Add `Kosmos-2` model (#24709)

* Add KOSMOS-2 model

* update

* update

* update

* address review comment - 001

* address review comment - 002

* address review comment - 003

* style

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* fix

* address review comment - 004

* address review comment - 005

* address review comment - 006

* address review comment - 007

* address review comment - 008

* address review comment - 009

* address review comment - 010

* address review comment - 011

* update readme

* fix

* fix

* fix

* [skip ci] fix

* revert the change in _decode

* fix docstring

* fix docstring

* Update docs/source/en/model_doc/kosmos-2.md
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* no more Kosmos2Tokenizer

* style

* remove "returned when being computed by the model"

* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* UTM5 Atten

* fix attn mask

* use present_key_value_states instead of next_decoder_cache

* style

* conversion scripts

* conversion scripts

* conversion scripts

* Add _reorder_cache

* fix doctest and copies

* rename 1

* rename 2

* rename 3

* make fixup

* fix table

* fix docstring

* rename 4

* change repo_id

* remove tip

* update md file

* make style

* update md file

* put docs/source/en/model_doc/kosmos-2.md to slow

* update conversion script

* Use CLIPImageProcessor in Kosmos2Processor

* Remove Kosmos2ImageProcessor

* Remove to_dict in Kosmos2Config

* Remove files

* fix import

* Update conversion

* normalized=False

* Not using hardcoded values like <image>

* elt --> element

* Apply suggestion

* Not using hardcoded values like </image>

* No assert

* No nested functions

* Fix md file

* copy

* update doc

* fix docstring

* fix name

* Remove _add_remove_spaces_around_tag_tokens

* Remove dummy docstring of _preprocess_single_example

* Use `BatchEncoding`

* temp

* temp

* temp

* Update

* Update

* Make Kosmos2ProcessorTest a bit pretty

* Update gradient checkpointing

* Fix gradient checkpointing test

* Remove one liner remove_special_fields

* Simplify conversion script

* fix add_eos_token

* update readme

* update tests

* Change to microsoft/kosmos-2-patch14-224

* style

* Fix doc

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

README.md

@@ -386,6 +386,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.

README_es.md

@@ -361,6 +361,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.

README_hd.md

@@ -335,6 +335,7 @@ conda install -c huggingface transformers
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce से) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. द्वाराअनुसंधान पत्र [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) के साथ जारी किया गया
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (माइक्रोसॉफ्ट रिसर्च एशिया से) साथ देने वाला पेपर [लेआउटएलएमवी3: यूनिफाइड टेक्स्ट और इमेज मास्किंग के साथ दस्तावेज़ एआई के लिए पूर्व-प्रशिक्षण](https://arxiv.org/abs/2204.08387) युपन हुआंग, टेंगचाओ लव, लेई कुई, युटोंग लू, फुरु वेई द्वारा पोस्ट किया गया।

README_ja.md

@@ -395,6 +395,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce から) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. から公開された研究論文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI から) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever から公開された研究論文: [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf)
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia から) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou から公開された研究論文: [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318)
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (Microsoft Research Asia から) Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou から公開された研究論文: [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740)
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (Microsoft Research Asia から) Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei から公開された研究論文: [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387)

README_ko.md

@@ -310,6 +310,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce 에서 제공)은 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.의 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)논문과 함께 발표했습니다.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI 에서) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever 의 [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) 논문과 함께 발표했습니다.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia 에서) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 의 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 논문과 함께 발표했습니다.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (Microsoft Research Asia 에서) Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou 의 [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) 논문과 함께 발표했습니다.
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (Microsoft Research Asia 에서) Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei 의 [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) 논문과 함께 발표했습니다.

README_zh-hans.md

@@ -334,6 +334,7 @@ conda install -c huggingface transformers
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (来自 Salesforce) 伴随论文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) 由 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi 发布。
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 由 Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 发布。
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) 由 Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou 发布。
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) 由 Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei 发布。

README_zh-hant.md

@@ -346,6 +346,7 @@ conda install -c huggingface transformers
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
 1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
+1. **[KOSMOS-2](https://huggingface.co/docs/transformers/main/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
 1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.
 1. **[LayoutLMv2](https://huggingface.co/docs/transformers/model_doc/layoutlmv2)** (from Microsoft Research Asia) released with the paper [LayoutLMv2: Multi-modal Pre-training for Visually-Rich Document Understanding](https://arxiv.org/abs/2012.14740) by Yang Xu, Yiheng Xu, Tengchao Lv, Lei Cui, Furu Wei, Guoxin Wang, Yijuan Lu, Dinei Florencio, Cha Zhang, Wanxiang Che, Min Zhang, Lidong Zhou.
 1. **[LayoutLMv3](https://huggingface.co/docs/transformers/model_doc/layoutlmv3)** (from Microsoft Research Asia) released with the paper [LayoutLMv3: Pre-training for Document AI with Unified Text and Image Masking](https://arxiv.org/abs/2204.08387) by Yupan Huang, Tengchao Lv, Lei Cui, Yutong Lu, Furu Wei.

