Integrate DeBERTa v2(the 1.5B model surpassed human performance on Su… (#10018)

* Integrate DeBERTa v2(the 1.5B model surpassed human performance on SuperGLUE); Add DeBERTa v2 900M,1.5B models;

* DeBERTa-v2

* Fix v2 model loading issue (#10129)

* Doc members

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

* Address Sylvain's comments

* Address Patrick's comments
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Style
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

README.md

@@ -201,6 +201,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[ConvBERT](https://huggingface.co/transformers/model_doc/convbert.html)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[CTRL](https://huggingface.co/transformers/model_doc/ctrl.html)** (from Salesforce) released with the paper [CTRL: A Conditional Transformer Language Model for Controllable Generation](https://arxiv.org/abs/1909.05858) by Nitish Shirish Keskar*, Bryan McCann*, Lav R. Varshney, Caiming Xiong and Richard Socher.
 1. **[DeBERTa](https://huggingface.co/transformers/model_doc/deberta.html)** (from Microsoft Research) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
+1. **[DeBERTa-v2](https://huggingface.co/transformers/model_doc/deberta_v2.html)** (from Microsoft Research) released with the paper [DeBERTa: Decoding-enhanced BERT with Disentangled Attention](https://arxiv.org/abs/2006.03654) by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen.
 1. **[DialoGPT](https://huggingface.co/transformers/model_doc/dialogpt.html)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 1. **[DistilBERT](https://huggingface.co/transformers/model_doc/distilbert.html)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/master/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/master/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/master/examples/distillation) and a German version of DistilBERT.
 1. **[DPR](https://huggingface.co/transformers/model_doc/dpr.html)** (from Facebook) released with the paper [Dense Passage Retrieval

docs/source/index.rst

@@ -117,95 +117,98 @@ and conversion utilities for the following models:
 12. :doc:`DeBERTa <model_doc/deberta>` (from Microsoft Research) released with the paper `DeBERTa: Decoding-enhanced
     BERT with Disentangled Attention <https://arxiv.org/abs/2006.03654>`__ by Pengcheng He, Xiaodong Liu, Jianfeng Gao,
     Weizhu Chen.
-13. :doc:`DialoGPT <model_doc/dialogpt>` (from Microsoft Research) released with the paper `DialoGPT: Large-Scale
+13. :doc:`DeBERTa-v2 <model_doc/deberta_v2>` (from Microsoft Research) released with the paper `DeBERTa:
+    Decoding-enhanced BERT with Disentangled Attention <https://arxiv.org/abs/2006.03654>`__ by Pengcheng He, Xiaodong
+    Liu, Jianfeng Gao, Weizhu Chen.
+14. :doc:`DialoGPT <model_doc/dialogpt>` (from Microsoft Research) released with the paper `DialoGPT: Large-Scale
     Generative Pre-training for Conversational Response Generation <https://arxiv.org/abs/1911.00536>`__ by Yizhe
     Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
-14. :doc:`DistilBERT <model_doc/distilbert>` (from HuggingFace), released together with the paper `DistilBERT, a
+15. :doc:`DistilBERT <model_doc/distilbert>` (from HuggingFace), released together with the paper `DistilBERT, a
     distilled version of BERT: smaller, faster, cheaper and lighter <https://arxiv.org/abs/1910.01108>`__ by Victor
     Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into `DistilGPT2
     <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, RoBERTa into `DistilRoBERTa
     <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__, Multilingual BERT into
     `DistilmBERT <https://github.com/huggingface/transformers/tree/master/examples/distillation>`__ and a German
     version of DistilBERT.
-15. :doc:`DPR <model_doc/dpr>` (from Facebook) released with the paper `Dense Passage Retrieval for Open-Domain
+16. :doc:`DPR <model_doc/dpr>` (from Facebook) released with the paper `Dense Passage Retrieval for Open-Domain
     Question Answering <https://arxiv.org/abs/2004.04906>`__ by Vladimir Karpukhin, Barlas Oğuz, Sewon Min, Patrick
     Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
-16. :doc:`ELECTRA <model_doc/electra>` (from Google Research/Stanford University) released with the paper `ELECTRA:
+17. :doc:`ELECTRA <model_doc/electra>` (from Google Research/Stanford University) released with the paper `ELECTRA:
     Pre-training text encoders as discriminators rather than generators <https://arxiv.org/abs/2003.10555>`__ by Kevin
     Clark, Minh-Thang Luong, Quoc V. Le, Christopher D. Manning.
-17. :doc:`FlauBERT <model_doc/flaubert>` (from CNRS) released with the paper `FlauBERT: Unsupervised Language Model
+18. :doc:`FlauBERT <model_doc/flaubert>` (from CNRS) released with the paper `FlauBERT: Unsupervised Language Model
     Pre-training for French <https://arxiv.org/abs/1912.05372>`__ by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne,
     Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab.
-18. :doc:`Funnel Transformer <model_doc/funnel>` (from CMU/Google Brain) released with the paper `Funnel-Transformer:
+19. :doc:`Funnel Transformer <model_doc/funnel>` (from CMU/Google Brain) released with the paper `Funnel-Transformer:
     Filtering out Sequential Redundancy for Efficient Language Processing <https://arxiv.org/abs/2006.03236>`__ by
     Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le.
-19. :doc:`GPT <model_doc/gpt>` (from OpenAI) released with the paper `Improving Language Understanding by Generative
+20. :doc:`GPT <model_doc/gpt>` (from OpenAI) released with the paper `Improving Language Understanding by Generative
     Pre-Training <https://blog.openai.com/language-unsupervised/>`__ by Alec Radford, Karthik Narasimhan, Tim Salimans
     and Ilya Sutskever.
-20. :doc:`GPT-2 <model_doc/gpt2>` (from OpenAI) released with the paper `Language Models are Unsupervised Multitask
+21. :doc:`GPT-2 <model_doc/gpt2>` (from OpenAI) released with the paper `Language Models are Unsupervised Multitask
     Learners <https://blog.openai.com/better-language-models/>`__ by Alec Radford*, Jeffrey Wu*, Rewon Child, David
     Luan, Dario Amodei** and Ilya Sutskever**.
-21. :doc:`LayoutLM <model_doc/layoutlm>` (from Microsoft Research Asia) released with the paper `LayoutLM: Pre-training
+22. :doc:`LayoutLM <model_doc/layoutlm>` (from Microsoft Research Asia) released with the paper `LayoutLM: Pre-training
     of Text and Layout for Document Image Understanding <https://arxiv.org/abs/1912.13318>`__ by Yiheng Xu, Minghao Li,
     Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
-22. :doc:`LED <model_doc/led>` (from AllenAI) released with the paper `Longformer: The Long-Document Transformer
+23. :doc:`LED <model_doc/led>` (from AllenAI) released with the paper `Longformer: The Long-Document Transformer
     <https://arxiv.org/abs/2004.05150>`__ by Iz Beltagy, Matthew E. Peters, Arman Cohan.
-23. :doc:`Longformer <model_doc/longformer>` (from AllenAI) released with the paper `Longformer: The Long-Document
+24. :doc:`Longformer <model_doc/longformer>` (from AllenAI) released with the paper `Longformer: The Long-Document
     Transformer <https://arxiv.org/abs/2004.05150>`__ by Iz Beltagy, Matthew E. Peters, Arman Cohan.
-24. :doc:`LXMERT <model_doc/lxmert>` (from UNC Chapel Hill) released with the paper `LXMERT: Learning Cross-Modality
+25. :doc:`LXMERT <model_doc/lxmert>` (from UNC Chapel Hill) released with the paper `LXMERT: Learning Cross-Modality
     Encoder Representations from Transformers for Open-Domain Question Answering <https://arxiv.org/abs/1908.07490>`__
     by Hao Tan and Mohit Bansal.
-25. :doc:`MarianMT <model_doc/marian>` Machine translation models trained using `OPUS <http://opus.nlpl.eu/>`__ data by
+26. :doc:`MarianMT <model_doc/marian>` Machine translation models trained using `OPUS <http://opus.nlpl.eu/>`__ data by
     Jörg Tiedemann. The `Marian Framework <https://marian-nmt.github.io/>`__ is being developed by the Microsoft
     Translator Team.
-26. :doc:`MBart <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Denoising Pre-training for
+27. :doc:`MBart <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Denoising Pre-training for
     Neural Machine Translation <https://arxiv.org/abs/2001.08210>`__ by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li,
     Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
-27. :doc:`MBart-50 <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Translation with Extensible
+28. :doc:`MBart-50 <model_doc/mbart>` (from Facebook) released with the paper `Multilingual Translation with Extensible
     Multilingual Pretraining and Finetuning <https://arxiv.org/abs/2008.00401>`__ by Yuqing Tang, Chau Tran, Xian Li,
     Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
-28. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
+29. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
     Pre-training for Language Understanding <https://arxiv.org/abs/2004.09297>`__ by Kaitao Song, Xu Tan, Tao Qin,
     Jianfeng Lu, Tie-Yan Liu.
-29. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
+30. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
     text-to-text transformer <https://arxiv.org/abs/2010.11934>`__ by Linting Xue, Noah Constant, Adam Roberts, Mihir
     Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
-30. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
+31. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
     Gap-sentences for Abstractive Summarization <https://arxiv.org/abs/1912.08777>`__> by Jingqing Zhang, Yao Zhao,
     Mohammad Saleh and Peter J. Liu.
-31. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
+32. :doc:`ProphetNet <model_doc/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.
-32. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
+33. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
     Transformer <https://arxiv.org/abs/2001.04451>`__ by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
-33. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
+34. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
     Pretraining Approach <https://arxiv.org/abs/1907.11692>`__ by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar
     Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
-34. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
+35. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
     about efficient neural networks? <https://arxiv.org/abs/2006.11316>`__ by Forrest N. Iandola, Albert E. Shaw, Ravi
     Krishna, and Kurt W. Keutzer.
-35. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
+36. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
     Unified Text-to-Text Transformer <https://arxiv.org/abs/1910.10683>`__ by Colin Raffel and Noam Shazeer and Adam
     Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
-36. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
+37. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
     Pre-training <https://arxiv.org/abs/2004.02349>`__ by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller,
     Francesco Piccinno and Julian Martin Eisenschlos.
-37. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+38. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
     Attentive Language Models Beyond a Fixed-Length Context <https://arxiv.org/abs/1901.02860>`__ by Zihang Dai*,
     Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
-38. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
+39. :doc:`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.
-39. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+40. :doc:`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.
-40. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+41. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (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.
-41. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+42. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (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.
-42. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
+43. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
     Pretraining for Language Understanding <https://arxiv.org/abs/1906.08237>`__ by Zhilin Yang*, Zihang Dai*, Yiming
     Yang, Jaime Carbonell, Ruslan Salakhutdinov, Quoc V. Le.
 
@@ -246,6 +249,8 @@ TensorFlow and/or Flax.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           DeBERTa           |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|         DeBERTa-v2          |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         DistilBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           ELECTRA           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
@@ -389,6 +394,7 @@ TensorFlow and/or Flax.
     model_doc/convbert
     model_doc/ctrl
     model_doc/deberta
+    model_doc/deberta_v2
     model_doc/dialogpt
     model_doc/distilbert
     model_doc/dpr

docs/source/model_doc/deberta.rst

@@ -60,7 +60,7 @@ DebertaModel
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.DebertaModel
-    :members:
+    :members: forward
 
 
 DebertaPreTrainedModel
@@ -74,25 +74,25 @@ DebertaForMaskedLM
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.DebertaForMaskedLM
-    :members:
+    :members: forward
 
 
 DebertaForSequenceClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.DebertaForSequenceClassification
-    :members:
+    :members: forward
 
 
 DebertaForTokenClassification
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.DebertaForTokenClassification
-    :members:
+    :members: forward
 
 
 DebertaForQuestionAnswering
 ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
 
