Add ALIGN to transformers (#21741)

Adds the ALIGN model to transformers. ALIGN is introduced in "Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision" by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
2025-07-31 02:02:21 +06:00 · 2023-03-01 21:23:31 +03:00 · 2023-03-01 21:23:31 +03:00 · 269b054939
commit 269b054939
parent f7c618e3b0
33 changed files with 3569 additions and 12 deletions
--- a/README.md
+++ b/README.md
@ -271,6 +271,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 🤗 Transformers currently provides the following architectures (see [here](https://huggingface.co/docs/transformers/model_summary) for a high-level summary of each them):

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
--- a/README_es.md
+++ b/README_es.md
@ -264,6 +264,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 🤗 Transformers actualmente proporciona las siguientes arquitecturas (ver [aquí](https://huggingface.co/docs/transformers/model_summary) para un resumen de alto nivel de cada uno de ellas.):

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
--- a/README_hd.md
+++ b/README_hd.md
@ -236,6 +236,7 @@ conda install -c huggingface transformers
 🤗 ट्रांसफॉर्मर वर्तमान में निम्नलिखित आर्किटेक्चर का समर्थन करते हैं (मॉडल के अवलोकन के लिए [यहां] देखें (https://huggingface.co/docs/transformers/model_summary))：

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (Google Research and the Toyota Technological Institute at Chicago) साथ थीसिस [ALBERT: A Lite BERT for Self-supervised भाषा प्रतिनिधित्व सीखना](https://arxiv.org/abs/1909.11942), झेंझोंग लैन, मिंगदा चेन, सेबेस्टियन गुडमैन, केविन गिम्पेल, पीयूष शर्मा, राडू सोरिकट
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research से) Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. द्वाराअनुसंधान पत्र [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) के साथ जारी किया गया
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (फेसबुक) साथ थीसिस [बार्ट: प्राकृतिक भाषा निर्माण, अनुवाद के लिए अनुक्रम-से-अनुक्रम पूर्व प्रशिक्षण , और समझ] (https://arxiv.org/pdf/1910.13461.pdf) पर निर्भर माइक लुईस, यिनहान लियू, नमन गोयल, मार्जन ग़ज़विनिनेजाद, अब्देलरहमान मोहम्मद, ओमर लेवी, वेस स्टोयानोव और ल्यूक ज़ेटलमॉयर
--- a/README_ja.md
+++ b/README_ja.md
@ -298,6 +298,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 🤗Transformersは現在、以下のアーキテクチャを提供しています（それぞれのハイレベルな要約は[こちら](https://huggingface.co/docs/transformers/model_summary)を参照してください）:

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (Google Research and the Toyota Technological Institute at Chicago から) Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut から公開された研究論文: [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942)
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research から) Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. から公開された研究論文 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918)
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (BAAI から) Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell から公開された研究論文: [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679)
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (MIT から) Yuan Gong, Yu-An Chung, James Glass から公開された研究論文: [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778)
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (Facebook から) Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer から公開された研究論文: [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461)
--- a/README_ko.md
+++ b/README_ko.md
@ -213,6 +213,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 🤗 Transformers는 다음 모델들을 제공합니다 (각 모델의 요약은 [여기](https://huggingface.co/docs/transformers/model_summary)서 확인하세요):

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (Google Research 에서 제공)은 Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.의 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918)논문과 함께 발표했습니다.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
--- a/README_zh-hans.md
+++ b/README_zh-hans.md
@ -237,6 +237,7 @@ conda install -c huggingface transformers
 🤗 Transformers 目前支持如下的架构（模型概述请阅[这里](https://huggingface.co/docs/transformers/model_summary)）：

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (来自 Google Research and the Toyota Technological Institute at Chicago) 伴随论文 [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), 由 Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut 发布。
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (来自 Google Research) 伴随论文 [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) 由 Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig 发布。
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (来自 BAAI) 伴随论文 [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) 由 Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell 发布。
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (来自 MIT) 伴随论文 [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) 由 Yuan Gong, Yu-An Chung, James Glass 发布。
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (来自 Facebook) 伴随论文 [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) 由 Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer 发布。
--- a/README_zh-hant.md
+++ b/README_zh-hant.md
@ -249,6 +249,7 @@ conda install -c huggingface transformers
 🤗 Transformers 目前支援以下的架構（模型概覽請參閱[這裡](https://huggingface.co/docs/transformers/model_summary)）：

 1. **[ALBERT](https://huggingface.co/docs/transformers/model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](https://huggingface.co/docs/transformers/main/model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
--- a/docs/source/de/index.mdx
+++ b/docs/source/de/index.mdx
@ -52,6 +52,7 @@ Die Bibliothek enthält derzeit JAX-, PyTorch- und TensorFlow-Implementierungen,
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->

 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
--- a/docs/source/en/_toctree.yml
+++ b/docs/source/en/_toctree.yml
@ -540,6 +540,8 @@
      title: Audio models
    - isExpanded: false
      sections:
+      - local: model_doc/align
+        title: ALIGN
      - local: model_doc/altclip
        title: AltCLIP
      - local: model_doc/blip
--- a/docs/source/en/index.mdx
+++ b/docs/source/en/index.mdx
@ -50,6 +50,7 @@ The documentation is organized into five sections:
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->

 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
@ -247,6 +248,7 @@ Flax), PyTorch, and/or TensorFlow.
 |             Model             | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support |
 |:-----------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:|
 |            ALBERT             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
+|             ALIGN             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            AltCLIP            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 | Audio Spectrogram Transformer |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             BART              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
--- a/docs/source/en/model_doc/align.mdx
+++ b/docs/source/en/model_doc/align.mdx
@ -0,0 +1,59 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ALIGN
+
+## Overview
+
+The ALIGN model was proposed in [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig. ALIGN features a dual-encoder architecture with [EfficientNet](efficientnet) as its vision encoder and [BERT](bert) as its text encoder, and learns to align visual and text representations with contrastive learning. Unlike previous work, ALIGN leverages a massive noisy dataset and shows that the scale of the corpus can be used to achieve SOTA representations with a simple recipe.
+
+The abstract from the paper is the following:
+
+*Pre-trained representations are becoming crucial for many NLP and perception tasks. While representation learning in NLP has transitioned to training on raw text without human annotations, visual and vision-language representations still rely heavily on curated training datasets that are expensive or require expert knowledge. For vision applications, representations are mostly learned using datasets with explicit class labels such as ImageNet or OpenImages. For vision-language, popular datasets like Conceptual Captions, MSCOCO, or CLIP all involve a non-trivial data collection (and cleaning) process. This costly curation process limits the size of datasets and hence hinders the scaling of trained models. In this paper, we leverage a noisy dataset of over one billion image alt-text pairs, obtained without expensive filtering or post-processing steps in the Conceptual Captions dataset. A simple dual-encoder architecture learns to align visual and language representations of the image and text pairs using a contrastive loss. We show that the scale of our corpus can make up for its noise and leads to state-of-the-art representations even with such a simple learning scheme. Our visual representation achieves strong performance when transferred to classification tasks such as ImageNet and VTAB. The aligned visual and language representations enables zero-shot image classification and also set new state-of-the-art results on Flickr30K and MSCOCO image-text retrieval benchmarks, even when compared with more sophisticated cross-attention models. The representations also enable cross-modality search with complex text and text + image queries.*
+
+This model was contributed by [Alara Dirik](https://huggingface.co/adirik).
+The original code is not released, this implementation is based on the Kakao Brain implementation based on the original paper.
+
+
+## AlignConfig
+
+[[autodoc]] AlignConfig
+    - from_text_vision_configs
+
+## AlignTextConfig
+
+[[autodoc]] AlignTextConfig
+
+## AlignVisionConfig
+
+[[autodoc]] AlignVisionConfig
+
+## AlignProcessor
+
+[[autodoc]] AlignProcessor
+
+## AlignModel
+
+[[autodoc]] AlignModel
+    - forward
+    - get_text_features
+    - get_image_features
+
+## AlignTextModel
+
+[[autodoc]] AlignTextModel
+    - forward
+
+## AlignVisionModel
+
+[[autodoc]] AlignVisionModel
+    - forward
--- a/docs/source/es/index.mdx
+++ b/docs/source/es/index.mdx
@ -46,6 +46,7 @@ La biblioteca actualmente contiene implementaciones de JAX, PyTorch y TensorFlow
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->

 1. **[ALBERT](model_doc/albert)** (de Google Research y el Instituto Tecnológico de Toyota en Chicago) publicado con el paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), por Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (de Google Research) publicado con el paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) por Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (de Facebook) publicado con el paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) por Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov y Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (de École polytechnique) publicado con el paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) por Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (de VinAI Research) publicado con el paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) por Nguyen Luong Tran, Duong Minh Le y Dat Quoc Nguyen.
