Add Nystromformer (#14659)

* Initial commit

* Config and modelling changes

Added Nystromformer-specific attributes to config and removed all decoder functionality from modelling.

* Modelling and test changes

Added Nystrom approximation and removed decoder tests.

* Code quality fixes

* Modeling changes and conversion script

Initial commits to conversion script, modeling changes.

* Minor modeling changes and conversion script

* Modeling changes

* Correct modeling, add tests and documentation

* Code refactor

* Remove tokenizers

* Code refactor

* Update __init__.py

* Fix bugs

* Update src/transformers/__init__.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/__init__.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/__init__.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/model_doc/nystromformer.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/configuration_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/configuration_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/configuration_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/configuration_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/convert_nystromformer_original_pytorch_checkpoint_to_pytorch.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/nystromformer/configuration_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update modeling and test_modeling

* Code refactor

* .rst to .mdx

* doc changes

* Doc changes

* Update modeling_nystromformer.py

* Doc changes

* Fix copies

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update configuration_nystromformer.py

* Fix copies

* Update tests/test_modeling_nystromformer.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update test_modeling_nystromformer.py

* Apply suggestions from code review
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Fix code style

* Update modeling_nystromformer.py

* Update modeling_nystromformer.py

* Fix code style

* Reformat modeling file

* Update modeling_nystromformer.py

* Modify NystromformerForMultipleChoice

* Fix code quality

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Code style changes and torch.no_grad()

* make style

* Apply suggestions from code review
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

README.md

@@ -285,6 +285,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
 1. **[Megatron-GPT2](https://huggingface.co/docs/transformers/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.
 1. **[MPNet](https://huggingface.co/docs/transformers/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.
 1. **[MT5](https://huggingface.co/docs/transformers/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.
+1. **[Nyströmformer](https://huggingface.co/docs/transformers/master/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[Pegasus](https://huggingface.co/docs/transformers/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.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.

README_ko.md

@@ -264,6 +264,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MPNet](https://huggingface.co/docs/transformers/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.
 1. **[MT5](https://huggingface.co/docs/transformers/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.
+1. **[Nyströmformer](https://huggingface.co/docs/transformers/master/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[Pegasus](https://huggingface.co/docs/transformers/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.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.

README_zh-hans.md

@@ -288,6 +288,7 @@ conda install -c huggingface transformers
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (来自 Studio Ousia) 伴随论文 [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) 由 Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka 发布。
 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (来自 Microsoft Research) 伴随论文 [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) 由 Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu 发布。
 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (来自 Google AI) 伴随论文 [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) 由 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel 发布。
+1. **[Nyströmformer](https://huggingface.co/docs/transformers/master/model_doc/nystromformer)** (来自 the University of Wisconsin - Madison) 伴随论文 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 由 Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 发布。
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (来自 Google) 伴随论文 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 由 Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 发布。
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (来自 Deepmind) 伴随论文 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 由 Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 发布。
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (来自 VinAI Research) 伴随论文 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 由 Dat Quoc Nguyen and Anh Tuan Nguyen 发布。

README_zh-hant.md

@@ -300,6 +300,7 @@ conda install -c huggingface transformers
 1. **[mLUKE](https://huggingface.co/docs/transformers/model_doc/mluke)** (from Studio Ousia) released with the paper [mLUKE: The Power of Entity Representations in Multilingual Pretrained Language Models](https://arxiv.org/abs/2110.08151) by Ryokan Ri, Ikuya Yamada, and Yoshimasa Tsuruoka.
 1. **[MPNet](https://huggingface.co/docs/transformers/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.
 1. **[MT5](https://huggingface.co/docs/transformers/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.
+1. **[Nyströmformer](https://huggingface.co/docs/transformers/master/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[Pegasus](https://huggingface.co/docs/transformers/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.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.

docs/source/_toctree.yml

@@ -214,6 +214,8 @@
       title: MPNet
     - local: model_doc/mt5
       title: MT5
+    - local: model_doc/nystromformer
+      title: Nyströmformer
     - local: model_doc/openai-gpt
       title: OpenAI GPT
     - local: model_doc/gpt2

docs/source/index.mdx

@@ -145,6 +145,7 @@ conversion utilities for the following models.
 1. **[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.
 1. **[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.
 1. **[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.
+1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
 1. **[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.
 1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
 1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
@@ -235,6 +236,7 @@ Flax), PyTorch, and/or TensorFlow.
 |         MobileBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |            MPNet            |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |             mT5             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|        Nystromformer        |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         OpenAI GPT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |        OpenAI GPT-2         |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |           Pegasus           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/model_doc/nystromformer.mdx