 .. autoclass:: transformers.DebertaForQuestionAnswering
-    :members:
+    :members: forward

docs/source/model_doc/deberta_v2.rst

+.. 
+    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.
+
+DeBERTa-v2
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The DeBERTa model was proposed in `DeBERTa: Decoding-enhanced BERT with Disentangled Attention
+<https://arxiv.org/abs/2006.03654>`__ by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen It is based on Google's
+BERT model released in 2018 and Facebook's RoBERTa model released in 2019.
+
+It builds on RoBERTa with disentangled attention and enhanced mask decoder training with half of the data used in
+RoBERTa.
+
+The abstract from the paper is the following:
+
+*Recent progress in pre-trained neural language models has significantly improved the performance of many natural
+language processing (NLP) tasks. In this paper we propose a new model architecture DeBERTa (Decoding-enhanced BERT with
+disentangled attention) that improves the BERT and RoBERTa models using two novel techniques. The first is the
+disentangled attention mechanism, where each word is represented using two vectors that encode its content and
+position, respectively, and the attention weights among words are computed using disentangled matrices on their
+contents and relative positions. Second, an enhanced mask decoder is used to replace the output softmax layer to
+predict the masked tokens for model pretraining. We show that these two techniques significantly improve the efficiency
+of model pretraining and performance of downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of
+the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9%
+(90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). The DeBERTa code and
+pre-trained models will be made publicly available at https://github.com/microsoft/DeBERTa.*
+
+
+The following information is visible directly on the [original implementation
+repository](https://github.com/microsoft/DeBERTa). DeBERTa v2 is the second version of the DeBERTa model. It includes
+the 1.5B model used for the SuperGLUE single-model submission and achieving 89.9, versus human baseline 89.8. You can
+find more details about this submission in the authors'
+[blog](https://www.microsoft.com/en-us/research/blog/microsoft-deberta-surpasses-human-performance-on-the-superglue-benchmark/)
+
+New in v2:
+
+- **Vocabulary** In v2 the tokenizer is changed to use a new vocabulary of size 128K built from the training data.
+  Instead of a GPT2-based tokenizer, the tokenizer is now
+  [sentencepiece-based](https://github.com/google/sentencepiece) tokenizer.
+- **nGiE(nGram Induced Input Encoding)** The DeBERTa-v2 model uses an additional convolution layer aside with the first
+  transformer layer to better learn the local dependency of input tokens.
+- **Sharing position projection matrix with content projection matrix in attention layer** Based on previous
+  experiments, this can save parameters without affecting the performance.
+- **Apply bucket to encode relative postions** The DeBERTa-v2 model uses log bucket to encode relative positions
+  similar to T5.
+- **900M model & 1.5B model** Two additional model sizes are available: 900M and 1.5B, which significantly improves the
+  performance of downstream tasks.
+
+The original code can be found `here <https://github.com/microsoft/DeBERTa>`__.
+
+
+DebertaV2Config
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2Config
+    :members:
+
+
+DebertaV2Tokenizer
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2Tokenizer
+    :members: build_inputs_with_special_tokens, get_special_tokens_mask,
+        create_token_type_ids_from_sequences, save_vocabulary
+
+
+DebertaV2Model
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2Model
+    :members: forward
+
+
+DebertaV2PreTrainedModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2PreTrainedModel
+    :members: forward
+
+
+DebertaV2ForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2ForMaskedLM
+    :members: forward
+
+
+DebertaV2ForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2ForSequenceClassification
+    :members: forward
+
+
+DebertaV2ForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2ForTokenClassification
+    :members: forward
+
+
+DebertaV2ForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.DebertaV2ForQuestionAnswering
+    :members: forward

docs/source/pretrained_models.rst

@@ -443,15 +443,30 @@ For the full list, refer to `https://huggingface.co/models <https://huggingface.
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/microsoft/unilm/tree/master/layoutlm>`__)                                                           |
 +--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-| DeBERTa            | ``microsoft/deberta-base``                                 | | 12-layer, 768-hidden, 12-heads, ~125M parameters                                                                                    |
+| DeBERTa            | ``microsoft/deberta-base``                                 | | 12-layer, 768-hidden, 12-heads, ~140M parameters                                                                                    |
 |                    |                                                            | | DeBERTa using the BERT-base architecture                                                                                            |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
 |                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
-|                    | ``microsoft/deberta-large``                                | | 24-layer, 1024-hidden, 16-heads, ~390M parameters                                                                                   |
+|                    | ``microsoft/deberta-large``                                | | 24-layer, 1024-hidden, 16-heads, ~400M parameters                                                                                   |
 |                    |                                                            | | DeBERTa using the BERT-large architecture                                                                                           |
 |                    |                                                            |                                                                                                                                       |
 |                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``microsoft/deberta-xlarge``                               | | 48-layer, 1024-hidden, 16-heads, ~750M parameters                                                                                   |
+|                    |                                                            | | DeBERTa XLarge with similar BERT architecture                                                                                       |
+|                    |                                                            |                                                                                                                                       |
+|                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``microsoft/deberta-xlarge-v2``                            | | 24-layer, 1536-hidden, 24-heads, ~900M parameters                                                                                   |
+|                    |                                                            | | DeBERTa XLarge V2 with similar BERT architecture                                                                                    |
+|                    |                                                            |                                                                                                                                       |
+|                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
+|                    +------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
+|                    | ``microsoft/deberta-xxlarge-v2``                           | | 48-layer, 1536-hidden, 24-heads, ~1.5B parameters                                                                                   |
+|                    |                                                            | | DeBERTa XXLarge V2 with similar BERT architecture                                                                                   |
+|                    |                                                            |                                                                                                                                       |
+|                    |                                                            | (see `details <https://github.com/microsoft/DeBERTa>`__)                                                                              |
 +--------------------+------------------------------------------------------------+---------------------------------------------------------------------------------------------------------------------------------------+
 | SqueezeBERT        | ``squeezebert/squeezebert-uncased``                        | | 12-layer, 768-hidden, 12-heads, 51M parameters, 4.3x faster than bert-base-uncased on a smartphone.                                 |
 |                    |                                                            | | SqueezeBERT architecture pretrained from scratch on masked language model (MLM) and sentence order prediction (SOP) tasks.          |

src/transformers/__init__.py

@@ -157,6 +157,7 @@ _import_structure = {
     "models.camembert": ["CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "CamembertConfig"],
     "models.ctrl": ["CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP", "CTRLConfig", "CTRLTokenizer"],
     "models.deberta": ["DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaConfig", "DebertaTokenizer"],
+    "models.deberta_v2": ["DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaV2Config", "DebertaV2Tokenizer"],
     "models.distilbert": ["DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "DistilBertConfig", "DistilBertTokenizer"],
     "models.dpr": [
         "DPR_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -515,6 +516,17 @@ if is_torch_available():
             "DebertaForQuestionAnswering",
         ]
     )
+    _import_structure["models.deberta_v2"].extend(
+        [
+            "DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "DebertaV2ForSequenceClassification",
+            "DebertaV2Model",
+            "DebertaV2ForMaskedLM",
+            "DebertaV2PreTrainedModel",
+            "DebertaV2ForTokenClassification",
+            "DebertaV2ForQuestionAnswering",
+        ]
+    )
     _import_structure["models.distilbert"].extend(
         [
             "DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -1287,6 +1299,7 @@ if TYPE_CHECKING:
     from .models.convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig, ConvBertTokenizer
     from .models.ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig, CTRLTokenizer
     from .models.deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig, DebertaTokenizer
+    from .models.deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config, DebertaV2Tokenizer
     from .models.distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig, DistilBertTokenizer
     from .models.dpr import (
         DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -1604,6 +1617,15 @@ if TYPE_CHECKING:
             DebertaModel,
             DebertaPreTrainedModel,
         )
+        from .models.deberta_v2 import (
+            DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DebertaV2ForMaskedLM,
+            DebertaV2ForQuestionAnswering,
+            DebertaV2ForSequenceClassification,
+            DebertaV2ForTokenClassification,
+            DebertaV2Model,
+            DebertaV2PreTrainedModel,
+        )
         from .models.distilbert import (
             DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST,
             DistilBertForMaskedLM,

src/transformers/models/auto/configuration_auto.py

@@ -31,6 +31,7 @@ from ..camembert.configuration_camembert import CAMEMBERT_PRETRAINED_CONFIG_ARCH
 from ..convbert.configuration_convbert import CONVBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, ConvBertConfig
 from ..ctrl.configuration_ctrl import CTRL_PRETRAINED_CONFIG_ARCHIVE_MAP, CTRLConfig
 from ..deberta.configuration_deberta import DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaConfig
+from ..deberta_v2.configuration_deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
 from ..distilbert.configuration_distilbert import DISTILBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, DistilBertConfig
 from ..dpr.configuration_dpr import DPR_PRETRAINED_CONFIG_ARCHIVE_MAP, DPRConfig
 from ..electra.configuration_electra import ELECTRA_PRETRAINED_CONFIG_ARCHIVE_MAP, ElectraConfig
@@ -103,6 +104,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP,
         DPR_PRETRAINED_CONFIG_ARCHIVE_MAP,
         DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP,
         SQUEEZEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         XLM_PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
         PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -138,6 +140,7 @@ CONFIG_MAPPING = OrderedDict(
         ("reformer", ReformerConfig),
         ("longformer", LongformerConfig),
         ("roberta", RobertaConfig),
+        ("deberta-v2", DebertaV2Config),
         ("deberta", DebertaConfig),
         ("flaubert", FlaubertConfig),
         ("fsmt", FSMTConfig),
@@ -199,6 +202,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("encoder-decoder", "Encoder decoder"),
         ("funnel", "Funnel Transformer"),
         ("lxmert", "LXMERT"),
+        ("deberta-v2", "DeBERTa-v2"),
         ("deberta", "DeBERTa"),
         ("layoutlm", "LayoutLM"),
         ("dpr", "DPR"),
@@ -366,7 +370,6 @@ class AutoConfig:
             {'foo': False}
         """
         config_dict, _ = PretrainedConfig.get_config_dict(pretrained_model_name_or_path, **kwargs)
-
         if "model_type" in config_dict:
             config_class = CONFIG_MAPPING[config_dict["model_type"]]
             return config_class.from_dict(config_dict, **kwargs)

src/transformers/models/auto/modeling_auto.py

@@ -84,6 +84,13 @@ from ..deberta.modeling_deberta import (
     DebertaForTokenClassification,
     DebertaModel,
 )
+from ..deberta_v2.modeling_deberta_v2 import (
+    DebertaV2ForMaskedLM,
+    DebertaV2ForQuestionAnswering,
+    DebertaV2ForSequenceClassification,
+    DebertaV2ForTokenClassification,
+    DebertaV2Model,
+)
 from ..distilbert.modeling_distilbert import (
     DistilBertForMaskedLM,
     DistilBertForMultipleChoice,
@@ -254,6 +261,7 @@ from .configuration_auto import (
     ConvBertConfig,
     CTRLConfig,
     DebertaConfig,
+    DebertaV2Config,
     DistilBertConfig,
     DPRConfig,
     ElectraConfig,
@@ -332,6 +340,7 @@ MODEL_MAPPING = OrderedDict(
         (LxmertConfig, LxmertModel),
         (BertGenerationConfig, BertGenerationEncoder),
         (DebertaConfig, DebertaModel),
+        (DebertaV2Config, DebertaV2Model),
         (DPRConfig, DPRQuestionEncoder),
         (XLMProphetNetConfig, XLMProphetNetModel),
         (ProphetNetConfig, ProphetNetModel),
@@ -408,6 +417,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (MPNetConfig, MPNetForMaskedLM),
         (TapasConfig, TapasForMaskedLM),
         (DebertaConfig, DebertaForMaskedLM),
+        (DebertaV2Config, DebertaV2ForMaskedLM),
     ]
 )
 
@@ -465,6 +475,7 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
         (MPNetConfig, MPNetForMaskedLM),
         (TapasConfig, TapasForMaskedLM),
         (DebertaConfig, DebertaForMaskedLM),
+        (DebertaV2Config, DebertaV2ForMaskedLM),
     ]
 )
 
@@ -510,6 +521,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (ElectraConfig, ElectraForSequenceClassification),
         (FunnelConfig, FunnelForSequenceClassification),
         (DebertaConfig, DebertaForSequenceClassification),
+        (DebertaV2Config, DebertaV2ForSequenceClassification),
         (GPT2Config, GPT2ForSequenceClassification),
         (OpenAIGPTConfig, OpenAIGPTForSequenceClassification),
         (ReformerConfig, ReformerForSequenceClassification),
@@ -545,6 +557,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
         (LxmertConfig, LxmertForQuestionAnswering),
         (MPNetConfig, MPNetForQuestionAnswering),
         (DebertaConfig, DebertaForQuestionAnswering),
+        (DebertaV2Config, DebertaV2ForQuestionAnswering),
     ]
 )
 
@@ -577,6 +590,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (FunnelConfig, FunnelForTokenClassification),
         (MPNetConfig, MPNetForTokenClassification),
         (DebertaConfig, DebertaForTokenClassification),
+        (DebertaV2Config, DebertaV2ForTokenClassification),
     ]
 )
 

src/transformers/models/auto/tokenization_auto.py

@@ -66,6 +66,7 @@ from .configuration_auto import (
     ConvBertConfig,
     CTRLConfig,
     DebertaConfig,
+    DebertaV2Config,
     DistilBertConfig,
     DPRConfig,
     ElectraConfig,
@@ -108,6 +109,7 @@ if is_sentencepiece_available():
     from ..barthez.tokenization_barthez import BarthezTokenizer
     from ..bert_generation.tokenization_bert_generation import BertGenerationTokenizer
     from ..camembert.tokenization_camembert import CamembertTokenizer
+    from ..deberta_v2.tokenization_deberta_v2 import DebertaV2Tokenizer
     from ..marian.tokenization_marian import MarianTokenizer
     from ..mbart.tokenization_mbart import MBartTokenizer
     from ..mt5 import MT5Tokenizer
@@ -122,6 +124,7 @@ else:
     BarthezTokenizer = None
     BertGenerationTokenizer = None
     CamembertTokenizer = None
+    DebertaV2Tokenizer = None
     MarianTokenizer = None
     MBartTokenizer = None
     MT5Tokenizer = None
@@ -233,6 +236,7 @@ TOKENIZER_MAPPING = OrderedDict(
         (FSMTConfig, (FSMTTokenizer, None)),
         (BertGenerationConfig, (BertGenerationTokenizer, None)),
         (DebertaConfig, (DebertaTokenizer, None)),
+        (DebertaV2Config, (DebertaV2Tokenizer, None)),
         (RagConfig, (RagTokenizer, None)),
         (XLMProphetNetConfig, (XLMProphetNetTokenizer, None)),
         (ProphetNetConfig, (ProphetNetTokenizer, None)),

src/transformers/models/deberta/configuration_deberta.py

@@ -23,6 +23,10 @@ logger = logging.get_logger(__name__)
 DEBERTA_PRETRAINED_CONFIG_ARCHIVE_MAP = {
     "microsoft/deberta-base": "https://huggingface.co/microsoft/deberta-base/resolve/main/config.json",
     "microsoft/deberta-large": "https://huggingface.co/microsoft/deberta-large/resolve/main/config.json",
+    "microsoft/deberta-xlarge": "https://huggingface.co/microsoft/deberta-xlarge/resolve/main/config.json",
+    "microsoft/deberta-base-mnli": "https://huggingface.co/microsoft/deberta-base-mnli/resolve/main/config.json",
+    "microsoft/deberta-large-mnli": "https://huggingface.co/microsoft/deberta-large-mnli/resolve/main/config.json",
+    "microsoft/deberta-xlarge-mnli": "https://huggingface.co/microsoft/deberta-xlarge-mnli/resolve/main/config.json",
 }
 
 

src/transformers/models/deberta/modeling_deberta.py

@@ -18,7 +18,6 @@ import math
 from collections.abc import Sequence
 
 import torch
-from packaging import version
 from torch import _softmax_backward_data, nn
 from torch.nn import CrossEntropyLoss
 
@@ -40,10 +39,15 @@ logger = logging.get_logger(__name__)
 
 _CONFIG_FOR_DOC = "DebertaConfig"
 _TOKENIZER_FOR_DOC = "DebertaTokenizer"
+_CHECKPOINT_FOR_DOC = "microsoft/deberta-base"
 
 DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST = [
     "microsoft/deberta-base",
     "microsoft/deberta-large",
+    "microsoft/deberta-xlarge",
+    "microsoft/deberta-base-mnli",
+    "microsoft/deberta-large-mnli",
+    "microsoft/deberta-xlarge-mnli",
 ]
 
 
@@ -54,7 +58,7 @@ class ContextPooler(nn.Module):
         self.dropout = StableDropout(config.pooler_dropout)
         self.config = config
 
-    def forward(self, hidden_states, mask=None):
+    def forward(self, hidden_states):
         # We "pool" the model by simply taking the hidden state corresponding
         # to the first token.
 