--- a/docs/source/fr/index.mdx
+++ b/docs/source/fr/index.mdx
@ -50,6 +50,7 @@ La documentation est organisée en 5 parties:
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->

 1. **[ALBERT](model_doc/albert)** (from Google Research and the Toyota Technological Institute at Chicago) released with the paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), by Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) released with the paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) by Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
--- a/docs/source/it/index.mdx
+++ b/docs/source/it/index.mdx
@ -51,6 +51,7 @@ La libreria attualmente contiene implementazioni in JAX, PyTorch e TensorFlow, p
 <!--This list is updated automatically from the README with _make fix-copies_. Do not update manually! -->

 1. **[ALBERT](model_doc/albert)** (da Google Research e l'Istituto Tecnologico di Chicago) rilasciato con il paper [ALBERT: A Lite BERT for Self-supervised Learning of Language Representations](https://arxiv.org/abs/1909.11942), da Zhenzhong Lan, Mingda Chen, Sebastian Goodman, Kevin Gimpel, Piyush Sharma, Radu Soricut.
+1. **[ALIGN](model_doc/align)** (from Google Research) rilasciato con il paper [Scaling Up Visual and Vision-Language Representation Learning With Noisy Text Supervision](https://arxiv.org/abs/2102.05918) da Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc V. Le, Yunhsuan Sung, Zhen Li, Tom Duerig.
 1. **[BART](model_doc/bart)** (da Facebook) rilasciato con il paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) da Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov e Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (da politecnico di École) rilasciato con il paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) da Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (da VinAI Research) rilasciato con il paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) da Nguyen Luong Tran, Duong Minh Le e Dat Quoc Nguyen.
--- a/src/transformers/init.py
+++ b/src/transformers/init.py
@ -123,6 +123,13 @@ _import_structure = {
    "models": [],
    # Models
    "models.albert": ["ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "AlbertConfig"],
+    "models.align": [
+        "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AlignConfig",
+        "AlignProcessor",
+        "AlignTextConfig",
+        "AlignVisionConfig",
+    ],
    "models.altclip": [
        "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP",
        "AltCLIPConfig",
@ -983,6 +990,15 @@ else:
            "load_tf_weights_in_albert",
        ]
    )
+    _import_structure["models.align"].extend(
+        [
+            "ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "AlignModel",
+            "AlignPreTrainedModel",
+            "AlignTextModel",
+            "AlignVisionModel",
+        ]
+    )
    _import_structure["models.altclip"].extend(
        [
            "ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
@ -3710,6 +3726,13 @@ if TYPE_CHECKING:
        load_tf2_weights_in_pytorch_model,
    )
    from .models.albert import ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, AlbertConfig
+    from .models.align import (
+        ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AlignConfig,
+        AlignProcessor,
+        AlignTextConfig,
+        AlignVisionConfig,
+    )
    from .models.altclip import (
        ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP,
        AltCLIPConfig,
@ -4474,6 +4497,13 @@ if TYPE_CHECKING:
            AlbertPreTrainedModel,
            load_tf_weights_in_albert,
        )
+        from .models.align import (
+            ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlignModel,
+            AlignPreTrainedModel,
+            AlignTextModel,
+            AlignVisionModel,
+        )
        from .models.altclip import (
            ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST,
            AltCLIPModel,
--- a/src/transformers/models/init.py
+++ b/src/transformers/models/init.py
@ -14,6 +14,7 @@

 from . import (
    albert,
+    align,
    altclip,
    audio_spectrogram_transformer,
    auto,
--- a/src/transformers/models/align/init.py
+++ b/src/transformers/models/align/init.py
@ -0,0 +1,73 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_align": [
+        "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "AlignConfig",
+        "AlignTextConfig",
+        "AlignVisionConfig",
+    ],
+    "processing_align": ["AlignProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_align"] = [
+        "ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "AlignModel",
+        "AlignPreTrainedModel",
+        "AlignTextModel",
+        "AlignVisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_align import (
+        ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        AlignConfig,
+        AlignTextConfig,
+        AlignVisionConfig,
+    )
+    from .processing_align import AlignProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_align import (
+            ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST,
+            AlignModel,
+            AlignPreTrainedModel,
+            AlignTextModel,
+            AlignVisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
--- a/src/transformers/models/align/configuration_align.py
+++ b/src/transformers/models/align/configuration_align.py
@ -0,0 +1,398 @@
+# coding=utf-8
+# Copyright 2023 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.
+""" ALIGN model configuration"""
+
+import copy
+import os
+from typing import TYPE_CHECKING, List, Union
+
+
+if TYPE_CHECKING:
+    pass
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "kakaobrain/align-base": "https://huggingface.co/kakaobrain/align-base/resolve/main/config.json",
+}
+
+
+class AlignTextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignTextModel`]. It is used to instantiate a
+    ALIGN text encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the text encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values here are
+    copied from BERT.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Align Text model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling [`AlignTextModel`].
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the encoder layers and the pooler layer.
+        num_hidden_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        intermediate_size (`int`, *optional*, defaults to 3072):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        hidden_act (`str` or `Callable`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        hidden_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_probs_dropout_prob (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        max_position_embeddings (`int`, *optional*, defaults to 512):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        type_vocab_size (`int`, *optional*, defaults to 2):
+            The vocabulary size of the `token_type_ids` passed when calling [`AlignTextModel`].
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-12):
+            The epsilon used by the layer normalization layers.
+        position_embedding_type (`str`, *optional*, defaults to `"absolute"`):
+            Type of position embedding. Choose one of `"absolute"`, `"relative_key"`, `"relative_key_query"`. For
+            positional embeddings use `"absolute"`. For more information on `"relative_key"`, please refer to
+            [Self-Attention with Relative Position Representations (Shaw et al.)](https://arxiv.org/abs/1803.02155).
+            For more information on `"relative_key_query"`, please refer to *Method 4* in [Improve Transformer Models
+            with Better Relative Position Embeddings (Huang et al.)](https://arxiv.org/abs/2009.13658).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models). Only
+            relevant if `config.is_decoder=True`.
+        pad_token_id (`int`, *optional*, defaults to 0)
+            Padding token id.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignTextConfig, AlignTextModel
+
+    >>> # Initializing a AlignTextConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignTextConfig()
+
+    >>> # Initializing a AlignTextModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignTextModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "align_text_model"
+
+    def __init__(
+        self,
+        vocab_size=30522,
+        hidden_size=768,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        intermediate_size=3072,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=512,
+        type_vocab_size=2,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        use_cache=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.type_vocab_size = type_vocab_size
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.use_cache = use_cache
+        self.pad_token_id = pad_token_id
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from AlignConfig
+        if config_dict.get("model_type") == "align":
+            config_dict = config_dict["text_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class AlignVisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`AlignVisionModel`]. It is used to instantiate a
+    ALIGN vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration with the defaults will yield a similar configuration to that of the vision encoder of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture. The default values are copied
+    from EfficientNet (efficientnet-b7)
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 600):
+            The input image size.
+        width_coefficient (`float`, *optional*, defaults to 2.0):
+            Scaling coefficient for network width at each stage.
+        depth_coefficient (`float`, *optional*, defaults to 3.1):
+            Scaling coefficient for network depth at each stage.
+        depth_divisor `int`, *optional*, defaults to 8):
+            A unit of network width.
+        kernel_sizes (`List[int]`, *optional*, defaults to `[3, 3, 5, 3, 5, 5, 3]`):
+            List of kernel sizes to be used in each block.
