fix converter

* add tokenizer test

* update checkpoint names

* fix tokenizer tests

* fix slow tests

* add copied from comments

* rst -> mdx

* flax model

* update flax tests

* quality

* style

* doc

* update index and readme

* fix copies

* fix doc

* update toctrr

* fix indent

* minor fixes

* fix config doc

* don't save embed_pos weights

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* address Sylvains commnets, few doc fixes

* fix check_repo

* align order of arguments

* fix copies

* fix labels

* remove unnecessary mapping

* fix saving tokenizer
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

README.md

@@ -319,6 +319,7 @@ AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Ch
 1. **[WavLM](https://huggingface.co/docs/transformers/master/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/master/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
+1. **[XGLM](https://huggingface.co/docs/master/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.

README_ko.md

@@ -297,6 +297,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/master/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
 1. **[WavLM](https://huggingface.co/docs/transformers/master/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
+1. **[XGLM](https://huggingface.co/docs/master/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.

README_zh-hans.md

@@ -321,6 +321,7 @@ conda install -c huggingface transformers
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/master/transformers/model_doc/wav2vec2_phoneme)** (来自 Facebook AI) 伴随论文 [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) 由 Qiantong Xu, Alexei Baevski, Michael Auli 发布。
 1. **[WavLM](https://huggingface.co/docs/transformers/master/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
+1. **[XGLM](https://huggingface.co/docs/master/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (来自 Facebook) 伴随论文 [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) 由 Guillaume Lample and Alexis Conneau 发布。
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-prophetnet)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
 1. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (来自 Facebook AI), 伴随论文 [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) 由 Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov 发布。

README_zh-hant.md

@@ -333,6 +333,7 @@ conda install -c huggingface transformers
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/master/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
 1. **[WavLM](https://huggingface.co/docs/transformers/master/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
+1. **[XGLM](https://huggingface.co/docs/master/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.

docs/source/_toctree.yml

@@ -304,6 +304,8 @@
       title: Wav2Vec2Phoneme
     - local: model_doc/wavlm
       title: WavLM
+    - local: model_doc/xglm
+      title: XGLM
     - local: model_doc/xlm
       title: XLM
     - local: model_doc/xlm-prophetnet

docs/source/index.mdx

@@ -178,6 +178,7 @@ conversion utilities for the following models.
 1. **[WavLM](model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
 1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2Phoneme](https://huggingface.co/docs/master/transformers/model_doc/wav2vec2_phoneme)** (from Facebook AI) released with the paper [Simple and Effective Zero-shot Cross-lingual Phoneme Recognition](https://arxiv.org/abs/2109.11680) by Qiantong Xu, Alexei Baevski, Michael Auli.
+1. **[XGLM](https://huggingface.co/docs/master/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
 1. **[XLM](model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](model_doc/xlm-prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[XLM-RoBERTa](model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
@@ -278,6 +279,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           ViTMAE            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
 |            WavLM            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            XGLM             |       ✅       |       ✅       |       ✅        |         ❌         |      ✅      |
 |             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |         XLM-RoBERTa         |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |        XLMProphetNet        |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |

docs/source/model_doc/xglm.mdx

+<!--Copyright 2021 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.
+-->
+
+# XGLM
+
+## Overview
+
+The XGLM model was proposed in [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668)
+by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, 
+Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, 
+Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
+
+The abstract from the paper is the following:
+
+*Large-scale autoregressive language models such as GPT-3 are few-shot learners that can perform a wide range of language 
+tasks without fine-tuning. While these models are known to be able to jointly represent many different languages, 
+their training data is dominated by English, potentially limiting their cross-lingual generalization. 
+In this work, we train multilingual autoregressive language models on a balanced corpus covering a diverse set of languages, 
+and study their few- and zero-shot learning capabilities in a wide range of tasks. Our largest model with 7.5 billion parameters 
+sets new state of the art in few-shot learning in more than 20 representative languages, outperforming GPT-3 of comparable size 
+in multilingual commonsense reasoning (with +7.4% absolute accuracy improvement in 0-shot settings and +9.4% in 4-shot settings) 
+and natural language inference (+5.4% in each of 0-shot and 4-shot settings). On the FLORES-101 machine translation benchmark, 
+our model outperforms GPT-3 on 171 out of 182 translation directions with 32 training examples, while surpassing the 
+official supervised baseline in 45 directions. We present a detailed analysis of where the model succeeds and fails, 
+showing in particular that it enables cross-lingual in-context learning on some tasks, while there is still room for improvement 
+on surface form robustness and adaptation to tasks that do not have a natural cloze form. Finally, we evaluate our models 
+in social value tasks such as hate speech detection in five languages and find it has limitations similar to comparable sized GPT-3 models.*
+
+
+This model was contributed by [Suraj](https://huggingface.co/valhalla). The original code can be found [here](https://github.com/pytorch/fairseq/tree/main/examples/xglm).
+
+## XGLMConfig
+
+[[autodoc]] XGLMConfig
+
+## XGLMTokenizer
+
+[[autodoc]] XGLMTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## XGLMTokenizerFast
+
+[[autodoc]] XGLMTokenizerFast
+
+## XGLMModel
+
+[[autodoc]] XGLMModel
+    - forward
+
+## XGLMForCausalLM
+
+[[autodoc]] XGLMForCausalLM
+    - forward
+
+## FlaxXGLMModel
+
+[[autodoc]] FlaxXGLMModel
+    - __call__
+
+## FlaxXGLMForCausalLM
+
+[[autodoc]] FlaxXGLMForCausalLM
+    - __call__
\ No newline at end of file

src/transformers/__init__.py

@@ -328,6 +328,7 @@ _import_structure = {
         "WAVLM_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "WavLMConfig",
     ],
+    "models.xglm": ["XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "XGLMConfig"],
     "models.xlm": ["XLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLMConfig", "XLMTokenizer"],
     "models.xlm_prophetnet": ["XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLMProphetNetConfig"],
     "models.xlm_roberta": ["XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "XLMRobertaConfig"],
@@ -408,6 +409,7 @@ if is_sentencepiece_available():
     _import_structure["models.rembert"].append("RemBertTokenizer")
     _import_structure["models.speech_to_text"].append("Speech2TextTokenizer")
     _import_structure["models.t5"].append("T5Tokenizer")
+    _import_structure["models.xglm"].append("XGLMTokenizer")
     _import_structure["models.xlm_prophetnet"].append("XLMProphetNetTokenizer")
     _import_structure["models.xlm_roberta"].append("XLMRobertaTokenizer")
     _import_structure["models.xlnet"].append("XLNetTokenizer")
@@ -422,6 +424,7 @@ else:
 if is_tokenizers_available():
     # Fast tokenizers
     _import_structure["models.realm"].append("RealmTokenizerFast")
+    _import_structure["models.xglm"].append("XGLMTokenizerFast")
     _import_structure["models.fnet"].append("FNetTokenizerFast")
     _import_structure["models.roformer"].append("RoFormerTokenizerFast")
     _import_structure["models.clip"].append("CLIPTokenizerFast")
@@ -1461,6 +1464,14 @@ if is_torch_available():
             "WavLMPreTrainedModel",
         ]
     )
+    _import_structure["models.xglm"].extend(
+        [
+            "XGLM_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "XGLMForCausalLM",
+            "XGLMModel",
+            "XGLMPreTrainedModel",
+        ]
+    )
     _import_structure["models.xlm"].extend(
         [
             "XLM_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2203,6 +2214,13 @@ if is_flax_available():
     _import_structure["models.wav2vec2"].extend(
         ["FlaxWav2Vec2ForCTC", "FlaxWav2Vec2ForPreTraining", "FlaxWav2Vec2Model", "FlaxWav2Vec2PreTrainedModel"]
     )
+    _import_structure["models.xglm"].extend(
+        [
+            "FlaxXGLMForCausalLM",
+            "FlaxXGLMModel",
+            "FlaxXGLMPreTrainedModel",
+        ]
+    )
 else:
     from .utils import dummy_flax_objects
 