@@ -79,22 +83,23 @@ class XSoftmax(torch.autograd.Function):
         dim (int): The dimension that will apply softmax
 
     Example::
-      import torch
-      from transformers.models.deberta import XSoftmax
-      # Make a tensor
-      x = torch.randn([4,20,100])
-      # Create a mask
-      mask = (x>0).int()
-      y = XSoftmax.apply(x, mask, dim=-1)
+
+          >>> import torch
+          >>> from transformers.models.deberta.modeling_deberta import XSoftmax
+
+          >>> # Make a tensor
+          >>> x = torch.randn([4,20,100])
+
+          >>> # Create a mask
+          >>> mask = (x>0).int()
+
+          >>> y = XSoftmax.apply(x, mask, dim=-1)
     """
 
     @staticmethod
     def forward(self, input, mask, dim):
         self.dim = dim
-        if version.Version(torch.__version__) >= version.Version("1.2.0a"):
         rmask = ~(mask.bool())
-        else:
-            rmask = (1 - mask).byte()  # This line is not supported by Onnx tracing.
 
         output = input.masked_fill(rmask, float("-inf"))
         output = torch.softmax(output, self.dim)
@@ -127,10 +132,7 @@ def get_mask(input, local_context):
         mask = local_context.mask if local_context.reuse_mask else None
 
     if dropout > 0 and mask is None:
-        if version.Version(torch.__version__) >= version.Version("1.2.0a"):
         mask = (1 - torch.empty_like(input).bernoulli_(1 - dropout)).bool()
-        else:
-            mask = (1 - torch.empty_like(input).bernoulli_(1 - dropout)).byte()
 
     if isinstance(local_context, DropoutContext):
         if local_context.mask is None:
@@ -166,9 +168,7 @@ class StableDropout(torch.nn.Module):
     Optimized dropout module for stabilizing the training
 
     Args:
-
         drop_prob (float): the dropout probabilities
-
     """
 
     def __init__(self, drop_prob):
@@ -183,8 +183,6 @@ class StableDropout(torch.nn.Module):
 
         Args:
             x (:obj:`torch.tensor`): The input tensor to apply dropout
-
-
         """
         if self.training and self.drop_prob > 0:
             return XDropout.apply(x, self.get_context())
@@ -302,7 +300,7 @@ class DebertaIntermediate(nn.Module):
 
 class DebertaOutput(nn.Module):
     def __init__(self, config):
-        super(DebertaOutput, self).__init__()
+        super().__init__()
         self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
         self.LayerNorm = DebertaLayerNorm(config.hidden_size, config.layer_norm_eps)
         self.dropout = StableDropout(config.hidden_dropout_prob)
@@ -317,7 +315,7 @@ class DebertaOutput(nn.Module):
 
 class DebertaLayer(nn.Module):
     def __init__(self, config):
-        super(DebertaLayer, self).__init__()
+        super().__init__()
         self.attention = DebertaAttention(config)
         self.intermediate = DebertaIntermediate(config)
         self.output = DebertaOutput(config)
@@ -701,7 +699,6 @@ class DebertaEmbeddings(nn.Module):
             self.embed_proj = nn.Linear(self.embedding_size, config.hidden_size, bias=False)
         self.LayerNorm = DebertaLayerNorm(config.hidden_size, config.layer_norm_eps)
         self.dropout = StableDropout(config.hidden_dropout_prob)
-        self.output_to_half = False
         self.config = config
 
         # position_ids (1, len position emb) is contiguous in memory and exported when serialized
@@ -763,6 +760,11 @@ class DebertaPreTrainedModel(PreTrainedModel):
     config_class = DebertaConfig
     base_model_prefix = "deberta"
     _keys_to_ignore_on_load_missing = ["position_ids"]
+    _keys_to_ignore_on_load_unexpected = ["position_embeddings"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self._register_load_state_dict_pre_hook(self._pre_load_hook)
 
     def _init_weights(self, module):
         """ Initialize the weights """
@@ -773,6 +775,25 @@ class DebertaPreTrainedModel(PreTrainedModel):
         if isinstance(module, nn.Linear) and module.bias is not None:
             module.bias.data.zero_()
 
+    def _pre_load_hook(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        """
+        Removes the classifier if it doesn't have the correct number of labels.
+        """
+        self_state = self.state_dict()
+        if (
+            ("classifier.weight" in self_state)
+            and ("classifier.weight" in state_dict)
+            and self_state["classifier.weight"].size() != state_dict["classifier.weight"].size()
+        ):
+            logger.warning(
+                f"The checkpoint classifier head has a shape {state_dict['classifier.weight'].size()} and this model "
+                f"classifier head has a shape {self_state['classifier.weight'].size()}. Ignoring the checkpoint "
+                f"weights. You should train your model on new data."
+            )
+            del state_dict["classifier.weight"]
+            if "classifier.bias" in state_dict:
+                del state_dict["classifier.bias"]
+
 
 DEBERTA_START_DOCSTRING = r"""
     The DeBERTa model was proposed in `DeBERTa: Decoding-enhanced BERT with Disentangled Attention
@@ -867,7 +888,7 @@ class DebertaModel(DebertaPreTrainedModel):
     @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
-        checkpoint="microsoft/deberta-base",
+        checkpoint=_CHECKPOINT_FOR_DOC,
         output_type=SequenceClassifierOutput,
         config_class=_CONFIG_FOR_DOC,
     )
@@ -953,7 +974,6 @@ class DebertaModel(DebertaPreTrainedModel):
 
 @add_start_docstrings("""DeBERTa Model with a `language modeling` head on top. """, DEBERTA_START_DOCSTRING)
 class DebertaForMaskedLM(DebertaPreTrainedModel):
-
     _keys_to_ignore_on_load_unexpected = [r"pooler"]
     _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
 
@@ -974,7 +994,7 @@ class DebertaForMaskedLM(DebertaPreTrainedModel):
     @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
-        checkpoint="microsoft/deberta-base",
+        checkpoint=_CHECKPOINT_FOR_DOC,
         output_type=MaskedLMOutput,
         config_class=_CONFIG_FOR_DOC,
     )
@@ -1114,7 +1134,7 @@ class DebertaForSequenceClassification(DebertaPreTrainedModel):
     @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
-        checkpoint="microsoft/deberta-base",
+        checkpoint=_CHECKPOINT_FOR_DOC,
         output_type=SequenceClassifierOutput,
         config_class=_CONFIG_FOR_DOC,
     )
@@ -1194,7 +1214,6 @@ class DebertaForSequenceClassification(DebertaPreTrainedModel):
     DEBERTA_START_DOCSTRING,
 )
 class DebertaForTokenClassification(DebertaPreTrainedModel):
-
     _keys_to_ignore_on_load_unexpected = [r"pooler"]
 
     def __init__(self, config):
@@ -1210,7 +1229,7 @@ class DebertaForTokenClassification(DebertaPreTrainedModel):
     @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
-        checkpoint="microsoft/deberta-base",
+        checkpoint=_CHECKPOINT_FOR_DOC,
         output_type=TokenClassifierOutput,
         config_class=_CONFIG_FOR_DOC,
     )
@@ -1283,7 +1302,6 @@ class DebertaForTokenClassification(DebertaPreTrainedModel):
     DEBERTA_START_DOCSTRING,
 )
 class DebertaForQuestionAnswering(DebertaPreTrainedModel):
-
     _keys_to_ignore_on_load_unexpected = [r"pooler"]
 
     def __init__(self, config):
@@ -1298,7 +1316,7 @@ class DebertaForQuestionAnswering(DebertaPreTrainedModel):
     @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
     @add_code_sample_docstrings(
         tokenizer_class=_TOKENIZER_FOR_DOC,
-        checkpoint="microsoft/deberta-base",
+        checkpoint=_CHECKPOINT_FOR_DOC,
         output_type=QuestionAnsweringModelOutput,
         config_class=_CONFIG_FOR_DOC,
     )

src/transformers/models/deberta/tokenization_deberta.py

@@ -44,12 +44,20 @@ PRETRAINED_VOCAB_FILES_MAP = {
     "vocab_file": {
         "microsoft/deberta-base": "https://huggingface.co/microsoft/deberta-base/resolve/main/bpe_encoder.bin",
         "microsoft/deberta-large": "https://huggingface.co/microsoft/deberta-large/resolve/main/bpe_encoder.bin",
+        "microsoft/deberta-xlarge": "https://huggingface.co/microsoft/deberta-xlarge/resolve/main/bpe_encoder.bin",
+        "microsoft/deberta-base-mnli": "https://huggingface.co/microsoft/deberta-base-mnli/resolve/main/bpe_encoder.bin",
+        "microsoft/deberta-large-mnli": "https://huggingface.co/microsoft/deberta-large-mnli/resolve/main/bpe_encoder.bin",
+        "microsoft/deberta-xlarge-mnli": "https://huggingface.co/microsoft/deberta-xlarge-mnli/resolve/main/bpe_encoder.bin",
     }
 }
 
 PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
     "microsoft/deberta-base": 512,
     "microsoft/deberta-large": 512,
+    "microsoft/deberta-xlarge": 512,
+    "microsoft/deberta-base-mnli": 512,
+    "microsoft/deberta-large-mnli": 512,
+    "microsoft/deberta-xlarge-mnli": 512,
 }
 
 PRETRAINED_INIT_CONFIGURATION = {

src/transformers/models/deberta_v2/__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
+
+from ...file_utils import _BaseLazyModule, is_torch_available
+
+
+_import_structure = {
+    "configuration_deberta_v2": ["DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP", "DebertaV2Config"],
+    "tokenization_deberta_v2": ["DebertaV2Tokenizer"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_deberta_v2"] = [
+        "DEBERTA_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "DebertaV2ForSequenceClassification",
+        "DebertaV2Model",
+        "DebertaV2ForMaskedLM",
+        "DebertaV2PreTrainedModel",
+        "DebertaV2ForTokenClassification",
+        "DebertaV2ForQuestionAnswering",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_deberta_v2 import DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP, DebertaV2Config
+    from .tokenization_deberta_v2 import DebertaV2Tokenizer
+
+    if is_torch_available():
+        from .modeling_deberta_v2 import (
+            DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            DebertaV2ForMaskedLM,
+            DebertaV2ForQuestionAnswering,
+            DebertaV2ForSequenceClassification,
+            DebertaV2ForTokenClassification,
+            DebertaV2Model,
+            DebertaV2PreTrainedModel,
+        )
+
+else:
+    import importlib
+    import os
+    import sys
+
+    class _LazyModule(_BaseLazyModule):
+        """
+        Module class that surfaces all objects but only performs associated imports when the objects are requested.
+        """
+
+        __file__ = globals()["__file__"]
+        __path__ = [os.path.dirname(__file__)]
+
+        def _get_module(self, module_name: str):
+            return importlib.import_module("." + module_name, self.__name__)
+
+    sys.modules[__name__] = _LazyModule(__name__, _import_structure)