+        in_channels (`List[int]`, *optional*, defaults to `[32, 16, 24, 40, 80, 112, 192]`):
+            List of input channel sizes to be used in each block for convolutional layers.
+        out_channels (`List[int]`, *optional*, defaults to `[16, 24, 40, 80, 112, 192, 320]`):
+            List of output channel sizes to be used in each block for convolutional layers.
+        depthwise_padding (`List[int]`, *optional*, defaults to `[]`):
+            List of block indices with square padding.
+        strides: (`List[int]`, *optional*, defaults to `[1, 2, 2, 2, 1, 2, 1]`):
+            List of stride sizes to be used in each block for convolutional layers.
+        num_block_repeats (`List[int]`, *optional*, defaults to `[1, 2, 2, 3, 3, 4, 1]`):
+            List of the number of times each block is to repeated.
+        expand_ratios (`List[int]`, *optional*, defaults to `[1, 6, 6, 6, 6, 6, 6]`):
+            List of scaling coefficient of each block.
+        squeeze_expansion_ratio (`float`, *optional*, defaults to 0.25):
+            Squeeze expansion ratio.
+        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+            The non-linear activation function (function or string) in each block. If string, `"gelu"`, `"relu"`,
+            `"selu", `"gelu_new"`, `"silu"` and `"mish"` are supported.
+        hiddem_dim (`int`, *optional*, defaults to 1280):
+            The hidden dimension of the layer before the classification head.
+        pooling_type (`str` or `function`, *optional*, defaults to `"mean"`):
+            Type of final pooling to be applied before the dense classification head. Available options are [`"mean"`,
+            `"max"`]
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        batch_norm_eps (`float`, *optional*, defaults to 1e-3):
+            The epsilon used by the batch normalization layers.
+        batch_norm_momentum (`float`, *optional*, defaults to 0.99):
+            The momentum used by the batch normalization layers.
+        dropout_rate (`float`, *optional*, defaults to 0.5):
+            The dropout rate to be applied before final classifier layer.
+        drop_connect_rate (`float`, *optional*, defaults to 0.2):
+            The drop rate for skip connections.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignVisionConfig, AlignVisionModel
+
+    >>> # Initializing a AlignVisionConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignVisionConfig()
+
+    >>> # Initializing a AlignVisionModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignVisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "align_vision_model"
+
+    def __init__(
+        self,
+        num_channels: int = 3,
+        image_size: int = 600,
+        width_coefficient: float = 2.0,
+        depth_coefficient: float = 3.1,
+        depth_divisor: int = 8,
+        kernel_sizes: List[int] = [3, 3, 5, 3, 5, 5, 3],
+        in_channels: List[int] = [32, 16, 24, 40, 80, 112, 192],
+        out_channels: List[int] = [16, 24, 40, 80, 112, 192, 320],
+        depthwise_padding: List[int] = [],
+        strides: List[int] = [1, 2, 2, 2, 1, 2, 1],
+        num_block_repeats: List[int] = [1, 2, 2, 3, 3, 4, 1],
+        expand_ratios: List[int] = [1, 6, 6, 6, 6, 6, 6],
+        squeeze_expansion_ratio: float = 0.25,
+        hidden_act: str = "swish",
+        hidden_dim: int = 2560,
+        pooling_type: str = "mean",
+        initializer_range: float = 0.02,
+        batch_norm_eps: float = 0.001,
+        batch_norm_momentum: float = 0.99,
+        dropout_rate: float = 0.5,
+        drop_connect_rate: float = 0.2,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.num_channels = num_channels
+        self.image_size = image_size
+        self.width_coefficient = width_coefficient
+        self.depth_coefficient = depth_coefficient
+        self.depth_divisor = depth_divisor
+        self.kernel_sizes = kernel_sizes
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.depthwise_padding = depthwise_padding
+        self.strides = strides
+        self.num_block_repeats = num_block_repeats
+        self.expand_ratios = expand_ratios
+        self.squeeze_expansion_ratio = squeeze_expansion_ratio
+        self.hidden_act = hidden_act
+        self.hidden_dim = hidden_dim
+        self.pooling_type = pooling_type
+        self.initializer_range = initializer_range
+        self.batch_norm_eps = batch_norm_eps
+        self.batch_norm_momentum = batch_norm_momentum
+        self.dropout_rate = dropout_rate
+        self.drop_connect_rate = drop_connect_rate
+        self.num_hidden_layers = sum(num_block_repeats) * 4
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from AlignConfig
+        if config_dict.get("model_type") == "align":
+            config_dict = config_dict["vision_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+class AlignConfig(PretrainedConfig):
+    r"""
+    [`AlignConfig`] is the configuration class to store the configuration of a [`AlignModel`]. It is used to
+    instantiate a ALIGN model according to the specified arguments, defining the text model and vision model configs.
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the ALIGN
+    [kakaobrain/align-base](https://huggingface.co/kakaobrain/align-base) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignTextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`AlignVisionConfig`].
+        projection_dim (`int`, *optional*, defaults to 512):
+            Dimentionality of text and vision projection layers.
+        temperature_init_value (`float`, *optional*, defaults to 1.0):
+            The inital value of the *temperature* paramter. Default is used as per the original ALIGN implementation.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import AlignConfig, AlignModel
+
+    >>> # Initializing a AlignConfig with kakaobrain/align-base style configuration
+    >>> configuration = AlignConfig()
+
+    >>> # Initializing a AlignModel (with random weights) from the kakaobrain/align-base style configuration
+    >>> model = AlignModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a AlignConfig from a AlignTextConfig and a AlignVisionConfig
+    >>> from transformers import AlignTextConfig, AlignVisionConfig
+
+    >>> # Initializing ALIGN Text and Vision configurations
+    >>> config_text = AlignTextConfig()
+    >>> config_vision = AlignVisionConfig()
+
+    >>> config = AlignConfig.from_text_vision_configs(config_text, config_vision)
+    ```"""
+
+    model_type = "align"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        projection_dim=640,
+        temperature_init_value=1.0,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the AlignTextConfig with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. Initializing the AlignVisionConfig with default values.")
+
+        self.text_config = AlignTextConfig(**text_config)
+        self.vision_config = AlignVisionConfig(**vision_config)
+
+        self.projection_dim = projection_dim
+        self.temperature_init_value = temperature_init_value
+        self.initializer_range = initializer_range
+
+    @classmethod
+    def from_text_vision_configs(cls, text_config: AlignTextConfig, vision_config: AlignVisionConfig, **kwargs):
+        r"""
+        Instantiate a [`AlignConfig`] (or a derived class) from align text model configuration and align vision model
+        configuration.
+
+        Returns:
+            [`AlignConfig`]: An instance of a configuration object
+        """
+
+        return cls(text_config=text_config.to_dict(), vision_config=vision_config.to_dict(), **kwargs)
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+        output["text_config"] = self.text_config.to_dict()
+        output["vision_config"] = self.vision_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output
--- a/src/transformers/models/align/convert_align_tf_to_hf.py
+++ b/src/transformers/models/align/convert_align_tf_to_hf.py
@ -0,0 +1,387 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Convert ALIGN checkpoints from the original repository."""