+<!--Copyright 2022 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.
+-->
+
+# Nyströmformer
+
+## Overview
+
+The Nyströmformer model was proposed in [*Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention*](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn
+Fung, Yin Li, and Vikas Singh.
+
+The abstract from the paper is the following:
+
+*Transformers have emerged as a powerful tool for a broad range of natural language processing tasks. A key component
+that drives the impressive performance of Transformers is the self-attention mechanism that encodes the influence or
+dependence of other tokens on each specific token. While beneficial, the quadratic complexity of self-attention on the
+input sequence length has limited its application to longer sequences -- a topic being actively studied in the
+community. To address this limitation, we propose Nyströmformer -- a model that exhibits favorable scalability as a
+function of sequence length. Our idea is based on adapting the Nyström method to approximate standard self-attention
+with O(n) complexity. The scalability of Nyströmformer enables application to longer sequences with thousands of
+tokens. We perform evaluations on multiple downstream tasks on the GLUE benchmark and IMDB reviews with standard
+sequence length, and find that our Nyströmformer performs comparably, or in a few cases, even slightly better, than
+standard self-attention. On longer sequence tasks in the Long Range Arena (LRA) benchmark, Nyströmformer performs
+favorably relative to other efficient self-attention methods. Our code is available at this https URL.*
+
+This model was contributed by [novice03](https://huggingface.co/novice03). The original code can be found [here](https://github.com/mlpen/Nystromformer).
+
+## NystromformerConfig
+
+[[autodoc]] NystromformerConfig
+
+## NystromformerModel
+
+[[autodoc]] NystromformerModel
+    - forward
+
+## NystromformerForMaskedLM
+
+[[autodoc]] NystromformerForMaskedLM
+    - forward
+
+## NystromformerForSequenceClassification
+
+[[autodoc]] NystromformerForSequenceClassification
+    - forward
+
+## NystromformerForMultipleChoice
+
+[[autodoc]] NystromformerForMultipleChoice
+    - forward
+
+## NystromformerForTokenClassification
+
+[[autodoc]] NystromformerForTokenClassification
+    - forward
+
+## NystromformerForQuestionAnswering
+
+[[autodoc]] NystromformerForQuestionAnswering
+    - forward

src/transformers/__init__.py

@@ -241,6 +241,10 @@ _import_structure = {
     "models.mobilebert": ["MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "MobileBertConfig", "MobileBertTokenizer"],
     "models.mpnet": ["MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP", "MPNetConfig", "MPNetTokenizer"],
     "models.mt5": ["MT5Config"],
+    "models.nystromformer": [
+        "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "NystromformerConfig",
+    ],
     "models.openai": ["OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "OpenAIGPTConfig", "OpenAIGPTTokenizer"],
     "models.pegasus": ["PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP", "PegasusConfig", "PegasusTokenizer"],
     "models.perceiver": ["PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP", "PerceiverConfig", "PerceiverTokenizer"],
@@ -1122,6 +1126,19 @@ if is_torch_available():
         ]
     )
     _import_structure["models.mt5"].extend(["MT5EncoderModel", "MT5ForConditionalGeneration", "MT5Model"])
+    _import_structure["models.nystromformer"].extend(
+        [
+            "NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "NystromformerForMaskedLM",
+            "NystromformerForMultipleChoice",
+            "NystromformerForQuestionAnswering",
+            "NystromformerForSequenceClassification",
+            "NystromformerForTokenClassification",
+            "NystromformerLayer",
+            "NystromformerModel",
+            "NystromformerPreTrainedModel",
+        ]
+    )
     _import_structure["models.openai"].extend(
         [
             "OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -2308,6 +2325,7 @@ if TYPE_CHECKING:
     from .models.mobilebert import MOBILEBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, MobileBertConfig, MobileBertTokenizer
     from .models.mpnet import MPNET_PRETRAINED_CONFIG_ARCHIVE_MAP, MPNetConfig, MPNetTokenizer
     from .models.mt5 import MT5Config
+    from .models.nystromformer import NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, NystromformerConfig
     from .models.openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig, OpenAIGPTTokenizer
     from .models.pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig, PegasusTokenizer
     from .models.perceiver import PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP, PerceiverConfig, PerceiverTokenizer
@@ -3041,6 +3059,17 @@ if TYPE_CHECKING:
             MPNetPreTrainedModel,
         )
         from .models.mt5 import MT5EncoderModel, MT5ForConditionalGeneration, MT5Model
+        from .models.nystromformer import (
+            NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            NystromformerForMaskedLM,
+            NystromformerForMultipleChoice,
+            NystromformerForQuestionAnswering,
+            NystromformerForSequenceClassification,
+            NystromformerForTokenClassification,
+            NystromformerLayer,
+            NystromformerModel,
+            NystromformerPreTrainedModel,
+        )
         from .models.openai import (
             OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST,
             OpenAIGPTDoubleHeadsModel,