src/transformers/models/deberta_v2/configuration_deberta_v2.py

+# coding=utf-8
+# Copyright 2020, Microsoft 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.
+""" DeBERTa-v2 model configuration """
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+DEBERTA_V2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/deberta-v2-xlarge": "https://huggingface.co/microsoft/deberta-v2-xlarge/resolve/main/config.json",
+    "microsoft/deberta-v2-xxlarge": "https://huggingface.co/microsoft/deberta-v2-xxlarge/resolve/main/config.json",
+    "microsoft/deberta-v2-xlarge-mnli": "https://huggingface.co/microsoft/deberta-v2-xlarge-mnli/resolve/main/config.json",
+    "microsoft/deberta-v2-xxlarge-mnli": "https://huggingface.co/microsoft/deberta-v2-xxlarge-mnli/resolve/main/config.json",
+}
+
+
+class DebertaV2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.DebertaV2Model`. It is used
+    to instantiate a DeBERTa-v2 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 DeBERTa
+    `microsoft/deberta-v2-xlarge <https://huggingface.co/microsoft/deberta-base>`__ architecture.
+
+    Configuration objects inherit from :class:`~transformers.PretrainedConfig` and can be used to control the model
+    outputs. Read the documentation from :class:`~transformers.PretrainedConfig` for more information.
+
+    Arguments:
+        vocab_size (:obj:`int`, `optional`, defaults to 128100):
+            Vocabulary size of the DeBERTa-v2 model. Defines the number of different tokens that can be represented by
+            the :obj:`inputs_ids` passed when calling :class:`~transformers.DebertaV2Model`.
+        hidden_size (:obj:`int`, `optional`, defaults to 1536):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (:obj:`int`, `optional`, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (:obj:`int`, `optional`, defaults to 24):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 6144):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (:obj:`str` or :obj:`Callable`, `optional`, defaults to :obj:`"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string,
+            :obj:`"gelu"`, :obj:`"relu"`, :obj:`"silu"`, :obj:`"gelu"`, :obj:`"tanh"`, :obj:`"gelu_fast"`,
+            :obj:`"mish"`, :obj:`"linear"`, :obj:`"sigmoid"` and :obj:`"gelu_new"` are supported.
+        hidden_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (:obj:`float`, `optional`, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (:obj:`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 (:obj:`int`, `optional`, defaults to 0):
+            The vocabulary size of the :obj:`token_type_ids` passed when calling :class:`~transformers.DebertaModel` or
+            :class:`~transformers.TFDebertaModel`.
+        initializer_range (:obj:`float`, `optional`, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (:obj:`float`, `optional`, defaults to 1e-7):
+            The epsilon used by the layer normalization layers.
+        relative_attention (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether use relative position encoding.
+        max_relative_positions (:obj:`int`, `optional`, defaults to -1):
+            The range of relative positions :obj:`[-max_position_embeddings, max_position_embeddings]`. Use the same
+            value as :obj:`max_position_embeddings`.
+        pad_token_id (:obj:`int`, `optional`, defaults to 0):
+            The value used to pad input_ids.
+        position_biased_input (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether add absolute position embedding to content embedding.
+        pos_att_type (:obj:`List[str]`, `optional`):
+            The type of relative position attention, it can be a combination of :obj:`["p2c", "c2p", "p2p"]`, e.g.
+            :obj:`["p2c"]`, :obj:`["p2c", "c2p"]`, :obj:`["p2c", "c2p", 'p2p"]`.
+        layer_norm_eps (:obj:`float`, optional, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+    """
+    model_type = "deberta-v2"
+
+    def __init__(
+        self,
+        vocab_size=128100,
+        hidden_size=1536,
+        num_hidden_layers=24,
+        num_attention_heads=24,
+        intermediate_size=6144,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=0,
+        initializer_range=0.02,
+        layer_norm_eps=1e-7,
+        relative_attention=False,
+        max_relative_positions=-1,
+        pad_token_id=0,
+        position_biased_input=True,
+        pos_att_type=None,
+        pooler_dropout=0,
+        pooler_hidden_act="gelu",
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+
+        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.initializer_range = initializer_range
+        self.relative_attention = relative_attention
+        self.max_relative_positions = max_relative_positions
+        self.pad_token_id = pad_token_id
+        self.position_biased_input = position_biased_input
+
+        # Backwards compatibility
+        if type(pos_att_type) == str:
+            pos_att_type = [x.strip() for x in pos_att_type.lower().split("|")]
+
+        self.pos_att_type = pos_att_type
+        self.vocab_size = vocab_size
+        self.layer_norm_eps = layer_norm_eps
+
+        self.pooler_hidden_size = kwargs.get("pooler_hidden_size", hidden_size)
+        self.pooler_dropout = pooler_dropout
+        self.pooler_hidden_act = pooler_hidden_act