+
+import argparse
+import os
+
+import align
+import numpy as np
+import requests
+import tensorflow as tf
+import torch
+from PIL import Image
+from tokenizer import Tokenizer
+
+from transformers import (
+    AlignConfig,
+    AlignModel,
+    AlignProcessor,
+    BertConfig,
+    BertTokenizer,
+    EfficientNetConfig,
+    EfficientNetImageProcessor,
+)
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+
+def preprocess(image):
+    image = tf.image.resize(image, (346, 346))
+    image = tf.image.crop_to_bounding_box(image, (346 - 289) // 2, (346 - 289) // 2, 289, 289)
+    return image
+
+
+def get_align_config():
+    vision_config = EfficientNetConfig.from_pretrained("google/efficientnet-b7")
+    vision_config.image_size = 289
+    vision_config.hidden_dim = 640
+    vision_config.id2label = {"0": "LABEL_0", "1": "LABEL_1"}
+    vision_config.label2id = {"LABEL_0": 0, "LABEL_1": 1}
+    vision_config.depthwise_padding = []
+
+    text_config = BertConfig()
+    config = AlignConfig.from_text_vision_configs(
+        text_config=text_config, vision_config=vision_config, projection_dim=640
+    )
+    return config
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+def get_processor():
+    image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        rescale_factor=1 / 127.5,
+        rescale_offset=True,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
+    tokenizer.model_max_length = 64
+    processor = AlignProcessor(image_processor=image_processor, tokenizer=tokenizer)
+    return processor
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def rename_keys(original_param_names):
+    # EfficientNet image encoder
+    block_names = [v.split("_")[0].split("block")[1] for v in original_param_names if v.startswith("block")]
+    block_names = list(set(block_names))
+    block_names = sorted(block_names)
+    num_blocks = len(block_names)
+    block_name_mapping = {b: str(i) for b, i in zip(block_names, range(num_blocks))}
+
+    rename_keys = []
+    rename_keys.append(("stem_conv/kernel:0", "embeddings.convolution.weight"))
+    rename_keys.append(("stem_bn/gamma:0", "embeddings.batchnorm.weight"))
+    rename_keys.append(("stem_bn/beta:0", "embeddings.batchnorm.bias"))
+    rename_keys.append(("stem_bn/moving_mean:0", "embeddings.batchnorm.running_mean"))
+    rename_keys.append(("stem_bn/moving_variance:0", "embeddings.batchnorm.running_var"))
+
+    for b in block_names:
+        hf_b = block_name_mapping[b]
+        rename_keys.append((f"block{b}_expand_conv/kernel:0", f"encoder.blocks.{hf_b}.expansion.expand_conv.weight"))
+        rename_keys.append((f"block{b}_expand_bn/gamma:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.weight"))
+        rename_keys.append((f"block{b}_expand_bn/beta:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_mean:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_expand_bn/moving_variance:0", f"encoder.blocks.{hf_b}.expansion.expand_bn.running_var")
+        )
+        rename_keys.append(
+            (f"block{b}_dwconv/depthwise_kernel:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_conv.weight")
+        )
+        rename_keys.append((f"block{b}_bn/gamma:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.weight"))
+        rename_keys.append((f"block{b}_bn/beta:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.bias"))
+        rename_keys.append(
+            (f"block{b}_bn/moving_mean:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_bn/moving_variance:0", f"encoder.blocks.{hf_b}.depthwise_conv.depthwise_norm.running_var")
+        )
+
+        rename_keys.append((f"block{b}_se_reduce/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.weight"))
+        rename_keys.append((f"block{b}_se_reduce/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.reduce.bias"))
+        rename_keys.append((f"block{b}_se_expand/kernel:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.weight"))
+        rename_keys.append((f"block{b}_se_expand/bias:0", f"encoder.blocks.{hf_b}.squeeze_excite.expand.bias"))
+        rename_keys.append(
+            (f"block{b}_project_conv/kernel:0", f"encoder.blocks.{hf_b}.projection.project_conv.weight")
+        )
+        rename_keys.append((f"block{b}_project_bn/gamma:0", f"encoder.blocks.{hf_b}.projection.project_bn.weight"))
+        rename_keys.append((f"block{b}_project_bn/beta:0", f"encoder.blocks.{hf_b}.projection.project_bn.bias"))
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_mean:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_mean")
+        )
+        rename_keys.append(
+            (f"block{b}_project_bn/moving_variance:0", f"encoder.blocks.{hf_b}.projection.project_bn.running_var")
+        )
+
+    key_mapping = {}
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = "vision_model." + item[1]
+
+    # BERT text encoder
+    rename_keys = []
+    old = "tf_bert_model/bert"
+    new = "text_model"
+    for i in range(12):
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/query/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.query.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/key/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.key.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/self/value/bias:0",
+                f"{new}.encoder.layer.{i}.attention.self.value.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/dense/bias:0",
+                f"{new}.encoder.layer.{i}.attention.output.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/gamma:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/attention/output/LayerNorm/beta:0",
+                f"{new}.encoder.layer.{i}.attention.output.LayerNorm.bias",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/kernel:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.weight",
+            )
+        )
+        rename_keys.append(
+            (
+                f"{old}/encoder/layer_._{i}/intermediate/dense/bias:0",
+                f"{new}.encoder.layer.{i}.intermediate.dense.bias",
+            )
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/kernel:0", f"{new}.encoder.layer.{i}.output.dense.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/dense/bias:0", f"{new}.encoder.layer.{i}.output.dense.bias")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/gamma:0", f"{new}.encoder.layer.{i}.output.LayerNorm.weight")
+        )
+        rename_keys.append(
+            (f"{old}/encoder/layer_._{i}/output/LayerNorm/beta:0", f"{new}.encoder.layer.{i}.output.LayerNorm.bias")
+        )
+
+    rename_keys.append((f"{old}/embeddings/word_embeddings/weight:0", f"{new}.embeddings.word_embeddings.weight"))
+    rename_keys.append(
+        (f"{old}/embeddings/position_embeddings/embeddings:0", f"{new}.embeddings.position_embeddings.weight")
+    )
+    rename_keys.append(
+        (f"{old}/embeddings/token_type_embeddings/embeddings:0", f"{new}.embeddings.token_type_embeddings.weight")
+    )
+    rename_keys.append((f"{old}/embeddings/LayerNorm/gamma:0", f"{new}.embeddings.LayerNorm.weight"))
+    rename_keys.append((f"{old}/embeddings/LayerNorm/beta:0", f"{new}.embeddings.LayerNorm.bias"))
+
+    rename_keys.append((f"{old}/pooler/dense/kernel:0", f"{new}.pooler.dense.weight"))
+    rename_keys.append((f"{old}/pooler/dense/bias:0", f"{new}.pooler.dense.bias"))
+    rename_keys.append(("dense/kernel:0", "text_projection.weight"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("dense/bias:0", "text_projection.bias"))
+    rename_keys.append(("temperature:0", "temperature"))
+
+    for item in rename_keys:
+        if item[0] in original_param_names:
+            key_mapping[item[0]] = item[1]
+    return key_mapping
+
+
+def replace_params(hf_params, tf_params, key_mapping):
+    list(hf_params.keys())
+
+    for key, value in tf_params.items():
+        if key not in key_mapping:
+            continue
+
+        hf_key = key_mapping[key]
+        if "_conv" in key and "kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(3, 2, 0, 1)
+        elif "embeddings" in key:
+            new_hf_value = torch.from_numpy(value)
+        elif "depthwise_kernel" in key:
+            new_hf_value = torch.from_numpy(value).permute(2, 3, 0, 1)
+        elif "kernel" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value))
+        elif "temperature" in key:
+            new_hf_value = value
+        elif "bn/gamma" or "bn/beta" in key:
+            new_hf_value = torch.from_numpy(np.transpose(value)).squeeze()
+        else:
+            new_hf_value = torch.from_numpy(value)
+
+        # Replace HF parameters with original TF model parameters
+        hf_params[hf_key].copy_(new_hf_value)
+
+
+@torch.no_grad()
+def convert_align_checkpoint(checkpoint_path, pytorch_dump_folder_path, save_model, push_to_hub):
+    """
+    Copy/paste/tweak model's weights to our ALIGN structure.
+    """
+    # Load original model
+    seq_length = 64
+    tok = Tokenizer(seq_length)
+    original_model = align.Align("efficientnet-b7", "bert-base", 640, seq_length, tok.get_vocab_size())
+    original_model.compile()
+    original_model.load_weights(checkpoint_path)
+
+    tf_params = original_model.trainable_variables
+    tf_non_train_params = original_model.non_trainable_variables
+    tf_params = {param.name: param.numpy() for param in tf_params}
+    for param in tf_non_train_params:
+        tf_params[param.name] = param.numpy()
+    tf_param_names = list(tf_params.keys())
+
+    # Load HuggingFace model
+    config = get_align_config()
+    hf_model = AlignModel(config).eval()
+    hf_params = hf_model.state_dict()
+
+    # Create src-to-dst parameter name mapping dictionary
+    print("Converting parameters...")