src/transformers/models/__init__.py

@@ -76,6 +76,7 @@ from . import (
     mobilebert,
     mpnet,
     mt5,
+    nystromformer,
     openai,
     pegasus,
     perceiver,

src/transformers/models/auto/configuration_auto.py

@@ -30,6 +30,7 @@ logger = logging.get_logger(__name__)
 CONFIG_MAPPING_NAMES = OrderedDict(
     [
         # Add configs here
+        ("nystromformer", "NystromformerConfig"),
         ("imagegpt", "ImageGPTConfig"),
         ("qdqbert", "QDQBertConfig"),
         ("vision-encoder-decoder", "VisionEncoderDecoderConfig"),
@@ -116,6 +117,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
 CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
     [
         # Add archive maps here
+        ("nystromformer", "NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("imagegpt", "IMAGEGPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("qdqbert", "QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("fnet", "FNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -190,6 +192,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
 MODEL_NAMES_MAPPING = OrderedDict(
     [
         # Add full (and cased) model names here
+        ("nystromformer", "Nystromformer"),
         ("imagegpt", "ImageGPT"),
         ("qdqbert", "QDQBert"),
         ("vision-encoder-decoder", "Vision Encoder decoder"),

src/transformers/models/auto/modeling_auto.py

@@ -28,6 +28,7 @@ logger = logging.get_logger(__name__)
 MODEL_MAPPING_NAMES = OrderedDict(
     [
         # Base model mapping
+        ("nystromformer", "NystromformerModel"),
         ("imagegpt", "ImageGPTModel"),
         ("qdqbert", "QDQBertModel"),
         ("fnet", "FNetModel"),
@@ -150,6 +151,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
 MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
     [
         # Model with LM heads mapping
+        ("nystromformer", "NystromformerForMaskedLM"),
         ("qdqbert", "QDQBertForMaskedLM"),
         ("fnet", "FNetForMaskedLM"),
         ("gptj", "GPTJForCausalLM"),
@@ -277,6 +279,7 @@ MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_MASKED_LM_MAPPING_NAMES = OrderedDict(
     [
         # Model for Masked LM mapping
+        ("nystromformer", "NystromformerForMaskedLM"),
         ("perceiver", "PerceiverForMaskedLM"),
         ("qdqbert", "QDQBertForMaskedLM"),
         ("fnet", "FNetForMaskedLM"),
@@ -349,6 +352,7 @@ MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Sequence Classification mapping
+        ("nystromformer", "NystromformerForSequenceClassification"),
         ("perceiver", "PerceiverForSequenceClassification"),
         ("qdqbert", "QDQBertForSequenceClassification"),
         ("fnet", "FNetForSequenceClassification"),
@@ -396,6 +400,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
     [
         # Model for Question Answering mapping
+        ("nystromformer", "NystromformerForQuestionAnswering"),
         ("qdqbert", "QDQBertForQuestionAnswering"),
         ("fnet", "FNetForQuestionAnswering"),
         ("gptj", "GPTJForQuestionAnswering"),
@@ -444,6 +449,7 @@ MODEL_FOR_TABLE_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
     [
         # Model for Token Classification mapping
+        ("nystromformer", "NystromformerForTokenClassification"),
         ("qdqbert", "QDQBertForTokenClassification"),
         ("fnet", "FNetForTokenClassification"),
         ("layoutlmv2", "LayoutLMv2ForTokenClassification"),
@@ -479,6 +485,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
 MODEL_FOR_MULTIPLE_CHOICE_MAPPING_NAMES = OrderedDict(
     [
         # Model for Multiple Choice mapping
+        ("nystromformer", "NystromformerForMultipleChoice"),
         ("qdqbert", "QDQBertForMultipleChoice"),
         ("fnet", "FNetForMultipleChoice"),
         ("rembert", "RemBertForMultipleChoice"),