src/transformers/models/deberta_v2/modeling_deberta_v2.py

+# coding=utf-8
+# Copyright 2020 Microsoft and the Hugging Face 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 DeBERTa-v2 model. """
+
+import math
+from collections.abc import Sequence
+
+import numpy as np
+import torch
+from torch import _softmax_backward_data, nn
+from torch.nn import CrossEntropyLoss, LayerNorm
+
+from ...activations import ACT2FN
+from ...file_utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward
+from ...modeling_outputs import (
+    BaseModelOutput,
+    MaskedLMOutput,
+    QuestionAnsweringModelOutput,
+    SequenceClassifierOutput,
+    TokenClassifierOutput,
+)
+from ...modeling_utils import PreTrainedModel
+from ...utils import logging
+from .configuration_deberta_v2 import DebertaV2Config
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "DebertaV2Config"
+_TOKENIZER_FOR_DOC = "DebertaV2Tokenizer"
+_CHECKPOINT_FOR_DOC = "microsoft/deberta-v2-xlarge"
+
+DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "microsoft/deberta-v2-xlarge",
+    "microsoft/deberta-v2-xxlarge",
+    "microsoft/deberta-v2-xlarge-mnli",
+    "microsoft/deberta-v2-xxlarge-mnli",
+]
+
+
+# Copied from transformers.models.deberta.modeling_deberta.ContextPooler
+class ContextPooler(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.pooler_hidden_size, config.pooler_hidden_size)
+        self.dropout = StableDropout(config.pooler_dropout)
+        self.config = config
+
+    def forward(self, hidden_states):
+        # We "pool" the model by simply taking the hidden state corresponding
+        # to the first token.
+
+        context_token = hidden_states[:, 0]
+        context_token = self.dropout(context_token)
+        pooled_output = self.dense(context_token)
+        pooled_output = ACT2FN[self.config.pooler_hidden_act](pooled_output)
+        return pooled_output
+
+    @property
+    def output_dim(self):
+        return self.config.hidden_size
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XSoftmax with deberta->deberta_v2
+class XSoftmax(torch.autograd.Function):
+    """
+    Masked Softmax which is optimized for saving memory
+
+    Args:
+        input (:obj:`torch.tensor`): The input tensor that will apply softmax.
+        mask (:obj:`torch.IntTensor`): The mask matrix where 0 indicate that element will be ignored in the softmax calculation.
+        dim (int): The dimension that will apply softmax
+
+    Example::
+
+          >>> import torch
+          >>> from transformers.models.deberta_v2.modeling_deberta_v2 import XSoftmax
+
+          >>> # Make a tensor
+          >>> x = torch.randn([4,20,100])
+
+          >>> # Create a mask
+          >>> mask = (x>0).int()
+
+          >>> y = XSoftmax.apply(x, mask, dim=-1)
+    """
+
+    @staticmethod
+    def forward(self, input, mask, dim):
+        self.dim = dim
+        rmask = ~(mask.bool())
+
+        output = input.masked_fill(rmask, float("-inf"))
+        output = torch.softmax(output, self.dim)
+        output.masked_fill_(rmask, 0)
+        self.save_for_backward(output)
+        return output
+
+    @staticmethod
+    def backward(self, grad_output):
+        (output,) = self.saved_tensors
+        inputGrad = _softmax_backward_data(grad_output, output, self.dim, output)
+        return inputGrad, None, None
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DropoutContext
+class DropoutContext(object):
+    def __init__(self):
+        self.dropout = 0
+        self.mask = None
+        self.scale = 1
+        self.reuse_mask = True
+
+
+# Copied from transformers.models.deberta.modeling_deberta.get_mask
+def get_mask(input, local_context):
+    if not isinstance(local_context, DropoutContext):
+        dropout = local_context
+        mask = None
+    else:
+        dropout = local_context.dropout
+        dropout *= local_context.scale
+        mask = local_context.mask if local_context.reuse_mask else None
+
+    if dropout > 0 and mask is None:
+        mask = (1 - torch.empty_like(input).bernoulli_(1 - dropout)).bool()
+
+    if isinstance(local_context, DropoutContext):
+        if local_context.mask is None:
+            local_context.mask = mask
+
+    return mask, dropout
+
+
+# Copied from transformers.models.deberta.modeling_deberta.XDropout
+class XDropout(torch.autograd.Function):
+    """Optimized dropout function to save computation and memory by using mask operation instead of multiplication."""
+
+    @staticmethod
+    def forward(ctx, input, local_ctx):
+        mask, dropout = get_mask(input, local_ctx)
+        ctx.scale = 1.0 / (1 - dropout)
+        if dropout > 0:
+            ctx.save_for_backward(mask)
+            return input.masked_fill(mask, 0) * ctx.scale
+        else:
+            return input
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        if ctx.scale > 1:
+            (mask,) = ctx.saved_tensors
+            return grad_output.masked_fill(mask, 0) * ctx.scale, None
+        else:
+            return grad_output, None
+
+
+# Copied from transformers.models.deberta.modeling_deberta.StableDropout
+class StableDropout(torch.nn.Module):
+    """
+    Optimized dropout module for stabilizing the training
+
+    Args:
+        drop_prob (float): the dropout probabilities
+    """
+
+    def __init__(self, drop_prob):
+        super().__init__()
+        self.drop_prob = drop_prob
+        self.count = 0
+        self.context_stack = None
+
+    def forward(self, x):
+        """
+        Call the module
+
+        Args:
+            x (:obj:`torch.tensor`): The input tensor to apply dropout
+        """
+        if self.training and self.drop_prob > 0:
+            return XDropout.apply(x, self.get_context())
+        return x
+
+    def clear_context(self):
+        self.count = 0
+        self.context_stack = None
+
+    def init_context(self, reuse_mask=True, scale=1):
+        if self.context_stack is None:
+            self.context_stack = []
+        self.count = 0
+        for c in self.context_stack:
+            c.reuse_mask = reuse_mask
+            c.scale = scale
+
+    def get_context(self):
+        if self.context_stack is not None:
+            if self.count >= len(self.context_stack):
+                self.context_stack.append(DropoutContext())
+            ctx = self.context_stack[self.count]
+            ctx.dropout = self.drop_prob
+            self.count += 1
+            return ctx
+        else:
+            return self.drop_prob
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaSelfOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2SelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaAttention with Deberta->DebertaV2
+class DebertaV2Attention(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.self = DisentangledSelfAttention(config)
+        self.output = DebertaV2SelfOutput(config)
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        return_att=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        self_output = self.self(
+            hidden_states,
+            attention_mask,
+            return_att,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if return_att:
+            self_output, att_matrix = self_output
+        if query_states is None:
+            query_states = hidden_states
+        attention_output = self.output(self_output, query_states)
+
+        if return_att:
+            return (attention_output, att_matrix)
+        else:
+            return attention_output
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->DebertaV2
+class DebertaV2Intermediate(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):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaOutput with DebertaLayerNorm->LayerNorm
+class DebertaV2Output(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, input_tensor):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaLayer with Deberta->DebertaV2
+class DebertaV2Layer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.attention = DebertaV2Attention(config)
+        self.intermediate = DebertaV2Intermediate(config)
+        self.output = DebertaV2Output(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        return_att=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        attention_output = self.attention(
+            hidden_states,
+            attention_mask,
+            return_att=return_att,
+            query_states=query_states,
+            relative_pos=relative_pos,
+            rel_embeddings=rel_embeddings,
+        )
+        if return_att:
+            attention_output, att_matrix = attention_output
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        if return_att:
+            return (layer_output, att_matrix)
+        else:
+            return layer_output
+
+
+class ConvLayer(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        kernel_size = getattr(config, "conv_kernel_size", 3)
+        groups = getattr(config, "conv_groups", 1)
+        self.conv_act = getattr(config, "conv_act", "tanh")
+        self.conv = torch.nn.Conv1d(
+            config.hidden_size, config.hidden_size, kernel_size, padding=(kernel_size - 1) // 2, groups=groups
+        )
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+    def forward(self, hidden_states, residual_states, input_mask):
+        out = self.conv(hidden_states.permute(0, 2, 1).contiguous()).permute(0, 2, 1).contiguous()
+        rmask = (1 - input_mask).bool()
+        out.masked_fill_(rmask.unsqueeze(-1).expand(out.size()), 0)
+        out = ACT2FN[self.conv_act](self.dropout(out))
+
+        layer_norm_input = residual_states + out
+        output = self.LayerNorm(layer_norm_input).to(layer_norm_input)
+
+        if input_mask is None:
+            output_states = output
+        else:
+            if input_mask.dim() != layer_norm_input.dim():
+                if input_mask.dim() == 4:
+                    input_mask = input_mask.squeeze(1).squeeze(1)
+                input_mask = input_mask.unsqueeze(2)
+
+            input_mask = input_mask.to(output.dtype)
+            output_states = output * input_mask
+
+        return output_states
+
+
+class DebertaV2Encoder(nn.Module):
+    """Modified BertEncoder with relative position bias support"""
+
+    def __init__(self, config):
+        super().__init__()
+
+        self.layer = nn.ModuleList([DebertaV2Layer(config) for _ in range(config.num_hidden_layers)])
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            pos_ebd_size = self.max_relative_positions * 2
+
+            if self.position_buckets > 0:
+                pos_ebd_size = self.position_buckets * 2
+
+            self.rel_embeddings = nn.Embedding(pos_ebd_size, config.hidden_size)
+
+        self.norm_rel_ebd = [x.strip() for x in getattr(config, "norm_rel_ebd", "none").lower().split("|")]
+
+        if "layer_norm" in self.norm_rel_ebd:
+            self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps, elementwise_affine=True)
+
+        self.conv = ConvLayer(config) if getattr(config, "conv_kernel_size", 0) > 0 else None
+
+    def get_rel_embedding(self):
+        rel_embeddings = self.rel_embeddings.weight if self.relative_attention else None
+        if rel_embeddings is not None and ("layer_norm" in self.norm_rel_ebd):
+            rel_embeddings = self.LayerNorm(rel_embeddings)
+        return rel_embeddings
+
+    def get_attention_mask(self, attention_mask):
+        if attention_mask.dim() <= 2:
+            extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+            attention_mask = extended_attention_mask * extended_attention_mask.squeeze(-2).unsqueeze(-1)
+            attention_mask = attention_mask.byte()
+        elif attention_mask.dim() == 3:
+            attention_mask = attention_mask.unsqueeze(1)
+
+        return attention_mask
+
+    def get_rel_pos(self, hidden_states, query_states=None, relative_pos=None):
+        if self.relative_attention and relative_pos is None:
+            q = query_states.size(-2) if query_states is not None else hidden_states.size(-2)
+            relative_pos = build_relative_position(
+                q, hidden_states.size(-2), bucket_size=self.position_buckets, max_position=self.max_relative_positions
+            )
+        return relative_pos
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        output_hidden_states=True,
+        output_attentions=False,
+        query_states=None,
+        relative_pos=None,
+        return_dict=True,
+    ):
+        if attention_mask.dim() <= 2:
+            input_mask = attention_mask
+        else:
+            input_mask = (attention_mask.sum(-2) > 0).byte()
+        attention_mask = self.get_attention_mask(attention_mask)
+        relative_pos = self.get_rel_pos(hidden_states, query_states, relative_pos)
+
+        all_hidden_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        if isinstance(hidden_states, Sequence):
+            next_kv = hidden_states[0]
+        else:
+            next_kv = hidden_states
+        rel_embeddings = self.get_rel_embedding()
+        for i, layer_module in enumerate(self.layer):
+
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            output_states = layer_module(
+                next_kv,
+                attention_mask,
+                output_attentions,
+                query_states=query_states,
+                relative_pos=relative_pos,
+                rel_embeddings=rel_embeddings,
+            )
+            if output_attentions:
+                output_states, att_m = output_states
+
+            if i == 0 and self.conv is not None:
+                output_states = self.conv(hidden_states, output_states, input_mask)
+
+            if query_states is not None:
+                query_states = output_states
+                if isinstance(hidden_states, Sequence):
+                    next_kv = hidden_states[i + 1] if i + 1 < len(self.layer) else None
+            else:
+                next_kv = output_states
+
+            if output_attentions:
+                all_attentions = all_attentions + (att_m,)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (output_states,)
+
+        if not return_dict:
+            return tuple(v for v in [output_states, all_hidden_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=output_states, hidden_states=all_hidden_states, attentions=all_attentions
+        )
+
+
+def make_log_bucket_position(relative_pos, bucket_size, max_position):
+    sign = np.sign(relative_pos)
+    mid = bucket_size // 2
+    abs_pos = np.where((relative_pos < mid) & (relative_pos > -mid), mid - 1, np.abs(relative_pos))
+    log_pos = np.ceil(np.log(abs_pos / mid) / np.log((max_position - 1) / mid) * (mid - 1)) + mid
+    bucket_pos = np.where(abs_pos <= mid, relative_pos, log_pos * sign).astype(np.int)
+    return bucket_pos
+
+
+def build_relative_position(query_size, key_size, bucket_size=-1, max_position=-1):
+    """
+    Build relative position according to the query and key
+
+    We assume the absolute position of query :math:`P_q` is range from (0, query_size) and the absolute position of key
+    :math:`P_k` is range from (0, key_size), The relative positions from query to key is :math:`R_{q \\rightarrow k} =
+    P_q - P_k`
+
+    Args:
+        query_size (int): the length of query
+        key_size (int): the length of key
+        bucket_size (int): the size of position bucket
+        max_position (int): the maxium allowed absolute positoin
+
+    Return:
+        :obj:`torch.LongTensor`: A tensor with shape [1, query_size, key_size]
+
+    """
+    q_ids = np.arange(0, query_size)
+    k_ids = np.arange(0, key_size)
+    rel_pos_ids = q_ids[:, None] - np.tile(k_ids, (q_ids.shape[0], 1))
+    if bucket_size > 0 and max_position > 0:
+        rel_pos_ids = make_log_bucket_position(rel_pos_ids, bucket_size, max_position)
+    rel_pos_ids = torch.tensor(rel_pos_ids, dtype=torch.long)
+    rel_pos_ids = rel_pos_ids[:query_size, :]
+    rel_pos_ids = rel_pos_ids.unsqueeze(0)
+    return rel_pos_ids
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.c2p_dynamic_expand
+def c2p_dynamic_expand(c2p_pos, query_layer, relative_pos):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), query_layer.size(2), relative_pos.size(-1)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.p2c_dynamic_expand
+def p2c_dynamic_expand(c2p_pos, query_layer, key_layer):
+    return c2p_pos.expand([query_layer.size(0), query_layer.size(1), key_layer.size(-2), key_layer.size(-2)])
+
+
+@torch.jit.script
+# Copied from transformers.models.deberta.modeling_deberta.pos_dynamic_expand
+def pos_dynamic_expand(pos_index, p2c_att, key_layer):
+    return pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2)))
+
+
+class DisentangledSelfAttention(torch.nn.Module):
+    """
+    Disentangled self-attention module
+
+    Parameters:
+        config (:obj:`DebertaV2Config`):
+            A model config class instance with the configuration to build a new model. The schema is similar to
+            `BertConfig`, for more details, please refer :class:`~transformers.DebertaV2Config`
+
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        if config.hidden_size % config.num_attention_heads != 0:
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention "
+                "heads (%d)" % (config.hidden_size, config.num_attention_heads)
+            )
+        self.num_attention_heads = config.num_attention_heads
+        _attention_head_size = config.hidden_size // config.num_attention_heads
+        self.attention_head_size = getattr(config, "attention_head_size", _attention_head_size)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+        self.query_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+        self.value_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+
+        self.share_att_key = getattr(config, "share_att_key", False)
+        self.pos_att_type = config.pos_att_type
+        self.relative_attention = getattr(config, "relative_attention", False)
+
+        if self.relative_attention:
+            self.position_buckets = getattr(config, "position_buckets", -1)
+            self.max_relative_positions = getattr(config, "max_relative_positions", -1)
+            if self.max_relative_positions < 1:
+                self.max_relative_positions = config.max_position_embeddings
+            self.pos_ebd_size = self.max_relative_positions
+            if self.position_buckets > 0:
+                self.pos_ebd_size = self.position_buckets
+
+            self.pos_dropout = StableDropout(config.hidden_dropout_prob)
+
+            if not self.share_att_key:
+                if "c2p" in self.pos_att_type or "p2p" in self.pos_att_type:
+                    self.pos_key_proj = nn.Linear(config.hidden_size, self.all_head_size, bias=True)
+                if "p2c" in self.pos_att_type or "p2p" in self.pos_att_type:
+                    self.pos_query_proj = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = StableDropout(config.attention_probs_dropout_prob)
+
+    def transpose_for_scores(self, x, attention_heads):
+        new_x_shape = x.size()[:-1] + (attention_heads, -1)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3).contiguous().view(-1, x.size(1), x.size(-1))
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask,
+        return_att=False,
+        query_states=None,
+        relative_pos=None,
+        rel_embeddings=None,
+    ):
+        """
+        Call the module
+
+        Args:
+            hidden_states (:obj:`torch.FloatTensor`):
+                Input states to the module usually the output from previous layer, it will be the Q,K and V in
+                `Attention(Q,K,V)`
+
+            attention_mask (:obj:`torch.ByteTensor`):
+                An attention mask matrix of shape [`B`, `N`, `N`] where `B` is the batch size, `N` is the maximum
+                sequence length in which element [i,j] = `1` means the `i` th token in the input can attend to the `j`
+                th token.
+
+            return_att (:obj:`bool`, optional):
+                Whether return the attention matrix.
+
+            query_states (:obj:`torch.FloatTensor`, optional):
+                The `Q` state in `Attention(Q,K,V)`.
+
+            relative_pos (:obj:`torch.LongTensor`):
+                The relative position encoding between the tokens in the sequence. It's of shape [`B`, `N`, `N`] with
+                values ranging in [`-max_relative_positions`, `max_relative_positions`].
+
+            rel_embeddings (:obj:`torch.FloatTensor`):
+                The embedding of relative distances. It's a tensor of shape [:math:`2 \\times
+                \\text{max_relative_positions}`, `hidden_size`].
+
+
+        """
+        if query_states is None:
+            query_states = hidden_states
+        query_layer = self.transpose_for_scores(self.query_proj(query_states), self.num_attention_heads)
+        key_layer = self.transpose_for_scores(self.key_proj(hidden_states), self.num_attention_heads)
+        value_layer = self.transpose_for_scores(self.value_proj(hidden_states), self.num_attention_heads)
+
+        rel_att = None
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        scale_factor = 1
+        if "c2p" in self.pos_att_type:
+            scale_factor += 1
+        if "p2c" in self.pos_att_type:
+            scale_factor += 1
+        if "p2p" in self.pos_att_type:
+            scale_factor += 1
+        scale = math.sqrt(query_layer.size(-1) * scale_factor)
+        attention_scores = torch.bmm(query_layer, key_layer.transpose(-1, -2)) / scale
+        if self.relative_attention:
+            rel_embeddings = self.pos_dropout(rel_embeddings)
+            rel_att = self.disentangled_attention_bias(
+                query_layer, key_layer, relative_pos, rel_embeddings, scale_factor
+            )
+
+        if rel_att is not None:
+            attention_scores = attention_scores + rel_att
+        attention_scores = attention_scores
+        attention_scores = attention_scores.view(
+            -1, self.num_attention_heads, attention_scores.size(-2), attention_scores.size(-1)
+        )
+
+        # bsz x height x length x dimension
+        attention_probs = XSoftmax.apply(attention_scores, attention_mask, -1)
+        attention_probs = self.dropout(attention_probs)
+        context_layer = torch.bmm(
+            attention_probs.view(-1, attention_probs.size(-2), attention_probs.size(-1)), value_layer
+        )
+        context_layer = (
+            context_layer.view(-1, self.num_attention_heads, context_layer.size(-2), context_layer.size(-1))
+            .permute(0, 2, 1, 3)
+            .contiguous()
+        )
+        new_context_layer_shape = context_layer.size()[:-2] + (-1,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+        if return_att:
+            return (context_layer, attention_probs)
+        else:
+            return context_layer
+