@@ -2463,6 +2481,7 @@ if TYPE_CHECKING:
     from .models.wav2vec2_phoneme import Wav2Vec2PhonemeCTCTokenizer
     from .models.wav2vec2_with_lm import Wav2Vec2ProcessorWithLM
     from .models.wavlm import WAVLM_PRETRAINED_CONFIG_ARCHIVE_MAP, WavLMConfig
+    from .models.xglm import XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XGLMConfig
     from .models.xlm import XLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMConfig, XLMTokenizer
     from .models.xlm_prophetnet import XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMProphetNetConfig
     from .models.xlm_roberta import XLM_ROBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, XLMRobertaConfig
@@ -2544,6 +2563,7 @@ if TYPE_CHECKING:
         from .models.rembert import RemBertTokenizer
         from .models.speech_to_text import Speech2TextTokenizer
         from .models.t5 import T5Tokenizer
+        from .models.xglm import XGLMTokenizer
         from .models.xlm_prophetnet import XLMProphetNetTokenizer
         from .models.xlm_roberta import XLMRobertaTokenizer
         from .models.xlnet import XLNetTokenizer
@@ -2591,6 +2611,7 @@ if TYPE_CHECKING:
         from .models.splinter import SplinterTokenizerFast
         from .models.squeezebert import SqueezeBertTokenizerFast
         from .models.t5 import T5TokenizerFast
+        from .models.xglm import XGLMTokenizerFast
         from .models.xlm_roberta import XLMRobertaTokenizerFast
         from .models.xlnet import XLNetTokenizerFast
         from .tokenization_utils_fast import PreTrainedTokenizerFast
@@ -3404,6 +3425,7 @@ if TYPE_CHECKING:
             WavLMModel,
             WavLMPreTrainedModel,
         )
+        from .models.xglm import XGLM_PRETRAINED_MODEL_ARCHIVE_LIST, XGLMForCausalLM, XGLMModel, XGLMPreTrainedModel
         from .models.xlm import (
             XLM_PRETRAINED_MODEL_ARCHIVE_LIST,
             XLMForMultipleChoice,
@@ -4018,6 +4040,7 @@ if TYPE_CHECKING:
             FlaxWav2Vec2Model,
             FlaxWav2Vec2PreTrainedModel,
         )
+        from .models.xglm import FlaxXGLMForCausalLM, FlaxXGLMModel, FlaxXGLMPreTrainedModel
     else:
         # Import the same objects as dummies to get them in the namespace.
         # They will raise an import error if the user tries to instantiate / use them.

src/transformers/convert_slow_tokenizer.py

@@ -910,6 +910,35 @@ class BlenderbotConverter(Converter):
         return tokenizer
 
 
+class XGLMConverter(SpmConverter):
+    def vocab(self, proto):
+        vocab = [
+            ("<s>", 0.0),
+            ("<pad>", 0.0),
+            ("</s>", 0.0),
+            ("<unk>", 0.0),
+        ]
+        vocab += [(piece.piece, piece.score) for piece in proto.pieces[3:]]
+        # fmt: off
+        vocab += [("<madeupword0>", 0.0), ("<madeupword1>", 0.0), ("<madeupword2>", 0.0), ("<madeupword3>", 0.0), ("<madeupword4>", 0.0), ("<madeupword5>", 0.0), ("<madeupword6>", 0.0)]
+        # fmt: on
+        return vocab
+
+    def unk_id(self, proto):
+        unk_id = 3
+        return unk_id
+
+    def post_processor(self):
+        return processors.TemplateProcessing(
+            single="</s> $A",
+            pair="</s> $A </s> </s> $B",
+            special_tokens=[
+                ("<s>", self.original_tokenizer.convert_tokens_to_ids("<s>")),
+                ("</s>", self.original_tokenizer.convert_tokens_to_ids("</s>")),
+            ],
+        )
+
+
 SLOW_TO_FAST_CONVERTERS = {
     "AlbertTokenizer": AlbertConverter,
     "BartTokenizer": RobertaConverter,
@@ -953,6 +982,7 @@ SLOW_TO_FAST_CONVERTERS = {
     "XLMRobertaTokenizer": XLMRobertaConverter,
     "XLNetTokenizer": XLNetConverter,
     "SplinterTokenizer": SplinterConverter,
+    "XGLMTokenizer": XGLMConverter,
 }
 