+    key_mapping = rename_keys(tf_param_names)
+    replace_params(hf_params, tf_params, key_mapping)
+
+    # Initialize processor
+    processor = get_processor()
+    inputs = processor(
+        images=prepare_img(), text="A picture of a cat", padding="max_length", max_length=64, return_tensors="pt"
+    )
+
+    # HF model inference
+    hf_model.eval()
+    with torch.no_grad():
+        outputs = hf_model(**inputs)
+
+    hf_image_features = outputs.image_embeds.detach().numpy()
+    hf_text_features = outputs.text_embeds.detach().numpy()
+
+    # Original model inference
+    original_model.trainable = False
+    tf_image_processor = EfficientNetImageProcessor(
+        do_center_crop=True,
+        do_rescale=False,
+        do_normalize=False,
+        include_top=False,
+        resample=Image.BILINEAR,
+    )
+    image = tf_image_processor(images=prepare_img(), return_tensors="tf", data_format="channels_last")["pixel_values"]
+    text = tok(tf.constant(["A picture of a cat"]))
+
+    image_features = original_model.image_encoder(image, training=False)
+    text_features = original_model.text_encoder(text, training=False)
+
+    image_features = tf.nn.l2_normalize(image_features, axis=-1)
+    text_features = tf.nn.l2_normalize(text_features, axis=-1)
+
+    # Check whether original and HF model outputs match  -> np.allclose
+    assert np.allclose(image_features, hf_image_features, atol=1e-3), "The predicted image features are not the same."
+    assert np.allclose(text_features, hf_text_features, atol=1e-3), "The predicted text features are not the same."
+    print("Model outputs match!")
+
+    if save_model:
+        # Create folder to save model
+        if not os.path.isdir(pytorch_dump_folder_path):
+            os.mkdir(pytorch_dump_folder_path)
+        # Save converted model and feature extractor
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+        processor.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        # Push model and feature extractor to hub
+        print("Pushing converted ALIGN to the hub...")
+        processor.push_to_hub("align-base")
+        hf_model.push_to_hub("align-base")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--checkpoint_path",
+        default="./weights/model-weights",
+        type=str,
+        help="Path to the pretrained TF ALIGN checkpoint.",
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path",
+        default="hf_model",
+        type=str,
+        help="Path to the output PyTorch model directory.",
+    )
+    parser.add_argument("--save_model", action="store_true", help="Save model to local")
+    parser.add_argument("--push_to_hub", action="store_true", help="Push model and feature extractor to the hub")
+
+    args = parser.parse_args()
+    convert_align_checkpoint(args.checkpoint_path, args.pytorch_dump_folder_path, args.save_model, args.push_to_hub)
--- a/src/transformers/models/align/modeling_align.py
+++ b/src/transformers/models/align/modeling_align.py
--- a/src/transformers/models/align/processing_align.py
+++ b/src/transformers/models/align/processing_align.py
@ -0,0 +1,122 @@
+# coding=utf-8
+# Copyright 2023 The HuggingFace Inc. team.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Image/Text processor class for ALIGN
+"""
+
+
+from ...processing_utils import ProcessorMixin
+from ...tokenization_utils_base import BatchEncoding
+
+
+class AlignProcessor(ProcessorMixin):
+    r"""
+    Constructs an ALIGN processor which wraps [`EfficientNetImageProcessor`] and
+    [`BertTokenizer`]/[`BertTokenizerFast`] into a single processor that interits both the image processor and
+    tokenizer functionalities. See the [`~AlignProcessor.__call__`] and [`~OwlViTProcessor.decode`] for more
+    information.
+
+    Args:
+        image_processor ([`EfficientNetImageProcessor`]):
+            The image processor is a required input.
+        tokenizer ([`BERTTokenizer`, `BertTokenizerFast`]):
+            The tokenizer is a required input.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    image_processor_class = "EfficientNetImageProcessor"
+    tokenizer_class = ("BertTokenizer", "BertTokenizerFast")
+
+    def __init__(self, image_processor, tokenizer):
+        super().__init__(image_processor, tokenizer)
+
+    def __call__(self, text=None, images=None, padding="max_length", max_length=64, return_tensors=None, **kwargs):
+        """
+        Main method to prepare text(s) and image(s) to be fed as input to the model. This method forwards the `text`
+        and `kwargs` arguments to BertTokenizerFast's [`~BertTokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the image(s), this method forwards the `images` and `kwargs` arguments to
+        EfficientNetImageProcessor's [`~EfficientNetImageProcessor.__call__`] if `images` is not `None`. Please refer
+        to the doctsring of the above two methods for more information.
+
+        Args:
+            text (`str`, `List[str]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. In case of a NumPy array/PyTorch tensor, each image should be of shape (C, H, W), where C is a
+                number of channels, H and W are image height and width.
+            padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `max_length`):
+                Activates and controls padding for tokenization of input text. Choose between [`True` or `'longest'`,
+                `'max_length'`, `False` or `'do_not_pad'`]
+            max_length (`int`, *optional*, defaults to `max_length`):
+                Maximum padding value to use to pad the input text during tokenization.
+
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchEncoding`]: A [`BatchEncoding`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+        """
+        if text is None and images is None:
+            raise ValueError("You have to specify either text or images. Both cannot be none.")
+
+        if text is not None:
+            encoding = self.tokenizer(
+                text, padding=padding, max_length=max_length, return_tensors=return_tensors, **kwargs
+            )
+
+        if images is not None:
+            image_features = self.image_processor(images, return_tensors=return_tensors, **kwargs)
+
+        if text is not None and images is not None:
+            encoding["pixel_values"] = image_features.pixel_values
+            return encoding
+        elif text is not None:
+            return encoding
+        else:
+            return BatchEncoding(data=dict(**image_features), tensor_type=return_tensors)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to BertTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
--- a/src/transformers/models/auto/configuration_auto.py
+++ b/src/transformers/models/auto/configuration_auto.py
@ -30,6 +30,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
    [
        # Add configs here
        ("albert", "AlbertConfig"),
+        ("align", "AlignConfig"),
        ("altclip", "AltCLIPConfig"),
        ("audio-spectrogram-transformer", "ASTConfig"),
        ("bart", "BartConfig"),
@ -207,6 +208,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
    [
        # Add archive maps here)
        ("albert", "ALBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("align", "ALIGN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("altclip", "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("audio-spectrogram-transformer", "AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
        ("bart", "BART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@ -366,6 +368,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
    [
        # Add full (and cased) model names here
        ("albert", "ALBERT"),
+        ("align", "ALIGN"),
        ("altclip", "AltCLIP"),
        ("audio-spectrogram-transformer", "Audio Spectrogram Transformer"),
        ("bart", "BART"),
--- a/src/transformers/models/auto/image_processing_auto.py
+++ b/src/transformers/models/auto/image_processing_auto.py
@ -37,6 +37,7 @@ logger = logging.get_logger(__name__)

 IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
    [
+        ("align", "EfficientNetImageProcessor"),
        ("beit", "BeitImageProcessor"),
        ("bit", "BitImageProcessor"),
        ("blip", "BlipImageProcessor"),
--- a/src/transformers/models/auto/modeling_auto.py
+++ b/src/transformers/models/auto/modeling_auto.py
@ -29,6 +29,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
    [
        # Base model mapping
        ("albert", "AlbertModel"),
+        ("align", "AlignModel"),
        ("altclip", "AltCLIPModel"),
        ("audio-spectrogram-transformer", "ASTModel"),
        ("bart", "BartModel"),
@ -922,6 +923,7 @@ MODEL_FOR_AUDIO_XVECTOR_MAPPING_NAMES = OrderedDict(
 _MODEL_FOR_ZERO_SHOT_IMAGE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
    [
        # Model for Zero Shot Image Classification mapping
+        ("align", "AlignModel"),
        ("altclip", "AltCLIPModel"),
        ("blip", "BlipModel"),
        ("chinese_clip", "ChineseCLIPModel"),
--- a/src/transformers/models/auto/processing_auto.py
+++ b/src/transformers/models/auto/processing_auto.py
@ -41,6 +41,7 @@ logger = logging.get_logger(__name__)

 PROCESSOR_MAPPING_NAMES = OrderedDict(
    [
+        ("align", "AlignProcessor"),
        ("altclip", "AltCLIPProcessor"),
        ("blip", "BlipProcessor"),
        ("blip-2", "Blip2Processor"),
--- a/src/transformers/models/auto/tokenization_auto.py
+++ b/src/transformers/models/auto/tokenization_auto.py
@ -53,6 +53,7 @@ else:
                    "AlbertTokenizerFast" if is_tokenizers_available() else None,
                ),
            ),
+            ("align", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
            ("bart", ("BartTokenizer", "BartTokenizerFast")),
            (
                "barthez",
--- a/src/transformers/models/efficientnet/image_processing_efficientnet.py
+++ b/src/transformers/models/efficientnet/image_processing_efficientnet.py
@ -199,7 +199,7 @@ class EfficientNetImageProcessor(BaseImageProcessor):
            rescaled_image = rescaled_image.astype(np.float32)
        else:
            rescaled_image = rescale(image, scale=scale, data_format=data_format, **kwargs)
-        return rescale(image, scale=scale, data_format=data_format, **kwargs)
+        return rescaled_image

    def normalize(
        self,
--- a/src/transformers/models/efficientnet/modeling_efficientnet.py
+++ b/src/transformers/models/efficientnet/modeling_efficientnet.py
@ -126,12 +126,12 @@ class EfficientNetEmbeddings(nn.Module):
    def __init__(self, config: EfficientNetConfig):
        super().__init__()

-        out_channels = round_filters(config, 32)
+        self.out_dim = round_filters(config, 32)
        self.padding = nn.ZeroPad2d(padding=(0, 1, 0, 1))
        self.convolution = nn.Conv2d(
-            config.num_channels, out_channels, kernel_size=3, stride=2, padding="valid", bias=False
+            config.num_channels, self.out_dim, kernel_size=3, stride=2, padding="valid", bias=False
        )
-        self.batchnorm = nn.BatchNorm2d(out_channels, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum)
+        self.batchnorm = nn.BatchNorm2d(self.out_dim, eps=config.batch_norm_eps, momentum=config.batch_norm_momentum)
        self.activation = ACT2FN[config.hidden_act]

    def forward(self, pixel_values: torch.Tensor) -> torch.Tensor:
@ -174,13 +174,7 @@ class EfficientNetExpansionLayer(nn.Module):
    This corresponds to the expansion phase of each block in the original implementation.