src/transformers/models/nystromformer/__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 2022 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+# rely on isort to merge the imports
+from ...file_utils import _LazyModule, is_tokenizers_available, is_torch_available
+
+
+_import_structure = {
+    "configuration_nystromformer": ["NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP", "NystromformerConfig"],
+}
+
+if is_torch_available():
+    _import_structure["modeling_nystromformer"] = [
+        "NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "NystromformerForMaskedLM",
+        "NystromformerForMultipleChoice",
+        "NystromformerForQuestionAnswering",
+        "NystromformerForSequenceClassification",
+        "NystromformerForTokenClassification",
+        "NystromformerLayer",
+        "NystromformerModel",
+        "NystromformerPreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_nystromformer import NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, NystromformerConfig
+
+    if is_torch_available():
+        from .modeling_nystromformer import (
+            NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            NystromformerForMaskedLM,
+            NystromformerForMultipleChoice,
+            NystromformerForQuestionAnswering,
+            NystromformerForSequenceClassification,
+            NystromformerForTokenClassification,
+            NystromformerLayer,
+            NystromformerModel,
+            NystromformerPreTrainedModel,
+        )
+
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/nystromformer/configuration_nystromformer.py

+# coding=utf-8
+# Copyright 2022 UW-Madison 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.
+""" Nystromformer model configuration"""
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "uw-madison/nystromformer-512": "https://huggingface.co/uw-madison/nystromformer-512/resolve/main/config.json",
+    # See all Nystromformer models at https://huggingface.co/models?filter=nystromformer
+}
+
+
+class NystromformerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`NystromformerModel`]. It is used to instantiate
+    an Nystromformer 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 Nystromformer
+    [uw-madison/nystromformer-512](https://huggingface.co/uw-madison/nystromformer-512) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30000):
+            Vocabulary size of the Nystromformer model. Defines the number of different tokens that can be represented
+            by the `inputs_ids` passed when calling [`NystromformerModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimension of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimension of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`NystromformerModel`].
+        segment_means_seq_len (`int`, *optional*, defaults to 64):
+            Sequence length used in segment-means.
+        num_landmarks (`int`, *optional*, defaults to 64):
+            The number of landmark (or Nystrom) points to use in Nystrom approximation of the softmax self-attention
+            matrix.
+        conv_kernel_size (`int`, *optional*, defaults to 65):
+            The kernel size of depthwise convolution used in Nystrom approximation.
+        inv_coeff_init_option (`bool`, *optional*, defaults to `False`):
+            Whether or not to use exact coefficient computation for the initial values for the iterative method of
+            calculating the Moore-Penrose inverse of a matrix.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import NystromformerModel, NystromformerConfig
+
+    >>> # Initializing a Nystromformer uw-madison/nystromformer-512 style configuration
+    >>> configuration = NystromformerConfig()
+
+    >>> # Initializing a model from the uw-madison/nystromformer-512 style configuration
+    >>> model = NystromformerModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "nystromformer"
+
+    def __init__(
+        self,
+        vocab_size=30000,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu_new",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=510,
+        type_vocab_size=2,
+        segment_means_seq_len=64,
+        num_landmarks=64,
+        conv_kernel_size=65,
+        inv_coeff_init_option=False,
+        initializer_range=0.02,
+        layer_norm_eps=1e-5,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.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.initializer_range = initializer_range
+        self.type_vocab_size = type_vocab_size
+        self.segment_means_seq_len = segment_means_seq_len
+        self.num_landmarks = num_landmarks
+        self.conv_kernel_size = conv_kernel_size
+        self.inv_coeff_init_option = inv_coeff_init_option
+        self.layer_norm_eps = layer_norm_eps
+        super().__init__(pad_token_id=pad_token_id, bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)