 

src/transformers/models/__init__.py

@@ -116,6 +116,7 @@ from . import (
     wav2vec2_phoneme,
     wav2vec2_with_lm,
     wavlm,
+    xglm,
     xlm,
     xlm_prophetnet,
     xlm_roberta,

src/transformers/models/auto/configuration_auto.py

@@ -36,6 +36,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("vit_mae", "ViTMAEConfig"),
         ("realm", "RealmConfig"),
         ("nystromformer", "NystromformerConfig"),
+        ("xglm", "XGLMConfig"),
         ("imagegpt", "ImageGPTConfig"),
         ("qdqbert", "QDQBertConfig"),
         ("vision-encoder-decoder", "VisionEncoderDecoderConfig"),
@@ -128,6 +129,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("vit_mae", "VIT_MAE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("realm", "REALM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("nystromformer", "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("xglm", "XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("imagegpt", "IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("qdqbert", "QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("fnet", "FNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -208,6 +210,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("vit_mae", "ViTMAE"),
         ("realm", "Realm"),
         ("nystromformer", "Nystromformer"),
+        ("xglm", "XGLM"),
         ("imagegpt", "ImageGPT"),
         ("qdqbert", "QDQBert"),
         ("vision-encoder-decoder", "Vision Encoder decoder"),

src/transformers/models/auto/modeling_auto.py

@@ -33,6 +33,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("vilt", "ViltModel"),
         ("vit_mae", "ViTMAEModel"),
         ("nystromformer", "NystromformerModel"),
+        ("xglm", "XGLMModel"),
         ("imagegpt", "ImageGPTModel"),
         ("qdqbert", "QDQBertModel"),
         ("fnet", "FNetModel"),
@@ -209,6 +210,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
     [
         # Model for Causal LM mapping
+        ("xglm", "XGLMForCausalLM"),
         ("qdqbert", "QDQBertLMHeadModel"),
         ("trocr", "TrOCRForCausalLM"),
         ("gptj", "GPTJForCausalLM"),

src/transformers/models/auto/modeling_flax_auto.py

@@ -28,6 +28,7 @@ logger = logging.get_logger(__name__)
 FLAX_MODEL_MAPPING_NAMES = OrderedDict(
     [
         # Base model mapping
+        ("xglm", "FlaxXGLMModel"),
         ("blenderbot-small", "FlaxBlenderbotSmallModel"),
         ("pegasus", "FlaxPegasusModel"),
         ("vision-text-dual-encoder", "FlaxVisionTextDualEncoderModel"),
@@ -121,6 +122,7 @@ FLAX_MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
         ("gpt2", "FlaxGPT2LMHeadModel"),
         ("gpt_neo", "FlaxGPTNeoForCausalLM"),
         ("gptj", "FlaxGPTJForCausalLM"),
+        ("xglm", "FlaxXGLMForCausalLM"),
     ]
 )
 

src/transformers/models/xglm/__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 2020 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
+
+# rely on isort to merge the imports
+from ...file_utils import _LazyModule, is_flax_available, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_xglm": ["XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "XGLMConfig"],
+    "tokenization_xglm": ["XGLMTokenizer"],
+}
+
+if is_tokenizers_available():
+    _import_structure["tokenization_xglm_fast"] = ["XGLMTokenizerFast"]
+
+if is_torch_available():
+    _import_structure["modeling_xglm"] = [
+        "XGLM_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "XGLMForCausalLM",
+        "XGLMModel",
+        "XGLMPreTrainedModel",
+    ]
+
+
+if is_flax_available():
+    _import_structure["modeling_flax_xglm"] = [
+        "FlaxXGLMForCausalLM",
+        "FlaxXGLMModel",
+        "FlaxXGLMPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_xglm import XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP, XGLMConfig
+    from .tokenization_xglm import XGLMTokenizer
+
+    if is_tokenizers_available():
+        from .tokenization_xglm_fast import XGLMTokenizerFast
+
+    if is_torch_available():
+        from .modeling_xglm import XGLM_PRETRAINED_MODEL_ARCHIVE_LIST, XGLMForCausalLM, XGLMModel, XGLMPreTrainedModel
+
+    if is_flax_available():
+        from .modeling_xglm import FlaxXGLMForCausalLM, FlaxXGLMModel, FlaxXGLMPreTrainedModel
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/xglm/configuration_xglm.py

