Add nvidia megatron models (#10911)

* Add support for NVIDIA Megatron models

* Add support for NVIDIA Megatron GPT2 and BERT

Add the megatron_gpt2 model. That model reuses the existing GPT2 model. This
commit includes a script to convert a Megatron-GPT2 checkpoint downloaded
from NVIDIA GPU Cloud. See examples/megatron-models/README.md for details.

Add the megatron_bert model. That model is implemented as a modification of
the existing BERT model in Transformers. This commit includes a script to
convert a Megatron-BERT checkpoint downloaded from NVIDIA GPU Cloud. See
examples/megatron-models/README.md for details.

* Update src/transformers/models/megatron_bert/configuration_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/configuration_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/configuration_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Remove model.half in tests + add "# Copied ..."

Remove the model.half() instruction which makes tests fail on the CPU.

Add a comment "# Copied ..." before many classes in the model to enable automatic
tracking in CI between the new Megatron classes and the original Bert ones.

* Fix issues

* Fix Flax/TF tests

* Fix copyright

* Update src/transformers/models/megatron_bert/configuration_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/configuration_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Update docs/source/model_doc/megatron_bert.rst

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update docs/source/model_doc/megatron_gpt2.rst

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/__init__.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/megatron_bert/modeling_megatron_bert.py

Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Resolve most of 'sgugger' comments

* Fix conversion issue + Run make fix-copies/quality/docs

* Apply suggestions from code review

* Causal LM & merge

* Fix init

* Add CausalLM to last auto class

Co-authored-by: Julien Demouth <jdemouth@nvidia.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

README.md

@@ -223,6 +223,8 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[MarianMT](https://huggingface.co/transformers/model_doc/marian.html)** 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.
 1. **[MBart](https://huggingface.co/transformers/model_doc/mbart.html)** (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.
 1. **[MBart-50](https://huggingface.co/transformers/model_doc/mbart.html)** (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.
+1. **[Megatron-BERT](https://huggingface.co/transformers/model_doc/megatron_bert.html)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
+1. **[Megatron-GPT2](https://huggingface.co/transformers/model_doc/megatron_gpt2.html)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
 1. **[MPNet](https://huggingface.co/transformers/model_doc/mpnet.html)** (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.
 1. **[MT5](https://huggingface.co/transformers/model_doc/mt5.html)** (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.
 1. **[Pegasus](https://huggingface.co/transformers/model_doc/pegasus.html)** (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.

docs/source/index.rst

@@ -178,58 +178,64 @@ and conversion utilities for the following models:
 32. :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.
-33. :doc:`MPNet <model_doc/mpnet>` (from Microsoft Research) released with the paper `MPNet: Masked and Permuted
+33. :doc:`Megatron-BERT <model_doc/megatron_bert>` (from NVIDIA) released with the paper `Megatron-LM: Training
+    Multi-Billion Parameter Language Models Using Model Parallelism <https://arxiv.org/abs/1909.08053>`__ by Mohammad
+    Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
+34. :doc:`Megatron-GPT2 <model_doc/megatron_gpt2>` (from NVIDIA) released with the paper `Megatron-LM: Training
+    Multi-Billion Parameter Language Models Using Model Parallelism <https://arxiv.org/abs/1909.08053>`__ by Mohammad
+    Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
+35. :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.
-34. :doc:`MT5 <model_doc/mt5>` (from Google AI) released with the paper `mT5: A massively multilingual pre-trained
+36. :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.
-35. :doc:`Pegasus <model_doc/pegasus>` (from Google) released with the paper `PEGASUS: Pre-training with Extracted
+37. :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.
-36. :doc:`ProphetNet <model_doc/prophetnet>` (from Microsoft Research) released with the paper `ProphetNet: Predicting
+38. :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.
-37. :doc:`Reformer <model_doc/reformer>` (from Google Research) released with the paper `Reformer: The Efficient
+39. :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.
-38. :doc:`RoBERTa <model_doc/roberta>` (from Facebook), released together with the paper a `Robustly Optimized BERT
+40. :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.
-39. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
+41. :doc:`SpeechToTextTransformer <model_doc/speech_to_text>` (from Facebook), released together with the paper
     `fairseq S2T: Fast Speech-to-Text Modeling with fairseq <https://arxiv.org/abs/2010.05171>`__ by Changhan Wang, Yun
     Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino.
-40. :doc:`SqueezeBert <model_doc/squeezebert>` released with the paper `SqueezeBERT: What can computer vision teach NLP
+42. :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.
-41. :doc:`T5 <model_doc/t5>` (from Google AI) released with the paper `Exploring the Limits of Transfer Learning with a
+43. :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.
-42. :doc:`TAPAS <model_doc/tapas>` (from Google AI) released with the paper `TAPAS: Weakly Supervised Table Parsing via
+44. :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.
-43. :doc:`Transformer-XL <model_doc/transformerxl>` (from Google/CMU) released with the paper `Transformer-XL:
+45. :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.
-44. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
+46. :doc:`Vision Transformer (ViT) <model_doc/vit>` (from Google AI) released with the paper `An Image is Worth 16x16
     Words: Transformers for Image Recognition at Scale <https://arxiv.org/abs/2010.11929>`__ by Alexey Dosovitskiy,
     Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias
     Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
-45. :doc:`Wav2Vec2 <model_doc/wav2vec2>` (from Facebook AI) released with the paper `wav2vec 2.0: A Framework for
+47. :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.
-46. :doc:`XLM <model_doc/xlm>` (from Facebook) released together with the paper `Cross-lingual Language Model
+48. :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.
-47. :doc:`XLM-ProphetNet <model_doc/xlmprophetnet>` (from Microsoft Research) released with the paper `ProphetNet:
+49. :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.
-48. :doc:`XLM-RoBERTa <model_doc/xlmroberta>` (from Facebook AI), released together with the paper `Unsupervised
+50. :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.
-49. :doc:`XLNet <model_doc/xlnet>` (from Google/CMU) released with the paper `​XLNet: Generalized Autoregressive
+51. :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.
-50. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
+52. :doc:`XLSR-Wav2Vec2 <model_doc/xlsr_wav2vec2>` (from Facebook AI) released with the paper `Unsupervised
     Cross-Lingual Representation Learning For Speech Recognition <https://arxiv.org/abs/2006.13979>`__ by Alexis
     Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
 