+    def disentangled_attention_bias(self, query_layer, key_layer, relative_pos, rel_embeddings, scale_factor):
+        if relative_pos is None:
+            q = query_layer.size(-2)
+            relative_pos = build_relative_position(
+                q, key_layer.size(-2), bucket_size=self.position_buckets, max_position=self.max_relative_positions
+            )
+        if relative_pos.dim() == 2:
+            relative_pos = relative_pos.unsqueeze(0).unsqueeze(0)
+        elif relative_pos.dim() == 3:
+            relative_pos = relative_pos.unsqueeze(1)
+        # bsz x height x query x key
+        elif relative_pos.dim() != 4:
+            raise ValueError(f"Relative postion ids must be of dim 2 or 3 or 4. {relative_pos.dim()}")
+
+        att_span = self.pos_ebd_size
+        relative_pos = relative_pos.long().to(query_layer.device)
+
+        rel_embeddings = rel_embeddings[self.pos_ebd_size - att_span : self.pos_ebd_size + att_span, :].unsqueeze(0)
+        if self.share_att_key:
+            pos_query_layer = self.transpose_for_scores(
+                self.query_proj(rel_embeddings), self.num_attention_heads
+            ).repeat(query_layer.size(0) // self.num_attention_heads, 1, 1)
+            pos_key_layer = self.transpose_for_scores(self.key_proj(rel_embeddings), self.num_attention_heads).repeat(
+                query_layer.size(0) // self.num_attention_heads, 1, 1
+            )
+        else:
+            if "c2p" in self.pos_att_type or "p2p" in self.pos_att_type:
+                pos_key_layer = self.transpose_for_scores(
+                    self.pos_key_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(
+                    query_layer.size(0) // self.num_attention_heads, 1, 1
+                )  # .split(self.all_head_size, dim=-1)
+            if "p2c" in self.pos_att_type or "p2p" in self.pos_att_type:
+                pos_query_layer = self.transpose_for_scores(
+                    self.pos_query_proj(rel_embeddings), self.num_attention_heads
+                ).repeat(
+                    query_layer.size(0) // self.num_attention_heads, 1, 1
+                )  # .split(self.all_head_size, dim=-1)
+
+        score = 0
+        # content->position
+        if "c2p" in self.pos_att_type:
+            scale = math.sqrt(pos_key_layer.size(-1) * scale_factor)
+            c2p_att = torch.bmm(query_layer, pos_key_layer.transpose(-1, -2))
+            c2p_pos = torch.clamp(relative_pos + att_span, 0, att_span * 2 - 1)
+            c2p_att = torch.gather(
+                c2p_att,
+                dim=-1,
+                index=c2p_pos.squeeze(0).expand([query_layer.size(0), query_layer.size(1), relative_pos.size(-1)]),
+            )
+            score += c2p_att / scale
+
+        # position->content
+        if "p2c" in self.pos_att_type or "p2p" in self.pos_att_type:
+            scale = math.sqrt(pos_query_layer.size(-1) * scale_factor)
+            if key_layer.size(-2) != query_layer.size(-2):
+                r_pos = build_relative_position(
+                    key_layer.size(-2),
+                    key_layer.size(-2),
+                    bucket_size=self.position_buckets,
+                    max_position=self.max_relative_positions,
+                ).to(query_layer.device)
+                r_pos = r_pos.unsqueeze(0)
+            else:
+                r_pos = relative_pos
+
+            p2c_pos = torch.clamp(-r_pos + att_span, 0, att_span * 2 - 1)
+            if query_layer.size(-2) != key_layer.size(-2):
+                pos_index = relative_pos[:, :, :, 0].unsqueeze(-1)
+
+        if "p2c" in self.pos_att_type:
+            p2c_att = torch.bmm(key_layer, pos_query_layer.transpose(-1, -2))
+            p2c_att = torch.gather(
+                p2c_att,
+                dim=-1,
+                index=p2c_pos.squeeze(0).expand([query_layer.size(0), key_layer.size(-2), key_layer.size(-2)]),
+            ).transpose(-1, -2)
+            if query_layer.size(-2) != key_layer.size(-2):
+                p2c_att = torch.gather(
+                    p2c_att,
+                    dim=-2,
+                    index=pos_index.expand(p2c_att.size()[:2] + (pos_index.size(-2), key_layer.size(-2))),
+                )
+            score += p2c_att / scale
+
+        # position->position
+        if "p2p" in self.pos_att_type:
+            pos_query = pos_query_layer[:, :, att_span:, :]
+            p2p_att = torch.matmul(pos_query, pos_key_layer.transpose(-1, -2))
+            p2p_att = p2p_att.expand(query_layer.size()[:2] + p2p_att.size()[2:])
+            if query_layer.size(-2) != key_layer.size(-2):
+                p2p_att = torch.gather(
+                    p2p_att,
+                    dim=-2,
+                    index=pos_index.expand(query_layer.size()[:2] + (pos_index.size(-2), p2p_att.size(-1))),
+                )
+            p2p_att = torch.gather(
+                p2p_att,
+                dim=-1,
+                index=c2p_pos.expand(
+                    [query_layer.size(0), query_layer.size(1), query_layer.size(2), relative_pos.size(-1)]
+                ),
+            )
+            score += p2p_att
+
+        return score
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaEmbeddings with DebertaLayerNorm->LayerNorm
+class DebertaV2Embeddings(nn.Module):
+    """Construct the embeddings from word, position and token_type embeddings."""
+
+    def __init__(self, config):
+        super().__init__()
+        pad_token_id = getattr(config, "pad_token_id", 0)
+        self.embedding_size = getattr(config, "embedding_size", config.hidden_size)
+        self.word_embeddings = nn.Embedding(config.vocab_size, self.embedding_size, padding_idx=pad_token_id)
+
+        self.position_biased_input = getattr(config, "position_biased_input", True)
+        if not self.position_biased_input:
+            self.position_embeddings = None
+        else:
+            self.position_embeddings = nn.Embedding(config.max_position_embeddings, self.embedding_size)
+
+        if config.type_vocab_size > 0:
+            self.token_type_embeddings = nn.Embedding(config.type_vocab_size, self.embedding_size)
+
+        if self.embedding_size != config.hidden_size:
+            self.embed_proj = nn.Linear(self.embedding_size, config.hidden_size, bias=False)
+        self.LayerNorm = LayerNorm(config.hidden_size, config.layer_norm_eps)
+        self.dropout = StableDropout(config.hidden_dropout_prob)
+        self.config = config
+
+        # position_ids (1, len position emb) is contiguous in memory and exported when serialized
+        self.register_buffer("position_ids", torch.arange(config.max_position_embeddings).expand((1, -1)))
+
+    def forward(self, input_ids=None, token_type_ids=None, position_ids=None, mask=None, inputs_embeds=None):
+        if input_ids is not None:
+            input_shape = input_ids.size()
+        else:
+            input_shape = inputs_embeds.size()[:-1]
+
+        seq_length = input_shape[1]
+
+        if position_ids is None:
+            position_ids = self.position_ids[:, :seq_length]
+
+        if token_type_ids is None:
+            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)
+
+        if self.position_embeddings is not None:
+            position_embeddings = self.position_embeddings(position_ids.long())
+        else:
+            position_embeddings = torch.zeros_like(inputs_embeds)
+
+        embeddings = inputs_embeds
+        if self.position_biased_input:
+            embeddings += position_embeddings
+        if self.config.type_vocab_size > 0:
+            token_type_embeddings = self.token_type_embeddings(token_type_ids)
+            embeddings += token_type_embeddings
+
+        if self.embedding_size != self.config.hidden_size:
+            embeddings = self.embed_proj(embeddings)
+
+        embeddings = self.LayerNorm(embeddings)
+
+        if mask is not None:
+            if mask.dim() != embeddings.dim():
+                if mask.dim() == 4:
+                    mask = mask.squeeze(1).squeeze(1)
+                mask = mask.unsqueeze(2)
+            mask = mask.to(embeddings.dtype)
+
+            embeddings = embeddings * mask
+
+        embeddings = self.dropout(embeddings)
+        return embeddings
+
+
+# Copied from transformers.models.deberta.modeling_deberta.DebertaPreTrainedModel with Deberta->DebertaV2
+class DebertaV2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = DebertaV2Config
+    base_model_prefix = "deberta"
+    _keys_to_ignore_on_load_missing = ["position_ids"]
+    _keys_to_ignore_on_load_unexpected = ["position_embeddings"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self._register_load_state_dict_pre_hook(self._pre_load_hook)
+
+    def _init_weights(self, module):
+        """ Initialize the weights """
+        if isinstance(module, (nn.Linear, nn.Embedding)):
+            # 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 isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _pre_load_hook(self, state_dict, prefix, local_metadata, strict, missing_keys, unexpected_keys, error_msgs):
+        """
+        Removes the classifier if it doesn't have the correct number of labels.
+        """
+        self_state = self.state_dict()
+        if (
+            ("classifier.weight" in self_state)
+            and ("classifier.weight" in state_dict)
+            and self_state["classifier.weight"].size() != state_dict["classifier.weight"].size()
+        ):
+            logger.warning(
+                f"The checkpoint classifier head has a shape {state_dict['classifier.weight'].size()} and this model "
+                f"classifier head has a shape {self_state['classifier.weight'].size()}. Ignoring the checkpoint "
+                f"weights. You should train your model on new data."
+            )
+            del state_dict["classifier.weight"]
+            if "classifier.bias" in state_dict:
+                del state_dict["classifier.bias"]
+
+
+DEBERTA_START_DOCSTRING = r"""
+    The DeBERTa model was proposed in `DeBERTa: Decoding-enhanced BERT with Disentangled Attention
+    <https://arxiv.org/abs/2006.03654>`_ by Pengcheng He, Xiaodong Liu, Jianfeng Gao, Weizhu Chen. It's build on top of
+    BERT/RoBERTa with two improvements, i.e. disentangled attention and enhanced mask decoder. With those two
+    improvements, it out perform BERT/RoBERTa on a majority of tasks with 80GB pretraining data.
+
+    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 (:class:`~transformers.DebertaV2Config`): 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 :meth:`~transformers.PreTrainedModel.from_pretrained` method to load the model
+            weights.
+"""
+
+DEBERTA_INPUTS_DOCSTRING = r"""
+    Args:
+        input_ids (:obj:`torch.LongTensor` of shape :obj:`{0}`):
+            Indices of input sequence tokens in the vocabulary.
+
+            Indices can be obtained using :class:`transformers.DebertaV2Tokenizer`. See
+            :func:`transformers.PreTrainedTokenizer.encode` and :func:`transformers.PreTrainedTokenizer.__call__` for
+            details.
+
+            `What are input IDs? <../glossary.html#input-ids>`__
+        attention_mask (:obj:`torch.FloatTensor` of shape :obj:`{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.html#attention-mask>`__
+        token_type_ids (:obj:`torch.LongTensor` of shape :obj:`{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.html#token-type-ids>`_
+        position_ids (:obj:`torch.LongTensor` of shape :obj:`{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.html#position-ids>`_
+        inputs_embeds (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`, `optional`):
+            Optionally, instead of passing :obj:`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 (:obj:`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 (:obj:`bool`, `optional`):
+            Whether or not to return the hidden states of all layers. See ``hidden_states`` under returned tensors for
+            more detail.
+        return_dict (:obj:`bool`, `optional`):
+            Whether or not to return a :class:`~transformers.file_utils.ModelOutput` instead of a plain tuple.
+"""
+
+
+@add_start_docstrings(
+    "The bare DeBERTa Model transformer outputting raw hidden-states without any specific head on top.",
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaModel with Deberta->DebertaV2
+class DebertaV2Model(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = DebertaV2Embeddings(config)
+        self.encoder = DebertaV2Encoder(config)
+        self.z_steps = 0
+        self.config = config
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, new_embeddings):
+        self.embeddings.word_embeddings = new_embeddings
+
+    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
+        """
+        raise NotImplementedError("The prune function is not implemented in DeBERTa model.")
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if input_ids is not None and inputs_embeds is not None:
+            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")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+        )
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask,
+            output_hidden_states=True,
+            output_attentions=output_attentions,
+            return_dict=return_dict,
+        )
+        encoded_layers = encoder_outputs[1]
+
+        if self.z_steps > 1:
+            hidden_states = encoded_layers[-2]
+            layers = [self.encoder.layer[-1] for _ in range(self.z_steps)]
+            query_states = encoded_layers[-1]
+            rel_embeddings = self.encoder.get_rel_embedding()
+            attention_mask = self.encoder.get_attention_mask(attention_mask)
+            rel_pos = self.encoder.get_rel_pos(embedding_output)
+            for layer in layers[1:]:
+                query_states = layer(
+                    hidden_states,
+                    attention_mask,
+                    return_att=False,
+                    query_states=query_states,
+                    relative_pos=rel_pos,
+                    rel_embeddings=rel_embeddings,
+                )
+                encoded_layers.append(query_states)
+
+        sequence_output = encoded_layers[-1]
+
+        if not return_dict:
+            return (sequence_output,) + encoder_outputs[(1 if output_hidden_states else 2) :]
+
+        return BaseModelOutput(
+            last_hidden_state=sequence_output,
+            hidden_states=encoder_outputs.hidden_states if output_hidden_states else None,
+            attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings("""DeBERTa Model with a `language modeling` head on top. """, DEBERTA_START_DOCSTRING)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForMaskedLM with Deberta->DebertaV2
+class DebertaV2ForMaskedLM(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+    _keys_to_ignore_on_load_missing = [r"position_ids", r"predictions.decoder.bias"]
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.deberta = DebertaV2Model(config)
+        self.cls = DebertaV2OnlyMLMHead(config)
+
+        self.init_weights()
+
+    def get_output_embeddings(self):
+        return self.cls.predictions.decoder
+
+    def set_output_embeddings(self, new_embeddings):
+        self.cls.predictions.decoder = new_embeddings
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=MaskedLMOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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]``
+        """
+
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        outputs = self.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        sequence_output = outputs[0]
+        prediction_scores = self.cls(sequence_output)
+
+        masked_lm_loss = None
+        if labels is not None:
+            loss_fct = CrossEntropyLoss()  # -100 index = padding token
+            masked_lm_loss = loss_fct(prediction_scores.view(-1, self.config.vocab_size), labels.view(-1))
+
+        if not return_dict:
+            output = (prediction_scores,) + outputs[1:]
+            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.bert.BertPredictionHeadTransform with bert -> deberta
+class DebertaV2PredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+# copied from transformers.models.bert.BertLMPredictionHead with bert -> deberta
+class DebertaV2LMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = DebertaV2PredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+# copied from transformers.models.bert.BertOnlyMLMHead with bert -> deberta
+class DebertaV2OnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = DebertaV2LMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
+@add_start_docstrings(
+    """
+    DeBERTa Model transformer with a sequence classification/regression head on top (a linear layer on top of the
+    pooled output) e.g. for GLUE tasks.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForSequenceClassification with Deberta->DebertaV2
+class DebertaV2ForSequenceClassification(DebertaV2PreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        num_labels = getattr(config, "num_labels", 2)
+        self.num_labels = num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.pooler = ContextPooler(config)
+        output_dim = self.pooler.output_dim
+
+        self.classifier = torch.nn.Linear(output_dim, num_labels)
+        drop_out = getattr(config, "cls_dropout", None)
+        drop_out = self.config.hidden_dropout_prob if drop_out is None else drop_out
+        self.dropout = StableDropout(drop_out)
+
+        self.init_weights()
+
+    def get_input_embeddings(self):
+        return self.deberta.get_input_embeddings()
+
+    def set_input_embeddings(self, new_embeddings):
+        self.deberta.set_input_embeddings(new_embeddings)
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=SequenceClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`):
+            Labels for computing the sequence classification/regression loss. Indices should be in :obj:`[0, ...,
+            config.num_labels - 1]`. If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`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.deberta(
+            input_ids,
+            token_type_ids=token_type_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            inputs_embeds=inputs_embeds,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        encoder_layer = outputs[0]
+        pooled_output = self.pooler(encoder_layer)
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        loss = None
+        if labels is not None:
+            if self.num_labels == 1:
+                # regression task
+                loss_fn = torch.nn.MSELoss()
+                logits = logits.view(-1).to(labels.dtype)
+                loss = loss_fn(logits, labels.view(-1))
+            elif labels.dim() == 1 or labels.size(-1) == 1:
+                label_index = (labels >= 0).nonzero()
+                labels = labels.long()
+                if label_index.size(0) > 0:
+                    labeled_logits = torch.gather(logits, 0, label_index.expand(label_index.size(0), logits.size(1)))
+                    labels = torch.gather(labels, 0, label_index.view(-1))
+                    loss_fct = CrossEntropyLoss()
+                    loss = loss_fct(labeled_logits.view(-1, self.num_labels).float(), labels.view(-1))
+                else:
+                    loss = torch.tensor(0).to(logits)
+            else:
+                log_softmax = torch.nn.LogSoftmax(-1)
+                loss = -((log_softmax(logits) * labels).sum(-1)).mean()
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+        else:
+            return SequenceClassifierOutput(
+                loss=loss,
+                logits=logits,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )
+
+
+@add_start_docstrings(
+    """
+    DeBERTa 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.
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForTokenClassification with Deberta->DebertaV2
+class DebertaV2ForTokenClassification(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=TokenClassifierOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        labels=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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()
+            # Only keep active parts of the loss
+            if attention_mask is not None:
+                active_loss = attention_mask.view(-1) == 1
+                active_logits = logits.view(-1, self.num_labels)
+                active_labels = torch.where(
+                    active_loss, labels.view(-1), torch.tensor(loss_fct.ignore_index).type_as(labels)
+                )
+                loss = loss_fct(active_logits, active_labels)
+            else:
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TokenClassifierOutput(
+            loss=loss,
+            logits=logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    DeBERTa 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`).
+    """,
+    DEBERTA_START_DOCSTRING,
+)
+# Copied from transformers.models.deberta.modeling_deberta.DebertaForQuestionAnswering with Deberta->DebertaV2
+class DebertaV2ForQuestionAnswering(DebertaV2PreTrainedModel):
+    _keys_to_ignore_on_load_unexpected = [r"pooler"]
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.deberta = DebertaV2Model(config)
+        self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_model_forward(DEBERTA_INPUTS_DOCSTRING.format("batch_size, sequence_length"))
+    @add_code_sample_docstrings(
+        tokenizer_class=_TOKENIZER_FOR_DOC,
+        checkpoint=_CHECKPOINT_FOR_DOC,
+        output_type=QuestionAnsweringModelOutput,
+        config_class=_CONFIG_FOR_DOC,
+    )
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        inputs_embeds=None,
+        start_positions=None,
+        end_positions=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        start_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`sequence_length`). Position outside of the
+            sequence are not taken into account for computing the loss.
+        end_positions (:obj:`torch.LongTensor` of shape :obj:`(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 (:obj:`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.deberta(
+            input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            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)
+        end_logits = end_logits.squeeze(-1)
+
+        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.clamp_(0, ignored_index)
+            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[1:]
+            return ((total_loss,) + output) if total_loss is not None else output
+
+        return QuestionAnsweringModelOutput(
+            loss=total_loss,
+            start_logits=start_logits,
+            end_logits=end_logits,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+        )