    """

-    def __init__(
-        self,
-        config: EfficientNetConfig,
-        in_dim: int,
-        out_dim: int,
-        stride: int,
-    ):
+    def __init__(self, config: EfficientNetConfig, in_dim: int, out_dim: int, stride: int):
        super().__init__()
        self.expand_conv = nn.Conv2d(
            in_channels=in_dim,
@ -189,7 +183,7 @@ class EfficientNetExpansionLayer(nn.Module):
            padding="same",
            bias=False,
        )
-        self.expand_bn = nn.BatchNorm2d(num_features=out_dim)
+        self.expand_bn = nn.BatchNorm2d(num_features=out_dim, eps=config.batch_norm_eps)
        self.expand_act = ACT2FN[config.hidden_act]

    def forward(self, hidden_states: torch.FloatTensor) -> torch.Tensor:
--- a/src/transformers/utils/dummy_pt_objects.py
+++ b/src/transformers/utils/dummy_pt_objects.py
@ -356,6 +356,37 @@ def load_tf_weights_in_albert(*args, **kwargs):
    requires_backends(load_tf_weights_in_albert, ["torch"])


+ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class AlignModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class AlignPreTrainedModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class AlignTextModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
+class AlignVisionModel(metaclass=DummyObject):
+    _backends = ["torch"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["torch"])
+
+
 ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST = None


--- a/tests/models/align/init.py
+++ b/tests/models/align/init.py
--- a/tests/models/align/test_modeling_align.py
+++ b/tests/models/align/test_modeling_align.py
@ -0,0 +1,590 @@
+# coding=utf-8
+# Copyright 2023 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 ALIGN model. """
+
+
+import inspect
+import os
+import tempfile
+import unittest
+
+import requests
+
+from transformers import AlignConfig, AlignProcessor, AlignTextConfig, AlignVisionConfig
+from transformers.testing_utils import (
+    is_flax_available,
+    require_torch,
+    require_vision,
+    slow,
+    torch_device,
+)
+from transformers.utils import is_torch_available, is_vision_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_common import (
+    ModelTesterMixin,
+    _config_zero_init,
+    floats_tensor,
+    ids_tensor,
+    random_attention_mask,
+)
+
+
+if is_torch_available():
+    import torch
+
+    from transformers import (
+        AlignModel,
+        AlignTextModel,
+        AlignVisionModel,
+    )
+    from transformers.models.align.modeling_align import ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+if is_flax_available():
+    pass
+
+
+class AlignVisionModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        image_size=32,
+        num_channels=3,
+        kernel_sizes=[3, 3, 5],
+        in_channels=[32, 16, 24],
+        out_channels=[16, 24, 30],
+        hidden_dim=64,
+        strides=[1, 1, 2],
+        num_block_repeats=[1, 1, 2],
+        expand_ratios=[1, 6, 6],
+        is_training=True,
+        hidden_act="gelu",
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.image_size = image_size
+        self.num_channels = num_channels
+        self.kernel_sizes = kernel_sizes
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.hidden_dim = hidden_dim
+        self.strides = strides
+        self.num_block_repeats = num_block_repeats
+        self.expand_ratios = expand_ratios
+        self.is_training = is_training
+        self.hidden_act = hidden_act
+
+    def prepare_config_and_inputs(self):
+        pixel_values = floats_tensor([self.batch_size, self.num_channels, self.image_size, self.image_size])
+        config = self.get_config()
+
+        return config, pixel_values
+
+    def get_config(self):
+        return AlignVisionConfig(
+            num_channels=self.num_channels,
+            kernel_sizes=self.kernel_sizes,
+            in_channels=self.in_channels,
+            out_channels=self.out_channels,
+            hidden_dim=self.hidden_dim,
+            strides=self.strides,
+            num_block_repeats=self.num_block_repeats,
+            expand_ratios=self.expand_ratios,
+            hidden_act=self.hidden_act,
+        )
+
+    def create_and_check_model(self, config, pixel_values):
+        model = AlignVisionModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            result = model(pixel_values)
+
+        patch_size = self.image_size // 4
+        self.parent.assertEqual(
+            result.last_hidden_state.shape, (self.batch_size, config.hidden_dim, patch_size, patch_size)
+        )
+        self.parent.assertEqual(result.pooler_output.shape, (self.batch_size, config.hidden_dim))
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, pixel_values = config_and_inputs
+        inputs_dict = {"pixel_values": pixel_values}
+        return config, inputs_dict
+
+
+@require_torch
+class AlignVisionModelTest(ModelTesterMixin, unittest.TestCase):
+    """
+    Here we also overwrite some of the tests of test_modeling_common.py, as ALIGN does not use input_ids, inputs_embeds,
+    attention_mask and seq_length.