@@ -304,6 +310,8 @@ TensorFlow and/or Flax.
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |           Marian            |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
+|        MegatronBert         |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
++-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         MobileBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 +-----------------------------+----------------+----------------+-----------------+--------------------+--------------+
 |         OpenAI GPT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
@@ -449,6 +457,8 @@ TensorFlow and/or Flax.
     model_doc/marian
     model_doc/m2m_100
     model_doc/mbart
+    model_doc/megatron_bert
+    model_doc/megatron_gpt2
     model_doc/mobilebert
     model_doc/mpnet
     model_doc/mt5

docs/source/model_doc/megatron_bert.rst

@@ -0,0 +1,153 @@
+.. 
+    Copyright 2021 NVIDIA Corporation and 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.
+
+MegatronBERT
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The MegatronBERT model was proposed in `Megatron-LM: Training Multi-Billion Parameter Language Models Using Model
+Parallelism <https://arxiv.org/abs/1909.08053>`__ by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley,
+Jared Casper and Bryan Catanzaro.
+
+The abstract from the paper is the following:
+
+*Recent work in language modeling demonstrates that training large transformer models advances the state of the art in
+Natural Language Processing applications. However, very large models can be quite difficult to train due to memory
+constraints. In this work, we present our techniques for training very large transformer models and implement a simple,
+efficient intra-layer model parallel approach that enables training transformer models with billions of parameters. Our
+approach does not require a new compiler or library changes, is orthogonal and complimentary to pipeline model
+parallelism, and can be fully implemented with the insertion of a few communication operations in native PyTorch. We
+illustrate this approach by converging transformer based models up to 8.3 billion parameters using 512 GPUs. We sustain
+15.1 PetaFLOPs across the entire application with 76% scaling efficiency when compared to a strong single GPU baseline
+that sustains 39 TeraFLOPs, which is 30% of peak FLOPs. To demonstrate that large language models can further advance
+the state of the art (SOTA), we train an 8.3 billion parameter transformer language model similar to GPT-2 and a 3.9
+billion parameter model similar to BERT. We show that careful attention to the placement of layer normalization in
+BERT-like models is critical to achieving increased performance as the model size grows. Using the GPT-2 model we
+achieve SOTA results on the WikiText103 (10.8 compared to SOTA perplexity of 15.8) and LAMBADA (66.5% compared to SOTA
+accuracy of 63.2%) datasets. Our BERT model achieves SOTA results on the RACE dataset (90.9% compared to SOTA accuracy
+of 89.4%).*
+
+Tips:
+
+We have provided pretrained `BERT-345M <https://ngc.nvidia.com/catalog/models/nvidia:megatron_bert_345m>`__ checkpoints
+for use to evaluate or finetuning downstream tasks.
+
+To access these checkpoints, first `sign up <https://ngc.nvidia.com/signup>`__ for and setup the NVIDIA GPU Cloud (NGC)
+Registry CLI. Further documentation for downloading models can be found in the `NGC documentation
+<https://docs.nvidia.com/dgx/ngc-registry-cli-user-guide/index.html#topic_6_4_1>`__.
+
+Alternatively, you can directly download the checkpoints using:
+
+BERT-345M-uncased::
+
+.. code-block:: bash
+
+  wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_bert_345m/versions/v0.1_uncased/zip
+  -O megatron_bert_345m_v0_1_uncased.zip
+
+BERT-345M-cased::
+
+.. code-block:: bash
+
+  wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_bert_345m/versions/v0.1_cased/zip -O
+  megatron_bert_345m_v0_1_cased.zip
+
+Once you have obtained the checkpoints from NVIDIA GPU Cloud (NGC), you have to convert them to a format that will
+easily be loaded by Hugging Face Transformers and our port of the BERT code.
+
+The following commands allow you to do the conversion. We assume that the folder ``models/megatron_bert`` contains
+``megatron_bert_345m_v0_1_{cased, uncased}.zip`` and that the commands are run from inside that folder::
+
+.. code-block:: bash
+
+  python3 $PATH_TO_TRANSFORMERS/models/megatron_bert/convert_megatron_bert_checkpoint.py megatron_bert_345m_v0_1_uncased.zip 
+
+.. code-block:: bash
+
+  python3 $PATH_TO_TRANSFORMERS/models/megatron_bert/convert_megatron_bert_checkpoint.py megatron_bert_345m_v0_1_cased.zip
+
+The original code can be found `here <https://github.com/NVIDIA/Megatron-LM>`__. That repository contains a multi-GPU
+and multi-node implementation of the Megatron Language models. In particular, it contains a hybrid model parallel
+approach using "tensor parallel" and "pipeline parallel" techniques.
+
+MegatronBertConfig
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertConfig
+    :members:
+
+
+MegatronBertModel
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertModel
+    :members: forward
+
+
+MegatronBertForMaskedLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForMaskedLM
+    :members: forward
+
+
+MegatronBertForCausalLM
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForCausalLM
+    :members: forward
+
+
+MegatronBertForNextSentencePrediction
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForNextSentencePrediction
+    :members: forward
+
+
+MegatronBertForPreTraining
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForPreTraining
+    :members: forward
+
+
+MegatronBertForSequenceClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForSequenceClassification
+    :members: forward
+
+
+MegatronBertForMultipleChoice
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForMultipleChoice
+    :members: forward
+
+
+MegatronBertForTokenClassification
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForTokenClassification
+    :members: forward
+
+
+MegatronBertForQuestionAnswering
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+.. autoclass:: transformers.MegatronBertForQuestionAnswering
+    :members: forward
+
+

docs/source/model_doc/megatron_gpt2.rst

@@ -0,0 +1,70 @@
+.. 
+    Copyright 2021 NVIDIA Corporation and 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.
+
+MegatronGPT2
+-----------------------------------------------------------------------------------------------------------------------
+
+Overview
+~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+
+The MegatronGPT2 model was proposed in `Megatron-LM: Training Multi-Billion Parameter Language Models Using Model
+Parallelism <https://arxiv.org/abs/1909.08053>`__ by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley,
+Jared Casper and Bryan Catanzaro.
+
+The abstract from the paper is the following:
+
+*Recent work in language modeling demonstrates that training large transformer models advances the state of the art in
+Natural Language Processing applications. However, very large models can be quite difficult to train due to memory
+constraints. In this work, we present our techniques for training very large transformer models and implement a simple,
+efficient intra-layer model parallel approach that enables training transformer models with billions of parameters. Our
+approach does not require a new compiler or library changes, is orthogonal and complimentary to pipeline model
+parallelism, and can be fully implemented with the insertion of a few communication operations in native PyTorch. We
+illustrate this approach by converging transformer based models up to 8.3 billion parameters using 512 GPUs. We sustain
+15.1 PetaFLOPs across the entire application with 76% scaling efficiency when compared to a strong single GPU baseline
+that sustains 39 TeraFLOPs, which is 30% of peak FLOPs. To demonstrate that large language models can further advance
+the state of the art (SOTA), we train an 8.3 billion parameter transformer language model similar to GPT-2 and a 3.9
+billion parameter model similar to BERT. We show that careful attention to the placement of layer normalization in
+BERT-like models is critical to achieving increased performance as the model size grows. Using the GPT-2 model we
+achieve SOTA results on the WikiText103 (10.8 compared to SOTA perplexity of 15.8) and LAMBADA (66.5% compared to SOTA
+accuracy of 63.2%) datasets. Our BERT model achieves SOTA results on the RACE dataset (90.9% compared to SOTA accuracy
+of 89.4%).*
+
+Tips:
+
+We have provided pretrained `GPT2-345M <https://ngc.nvidia.com/catalog/models/nvidia:megatron_lm_345m>`__ checkpoints
+for use to evaluate or finetuning downstream tasks.
+
+To access these checkpoints, first `sign up <https://ngc.nvidia.com/signup>`__ for and setup the NVIDIA GPU Cloud (NGC)
+Registry CLI. Further documentation for downloading models can be found in the `NGC documentation
+<https://docs.nvidia.com/dgx/ngc-registry-cli-user-guide/index.html#topic_6_4_1>`__.
+
+Alternatively, you can directly download the checkpoints using::
+
+.. code-block:: bash
+
+  wget --content-disposition https://api.ngc.nvidia.com/v2/models/nvidia/megatron_lm_345m/versions/v0.0/zip -O
+  megatron_gpt2_345m_v0_0.zip
+
+Once you have obtained the checkpoint from NVIDIA GPU Cloud (NGC), you have to convert it to a format that will easily
+be loaded by Hugging Face Transformers GPT2 implementation.
+
+The following command allows you to do the conversion. We assume that the folder ``models/megatron_gpt2`` contains
+``megatron_gpt2_345m_v0_0.zip`` and that the command is run from that folder::
+
+.. code-block:: bash
+
+  python3 $PATH_TO_TRANSFORMERS/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py megatron_gpt2_345m_v0_0.zip
+
+The original code can be found `here <https://github.com/NVIDIA/Megatron-LM>`__. That repository contains a multi-GPU
+and multi-node implementation of the Megatron Language models. In particular, it contains a hybrid model parallel
+approach using "tensor parallel" and "pipeline parallel" techniques.
+