src/transformers/models/nystromformer/convert_nystromformer_original_pytorch_checkpoint_to_pytorch.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+"""Convert Nystromformer checkpoints from the original repository."""
+
+import argparse
+
+import torch
+
+from transformers import NystromformerConfig, NystromformerForMaskedLM
+
+
+def rename_key(orig_key):
+    if "model" in orig_key:
+        orig_key = orig_key.replace("model.", "")
+    if "norm1" in orig_key:
+        orig_key = orig_key.replace("norm1", "attention.output.LayerNorm")
+    if "norm2" in orig_key:
+        orig_key = orig_key.replace("norm2", "output.LayerNorm")
+    if "norm" in orig_key:
+        orig_key = orig_key.replace("norm", "LayerNorm")
+    if "transformer" in orig_key:
+        layer_num = orig_key.split(".")[0].split("_")[-1]
+        orig_key = orig_key.replace(f"transformer_{layer_num}", f"encoder.layer.{layer_num}")
+    if "mha.attn" in orig_key:
+        orig_key = orig_key.replace("mha.attn", "attention.self")
+    if "mha" in orig_key:
+        orig_key = orig_key.replace("mha", "attention")
+    if "W_q" in orig_key:
+        orig_key = orig_key.replace("W_q", "self.query")
+    if "W_k" in orig_key:
+        orig_key = orig_key.replace("W_k", "self.key")
+    if "W_v" in orig_key:
+        orig_key = orig_key.replace("W_v", "self.value")
+    if "ff1" in orig_key:
+        orig_key = orig_key.replace("ff1", "intermediate.dense")
+    if "ff2" in orig_key:
+        orig_key = orig_key.replace("ff2", "output.dense")
+    if "ff" in orig_key:
+        orig_key = orig_key.replace("ff", "output.dense")
+    if "mlm_class" in orig_key:
+        orig_key = orig_key.replace("mlm.mlm_class", "cls.predictions.decoder")
+    if "mlm" in orig_key:
+        orig_key = orig_key.replace("mlm", "cls.predictions.transform")
+    if "cls" not in orig_key:
+        orig_key = "nystromformer." + orig_key
+
+    return orig_key
+
+
+def convert_checkpoint_helper(config, orig_state_dict):
+    for key in orig_state_dict.copy().keys():
+        val = orig_state_dict.pop(key)
+
+        if ("pooler" in key) or ("sen_class" in key) or ("conv.bias" in key):
+            continue
+        else:
+            orig_state_dict[rename_key(key)] = val
+
+    orig_state_dict["cls.predictions.bias"] = orig_state_dict["cls.predictions.decoder.bias"]
+    orig_state_dict["nystromformer.embeddings.position_ids"] = (
+        torch.arange(config.max_position_embeddings).expand((1, -1)) + 2
+    )
+
+    return orig_state_dict
+
+
+def convert_nystromformer_checkpoint(checkpoint_path, nystromformer_config_file, pytorch_dump_path):
+
+    orig_state_dict = torch.load(checkpoint_path, map_location="cpu")["model_state_dict"]
+    config = NystromformerConfig.from_json_file(nystromformer_config_file)
+    model = NystromformerForMaskedLM(config)
+
+    new_state_dict = convert_checkpoint_helper(config, orig_state_dict)
+
+    model.load_state_dict(new_state_dict)
+    model.eval()
+    model.save_pretrained(pytorch_dump_path)
+
+    print(f"Checkpoint successfuly converted. Model saved at {pytorch_dump_path}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--pytorch_model_path", default=None, type=str, required=True, help="Path to Nystromformer pytorch checkpoint."
+    )
+    parser.add_argument(
+        "--config_file",
+        default=None,
+        type=str,
+        required=True,
+        help="The json file for Nystromformer model config.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_nystromformer_checkpoint(args.pytorch_model_path, args.config_file, args.pytorch_dump_path)

src/transformers/utils/dummy_pt_objects.py

@@ -3674,6 +3674,98 @@ class MT5Model:
         requires_backends(self, ["torch"])
 
 
+NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class NystromformerForMaskedLM:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerForMultipleChoice:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerForQuestionAnswering:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerForSequenceClassification:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerForTokenClassification:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerLayer:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerModel:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class NystromformerPreTrainedModel:
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+    @classmethod
+    def from_pretrained(cls, *args, **kwargs):
+        requires_backends(cls, ["torch"])
+
+    def forward(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 OPENAI_GPT_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/test_modeling_nystromformer.py