+    """
+
+    all_model_classes = (AlignVisionModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_head_masking = False
+    has_attentions = False
+
+    def setUp(self):
+        self.model_tester = AlignVisionModelTester(self)
+        self.config_tester = ConfigTester(
+            self, config_class=AlignVisionConfig, has_text_modality=False, hidden_size=37
+        )
+
+    def test_config(self):
+        self.create_and_test_config_common_properties()
+        self.config_tester.create_and_test_config_to_json_string()
+        self.config_tester.create_and_test_config_to_json_file()
+        self.config_tester.create_and_test_config_from_and_save_pretrained()
+        self.config_tester.create_and_test_config_with_num_labels()
+        self.config_tester.check_config_can_be_init_without_params()
+        self.config_tester.check_config_arguments_init()
+
+    def create_and_test_config_common_properties(self):
+        return
+
+    @unittest.skip(reason="AlignVisionModel does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="AlignVisionModel does not support input and output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.forward)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = ["pixel_values"]
+            self.assertListEqual(arg_names[:1], expected_arg_names)
+
+    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_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+            model.to(torch_device)
+            model.eval()
+
+            with torch.no_grad():
+                outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+            num_blocks = sum(config.num_block_repeats) * 4
+            self.assertEqual(len(hidden_states), num_blocks)
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [self.model_tester.image_size // 2, self.model_tester.image_size // 2],
+            )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    def test_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = AlignVisionModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class AlignTextModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=7,
+        is_training=True,
+        use_input_mask=True,
+        use_token_type_ids=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,
+        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.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.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)
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask
+
+    def get_config(self):
+        return AlignTextConfig(
+            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):
+        model = AlignTextModel(config=config)
+        model.to(torch_device)
+        model.eval()
+        with torch.no_grad():
+            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 prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        (
+            config,
+            input_ids,
+            token_type_ids,
+            input_mask,
+        ) = 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 AlignTextModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (AlignTextModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_pruning = False
+    test_head_masking = False
+
+    def setUp(self):
+        self.model_tester = AlignTextModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=AlignTextConfig, 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_training(self):
+        pass
+
+    def test_training_gradient_checkpointing(self):
+        pass
+
+    @unittest.skip(reason="ALIGN does not use inputs_embeds")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="AlignTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_from_base(self):
+        pass
+
+    @unittest.skip(reason="AlignTextModel has no base class and is not available in MODEL_MAPPING")
+    def test_save_load_fast_init_to_base(self):
+        pass
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = AlignTextModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+class AlignModelTester:
+    def __init__(self, parent, text_kwargs=None, vision_kwargs=None, is_training=True):
+        if text_kwargs is None:
+            text_kwargs = {}
+        if vision_kwargs is None:
+            vision_kwargs = {}
+
+        self.parent = parent
+        self.text_model_tester = AlignTextModelTester(parent, **text_kwargs)
+        self.vision_model_tester = AlignVisionModelTester(parent, **vision_kwargs)
+        self.is_training = is_training
+
+    def prepare_config_and_inputs(self):
+        test_config, input_ids, token_type_ids, input_mask = self.text_model_tester.prepare_config_and_inputs()
+        vision_config, pixel_values = self.vision_model_tester.prepare_config_and_inputs()
+
+        config = self.get_config()
+
+        return config, input_ids, token_type_ids, input_mask, pixel_values
+
+    def get_config(self):
+        return AlignConfig.from_text_vision_configs(
+            self.text_model_tester.get_config(), self.vision_model_tester.get_config(), projection_dim=64
+        )
+
+    def create_and_check_model(self, config, input_ids, token_type_ids, attention_mask, pixel_values):
+        model = AlignModel(config).to(torch_device).eval()
+        with torch.no_grad():
+            result = model(input_ids, pixel_values, attention_mask, token_type_ids)
+        self.parent.assertEqual(
+            result.logits_per_image.shape, (self.vision_model_tester.batch_size, self.text_model_tester.batch_size)
+        )
+        self.parent.assertEqual(
+            result.logits_per_text.shape, (self.text_model_tester.batch_size, self.vision_model_tester.batch_size)
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config_and_inputs = self.prepare_config_and_inputs()
+        config, input_ids, token_type_ids, input_mask, pixel_values = config_and_inputs
+        inputs_dict = {
+            "input_ids": input_ids,
+            "token_type_ids": token_type_ids,
+            "attention_mask": input_mask,
+            "pixel_values": pixel_values,
+            "return_loss": True,
+        }
+        return config, inputs_dict
+
+
+@require_torch
+class AlignModelTest(ModelTesterMixin, unittest.TestCase):
+    all_model_classes = (AlignModel,) if is_torch_available() else ()
+    fx_compatible = False
+    test_head_masking = False
+    test_pruning = False
+    test_resize_embeddings = False
+    test_attention_outputs = False
+
+    def setUp(self):
+        self.model_tester = AlignModelTester(self)
+
+    def test_model(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model(*config_and_inputs)
+
+    @unittest.skip(reason="Hidden_states is tested in individual model tests")
+    def test_hidden_states_output(self):
+        pass
+
+    @unittest.skip(reason="Inputs_embeds is tested in individual model tests")
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip(reason="Retain_grad is tested in individual model tests")
+    def test_retain_grad_hidden_states_attentions(self):
+        pass
+
+    @unittest.skip(reason="AlignModel does not have input/output embeddings")
+    def test_model_common_attributes(self):
+        pass
+
+    # override as the `temperature` parameter initilization is different for ALIGN
+    def test_initialization(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        configs_no_init = _config_zero_init(config)
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            for name, param in model.named_parameters():
+                if param.requires_grad:
+                    # check if `temperature` is initilized as per the original implementation
+                    if name == "temperature":
+                        self.assertAlmostEqual(
+                            param.data.item(),
+                            1.0,
+                            delta=1e-3,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    elif name == "text_projection.weight":
+                        self.assertTrue(
+                            -1.0 <= ((param.data.mean() * 1e9).round() / 1e9).item() <= 1.0,
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+                    else:
+                        self.assertIn(
+                            ((param.data.mean() * 1e9).round() / 1e9).item(),
+                            [0.0, 1.0],
+                            msg=f"Parameter {name} of model {model_class} seems not properly initialized",
+                        )
+
+    def _create_and_check_torchscript(self, config, inputs_dict):
+        if not self.test_torchscript:
+            return
+
+        configs_no_init = _config_zero_init(config)  # To be sure we have no Nan
+        configs_no_init.torchscript = True
+        configs_no_init.return_dict = False
+        for model_class in self.all_model_classes:
+            model = model_class(config=configs_no_init)
+            model.to(torch_device)
+            model.eval()
+
+            try:
+                input_ids = inputs_dict["input_ids"]
+                pixel_values = inputs_dict["pixel_values"]  # ALIGN needs pixel_values
+                traced_model = torch.jit.trace(model, (input_ids, pixel_values))
+            except RuntimeError:
+                self.fail("Couldn't trace module.")
+
+            with tempfile.TemporaryDirectory() as tmp_dir_name:
+                pt_file_name = os.path.join(tmp_dir_name, "traced_model.pt")
+
+                try:
+                    torch.jit.save(traced_model, pt_file_name)
+                except Exception:
+                    self.fail("Couldn't save module.")
+
+                try:
+                    loaded_model = torch.jit.load(pt_file_name)
+                except Exception:
+                    self.fail("Couldn't load module.")