src/transformers/__init__.py

@@ -191,6 +191,7 @@ _import_structure = {
     "models.m2m_100": ["M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP", "M2M100Config"],
     "models.marian": ["MarianConfig"],
     "models.mbart": ["MBartConfig"],
+    "models.megatron_bert": ["MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MegatronBertConfig"],
     "models.mmbt": ["MMBTConfig"],
     "models.mobilebert": ["MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MobileBertConfig", "MobileBertTokenizer"],
     "models.mpnet": ["MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "MPNetConfig", "MPNetTokenizer"],
@@ -765,6 +766,20 @@ if is_torch_available():
             "MBartModel",
         ]
     )
+    _import_structure["models.megatron_bert"].extend(
+        [
+            "MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "MegatronBertForCausalLM",
+            "MegatronBertForMaskedLM",
+            "MegatronBertForMultipleChoice",
+            "MegatronBertForNextSentencePrediction",
+            "MegatronBertForPreTraining",
+            "MegatronBertForQuestionAnswering",
+            "MegatronBertForSequenceClassification",
+            "MegatronBertForTokenClassification",
+            "MegatronBertModel",
+        ]
+    )
     _import_structure["models.mmbt"].extend(["MMBTForClassification", "MMBTModel", "ModalEmbeddings"])
     _import_structure["models.mobilebert"].extend(
         [
@@ -1514,6 +1529,7 @@ if TYPE_CHECKING:
     from .models.m2m_100 import M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP, M2M100Config
     from .models.marian import MarianConfig
     from .models.mbart import MBartConfig
+    from .models.megatron_bert import MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MegatronBertConfig
     from .models.mmbt import MMBTConfig
     from .models.mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig, MobileBertTokenizer
     from .models.mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig, MPNetTokenizer
@@ -1999,6 +2015,18 @@ if TYPE_CHECKING:
             MBartForSequenceClassification,
             MBartModel,
         )
+        from .models.megatron_bert import (
+            MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            MegatronBertForCausalLM,
+            MegatronBertForMaskedLM,
+            MegatronBertForMultipleChoice,
+            MegatronBertForNextSentencePrediction,
+            MegatronBertForPreTraining,
+            MegatronBertForQuestionAnswering,
+            MegatronBertForSequenceClassification,
+            MegatronBertForTokenClassification,
+            MegatronBertModel,
+        )
         from .models.mmbt import MMBTForClassification, MMBTModel, ModalEmbeddings
         from .models.mobilebert import (
             MOBILEBERT_PRETRAINED_MODEL_ARCHIVE_LIST,

src/transformers/models/__init__.py

@@ -50,6 +50,7 @@ from . import (
     m2m_100,
     marian,
     mbart,
+    megatron_bert,
     mmbt,
     mobilebert,
     mpnet,

src/transformers/models/auto/configuration_auto.py

@@ -50,6 +50,7 @@ from ..lxmert.configuration_lxmert import LXMERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
 from ..m2m_100.configuration_m2m_100 import M2M_100_PRETRAINED_CONFIG_ARCHIVE_MAP, M2M100Config
 from ..marian.configuration_marian import MarianConfig
 from ..mbart.configuration_mbart import MBART_PRETRAINED_CONFIG_ARCHIVE_MAP, MBartConfig
+from ..megatron_bert.configuration_megatron_bert import MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MegatronBertConfig
 from ..mobilebert.configuration_mobilebert import MobileBertConfig
 from ..mpnet.configuration_mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig
 from ..mt5.configuration_mt5 import MT5Config
@@ -85,6 +86,7 @@ ALL_PRETRAINED_CONFIG_ARCHIVE_MAP = dict(
         # Add archive maps here
         GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP,
         BIG_BIRD_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         SPEECH_TO_TEXT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         VIT_PRETRAINED_CONFIG_ARCHIVE_MAP,
         WAV_2_VEC_2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -155,6 +157,7 @@ CONFIG_MAPPING = OrderedDict(
         ("pegasus", PegasusConfig),
         ("marian", MarianConfig),
         ("mbart", MBartConfig),
+        ("megatron_bert", MegatronBertConfig),
         ("mpnet", MPNetConfig),
         ("bart", BartConfig),
         ("blenderbot", BlenderbotConfig),
@@ -211,6 +214,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("blenderbot", "Blenderbot"),
         ("marian", "Marian"),
         ("mbart", "mBART"),
+        ("megatron_bert", "MegatronBert"),
         ("bart", "BART"),
         ("reformer", "Reformer"),
         ("longformer", "Longformer"),