+# coding=utf-8
+# Copyright The HuggingFace Inc. 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.
+""" XGLM model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+XGLM_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "facebook/xglm-564M": "https://huggingface.co/facebook/xglm-564M/resolve/main/config.json",
+    # See all XGLM models at https://huggingface.co/models?filter=xglm
+}
+
+
+class XGLMConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`XGLMModel`]. It is used to instantiate an XGLM
+    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 XGLM
+    [facebook/xglm-564M](https://huggingface.co/facebook/xglm-564M) 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 256008):
+            Vocabulary size of the XGLM model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`XGLMModel`] or [`FlaxXGLMModel`].
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            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).
+        d_model (`int`, *optional*, defaults to 1024):
+            Dimension of the layers and the pooler layer.
+        ffn_dim (`int`, *optional*, defaults to 4096):
+            Dimension of the "intermediate" (often named feed-forward) layer in decoder.
+        num_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers Transformer decoder.
+        attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer decoder.
+        activation_function (`str` or `function`, *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.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, dencoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        layerdrop: (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the encoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        scale_embedding (`bool`, *optional*, defaults to `True`):
+            Scale embeddings by diving by sqrt(d_model).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+
+    Example:
+
+    ```python
+    >>> from transformers import XGLMModel, XGLMConfig
+
+    >>> # Initializing a XGLM facebook/xglm-564M style configuration
+    >>> configuration = XGLMConfig()
+
+    >>> # Initializing a model from the facebook/xglm-564M style configuration
+    >>> model = XGLMModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "xglm"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    attribute_map = {
+        "num_attention_heads": "attention_heads",
+        "hidden_size": "d_model",
+        "num_hidden_layers": "num_layers",
+    }
+
+    def __init__(
+        self,
+        vocab_size=256008,
+        max_position_embeddings=2048,
+        d_model=1024,
+        ffn_dim=4096,
+        num_layers=24,
+        attention_heads=16,
+        activation_function="gelu",
+        dropout=0.1,
+        attention_dropout=0.1,
+        activation_dropout=0.0,
+        layerdrop=0.0,
+        init_std=0.02,
+        scale_embedding=True,
+        use_cache=True,
+        decoder_start_token_id=2,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.d_model = d_model
+        self.ffn_dim = ffn_dim
+        self.num_layers = num_layers
+        self.attention_heads = attention_heads
+        self.activation_function = activation_function
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.layerdrop = layerdrop
+        self.init_std = init_std
+        self.scale_embedding = scale_embedding  # scale factor will be sqrt(d_model) if True
+        self.use_cache = use_cache
+
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            decoder_start_token_id=decoder_start_token_id,
+            **kwargs,
+        )

src/transformers/models/xglm/tokenization_xglm.py

+# coding=utf-8
+# Copyright The HuggingFace Team and The HuggingFace Inc. 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.
+"""Tokenization classes for ."""
+import os
+from shutil import copyfile
+from typing import Any, Dict, List, Optional, Tuple
+
+import sentencepiece as spm
+
+from ...tokenization_utils import PreTrainedTokenizer
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+SPIECE_UNDERLINE = "▁"
+
+VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "facebook/xglm-564M": "https://huggingface.co/facebook/xglm-564M/resolve/main/sentencepiece.bpe.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "facebook/xglm-564M": 1024,
+}
+
+
+class XGLMTokenizer(PreTrainedTokenizer):
+    """
+    Adapted from [`RobertaTokenizer`] and [`XLNetTokenizer`]. Based on
+    [SentencePiece](https://github.com/google/sentencepiece).
+
+    This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
+    this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"</s>"`):
+            The end of sequence token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the end of sequence.
+            The token used is the `sep_token`.