src/transformers/models/deberta_v2/tokenization_deberta_v2.py

+# coding=utf-8
+# Copyright 2020 Microsoft 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.
+""" Tokenization class for model DeBERTa."""
+
+import os
+import unicodedata
+from typing import Optional, Tuple
+
+import sentencepiece as sp
+import six
+
+from ...tokenization_utils import PreTrainedTokenizer
+
+
+PRETRAINED_VOCAB_FILES_MAP = {
+    "vocab_file": {
+        "microsoft/deberta-v2-xlarge": "https://huggingface.co/microsoft/deberta-v2-xlarge/resolve/main/spm.model",
+        "microsoft/deberta-v2-xxlarge": "https://huggingface.co/microsoft/deberta-v2-xxlarge/resolve/main/spm.model",
+        "microsoft/deberta-v2-xlarge-mnli": "https://huggingface.co/microsoft/deberta-v2-xlarge-mnli/resolve/main/spm.model",
+        "microsoft/deberta-v2-xxlarge-mnli": "https://huggingface.co/microsoft/deberta-v2-xxlarge-mnli/resolve/main/spm.model",
+    }
+}
+
+PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES = {
+    "microsoft/deberta-v2-xlarge": 512,
+    "microsoft/deberta-v2-xxlarge": 512,
+    "microsoft/deberta-v2-xlarge-mnli": 512,
+    "microsoft/deberta-v2-xxlarge-mnli": 512,
+}
+
+PRETRAINED_INIT_CONFIGURATION = {
+    "microsoft/deberta-v2-xlarge": {"do_lower_case": False},
+    "microsoft/deberta-v2-xxlarge": {"do_lower_case": False},
+    "microsoft/deberta-v2-xlarge-mnli": {"do_lower_case": False},
+    "microsoft/deberta-v2-xxlarge-mnli": {"do_lower_case": False},
+}
+
+VOCAB_FILES_NAMES = {"vocab_file": "spm.model"}
+
+
+class DebertaV2Tokenizer(PreTrainedTokenizer):
+    r"""
+    Constructs a DeBERTa-v2 tokenizer. Based on `SentencePiece <https://github.com/google/sentencepiece>`__.
+
+    Args:
+        vocab_file (:obj:`str`):
+            `SentencePiece <https://github.com/google/sentencepiece>`__ file (generally has a `.spm` extension) that
+            contains the vocabulary necessary to instantiate a tokenizer.
+        do_lower_case (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            Whether or not to lowercase the input when tokenizing.
+        unk_token (:obj:`str`, `optional`, defaults to :obj:`"[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.
+        sep_token (:obj:`str`, `optional`, defaults to :obj:`"[SEP]"`):
+            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.
+        pad_token (:obj:`str`, `optional`, defaults to :obj:`"[PAD]"`):
+            The token used for padding, for example when batching sequences of different lengths.
+        cls_token (:obj:`str`, `optional`, defaults to :obj:`"[CLS]"`):
+            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.
+        mask_token (:obj:`str`, `optional`, defaults to :obj:`"[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.
+    """
+
+    vocab_files_names = VOCAB_FILES_NAMES
+    pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
+    pretrained_init_configuration = PRETRAINED_INIT_CONFIGURATION
+    max_model_input_sizes = PRETRAINED_POSITIONAL_EMBEDDINGS_SIZES
+
+    def __init__(
+        self,
+        vocab_file,
+        do_lower_case=False,
+        split_by_punct=False,
+        unk_token="[UNK]",
+        sep_token="[SEP]",
+        pad_token="[PAD]",
+        cls_token="[CLS]",
+        mask_token="[MASK]",
+        **kwargs
+    ):
+        super().__init__(
+            do_lower_case=do_lower_case,
+            unk_token=unk_token,
+            sep_token=sep_token,
+            pad_token=pad_token,
+            cls_token=cls_token,
+            mask_token=mask_token,
+            split_by_punct=split_by_punct,
+            **kwargs,
+        )
+
+        if not os.path.isfile(vocab_file):
+            raise ValueError(
+                "Can't find a vocabulary file at path '{}'. To load the vocabulary from a Google pretrained "
+                "model use `tokenizer = DebertaV2Tokenizer.from_pretrained(PRETRAINED_MODEL_NAME)`".format(vocab_file)
+            )
+        self.do_lower_case = do_lower_case
+        self.split_by_punct = split_by_punct
+        self._tokenizer = SPMTokenizer(vocab_file, split_by_punct=split_by_punct)
+
+    @property
+    def vocab_size(self):
+        return len(self.vocab)
+
+    @property
+    def vocab(self):
+        return self._tokenizer.vocab
+
+    def get_vocab(self):
+        vocab = self.vocab.copy()
+        vocab.update(self.get_added_vocab())
+        return vocab
+
+    def _tokenize(self, text):
+        """Take as input a string and return a list of strings (tokens) for words/sub-words"""
+        if self.do_lower_case:
+            text = text.lower()
+        return self._tokenizer.tokenize(text)
+
+    def _convert_token_to_id(self, token):
+        """ Converts a token (str) in an id using the vocab. """
+        return self._tokenizer.spm.PieceToId(token)
+
+    def _convert_id_to_token(self, index):
+        """Converts an index (integer) in a token (str) using the vocab."""
+        return self._tokenizer.spm.IdToPiece(index) if index < self.vocab_size else self.unk_token
+
+    def convert_tokens_to_string(self, tokens):
+        """ Converts a sequence of tokens (string) in a single string. """
+        return self._tokenizer.decode(tokens)
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        """
+        Build model inputs from a sequence or a pair of sequence for sequence classification tasks by concatenating and
+        adding special tokens. A DeBERTa sequence has the following format:
+
+        - single sequence: [CLS] X [SEP]
+        - pair of sequences: [CLS] A [SEP] B [SEP]
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs to which the special tokens will be added.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `input IDs <../glossary.html#input-ids>`__ with the appropriate special tokens.
+        """
+
+        if token_ids_1 is None:
+            return [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+        cls = [self.cls_token_id]
+        sep = [self.sep_token_id]
+        return cls + token_ids_0 + sep + token_ids_1 + sep
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """
+        Retrieves 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`` or ``encode_plus`` methods.
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+            already_has_special_tokens (:obj:`bool`, `optional`, defaults to :obj:`False`):
+                Whether or not the token list is already formatted with special tokens for the model.
+
+        Returns:
+            :obj:`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:
+            if token_ids_1 is not None:
+                raise ValueError(
+                    "You should not supply a second sequence if the provided sequence of "
+                    "ids is already formatted with special tokens for the model."
+                )
+            return list(
+                map(
+                    lambda x: 1 if x in [self.sep_token_id, self.cls_token_id] else 0,
+                    token_ids_0,
+                )
+            )
+
+        if token_ids_1 is not None:
+            return [1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+        return [1] + ([0] * len(token_ids_0)) + [1]
+
+    def create_token_type_ids_from_sequences(self, token_ids_0, token_ids_1=None):
+        """
+        Create a mask from the two sequences passed to be used in a sequence-pair classification task. A DeBERTa
+        sequence pair mask has the following format:
+
+        ::
+
+            0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 1
+            | first sequence    | second sequence |
+
+        If :obj:`token_ids_1` is :obj:`None`, this method only returns the first portion of the mask (0s).
+
+        Args:
+            token_ids_0 (:obj:`List[int]`):
+                List of IDs.
+            token_ids_1 (:obj:`List[int]`, `optional`):
+                Optional second list of IDs for sequence pairs.
+
+        Returns:
+            :obj:`List[int]`: List of `token type IDs <../glossary.html#token-type-ids>`_ according to the given
+            sequence(s).
+        """
+        sep = [self.sep_token_id]
+        cls = [self.cls_token_id]
+        if token_ids_1 is None:
+            return len(cls + token_ids_0 + sep) * [0]
+        return len(cls + token_ids_0 + sep) * [0] + len(token_ids_1 + sep) * [1]
+
+    def prepare_for_tokenization(self, text, is_split_into_words=False, **kwargs):
+        add_prefix_space = kwargs.pop("add_prefix_space", False)
+        if is_split_into_words or add_prefix_space:
+            text = " " + text
+        return (text, kwargs)
+
+    def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        return self._tokenizer.save_pretrained(save_directory, filename_prefix=filename_prefix)
+
+
+class SPMTokenizer:
+    def __init__(self, vocab_file, split_by_punct=False):
+        self.split_by_punct = split_by_punct
+        self.vocab_file = vocab_file
+        spm = sp.SentencePieceProcessor()
+        assert os.path.exists(vocab_file)
+        spm.load(vocab_file)
+        bpe_vocab_size = spm.GetPieceSize()
+        # Token map
+        # <unk> 0+1
+        # <s> 1+1
+        # </s> 2+1
+        self.vocab = {spm.IdToPiece(i): i for i in range(bpe_vocab_size)}
+        self.id_to_tokens = [spm.IdToPiece(i) for i in range(bpe_vocab_size)]
+        # self.vocab['[PAD]'] = 0
+        # self.vocab['[CLS]'] = 1
+        # self.vocab['[SEP]'] = 2
+        # self.vocab['[UNK]'] = 3
+
+        self.spm = spm
+
+    def __getstate__(self):
+        state = self.__dict__.copy()
+        state["spm"] = None
+        return state
+
+    def __setstate__(self, d):
+        self.__dict__ = d
+        self.spm = sp.SentencePieceProcessor()
+        self.spm.Load(self.vocab_file)
+
+    def tokenize(self, text):
+        pieces = self._encode_as_pieces(text)
+
+        def _norm(x):
+            if x not in self.vocab or x == "<unk>":
+                return "[UNK]"
+            else:
+                return x
+
+        pieces = [_norm(p) for p in pieces]
+        return pieces
+
+    def convert_ids_to_tokens(self, ids):
+        tokens = []
+        for i in ids:
+            tokens.append(self.ids_to_tokens[i])
+        return tokens
+
+    def decode(self, tokens, start=-1, end=-1, raw_text=None):
+        if raw_text is None:
+            return self.spm.decode_pieces([t for t in tokens])
+        else:
+            words = self.split_to_words(raw_text)
+            word_tokens = [self.tokenize(w) for w in words]
+            token2words = [0] * len(tokens)
+            tid = 0
+            for i, w in enumerate(word_tokens):
+                for k, t in enumerate(w):
+                    token2words[tid] = i
+                    tid += 1
+            word_start = token2words[start]
+            word_end = token2words[end] if end < len(tokens) else len(words)
+            text = "".join(words[word_start:word_end])
+            return text
+
+    def add_special_token(self, token):
+        if token not in self.special_tokens:
+            self.special_tokens.append(token)
+            if token not in self.vocab:
+                self.vocab[token] = len(self.vocab) - 1
+                self.id_to_tokens.append(token)
+        return self.id(token)
+
+    def part_of_whole_word(self, token, is_bos=False):
+        if is_bos:
+            return True
+        if (
+            len(token) == 1
+            and (_is_whitespace(list(token)[0]) or _is_control(list(token)[0]) or _is_punctuation(list(token)[0]))
+        ) or token in self.special_tokens:
+            return False
+
+        word_start = b"\xe2\x96\x81".decode("utf-8")
+        return not token.startswith(word_start)
+
+    def pad(self):
+        return "[PAD]"
+
+    def bos(self):
+        return "[CLS]"
+
+    def eos(self):
+        return "[SEP]"
+
+    def unk(self):
+        return "[UNK]"
+
+    def mask(self):
+        return "[MASK]"
+
+    def sym(self, id):
+        return self.ids_to_tokens[id]
+
+    def id(self, sym):
+        return self.vocab[sym] if sym in self.vocab else 1
+
+    def _encode_as_pieces(self, text):
+        text = convert_to_unicode(text)
+        if self.split_by_punct:
+            words = self._run_split_on_punc(text)
+            pieces = [self.spm.encode_as_pieces(w) for w in words]
+            return [p for w in pieces for p in w]
+        else:
+            return self.spm.encode_as_pieces(text)
+
+    def split_to_words(self, text):
+        pieces = self._encode_as_pieces(text)
+        word_start = b"\xe2\x96\x81".decode("utf-8")
+        words = []
+        offset = 0
+        prev_end = 0
+        for i, p in enumerate(pieces):
+            if p.startswith(word_start):
+                if offset > prev_end:
+                    words.append(text[prev_end:offset])
+                prev_end = offset
+                w = p.replace(word_start, "")
+            else:
+                w = p
+            try:
+                s = text.index(w, offset)
+                pn = ""
+                k = i + 1
+                while k < len(pieces):
+                    pn = pieces[k].replace(word_start, "")
+                    if len(pn) > 0:
+                        break
+                    k += 1
+
+                if len(pn) > 0 and pn in text[offset:s]:
+                    offset = offset + 1
+                else:
+                    offset = s + len(w)
+            except Exception:
+                offset = offset + 1
+
+        if prev_end < offset:
+            words.append(text[prev_end:offset])
+
+        return words
+
+    def _run_strip_accents(self, text):
+        """Strips accents from a piece of text."""
+        text = unicodedata.normalize("NFD", text)
+        output = []
+        for char in text:
+            cat = unicodedata.category(char)
+            if cat == "Mn":
+                continue
+            output.append(char)
+        return "".join(output)
+
+    def _run_split_on_punc(self, text):
+        """Splits punctuation on a piece of text."""
+        chars = list(text)
+        i = 0
+        start_new_word = True
+        output = []
+        while i < len(chars):
+            char = chars[i]
+            if _is_punctuation(char):
+                output.append([char])
+                start_new_word = True
+            else:
+                if start_new_word:
+                    output.append([])
+                start_new_word = False
+                output[-1].append(char)
+            i += 1
+
+        return ["".join(x) for x in output]
+
+    def save_pretrained(self, path: str, filename_prefix: str = None):
+        filename = VOCAB_FILES_NAMES[list(VOCAB_FILES_NAMES.keys())[0]]
+        if filename_prefix is not None:
+            filename = filename_prefix + "-" + filename
+        full_path = os.path.join(path, filename)
+        with open(full_path, "wb") as fs:
+            fs.write(self.spm.serialized_model_proto())
+        return (full_path,)
+
+
+def _is_whitespace(char):
+    """Checks whether `chars` is a whitespace character."""
+    # \t, \n, and \r are technically contorl characters but we treat them
+    # as whitespace since they are generally considered as such.
+    if char == " " or char == "\t" or char == "\n" or char == "\r":
+        return True
+    cat = unicodedata.category(char)
+    if cat == "Zs":
+        return True
+    return False
+
+
+def _is_control(char):
+    """Checks whether `chars` is a control character."""
+    # These are technically control characters but we count them as whitespace
+    # characters.
+    if char == "\t" or char == "\n" or char == "\r":
+        return False
+    cat = unicodedata.category(char)
+    if cat.startswith("C"):
+        return True
+    return False
+
+
+def _is_punctuation(char):
+    """Checks whether `chars` is a punctuation character."""
+    cp = ord(char)
+    # We treat all non-letter/number ASCII as punctuation.
+    # Characters such as "^", "$", and "`" are not in the Unicode
+    # Punctuation class but we treat them as punctuation anyways, for
+    # consistency.
+    if (cp >= 33 and cp <= 47) or (cp >= 58 and cp <= 64) or (cp >= 91 and cp <= 96) or (cp >= 123 and cp <= 126):
+        return True
+    cat = unicodedata.category(char)
+    if cat.startswith("P"):
+        return True
+    return False
+
+
+def convert_to_unicode(text):
+    """Converts `text` to Unicode (if it's not already), assuming utf-8 input."""
+    if six.PY3:
+        if isinstance(text, str):
+            return text
+        elif isinstance(text, bytes):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    elif six.PY2:
+        if isinstance(text, str):
+            return text.decode("utf-8", "ignore")
+        else:
+            raise ValueError("Unsupported string type: %s" % (type(text)))
+    else:
+        raise ValueError("Not running on Python2 or Python 3?")