src/transformers/models/auto/modeling_auto.py

@@ -174,6 +174,17 @@ from ..mbart.modeling_mbart import (
     MBartForSequenceClassification,
     MBartModel,
 )
+from ..megatron_bert.modeling_megatron_bert import (
+    MegatronBertForCausalLM,
+    MegatronBertForMaskedLM,
+    MegatronBertForMultipleChoice,
+    MegatronBertForNextSentencePrediction,
+    MegatronBertForPreTraining,
+    MegatronBertForQuestionAnswering,
+    MegatronBertForSequenceClassification,
+    MegatronBertForTokenClassification,
+    MegatronBertModel,
+)
 from ..mobilebert.modeling_mobilebert import (
     MobileBertForMaskedLM,
     MobileBertForMultipleChoice,
@@ -298,6 +309,7 @@ from .configuration_auto import (
     M2M100Config,
     MarianConfig,
     MBartConfig,
+    MegatronBertConfig,
     MobileBertConfig,
     MPNetConfig,
     MT5Config,
@@ -355,6 +367,7 @@ MODEL_MAPPING = OrderedDict(
         (BertConfig, BertModel),
         (OpenAIGPTConfig, OpenAIGPTModel),
         (GPT2Config, GPT2Model),
+        (MegatronBertConfig, MegatronBertModel),
         (MobileBertConfig, MobileBertModel),
         (TransfoXLConfig, TransfoXLModel),
         (XLNetConfig, XLNetModel),
@@ -398,6 +411,7 @@ MODEL_FOR_PRETRAINING_MAPPING = OrderedDict(
         (BigBirdConfig, BigBirdForPreTraining),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
+        (MegatronBertConfig, MegatronBertForPreTraining),
         (MobileBertConfig, MobileBertForPreTraining),
         (TransfoXLConfig, TransfoXLLMHeadModel),
         (XLNetConfig, XLNetLMHeadModel),
@@ -441,6 +455,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (BertConfig, BertForMaskedLM),
         (OpenAIGPTConfig, OpenAIGPTLMHeadModel),
         (GPT2Config, GPT2LMHeadModel),
+        (MegatronBertConfig, MegatronBertForMaskedLM),
         (MobileBertConfig, MobileBertForMaskedLM),
         (TransfoXLConfig, TransfoXLLMHeadModel),
         (XLNetConfig, XLNetLMHeadModel),
@@ -456,6 +471,7 @@ MODEL_WITH_LM_HEAD_MAPPING = OrderedDict(
         (DebertaConfig, DebertaForMaskedLM),
         (DebertaV2Config, DebertaV2ForMaskedLM),
         (IBertConfig, IBertForMaskedLM),
+        (MegatronBertConfig, MegatronBertForCausalLM),
     ]
 )
 
@@ -487,6 +503,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING = OrderedDict(
         (MarianConfig, MarianForCausalLM),
         (BlenderbotConfig, BlenderbotForCausalLM),
         (BlenderbotSmallConfig, BlenderbotSmallForCausalLM),
+        (MegatronBertConfig, MegatronBertForCausalLM),
     ]
 )
 
@@ -514,6 +531,7 @@ MODEL_FOR_MASKED_LM_MAPPING = OrderedDict(
         (RobertaConfig, RobertaForMaskedLM),
         (SqueezeBertConfig, SqueezeBertForMaskedLM),
         (BertConfig, BertForMaskedLM),
+        (MegatronBertConfig, MegatronBertForMaskedLM),
         (MobileBertConfig, MobileBertForMaskedLM),
         (FlaubertConfig, FlaubertWithLMHeadModel),
         (XLMConfig, XLMWithLMHeadModel),
@@ -566,6 +584,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING = OrderedDict(
         (LayoutLMConfig, LayoutLMForSequenceClassification),
         (BertConfig, BertForSequenceClassification),
         (XLNetConfig, XLNetForSequenceClassification),
+        (MegatronBertConfig, MegatronBertForSequenceClassification),
         (MobileBertConfig, MobileBertForSequenceClassification),
         (FlaubertConfig, FlaubertForSequenceClassification),
         (XLMConfig, XLMForSequenceClassification),
@@ -602,6 +621,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING = OrderedDict(
         (BertConfig, BertForQuestionAnswering),
         (XLNetConfig, XLNetForQuestionAnsweringSimple),
         (FlaubertConfig, FlaubertForQuestionAnsweringSimple),
+        (MegatronBertConfig, MegatronBertForQuestionAnswering),
         (MobileBertConfig, MobileBertForQuestionAnswering),
         (XLMConfig, XLMForQuestionAnsweringSimple),
         (ElectraConfig, ElectraForQuestionAnswering),
@@ -637,6 +657,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING = OrderedDict(
         (RobertaConfig, RobertaForTokenClassification),
         (SqueezeBertConfig, SqueezeBertForTokenClassification),
         (BertConfig, BertForTokenClassification),
+        (MegatronBertConfig, MegatronBertForTokenClassification),
         (MobileBertConfig, MobileBertForTokenClassification),
         (XLNetConfig, XLNetForTokenClassification),
         (AlbertConfig, AlbertForTokenClassification),
@@ -663,6 +684,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
         (SqueezeBertConfig, SqueezeBertForMultipleChoice),
         (BertConfig, BertForMultipleChoice),
         (DistilBertConfig, DistilBertForMultipleChoice),
+        (MegatronBertConfig, MegatronBertForMultipleChoice),
         (MobileBertConfig, MobileBertForMultipleChoice),
         (XLNetConfig, XLNetForMultipleChoice),
         (AlbertConfig, AlbertForMultipleChoice),
@@ -677,6 +699,7 @@ MODEL_FOR_MULTIPLE_CHOICE_MAPPING = OrderedDict(
 MODEL_FOR_NEXT_SENTENCE_PREDICTION_MAPPING = OrderedDict(
     [
         (BertConfig, BertForNextSentencePrediction),
+        (MegatronBertConfig, MegatronBertForNextSentencePrediction),
         (MobileBertConfig, MobileBertForNextSentencePrediction),
     ]
 )

src/transformers/models/megatron_bert/__init__.py

@@ -0,0 +1,74 @@
+# 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 2021  NVIDIA Corporation and 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_megatron_bert": ["MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MegatronBertConfig"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_megatron_bert"] = [
+        "MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "MegatronBertForCausalLM",
+        "MegatronBertForMaskedLM",
+        "MegatronBertForMultipleChoice",
+        "MegatronBertForNextSentencePrediction",
+        "MegatronBertForPreTraining",
+        "MegatronBertForQuestionAnswering",
+        "MegatronBertForSequenceClassification",
+        "MegatronBertForTokenClassification",
+        "MegatronBertModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_megatron_bert import MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MegatronBertConfig
+
+    if is_torch_available():
+        from .modeling_megatron_bert import (
+            MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST,
+            MegatronBertForCausalLM,
+            MegatronBertForMaskedLM,
+            MegatronBertForMultipleChoice,
+            MegatronBertForNextSentencePrediction,
+            MegatronBertForPreTraining,
+            MegatronBertForQuestionAnswering,
+            MegatronBertForSequenceClassification,
+            MegatronBertForTokenClassification,
+            MegatronBertModel,
+        )
+
+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/megatron_bert/configuration_megatron_bert.py