+
+            </Tip>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (`str`, *optional*, defaults to `"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        mask_token (`str`, *optional*, defaults to `"<mask>"`):
+            The token used for masking values. This is the token used when training this model with masked language
+            modeling. This is the token which the model will try to predict.
+        additional_special_tokens (`List[str]`, *optional*, defaults to `["<s>NOTUSED", "</s>NOTUSED"]`):
+            Additional special tokens used by the tokenizer.
+        sp_model_kwargs (`dict`, *optional*):
+            Will be passed to the `SentencePieceProcessor.__init__()` method. The [Python wrapper for
+            SentencePiece](https://github.com/google/sentencepiece/tree/master/python) can be used, among other things,
+            to set:
+
+            - `enable_sampling`: Enable subword regularization.
+            - `nbest_size`: Sampling parameters for unigram. Invalid for BPE-Dropout.
+
+              - `nbest_size = {0,1}`: No sampling is performed.
+              - `nbest_size > 1`: samples from the nbest_size results.
+              - `nbest_size < 0`: assuming that nbest_size is infinite and samples from the all hypothesis (lattice)
+                using forward-filtering-and-backward-sampling algorithm.
+
+            - `alpha`: Smoothing parameter for unigram sampling, and dropout probability of merge operations for
+              BPE-dropout.
+
+    Attributes:
+        sp_model (`SentencePieceProcessor`):
+            The *SentencePiece* processor that is used for every conversion (string, tokens and IDs).
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+
+    def __init__(
+        self,
+        vocab_file,
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        sp_model_kwargs: Optional[Dict[str, Any]] = None,
+        **kwargs
+    ) -> None:
+        self.sp_model_kwargs = {} if sp_model_kwargs is None else sp_model_kwargs
+
+        # Compatibility with the original tokenizer
+        self.num_madeup_words = 7
+        madeup_words = [f"<madeupword{i}>" for i in range(self.num_madeup_words)]
+
+        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", [])
+        kwargs["additional_special_tokens"] += [
+            word for word in madeup_words if word not in kwargs["additional_special_tokens"]
+        ]
+
+        super().__init__(
+            bos_token=bos_token,
+            eos_token=eos_token,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            pad_token=pad_token,
+            sp_model_kwargs=self.sp_model_kwargs,
+            **kwargs,
+        )
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.Load(str(vocab_file))
+        self.vocab_file = vocab_file
+
+        # Original fairseq vocab and spm vocab must be "aligned":
+        # Vocab    |    0    |    1    |   2    |    3    |  4  |  5  |  6  |   7   |   8   |  9
+        # -------- | ------- | ------- | ------ | ------- | --- | --- | --- | ----- | ----- | ----
+        # fairseq  | '<s>'   | '<pad>' | '</s>' | '<unk>' | ',' | '.' | '▁' | 's'   | '▁de' | '-'
+        # spm      | '<unk>' | '<s>'   | '</s>' | ','     | '.' | '▁' | 's' | '▁de' | '-'   | '▁a'
+
+        # The first "real" token "," has position 4 in the original fairseq vocab and position 3 in the spm vocab
+        self.fairseq_offset = 1
+
+        # Mimic fairseq token-to-id alignment for the first 4 token
+        self.fairseq_tokens_to_ids = {"<s>": 0, "<pad>": 1, "</s>": 2, "<unk>": 3}
+
+        sp_size = len(self.sp_model)
+        madeup_words = {f"<madeupword{i}>": sp_size + i + self.fairseq_offset for i in range(self.num_madeup_words)}
+        self.fairseq_tokens_to_ids.update(madeup_words)
+
+        self.fairseq_ids_to_tokens = {v: k for k, v in self.fairseq_tokens_to_ids.items()}
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["sp_model"] = None
+        state["sp_model_proto"] = self.sp_model.serialized_model_proto()
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+
+        # for backward compatibility
+        if not hasattr(self, "sp_model_kwargs"):
+            self.sp_model_kwargs = {}
+
+        self.sp_model = spm.SentencePieceProcessor(**self.sp_model_kwargs)
+        self.sp_model.LoadFromSerializedProto(self.sp_model_proto)
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. An XLM-RoBERTa sequence has the following format:
+
+        - single sequence: `<s> X </s>`
+        - pair of sequences: `<s> A </s></s> B </s>`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+
+        if token_ids_1 is None:
+            return [self.sep_token_id] + token_ids_0
+        sep = [self.sep_token_id]
+        return sep + token_ids_0 + sep + sep + token_ids_1