src/transformers/utils/dummy_pt_objects.py

@@ -883,6 +883,63 @@ class DebertaPreTrainedModel:
         requires_pytorch(self)
 
 
+DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class DebertaV2ForMaskedLM:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
+class DebertaV2ForQuestionAnswering:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
+class DebertaV2ForSequenceClassification:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
+class DebertaV2ForTokenClassification:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
+class DebertaV2Model:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
+class DebertaV2PreTrainedModel:
+    def __init__(self, *args, **kwargs):
+        requires_pytorch(self)
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_pytorch(self)
+
+
 DISTILBERT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/test_modeling_deberta_v2.py

+# coding=utf-8
+# Copyright 2018 Microsoft Authors 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.
+
+
+import random
+import unittest
+
+import numpy as np
+
+from transformers import is_torch_available
+from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
+
+from .test_configuration_common import ConfigTester
+from .test_modeling_common import ModelTesterMixin, ids_tensor
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        DebertaV2Config,
+        DebertaV2ForMaskedLM,
+        DebertaV2ForQuestionAnswering,
+        DebertaV2ForSequenceClassification,
+        DebertaV2ForTokenClassification,
+        DebertaV2Model,
+    )
+    from transformers.models.deberta_v2.modeling_deberta_v2 import DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+@require_torch
+class DebertaV2ModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            DebertaV2Model,
+            DebertaV2ForMaskedLM,
+            DebertaV2ForSequenceClassification,
+            DebertaV2ForTokenClassification,
+            DebertaV2ForQuestionAnswering,
+        )
+        if is_torch_available()
+        else ()
+    )
+
+    test_torchscript = False
+    test_pruning = False
+    test_head_masking = False
+    is_encoder_decoder = False
+
+    class DebertaV2ModelTester(object):
+        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,
+            relative_attention=False,
+            position_biased_input=True,
+            pos_att_type="None",
+            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.relative_attention = relative_attention
+            self.position_biased_input = position_biased_input
+            self.pos_att_type = pos_att_type
+            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 = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
+
+            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 = DebertaV2Config(
+                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,
+                relative_attention=self.relative_attention,
+                position_biased_input=self.position_biased_input,
+                pos_att_type=self.pos_att_type,
+            )
+
+            return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+        def check_loss_output(self, result):
+            self.parent.assertListEqual(list(result.loss.size()), [])
+
+        def create_and_check_deberta_model(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            model = DebertaV2Model(config=config)
+            model.to(torch_device)
+            model.eval()
+            sequence_output = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)[0]
+            sequence_output = model(input_ids, token_type_ids=token_type_ids)[0]
+            sequence_output = model(input_ids)[0]
+
+            self.parent.assertListEqual(
+                list(sequence_output.size()), [self.batch_size, self.seq_length, self.hidden_size]
+            )
+
+        def create_and_check_deberta_for_masked_lm(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            model = DebertaV2ForMaskedLM(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_deberta_for_sequence_classification(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            config.num_labels = self.num_labels
+            model = DebertaV2ForSequenceClassification(config)
+            model.to(torch_device)
+            model.eval()
+            result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
+            self.parent.assertListEqual(list(result.logits.size()), [self.batch_size, self.num_labels])
+            self.check_loss_output(result)
+
+        def create_and_check_deberta_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 = DebertaV2ForTokenClassification(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_deberta_for_question_answering(
+            self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+        ):
+            model = DebertaV2ForQuestionAnswering(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
+
+    def setUp(self):
+        self.model_tester = DebertaV2ModelTest.DebertaV2ModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=DebertaV2Config, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_deberta_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deberta_model(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_deberta_for_sequence_classification(*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_deberta_for_masked_lm(*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_deberta_for_question_answering(*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_deberta_for_token_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in DEBERTA_V2_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = DebertaV2Model.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class DebertaV2ModelIntegrationTest(unittest.TestCase):
+    @unittest.skip(reason="Model not available yet")
+    def test_inference_masked_lm(self):
+        pass
+
+    @slow
+    def test_inference_no_head(self):
+        random.seed(0)
+        np.random.seed(0)
+        torch.manual_seed(0)
+        torch.cuda.manual_seed_all(0)
+        model = DebertaV2Model.from_pretrained("microsoft/deberta-v2-xlarge")
+
+        input_ids = torch.tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
+        output = model(input_ids)[0]
+        # compare the actual values for a slice.
+        expected_slice = torch.tensor(
+            [[[-0.2913, 0.2647, 0.5627], [-0.4318, 0.1389, 0.3881], [-0.2929, -0.2489, 0.3452]]]
+        )
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4), f"{output[:, :3, :3]}")

tests/test_tokenization_deberta_v2.py

+# coding=utf-8
+# Copyright 2019 Hugging Face inc.
+#
+# 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 os
+import unittest
+
+from transformers import DebertaV2Tokenizer
+from transformers.testing_utils import require_sentencepiece, require_tokenizers
+
+from .test_tokenization_common import TokenizerTesterMixin
+
+
+SAMPLE_VOCAB = os.path.join(os.path.dirname(os.path.abspath(__file__)), "fixtures/spiece.model")
+
+
+@require_sentencepiece
+@require_tokenizers
+class DebertaV2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
+
+    tokenizer_class = DebertaV2Tokenizer
+    rust_tokenizer_class = None
+    test_rust_tokenizer = False
+
+    def setUp(self):
+        super().setUp()
+
+        # We have a SentencePiece fixture for testing
+        tokenizer = DebertaV2Tokenizer(SAMPLE_VOCAB)
+        tokenizer.save_pretrained(self.tmpdirname)
+
+    def get_input_output_texts(self, tokenizer):
+        input_text = "this is a test"
+        output_text = "this is a test"
+        return input_text, output_text
+
+    def test_rust_and_python_full_tokenizers(self):
+        if not self.test_rust_tokenizer:
+            return
+
+        tokenizer = self.get_tokenizer()
+        rust_tokenizer = self.get_rust_tokenizer()
+
+        sequence = "I was born in 92000, and this is falsé."
+
+        tokens = tokenizer.tokenize(sequence)
+        rust_tokens = rust_tokenizer.tokenize(sequence)
+        self.assertListEqual(tokens, rust_tokens)
+
+        ids = tokenizer.encode(sequence, add_special_tokens=False)
+        rust_ids = rust_tokenizer.encode(sequence, add_special_tokens=False)
+        self.assertListEqual(ids, rust_ids)
+
+        rust_tokenizer = self.get_rust_tokenizer()
+        ids = tokenizer.encode(sequence)
+        rust_ids = rust_tokenizer.encode(sequence)
+        self.assertListEqual(ids, rust_ids)
+
+    def test_full_tokenizer(self):
+        tokenizer = DebertaV2Tokenizer(SAMPLE_VOCAB, keep_accents=True)
+
+        tokens = tokenizer.tokenize("This is a test")
+        self.assertListEqual(tokens, ["▁", "[UNK]", "his", "▁is", "▁a", "▁test"])
+
+        self.assertListEqual(tokenizer.convert_tokens_to_ids(tokens), [13, 1, 4398, 25, 21, 1289])
+
+        tokens = tokenizer.tokenize("I was born in 92000, and this is falsé.")
+        # fmt: off
+        self.assertListEqual(
+            tokens,
+            ["▁", "[UNK]", "▁was", "▁born", "▁in", "▁9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "[UNK]", "."],
+        )
+        ids = tokenizer.convert_tokens_to_ids(tokens)
+        self.assertListEqual(ids, [13, 1, 23, 386, 19, 561, 3050, 15, 17, 48, 25, 8256, 18, 1, 9])
+
+        back_tokens = tokenizer.convert_ids_to_tokens(ids)
+        self.assertListEqual(
+            back_tokens,
+            ["▁", "<unk>", "▁was", "▁born", "▁in", "▁9", "2000", ",", "▁and", "▁this", "▁is", "▁fal", "s", "<unk>", "."],
+        )
+        # fmt: on
+
+    def test_sequence_builders(self):
+        tokenizer = DebertaV2Tokenizer(SAMPLE_VOCAB)
+
+        text = tokenizer.encode("sequence builders")
+        text_2 = tokenizer.encode("multi-sequence build")
+
+        encoded_sentence = tokenizer.build_inputs_with_special_tokens(text)
+        encoded_pair = tokenizer.build_inputs_with_special_tokens(text, text_2)
+
+        assert encoded_sentence == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id]
+        assert encoded_pair == [tokenizer.cls_token_id] + text + [tokenizer.sep_token_id] + text_2 + [
+            tokenizer.sep_token_id
+        ]
+
+    def test_tokenizer_integration(self):
+        tokenizer_classes = [self.tokenizer_class]
+        if self.test_rust_tokenizer:
+            tokenizer_classes.append(self.rust_tokenizer_class)
+
+        for tokenizer_class in tokenizer_classes:
+            tokenizer = tokenizer_class.from_pretrained("microsoft/deberta-xlarge-v2")
+
+            sequences = [
+                [
+                    "DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
+                    "DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
+                ],
+                [
+                    "Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks.",
+                    "DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
+                ],
+                [
+                    "In this paper we propose a new model architecture DeBERTa",
+                    "DeBERTa: Decoding-enhanced BERT with Disentangled Attention",
+                ],
+            ]
+
+            encoding = tokenizer(sequences, padding=True)
+            decoded_sequences = [tokenizer.decode(seq, skip_special_tokens=True) for seq in encoding["input_ids"]]
+
+            # fmt: off
+            expected_encoding = {
+                'input_ids': [
+                    [1, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    [1, 9755, 1944, 11, 1053, 18, 16899, 12730, 1072, 1506, 45, 2497, 2510, 5, 610, 9, 127, 699, 1072, 2101, 36, 99388, 53, 2930, 4, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2],
+                    [1, 84, 32, 778, 42, 9441, 10, 94, 735, 3372, 1804, 69418, 191, 2, 1804, 69418, 191, 43, 117056, 18, 44596, 448, 37132, 19, 8655, 10625, 69860, 21149, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
+                'token_type_ids': [
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],
+                'attention_mask': [
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1],
+                    [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
+                ]
+            }
+
+            expected_decoded_sequences = [
+                'DeBERTa: Decoding-enhanced BERT with Disentangled Attention DeBERTa: Decoding-enhanced BERT with Disentangled Attention',
+                'Recent progress in pre-trained neural language models has significantly improved the performance of many natural language processing (NLP) tasks. DeBERTa: Decoding-enhanced BERT with Disentangled Attention',
+                'In this paper we propose a new model architecture DeBERTa DeBERTa: Decoding-enhanced BERT with Disentangled Attention'
+            ]
+            # fmt: on
+
+            self.assertDictEqual(encoding.data, expected_encoding)
+
+            for expected, decoded in zip(expected_decoded_sequences, decoded_sequences):
+                self.assertEqual(expected, decoded)