@@ -0,0 +1,132 @@
+# coding=utf-8
+# Copyright 2021- NVIDIA Corporation and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" MEGATRON_BERT model configuration """
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+MEGATRON_BERT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    # See all MEGATRON_BERT models at https://huggingface.co/models?filter=bert
+}
+
+
+class MegatronBertConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a :class:`~transformers.MegatronBertModel`. It is
+    used to instantiate a MEGATRON_BERT 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 MEGATRON_BERT
+    `megatron-bert-uncased-345m <https://huggingface.co/nvidia/megatron-bert-uncased-345m>`__ 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.
+
+
+    Args:
+        vocab_size (:obj:`int`, `optional`, defaults to 29056):
+            Vocabulary size of the MEGATRON_BERT model. Defines the number of different tokens that can be represented
+            by the :obj:`inputs_ids` passed when calling :class:`~transformers.MegatronBertModel`.
+        hidden_size (:obj:`int`, `optional`, defaults to 1024):
+            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 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (:obj:`int`, `optional`, defaults to 4096):
+            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"` 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 2):
+            The vocabulary size of the :obj:`token_type_ids` passed when calling
+            :class:`~transformers.MegatronBertModel`.
+        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-12):
+            The epsilon used by the layer normalization layers.
+        gradient_checkpointing (:obj:`bool`, `optional`, defaults to :obj:`False`):
+            If True, use gradient checkpointing to save memory at the expense of slower backward pass.
+        position_embedding_type (:obj:`str`, `optional`, defaults to :obj:`"absolute"`):
+            Type of position embedding. Choose one of :obj:`"absolute"`, :obj:`"relative_key"`,
+            :obj:`"relative_key_query"`. For positional embeddings use :obj:`"absolute"`. For more information on
+            :obj:`"relative_key"`, please refer to `Self-Attention with Relative Position Representations (Shaw et al.)
+            <https://arxiv.org/abs/1803.02155>`__. For more information on :obj:`"relative_key_query"`, please refer to
+            `Method 4` in `Improve Transformer Models with Better Relative Position Embeddings (Huang et al.)
+            <https://arxiv.org/abs/2009.13658>`__.
+        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if ``config.is_decoder=True``.
+
+    Examples::
+
+        >>> from transformers import MegatronBertModel, MegatronBertConfig
+
+        >>> # Initializing a MEGATRON_BERT bert-base-uncased style configuration
+        >>> configuration = MegatronBertConfig()
+
+        >>> # Initializing a model from the bert-base-uncased style configuration
+        >>> model = MegatronBertModel(configuration)
+
+        >>> # Accessing the model configuration
+        >>> configuration = model.config
+    """
+    model_type = "megatron-bert"
+
+    def __init__(
+        self,
+        vocab_size=29056,
+        hidden_size=1024,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        intermediate_size=4096,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        gradient_checkpointing=False,
+        position_embedding_type="absolute",
+        use_cache=True,
+        **kwargs
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.gradient_checkpointing = gradient_checkpointing
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache

src/transformers/models/megatron_bert/convert_megatron_bert_checkpoint.py

@@ -0,0 +1,265 @@
+####################################################################################################
+
+# Copyright (c) 2021-, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+####################################################################################################
+
+import argparse
+import json
+import os
+import re
+import zipfile
+
+import torch
+
+
+####################################################################################################
+
+
+def recursive_print(name, val, spaces=0):
+    # Format the message.
+    if name is None:
+        msg = None
+    else:
+        fmt = "." * max(0, spaces - 2) + "# {:" + str(50 - spaces) + "s}"
+        msg = fmt.format(name)
+
+    # Print and recurse (if needed).
+    if isinstance(val, dict):
+        if msg is not None:
+            print(msg)
+        for k in val.keys():
+            recursive_print(k, val[k], spaces + 2)
+    elif isinstance(val, torch.Tensor):
+        print(msg, ":", val.size())
+    else:
+        print(msg, ":", val)
+
+
+####################################################################################################
+
+
+def convert_megatron_checkpoint(args, input_state_dict):
+    # The converted output model.
+    output_state_dict = {}
+
+    # The model.
+    model = input_state_dict["model"]
+    # The language model.
+    lm = model["language_model"]
+    # The embeddings.
+    embeddings = lm["embedding"]
+
+    # The word embeddings.
+    word_embeddings = embeddings["word_embeddings"]["weight"]
+    # Store the word embeddings.
+    output_state_dict["bert.embeddings.word_embeddings.weight"] = word_embeddings
+
+    # The position embeddings.
+    pos_embeddings = embeddings["position_embeddings"]["weight"]
+    # Trained for 512 x 1024.
+    assert pos_embeddings.size(0) == 512 and pos_embeddings.size(1) == 1024
+    # Store the position embeddings.
+    output_state_dict["bert.embeddings.position_embeddings.weight"] = pos_embeddings
+
+    # The token-type embeddings.
+    tokentype_embeddings = embeddings["tokentype_embeddings"]["weight"]
+    # Store the position embeddings.
+    output_state_dict["bert.embeddings.token_type_embeddings.weight"] = tokentype_embeddings
+
+    # The transformer.
+    transformer = lm["transformer"]
+
+    # The regex to extract layer names.
+    layer_re = re.compile("layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+
+    # The simple map of names for "automated" rules.
+    megatron_to_transformers = {
+        "attention.dense": ".attention.output.dense.",
+        "mlp.dense_h_to_4h": ".intermediate.dense.",
+        "mlp.dense_4h_to_h": ".output.dense.",
+    }
+
+    # Keep track of the attention/query/value tensor.
+    attention_qkv_weight = None
+
+    # Extract the layers.
+    for key, val in transformer.items():
+        # Match the name.
+        m = layer_re.match(key)
+
+        # Stop if that's not a layer
+        if m is None:
+            break
+
+        # The index of the layer.
+        layer_idx = int(m.group(1))
+        # The name of the operation.
+        op_name = m.group(2)
+        # Is it a weight or a bias?
+        weight_or_bias = m.group(3)
+
+        # The name of the layer.
+        layer_name = f"bert.encoder.layer.{layer_idx}"
+
+        # For layernorm(s), simply store the layer norm.
+        if op_name.endswith("layernorm"):
+
+            ln_name = "attention.ln" if op_name.startswith("input") else "ln"
+            output_state_dict[layer_name + "." + ln_name + "." + weight_or_bias] = val
+
+        # Transpose the QKV matrix.
+        elif op_name == "attention.query_key_value" and weight_or_bias == "weight":
+
+            # Make sure the QKV pointer is nil.
+            assert attention_qkv_weight is None, ""
+
+            # Store the tensor as we need the bias as well to interleave QKV and biases.
+            attention_qkv_weight = val
+
+        # Transpose the bias.
+        elif op_name == "attention.query_key_value" and weight_or_bias == "bias":
+
+            # Make sure we read the weight tensor.
+            assert attention_qkv_weight is not None, ""
+
+            # Split the QKV matrix into Q, K and V. Megatron stores Q,K,V interleaved.
+            q = attention_qkv_weight[0 * 1024 : 1 * 1024, :]
+            k = attention_qkv_weight[1 * 1024 : 2 * 1024, :]
+            v = attention_qkv_weight[2 * 1024 : 3 * 1024, :]
+
+            # Split the bias.
+            q_bias = val[0 * 1024 : 1 * 1024]
+            k_bias = val[1 * 1024 : 2 * 1024]
+            v_bias = val[2 * 1024 : 3 * 1024]
+
+            # Store.
+            output_state_dict[f"{layer_name}.attention.self.query.weight"] = q
+            output_state_dict[f"{layer_name}.attention.self.query.bias"] = q_bias
+            output_state_dict[f"{layer_name}.attention.self.key.weight"] = k
+            output_state_dict[f"{layer_name}.attention.self.key.bias"] = k_bias
+            output_state_dict[f"{layer_name}.attention.self.value.weight"] = v
+            output_state_dict[f"{layer_name}.attention.self.value.bias"] = v_bias
+
+            # Clear the stored tensor.
+            attention_qkv_weight = None
+
+        # Copy weights and biases as is.
+        elif weight_or_bias in ["weight", "bias"]:
+
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + weight_or_bias] = val
+
+    # The final layernorm.
+    output_state_dict["bert.encoder.ln.weight"] = transformer["final_layernorm.weight"]
+    output_state_dict["bert.encoder.ln.bias"] = transformer["final_layernorm.bias"]
+
+    # The config.
+    output_config = {
+        "vocab_size": word_embeddings.size(0),
+        "hidden_size": 1024,
+        "num_hidden_layers": 24,
+        "num_attention_heads": 16,
+        "hidden_act": "gelu_new",
+        "intermediate_size": 4096,
+        "hidden_dropout_prob": 0.1,
+        "attention_probs_dropout_prob": 0.1,
+        "max_position_embeddings": 512,
+        "type_vocab_size": 2,
+        "initializer_range": 0.2,
+        "layer_norm_eps": 1e-12,
+        "gradient_checkpointing": False,
+        "position_embedding_type": "absolute",
+        "use_cache": False,
+    }
+
+    # The pooler.
+    pooler = lm["pooler"]
+
+    # Store the matrix and the bias.
+    output_state_dict["bert.pooler.dense.weight"] = pooler["dense.weight"]
+    output_state_dict["bert.pooler.dense.bias"] = pooler["dense.bias"]
+
+    # The LM head from Megatron (for RACE).
+    lm_head = model["lm_head"]
+
+    # The transform matrix.
+    output_state_dict["cls.predictions.transform.dense.weight"] = lm_head["dense.weight"]
+    output_state_dict["cls.predictions.transform.dense.bias"] = lm_head["dense.bias"]
+
+    # The transform LN.
+    output_state_dict["cls.predictions.transform.LayerNorm.weight"] = lm_head["layernorm.weight"]
+    output_state_dict["cls.predictions.transform.LayerNorm.bias"] = lm_head["layernorm.bias"]
+
+    # For the decoder, we replicate the weights.
+    output_state_dict["cls.predictions.decoder.weight"] = word_embeddings
+    output_state_dict["cls.predictions.bias"] = lm_head["bias"]
+
+    # The classifier from Megatron (for MLNI).
+    binary_head = model["binary_head"]
+
+    # Store the classifier.
+    output_state_dict["cls.seq_relationship.weight"] = binary_head["weight"]
+    output_state_dict["cls.seq_relationship.bias"] = binary_head["bias"]
+
+    # It should be done!
+    return output_state_dict, output_config
+
+
+####################################################################################################
+
+
+def main():
+    # Create the argument parser.
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--print-checkpoint-structure", action="store_true")
+    parser.add_argument("path_to_checkpoint", type=str, help="Path to the ZIP file containing the checkpoint")
+    args = parser.parse_args()
+
+    # Extract the basename.
+    basename = os.path.dirname(args.path_to_checkpoint)
+
+    # Load the model.
+    print(f'Extracting PyTorch state dictionary from "{args.path_to_checkpoint}"')
+    with zipfile.ZipFile(args.path_to_checkpoint, "r") as checkpoint:
+        with checkpoint.open("release/mp_rank_00/model_optim_rng.pt") as pytorch_dict:
+            input_state_dict = torch.load(pytorch_dict, map_location="cpu")
+
+    # Convert.
+    print("Converting")
+    output_state_dict, output_config = convert_megatron_checkpoint(args, input_state_dict)
+
+    # Print the structure of converted state dict.
+    if args.print_checkpoint_structure:
+        recursive_print(None, output_state_dict)
+
+    # Store the config to file.
+    output_config_file = os.path.join(basename, "config.json")
+    print(f'Saving config to "{output_config_file}"')
+    with open(output_config_file, "w") as f:
+        json.dump(output_config, f)
+
+    # Store the state_dict to file.
+    output_checkpoint_file = os.path.join(basename, "pytorch_model.bin")
+    print(f'Saving checkpoint to "{output_checkpoint_file}"')
+    torch.save(output_state_dict, output_checkpoint_file)
+
+
+####################################################################################################
+
+if __name__ == "__main__":
+    main()
+
+####################################################################################################