+
+    def get_special_tokens_mask(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None, already_has_special_tokens: bool = False
+    ) -> List[int]:
+        """
+        Retrieve sequence ids from a token list that has no special tokens added. This method is called when adding
+        special tokens using the tokenizer `prepare_for_model` method.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (`bool`, *optional*, defaults to `False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            `List[int]`: A list of integers in the range [0, 1]: 1 for a special token, 0 for a sequence token.
+        """
+
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(
+                token_ids_0=token_ids_0, token_ids_1=token_ids_1, already_has_special_tokens=True
+            )
+
+        if token_ids_1 is None:
+            return [1] + ([0] * len(token_ids_0))
+        return [1] + ([0] * len(token_ids_0)) + [1, 1] + ([0] * len(token_ids_1))
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. XLM-RoBERTa does
+        not make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+
+        """
+
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(sep + token_ids_0) * [0]
+        return len(sep + token_ids_0 + sep + sep + token_ids_1) * [0]
+
+    @property
+    def vocab_size(self):
+        return len(self.sp_model) + self.fairseq_offset + self.num_madeup_words
+
+    def get_vocab(self):
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text: str) -> List[str]:
+        return self.sp_model.encode(text, out_type=str)
+
+    def _convert_token_to_id(self, token):
+        """Converts a token (str) in an id using the vocab."""
+        if token in self.fairseq_tokens_to_ids:
+            return self.fairseq_tokens_to_ids[token]
+        spm_id = self.sp_model.PieceToId(token)
+
+        # Need to return unknown token if the SP model returned 0
+        return spm_id + self.fairseq_offset if spm_id else self.unk_token_id
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        if index in self.fairseq_ids_to_tokens:
+            return self.fairseq_ids_to_tokens[index]
+        return self.sp_model.IdToPiece(index - self.fairseq_offset)
+
+    def convert_tokens_to_string(self, tokens):
+        """Converts a sequence of tokens (strings for sub-words) in a single string."""
+        out_string = "".join(tokens).replace(SPIECE_UNDERLINE, " ").strip()
+        return out_string
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)

src/transformers/models/xglm/tokenization_xglm_fast.py

+# coding=utf-8
+# Copyright The HuggingFace Team and The HuggingFace Inc. 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.
+"""Tokenization classes for XGLM."""
+
+import os
+from shutil import copyfile
+from typing import List, Optional, Tuple
+
+from ...file_utils import is_sentencepiece_available
+from ...tokenization_utils_fast import PreTrainedTokenizerFast
+from ...utils import logging
+
+
+if is_sentencepiece_available():
+    from .tokenization_xglm import XGLMTokenizer
+else:
+    XGLMTokenizer = None
+
+
+logger = logging.get_logger(__name__)
+
+VOCAB_FILES_NAMES = {"vocab_file": "sentencepiece.bpe.model", "tokenizer_file": "tokenizer.json"}
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "facebook/xglm-564M": "https://huggingface.co/facebook/xglm-564M/resolve/main/sentencepiece.bpe.model",
+    },
+    "tokenizer_file": {
+        "facebook/xglm-564M": "https://huggingface.co/facebook/xglm-564M/resolve/main/tokenizer.json",
+    },
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "facebook/xglm-564M": 1024,
+}
+
+
+class XGLMTokenizerFast(PreTrainedTokenizerFast):
+    """
+    Construct a "fast" XGLM tokenizer (backed by HuggingFace's *tokenizers* library). Adapted from [`RobertaTokenizer`]
+    and [`XLNetTokenizer`]. Based on
+    [BPE](https://huggingface.co/docs/tokenizers/python/latest/components.html?highlight=BPE#models).
+
+    This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
+    refer to this superclass for more information regarding those methods.
+
+    Args:
+        vocab_file (`str`):
+            Path to the vocabulary file.
+        bos_token (`str`, *optional*, defaults to `"<s>"`):
+            The beginning of sequence token that was used during pretraining. Can be used a sequence classifier token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the beginning of
+            sequence. The token used is the `cls_token`.
+
+            </Tip>
+
+        eos_token (`str`, *optional*, defaults to `"</s>"`):
+            The end of sequence token.
+
+            <Tip>
+
+            When building a sequence using special tokens, this is not the token that is used for the end of sequence.
+            The token used is the `sep_token`.