+
+            model.to(torch_device)
+            model.eval()
+
+            loaded_model.to(torch_device)
+            loaded_model.eval()
+
+            model_state_dict = model.state_dict()
+            loaded_model_state_dict = loaded_model.state_dict()
+
+            self.assertEqual(set(model_state_dict.keys()), set(loaded_model_state_dict.keys()))
+
+            models_equal = True
+            for layer_name, p1 in model_state_dict.items():
+                p2 = loaded_model_state_dict[layer_name]
+                if p1.data.ne(p2.data).sum() > 0:
+                    models_equal = False
+
+            self.assertTrue(models_equal)
+
+    def test_load_vision_text_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        # Save AlignConfig and check if we can load AlignVisionConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            vision_config = AlignVisionConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.vision_config.to_dict(), vision_config.to_dict())
+
+        # Save AlignConfig and check if we can load AlignTextConfig from it
+        with tempfile.TemporaryDirectory() as tmp_dir_name:
+            config.save_pretrained(tmp_dir_name)
+            text_config = AlignTextConfig.from_pretrained(tmp_dir_name)
+            self.assertDictEqual(config.text_config.to_dict(), text_config.to_dict())
+
+    @slow
+    def test_model_from_pretrained(self):
+        for model_name in ALIGN_PRETRAINED_MODEL_ARCHIVE_LIST[:1]:
+            model = AlignModel.from_pretrained(model_name)
+            self.assertIsNotNone(model)
+
+
+# We will verify our results on an image of cute cats
+def prepare_img():
+    url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+    im = Image.open(requests.get(url, stream=True).raw)
+    return im
+
+
+@require_vision
+@require_torch
+class AlignModelIntegrationTest(unittest.TestCase):
+    @slow
+    def test_inference(self):
+        model_name = "kakaobrain/align-base"
+        model = AlignModel.from_pretrained(model_name).to(torch_device)
+        processor = AlignProcessor.from_pretrained(model_name)
+
+        image = prepare_img()
+        texts = ["a photo of a cat", "a photo of a dog"]
+        inputs = processor(text=texts, images=image, return_tensors="pt").to(torch_device)
+
+        # forward pass
+        with torch.no_grad():
+            outputs = model(**inputs)
+
+        # verify the logits
+        self.assertEqual(
+            outputs.logits_per_image.shape,
+            torch.Size((inputs.pixel_values.shape[0], inputs.input_ids.shape[0])),
+        )
+        self.assertEqual(
+            outputs.logits_per_text.shape,
+            torch.Size((inputs.input_ids.shape[0], inputs.pixel_values.shape[0])),
+        )
+        expected_logits = torch.tensor([[9.7093, 3.4679]], device=torch_device)
+        self.assertTrue(torch.allclose(outputs.logits_per_image, expected_logits, atol=1e-3))
--- a/tests/models/align/test_processor_align.py
+++ b/tests/models/align/test_processor_align.py
@ -0,0 +1,207 @@
+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+import os
+import shutil
+import tempfile
+import unittest
+
+import numpy as np
+import pytest
+
+from transformers import BertTokenizer, BertTokenizerFast
+from transformers.models.bert.tokenization_bert import VOCAB_FILES_NAMES
+from transformers.testing_utils import require_vision
+from transformers.utils import IMAGE_PROCESSOR_NAME, is_vision_available
+
+
+if is_vision_available():
+    from PIL import Image
+
+    from transformers import AlignProcessor, EfficientNetImageProcessor
+
+
+@require_vision
+class AlignProcessorTest(unittest.TestCase):
+    def setUp(self):
+        self.tmpdirname = tempfile.mkdtemp()
+
+        vocab_tokens = [
+            "[UNK]",
+            "[CLS]",
+            "[SEP]",
+            "[PAD]",
+            "[MASK]",
+            "want",
+            "##want",
+            "##ed",
+            "wa",
+            "un",
+            "runn",
+            "##ing",
+            ",",
+            "low",
+            "lowest",
+        ]
+        self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
+        with open(self.vocab_file, "w", encoding="utf-8") as vocab_writer:
+            vocab_writer.write("".join([x + "\n" for x in vocab_tokens]))
+
+        image_processor_map = {
+            "do_resize": True,
+            "size": 20,
+            "do_center_crop": True,
+            "crop_size": 18,
+            "do_normalize": True,
+            "image_mean": [0.48145466, 0.4578275, 0.40821073],
+            "image_std": [0.26862954, 0.26130258, 0.27577711],
+        }
+        self.image_processor_file = os.path.join(self.tmpdirname, IMAGE_PROCESSOR_NAME)
+        with open(self.image_processor_file, "w", encoding="utf-8") as fp:
+            json.dump(image_processor_map, fp)
+
+    def get_tokenizer(self, **kwargs):
+        return BertTokenizer.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_rust_tokenizer(self, **kwargs):
+        return BertTokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
+
+    def get_image_processor(self, **kwargs):
+        return EfficientNetImageProcessor.from_pretrained(self.tmpdirname, **kwargs)
+
+    def tearDown(self):
+        shutil.rmtree(self.tmpdirname)
+
+    def prepare_image_inputs(self):
+        """This function prepares a list of PIL images, or a list of numpy arrays if one specifies numpify=True,
+        or a list of PyTorch tensors if one specifies torchify=True.
+        """
+
+        image_inputs = [np.random.randint(255, size=(3, 30, 400), dtype=np.uint8)]
+        image_inputs = [Image.fromarray(np.moveaxis(x, 0, -1)) for x in image_inputs]
+        return image_inputs
+
+    def test_save_load_pretrained_default(self):
+        tokenizer_slow = self.get_tokenizer()
+        tokenizer_fast = self.get_rust_tokenizer()
+        image_processor = self.get_image_processor()
+
+        processor_slow = AlignProcessor(tokenizer=tokenizer_slow, image_processor=image_processor)
+        processor_slow.save_pretrained(self.tmpdirname)
+        processor_slow = AlignProcessor.from_pretrained(self.tmpdirname, use_fast=False)
+
+        processor_fast = AlignProcessor(tokenizer=tokenizer_fast, image_processor=image_processor)
+        processor_fast.save_pretrained(self.tmpdirname)
+        processor_fast = AlignProcessor.from_pretrained(self.tmpdirname)
+
+        self.assertEqual(processor_slow.tokenizer.get_vocab(), tokenizer_slow.get_vocab())
+        self.assertEqual(processor_fast.tokenizer.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertEqual(tokenizer_slow.get_vocab(), tokenizer_fast.get_vocab())
+        self.assertIsInstance(processor_slow.tokenizer, BertTokenizer)
+        self.assertIsInstance(processor_fast.tokenizer, BertTokenizerFast)
+
+        self.assertEqual(processor_slow.image_processor.to_json_string(), image_processor.to_json_string())
+        self.assertEqual(processor_fast.image_processor.to_json_string(), image_processor.to_json_string())
+        self.assertIsInstance(processor_slow.image_processor, EfficientNetImageProcessor)
+        self.assertIsInstance(processor_fast.image_processor, EfficientNetImageProcessor)
+
+    def test_save_load_pretrained_additional_features(self):
+        processor = AlignProcessor(tokenizer=self.get_tokenizer(), image_processor=self.get_image_processor())
+        processor.save_pretrained(self.tmpdirname)
+
+        tokenizer_add_kwargs = self.get_tokenizer(bos_token="(BOS)", eos_token="(EOS)")
+        image_processor_add_kwargs = self.get_image_processor(do_normalize=False, padding_value=1.0)
+
+        processor = AlignProcessor.from_pretrained(
+            self.tmpdirname, bos_token="(BOS)", eos_token="(EOS)", do_normalize=False, padding_value=1.0
+        )
+
+        self.assertEqual(processor.tokenizer.get_vocab(), tokenizer_add_kwargs.get_vocab())
+        self.assertIsInstance(processor.tokenizer, BertTokenizerFast)
+
+        self.assertEqual(processor.image_processor.to_json_string(), image_processor_add_kwargs.to_json_string())
+        self.assertIsInstance(processor.image_processor, EfficientNetImageProcessor)
+
+    def test_image_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = AlignProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        image_input = self.prepare_image_inputs()
+
+        input_image_proc = image_processor(image_input, return_tensors="np")
+        input_processor = processor(images=image_input, return_tensors="np")
+
+        for key in input_image_proc.keys():
+            self.assertAlmostEqual(input_image_proc[key].sum(), input_processor[key].sum(), delta=1e-2)
+
+    def test_tokenizer(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = AlignProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+
+        encoded_processor = processor(text=input_str)
+
+        encoded_tok = tokenizer(input_str, padding="max_length", max_length=64)
+
+        for key in encoded_tok.keys():
+            self.assertListEqual(encoded_tok[key], encoded_processor[key])
+
+    def test_processor(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = AlignProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), ["input_ids", "token_type_ids", "attention_mask", "pixel_values"])
+
+        # test if it raises when no input is passed
+        with pytest.raises(ValueError):
+            processor()
+
+    def test_tokenizer_decode(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = AlignProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        predicted_ids = [[1, 4, 5, 8, 1, 0, 8], [3, 4, 3, 1, 1, 8, 9]]
+
+        decoded_processor = processor.batch_decode(predicted_ids)
+        decoded_tok = tokenizer.batch_decode(predicted_ids)
+
+        self.assertListEqual(decoded_tok, decoded_processor)
+
+    def test_model_input_names(self):
+        image_processor = self.get_image_processor()
+        tokenizer = self.get_tokenizer()
+
+        processor = AlignProcessor(tokenizer=tokenizer, image_processor=image_processor)
+
+        input_str = "lower newer"
+        image_input = self.prepare_image_inputs()
+
+        inputs = processor(text=input_str, images=image_input)
+
+        self.assertListEqual(list(inputs.keys()), processor.model_input_names)
--- a/utils/check_repo.py
+++ b/utils/check_repo.py
@ -182,6 +182,8 @@ TEST_FILES_WITH_NO_COMMON_TESTS = [
 # should **not** be the rule.
 IGNORE_NON_AUTO_CONFIGURED = PRIVATE_MODELS.copy() + [
    # models to ignore for model xxx mapping
+    "AlignTextModel",
+    "AlignVisionModel",
    "ClapTextModel",
    "ClapTextModelWithProjection",
    "ClapAudioModel",