src/transformers/models/megatron_gpt2/convert_megatron_gpt2_checkpoint.py

@@ -0,0 +1,238 @@
+####################################################################################################
+
+# Copyright (c) 2021-, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+####################################################################################################
+
+import argparse
+import json
+import os
+import re
+import zipfile
+
+import torch
+
+
+####################################################################################################
+
+
+def recursive_print(name, val, spaces=0):
+    # Format the message.
+    if name is None:
+        msg = None
+    else:
+        fmt = "." * max(0, spaces - 2) + "# {:" + str(50 - spaces) + "s}"
+        msg = fmt.format(name)
+
+    # Print and recurse (if needed).
+    if isinstance(val, dict):
+        if msg is not None:
+            print(msg)
+        for k in val.keys():
+            recursive_print(k, val[k], spaces + 2)
+    elif isinstance(val, torch.Tensor):
+        print(msg, ":", val.size())
+    else:
+        print(msg, ":", val)
+
+
+####################################################################################################
+
+
+def convert_megatron_checkpoint(args, input_state_dict):
+    # The converted output model.
+    output_state_dict = {}
+
+    # The number of heads.
+    heads = 16
+    # The hidden_size per head.
+    hidden_size_per_head = 64
+
+    # The model.
+    model = input_state_dict["model"]
+    # The language model.
+    lm = model["language_model"]
+    # The embeddings.
+    embeddings = lm["embedding"]
+
+    # The word embeddings.
+    word_embeddings = embeddings["word_embeddings"]["weight"]
+    # Truncate the embedding table to 50257 rows.
+    word_embeddings = word_embeddings[:50257, :]
+    # Truncate the embedding table to 50257 rows.
+    output_state_dict["transformer.wte.weight"] = word_embeddings
+
+    # The position embeddings.
+    pos_embeddings = embeddings["position_embeddings"]["weight"]
+    # Read the hidden dimension.
+    hidden_size = pos_embeddings.size(0)
+    # DEBUG.
+    assert hidden_size == heads * hidden_size_per_head
+    # Store the position embeddings.
+    output_state_dict["transformer.wpe.weight"] = pos_embeddings
+
+    # The transformer.
+    transformer = lm["transformer"]
+
+    # The regex to extract layer names.
+    layer_re = re.compile("layers\.(\d+)\.([a-z0-9_.]+)\.([a-z]+)")
+
+    # The simple map of names for "automated" rules.
+    megatron_to_transformers = {
+        "attention.dense": ".attn.c_proj.",
+        "mlp.dense_h_to_4h": ".mlp.c_fc.",
+        "mlp.dense_4h_to_h": ".mlp.c_proj.",
+    }
+
+    # Extract the layers.
+    for key, val in transformer.items():
+        # Match the name.
+        m = layer_re.match(key)
+
+        # Stop if that's not a layer
+        if m is None:
+            break
+
+        # The index of the layer.
+        layer_idx = int(m.group(1))
+        # The name of the operation.
+        op_name = m.group(2)
+        # Is it a weight or a bias?
+        weight_or_bias = m.group(3)
+
+        # The name of the layer.
+        layer_name = f"transformer.h.{layer_idx}"
+
+        # For layernorm(s), simply store the layer norm.
+        if op_name.endswith("layernorm"):
+
+            ln_name = "ln_1" if op_name.startswith("input") else "ln_2"
+            output_state_dict[layer_name + "." + ln_name + "." + weight_or_bias] = val
+
+        # Transpose the QKV matrix.
+        elif op_name == "attention.query_key_value" and weight_or_bias == "weight":
+
+            # Insert a tensor of 1x1xDxD bias.
+            zeros = torch.ones(1, 1, hidden_size, hidden_size)
+            output_state_dict[layer_name + ".attn.bias"] = zeros
+
+            # Insert a "dummy" tensor for masked_bias.
+            masked_bias = torch.tensor(-1e4)
+            output_state_dict[layer_name + ".attn.masked_bias"] = masked_bias
+
+            # Megatron stores (3*D) x D but transformers-GPT2 expects D x 3*D.
+            out_val = val.transpose(0, 1)
+            # Store.
+            output_state_dict[layer_name + ".attn.c_attn.weight"] = out_val
+
+        # Transpose the bias.
+        elif op_name == "attention.query_key_value" and weight_or_bias == "bias":
+
+            # Store. No change of shape.
+            output_state_dict[layer_name + ".attn.c_attn.bias"] = val
+
+        # Transpose the weights.
+        elif weight_or_bias == "weight":
+
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "weight"] = val.transpose(0, 1)
+
+        # Copy the bias.
+        elif weight_or_bias == "bias":
+
+            out_name = megatron_to_transformers[op_name]
+            output_state_dict[layer_name + out_name + "bias"] = val
+
+    # The final layernorm.
+    output_state_dict["transformer.ln_f.weight"] = transformer["final_layernorm.weight"]
+    output_state_dict["transformer.ln_f.bias"] = transformer["final_layernorm.bias"]
+
+    # For LM head, transformers' wants the matrix to weight embeddings.
+    output_state_dict["lm_head.weight"] = word_embeddings
+
+    # The config.
+    output_config = {
+        "activation_function": "gelu_new",
+        "architectures": ["GPT2LMHeadModel"],
+        "attn_pdrop": 0.1,
+        "bos_token_id": 50256,
+        "embd_pdrop": 0.1,
+        "eos_token_id": 50256,
+        "initializer_range": 0.02,
+        "layer_norm_epsilon": 1e-05,
+        "model_type": "gpt2",
+        "n_ctx": 1024,
+        "n_embd": 1024,
+        "n_head": 16,
+        "n_layer": 24,
+        "n_positions": 1024,
+        "resid_pdrop": 0.1,
+        "summary_activation": None,
+        "summary_first_dropout": 0.1,
+        "summary_proj_to_labels": True,
+        "summary_type": "cls_index",
+        "summary_use_proj": True,
+        "vocab_size": 50257,
+    }
+
+    # It should be done!
+    return output_state_dict, output_config
+
+
+####################################################################################################
+
+
+def main():
+    # Create the argument parser.
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--print-checkpoint-structure", action="store_true")
+    parser.add_argument("path_to_checkpoint", type=str, help="Path to the ZIP file containing the checkpoint")
+    args = parser.parse_args()
+
+    # Extract the basename.
+    basename = os.path.dirname(args.path_to_checkpoint)
+
+    # Load the model.
+    print('Extracting PyTorch state dictionary from "{}"'.format(args.path_to_checkpoint))
+    with zipfile.ZipFile(args.path_to_checkpoint, "r") as checkpoint:
+        with checkpoint.open("release/mp_rank_00/model_optim_rng.pt") as pytorch_dict:
+            input_state_dict = torch.load(pytorch_dict, map_location="cpu")
+
+    # Convert.
+    print("Converting")
+    output_state_dict, output_config = convert_megatron_checkpoint(args, input_state_dict)
+
+    # Print the structure of converted state dict.
+    if args.print_checkpoint_structure:
+        recursive_print(None, output_state_dict)
+
+    # Store the config to file.
+    output_config_file = os.path.join(basename, "config.json")
+    print(f'Saving config to "{output_config_file}"')
+    with open(output_config_file, "w") as f:
+        json.dump(output_config, f)
+
+    # Store the state_dict to file.
+    output_checkpoint_file = os.path.join(basename, "pytorch_model.bin")
+    print(f'Saving checkpoint to "{output_checkpoint_file}"')
+    torch.save(output_state_dict, output_checkpoint_file)
+
+
+####################################################################################################
+
+if __name__ == "__main__":
+    main()
+
+####################################################################################################

src/transformers/utils/dummy_pt_objects.py

@@ -1840,6 +1840,78 @@ class MBartModel:
         requires_backends(self, ["torch"])
 