+
+            </Tip>
+
+        sep_token (`str`, *optional*, defaults to `"</s>"`):
+            The separator token, which is used when building a sequence from multiple sequences, e.g. two sequences for
+            sequence classification or for a text and a question for question answering. It is also used as the last
+            token of a sequence built with special tokens.
+        cls_token (`str`, *optional*, defaults to `"<s>"`):
+            The classifier token which is used when doing sequence classification (classification of the whole sequence
+            instead of per-token classification). It is the first token of the sequence when built with special tokens.
+        unk_token (`str`, *optional*, defaults to `"<unk>"`):
+            The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
+            token instead.
+        pad_token (`str`, *optional*, defaults to `"<pad>"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        additional_special_tokens (`List[str]`, *optional*, defaults to `["<s>NOTUSED", "</s>NOTUSED"]`):
+            Additional special tokens used by the tokenizer.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+    model_input_names = ["input_ids", "attention_mask"]
+    slow_tokenizer_class = XGLMTokenizer
+
+    def __init__(
+        self,
+        vocab_file=None,
+        tokenizer_file=None,
+        bos_token="<s>",
+        eos_token="</s>",
+        sep_token="</s>",
+        cls_token="<s>",
+        unk_token="<unk>",
+        pad_token="<pad>",
+        **kwargs
+    ):
+        # Compatibility with the original tokenizer
+        self.num_madeup_words = 7
+        madeup_words = [f"<madeupword{i}>" for i in range(self.num_madeup_words)]
+
+        kwargs["additional_special_tokens"] = kwargs.get("additional_special_tokens", [])
+        kwargs["additional_special_tokens"] += [
+            word for word in madeup_words if word not in kwargs["additional_special_tokens"]
+        ]
+
+        super().__init__(
+            vocab_file,
+            tokenizer_file=tokenizer_file,
+            bos_token=bos_token,
+            eos_token=eos_token,
+            sep_token=sep_token,
+            cls_token=cls_token,
+            unk_token=unk_token,
+            pad_token=pad_token,
+            **kwargs,
+        )
+
+        self.vocab_file = vocab_file
+        self.can_save_slow_tokenizer = False if not self.vocab_file else True
+
+    def build_inputs_with_special_tokens(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. An XLM-RoBERTa sequence has the following format:
+
+        - single sequence: `<s> X </s>`
+        - pair of sequences: `<s> A </s></s> B </s>`
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of [input IDs](../glossary#input-ids) with the appropriate special tokens.
+        """
+
+        if token_ids_1 is None:
+            return [self.sep_token_id] + token_ids_0
+        sep = [self.sep_token_id]
+        return sep + token_ids_0 + sep + sep + token_ids_1
+
+    def create_token_type_ids_from_sequences(
+        self, token_ids_0: List[int], token_ids_1: Optional[List[int]] = None
+    ) -> List[int]:
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. XLM-RoBERTa does
+        not make use of token type ids, therefore a list of zeros is returned.
+
+        Args:
+            token_ids_0 (`List[int]`):
+                List of IDs.
+            token_ids_1 (`List[int]`, *optional*):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            `List[int]`: List of zeros.
+
+        """
+
+        sep = [self.sep_token_id]
+
+        if token_ids_1 is None:
+            return len(sep + token_ids_0) * [0]
+        return len(sep + token_ids_0 + sep + sep + token_ids_1) * [0]
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        if not self.can_save_slow_tokenizer:
+            raise ValueError(
+                "Your fast tokenizer does not have the necessary information to save the vocabulary for a slow "
+                "tokenizer."
+            )
+
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory.")
+            return
+        out_vocab_file = os.path.join(
+            save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+
+        return (out_vocab_file,)

src/transformers/utils/dummy_flax_objects.py

@@ -961,3 +961,24 @@ class FlaxWav2Vec2PreTrainedModel(metaclass=DummyObject):
 
     def __init__(self, *args, **kwargs):
         requires_backends(self, ["flax"])
+
+
+class FlaxXGLMForCausalLM(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxXGLMModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])
+
+
+class FlaxXGLMPreTrainedModel(metaclass=DummyObject):
+    _backends = ["flax"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["flax"])