+# coding=utf-8
+# Copyright 2021 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.
+""" Testing suite for the PyTorch Nystromformer model. """
+
+
+import unittest
+
+from transformers import AutoTokenizer, NystromformerConfig, is_torch_available
+from transformers.testing_utils import 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 (
+        NystromformerForMaskedLM,
+        NystromformerForMultipleChoice,
+        NystromformerForQuestionAnswering,
+        NystromformerForSequenceClassification,
+        NystromformerForTokenClassification,
+        NystromformerModel,
+    )
+    from transformers.models.nystromformer.modeling_nystromformer import NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+class NystromformerModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=True,
+        use_labels=True,
+        vocab_size=99,
+        hidden_size=32,
+        num_hidden_layers=5,
+        num_attention_heads=4,
+        intermediate_size=37,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=16,
+        type_sequence_label_size=2,
+        initializer_range=0.02,
+        num_labels=3,
+        num_choices=4,
+        scope=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_input_mask = use_input_mask
+        self.use_token_type_ids = use_token_type_ids
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.intermediate_size = intermediate_size
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.type_sequence_label_size = type_sequence_label_size
+        self.initializer_range = initializer_range
+        self.num_labels = num_labels
+        self.num_choices = num_choices
+        self.scope = scope
+
+    def prepare_config_and_inputs(self):
+        input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        input_mask = None
+        if self.use_input_mask:
+            input_mask = random_attention_mask([self.batch_size, self.seq_length])
+
+        token_type_ids = None
+        if self.use_token_type_ids:
+            token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
+
+        sequence_labels = None
+        token_labels = None
+        choice_labels = None
+        if self.use_labels:
+            sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
+            token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
+            choice_labels = ids_tensor([self.batch_size], self.num_choices)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+
+    def get_config(self):
+        return NystromformerConfig(
+            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,
+            is_decoder=False,
+            initializer_range=self.initializer_range,
+        )
+
+    def create_and_check_model(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = NystromformerModel(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))
+
+    def create_and_check_for_masked_lm(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = NystromformerForMaskedLM(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_question_answering(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        model = NystromformerForQuestionAnswering(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_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 = NystromformerForSequenceClassification(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_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 = NystromformerForTokenClassification(config=config)
+        model.to(torch_device)
+        model.eval()
+        result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=token_labels)
+        self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
+
+    def create_and_check_for_multiple_choice(
+        self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
+    ):
+        config.num_choices = self.num_choices
+        model = NystromformerForMultipleChoice(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 NystromformerModelTest(ModelTesterMixin, unittest.TestCase):
+
+    all_model_classes = (
+        (
+            NystromformerModel,
+            NystromformerForMaskedLM,
+            NystromformerForMultipleChoice,
+            NystromformerForQuestionAnswering,
+            NystromformerForSequenceClassification,
+            NystromformerForTokenClassification,
+        )
+        if is_torch_available()
+        else ()
+    )
+    test_pruning = False
+    test_headmasking = False
+
+    def setUp(self):
+        self.model_tester = NystromformerModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=NystromformerConfig, hidden_size=37)
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_model_various_embeddings(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        for type in ["absolute", "relative_key", "relative_key_query"]:
+            config_and_inputs[0].position_embedding_type = type
+            self.model_tester.create_and_check_model(*config_and_inputs)
+
+    def test_for_masked_lm(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_masked_lm(*config_and_inputs)
+
+    def test_for_multiple_choice(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_multiple_choice(*config_and_inputs)
+
+    def test_for_question_answering(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_for_question_answering(*config_and_inputs)
+
+    def test_for_sequence_classification(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_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_for_token_classification(*config_and_inputs)
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in NYSTROMFORMER_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = NystromformerModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+@require_torch
+class NystromformerModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference_no_head(self):
+        model = NystromformerModel.from_pretrained("uw-madison/nystromformer-512")
+        input_ids = torch.tensor([[0, 1, 2, 3, 4, 5]])
+
+        with torch.no_grad():
+            output = model(input_ids)[0]
+
+        expected_shape = torch.Size((1, 6, 768))
+        self.assertEqual(output.shape, expected_shape)
+
+        expected_slice = torch.tensor(
+            [[[-0.4532, -0.0936, 0.5137], [-0.2676, 0.0628, 0.6186], [-0.3629, -0.1726, 0.4716]]]
+        )
+
+        self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-4))
+
+    @slow
+    def test_masked_lm_end_to_end(self):
+        sentence = "the [MASK] of Belgium is Brussels"
+
+        tokenizer = AutoTokenizer.from_pretrained("uw-madison/nystromformer-512")
+        model = NystromformerForMaskedLM.from_pretrained("uw-madison/nystromformer-512")
+
+        encoding = tokenizer(sentence, return_tensors="pt")
+
+        with torch.no_grad():
+            token_logits = model(encoding.input_ids).logits
+
+        prediction = token_logits[:, 2, :].argmax(-1)[0]
+
+        self.assertEqual(tokenizer.decode(prediction), "capital")