 
+MEGATRON_BERT_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class MegatronBertForCausalLM:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForMaskedLM:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForMultipleChoice:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForNextSentencePrediction:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForPreTraining:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForQuestionAnswering:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForSequenceClassification:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertForTokenClassification:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class MegatronBertModel:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 class MMBTForClassification:
     def __init__(self, *args, **kwargs):
         requires_backends(self, ["torch"])

src/transformers/utils/modeling_auto_mapping.py

@@ -21,6 +21,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
         ("BertConfig", "BertForQuestionAnswering"),
         ("XLNetConfig", "XLNetForQuestionAnsweringSimple"),
         ("FlaubertConfig", "FlaubertForQuestionAnsweringSimple"),
+        ("MegatronBertConfig", "MegatronBertForQuestionAnswering"),
         ("MobileBertConfig", "MobileBertForQuestionAnswering"),
         ("XLMConfig", "XLMForQuestionAnsweringSimple"),
         ("ElectraConfig", "ElectraForQuestionAnswering"),

tests/test_modeling_megatron_bert.py

@@ -0,0 +1,377 @@
+# coding=utf-8
+# Copyright 2021 The HuggingFace Inc. team. All rights reserved.
+# Copyright 2021 NVIDIA Corporation. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the PyTorch MegatronBERT model. """
+
+
+import math
+import os
+import unittest
+
+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, random_attention_mask
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        MODEL_FOR_PRETRAINING_MAPPING,
+        MegatronBertConfig,
+        MegatronBertForCausalLM,
+        MegatronBertForMaskedLM,
+        MegatronBertForMultipleChoice,
+        MegatronBertForNextSentencePrediction,
+        MegatronBertForPreTraining,
+        MegatronBertForQuestionAnswering,
+        MegatronBertForSequenceClassification,
+        MegatronBertForTokenClassification,
+        MegatronBertModel,
+    )
+
+
+class MegatronBertModelTester:
+    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=64,
+        embedding_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.embedding_size = embedding_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = MegatronBertConfig(
+            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,
+            embedding_size=self.embedding_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,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def create_and_check_megatron_bert_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
+        result = model(input_ids, token_type_ids=token_type_ids)
+        result = model(input_ids)
+
+        self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, self.hidden_size))
+
+    def create_and_check_megatron_bert_for_masked_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertForMaskedLM(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+
+    def create_and_check_for_causal_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertForCausalLM(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_megatron_bert_for_next_sequence_prediction(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertForNextSentencePrediction(config=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.assertEqual(result.logits.shape, (self.batch_size, 2))
+
+    def create_and_check_megatron_bert_for_pretraining(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertForPreTraining(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,
+            next_sentence_label=sequence_labels,
+        )
+        self.parent.assertEqual(result.prediction_logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
+        self.parent.assertEqual(result.seq_relationship_logits.shape, (self.batch_size, 2))
+
+    def create_and_check_megatron_bert_for_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = MegatronBertForQuestionAnswering(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 create_and_check_megatron_bert_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 = MegatronBertForSequenceClassification(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.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
+
+    def create_and_check_megatron_bert_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 = MegatronBertForTokenClassification(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_megatron_bert_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = MegatronBertForMultipleChoice(config=config)
+        model.to(torch_device)
+        model.eval()
+        multiple_choice_inputs_ids = input_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_token_type_ids = token_type_ids.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        multiple_choice_input_mask = input_mask.unsqueeze(1).expand(-1, self.num_choices, -1).contiguous()
+        result = model(
+            multiple_choice_inputs_ids,
+            attention_mask=multiple_choice_input_mask,
+            token_type_ids=multiple_choice_token_type_ids,
+            labels=choice_labels,
+        )
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_choices))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+            sequence_labels,
+            token_labels,
+            choice_labels,
+        ) = config_and_inputs
+        inputs_dict = {"input_ids": input_ids, "token_type_ids": token_type_ids, "attention_mask": input_mask}
+        return config, inputs_dict
+
+
+@require_torch
+class MegatronBertModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            MegatronBertModel,
+            MegatronBertForMaskedLM,
+            MegatronBertForCausalLM,
+            MegatronBertForMultipleChoice,
+            MegatronBertForNextSentencePrediction,
+            MegatronBertForPreTraining,
+            MegatronBertForQuestionAnswering,
+            MegatronBertForSequenceClassification,
+            MegatronBertForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+
+    # test_resize_embeddings = False
+    test_head_masking = False
+
+    # special case for ForPreTraining model
+    def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
+        inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
+
+        if return_labels:
+            if model_class in MODEL_FOR_PRETRAINING_MAPPING.values():
+                inputs_dict["labels"] = torch.zeros(
+                    (self.model_tester.batch_size, self.model_tester.seq_length), dtype=torch.long, device=torch_device
+                )
+                inputs_dict["next_sentence_label"] = torch.zeros(
+                    self.model_tester.batch_size, dtype=torch.long, device=torch_device
+                )
+        return inputs_dict
+
+    def setUp(self):
+        self.model_tester = MegatronBertModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=MegatronBertConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_megatron_bert_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_megatron_bert_model(*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_megatron_bert_for_masked_lm(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_megatron_bert_for_multiple_choice(*config_and_inputs)
+
+    def test_for_next_sequence_prediction(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_megatron_bert_for_next_sequence_prediction(*config_and_inputs)
+
+    def test_for_pretraining(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_megatron_bert_for_pretraining(*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_megatron_bert_for_question_answering(*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_megatron_bert_for_sequence_classification(*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_megatron_bert_for_token_classification(*config_and_inputs)
+
+
+def _long_tensor(tok_lst):
+    return torch.tensor(
+        tok_lst,
+        dtype=torch.long,
+        device=torch_device,
+    )
+
+
+TOLERANCE = 1e-4
+
+
+@require_torch
+@require_sentencepiece
+@require_tokenizers
+class MegatronBertModelIntegrationTests(unittest.TestCase):
+    @slow
+    def test_inference_no_head(self):
+        directory = "nvidia/megatron-bert-uncased-345m"
+        if "MYDIR" in os.environ:
+            directory = os.path.join(os.environ["MYDIR"], directory)
+        model = MegatronBertModel.from_pretrained(directory)
+        model.to(torch_device)
+        model.half()
+        input_ids = _long_tensor([[101, 7110, 1005, 1056, 2023, 11333, 17413, 1029, 102]])
+        with torch.no_grad():
+            output = model(input_ids)[0]
+        expected_shape = torch.Size((1, 9, 1024))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected = [-0.6040, -0.2517, -0.1025, 0.3420, -0.6758, -0.0017, -0.1089, -0.1990, 0.5728]
+        for ii in range(3):
+            for jj in range(3):
+                a = output[0, ii, jj]
+                b = expected[3 * ii + jj]
+                msg = "ii={} jj={} a={} b={}".format(ii, jj, a, b)
+                self.assertTrue(math.isclose(a, b, rel_tol=TOLERANCE, abs_tol=TOLERANCE), msg=msg)

utils/check_repo.py

@@ -45,6 +45,10 @@ IGNORE_NON_TESTED = [
     "BlenderbotDecoderWrapper",  # Building part of bigger (tested) model.
     "MBartEncoder",  # Building part of bigger (tested) model.
     "MBartDecoderWrapper",  # Building part of bigger (tested) model.
+    "MegatronBertLMHeadModel",  # Building part of bigger (tested) model.
+    "MegatronBertEncoder",  # Building part of bigger (tested) model.
+    "MegatronBertDecoder",  # Building part of bigger (tested) model.
+    "MegatronBertDecoderWrapper",  # Building part of bigger (tested) model.
     "PegasusEncoder",  # Building part of bigger (tested) model.
     "PegasusDecoderWrapper",  # Building part of bigger (tested) model.
     "DPREncoder",  # Building part of bigger (tested) model.