docs/source/en/_toctree.yml

@@ -368,6 +368,8 @@
         title: I-BERT
       - local: model_doc/jukebox
         title: Jukebox
+      - local: model_doc/kosmos-2
+        title: KOSMOS-2
       - local: model_doc/led
         title: LED
       - local: model_doc/llama

docs/source/en/index.md

@@ -158,6 +158,7 @@ Flax), PyTorch, and/or TensorFlow.
 |                      [Informer](model_doc/informer)                      |       ✅        |         ❌         |      ❌      |
 |                  [InstructBLIP](model_doc/instructblip)                  |       ✅        |         ❌         |      ❌      |
 |                       [Jukebox](model_doc/jukebox)                       |       ✅        |         ❌         |      ❌      |
+|                      [KOSMOS-2](model_doc/kosmos-2)                      |       ✅        |         ❌         |      ❌      |
 |                      [LayoutLM](model_doc/layoutlm)                      |       ✅        |         ✅         |      ❌      |
 |                    [LayoutLMv2](model_doc/layoutlmv2)                    |       ✅        |         ❌         |      ❌      |
 |                    [LayoutLMv3](model_doc/layoutlmv3)                    |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/kosmos-2.md

+<!--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.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# KOSMOS-2
+
+## Overview
+
+The KOSMOS-2 model was proposed in [Kosmos-2: Grounding Multimodal Large Language Models to the World]
+(https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei
+
+KOSMOS-2 is a Transformer-based causal language model and is trained using the next-word prediction task on a web-scale
+dataset of grounded image-text pairs [GRIT](https://huggingface.co/datasets/zzliang/GRIT). The spatial coordinates of
+the bounding boxes in the dataset are converted to a sequence of location tokens, which are appended to their respective
+entity text spans (for example, `a snowman` followed by `<patch_index_0044><patch_index_0863>`). The data format is
+similar to “hyperlinks” that connect the object regions in an image to their text span in the corresponding caption.
+
+The abstract from the paper is the following:
+
+*We introduce Kosmos-2, a Multimodal Large Language Model (MLLM), enabling new capabilities of perceiving object descriptions (e.g., bounding boxes) and grounding text to the visual world. Specifically, we represent refer expressions as links in Markdown, i.e., ``[text span](bounding boxes)'', where object descriptions are sequences of location tokens. Together with multimodal corpora, we construct large-scale data of grounded image-text pairs (called GrIT) to train the model. In addition to the existing capabilities of MLLMs (e.g., perceiving general modalities, following instructions, and performing in-context learning), Kosmos-2 integrates the grounding capability into downstream applications. We evaluate Kosmos-2 on a wide range of tasks, including (i) multimodal grounding, such as referring expression comprehension, and phrase grounding, (ii) multimodal referring, such as referring expression generation, (iii) perception-language tasks, and (iv) language understanding and generation. This work lays out the foundation for the development of Embodiment AI and sheds light on the big convergence of language, multimodal perception, action, and world modeling, which is a key step toward artificial general intelligence. Code and pretrained models are available at https://aka.ms/kosmos-2.*
+
+## Example
+
+```python
+>>> from PIL import Image
+>>> import requests
+>>> from transformers import AutoProcessor, Kosmos2ForConditionalGeneration
+
+>>> model = Kosmos2ForConditionalGeneration.from_pretrained("microsoft/kosmos-2-patch14-224")
+>>> processor = AutoProcessor.from_pretrained("microsoft/kosmos-2-patch14-224")
+
+>>> url = "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/snowman.jpg"
+>>> image = Image.open(requests.get(url, stream=True).raw)
+
+>>> prompt = "<grounding> An image of"
+
+>>> inputs = processor(text=prompt, images=image, return_tensors="pt")
+
+>>> generated_ids = model.generate(
+...     pixel_values=inputs["pixel_values"],
+...     input_ids=inputs["input_ids"],
+...     attention_mask=inputs["attention_mask"],
+...     image_embeds=None,
+...     image_embeds_position_mask=inputs["image_embeds_position_mask"],
+...     use_cache=True,
+...     max_new_tokens=64,
+... )
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+>>> processed_text = processor.post_process_generation(generated_text, cleanup_and_extract=False)
+>>> processed_text
+'<grounding> An image of<phrase> a snowman</phrase><object><patch_index_0044><patch_index_0863></object> warming himself by<phrase> a fire</phrase><object><patch_index_0005><patch_index_0911></object>.'
+
+>>> caption, entities = processor.post_process_generation(generated_text)
+>>> caption
+'An image of a snowman warming himself by a fire.'
+
+>>> entities
+[('a snowman', (12, 21), [(0.390625, 0.046875, 0.984375, 0.828125)]), ('a fire', (41, 47), [(0.171875, 0.015625, 0.484375, 0.890625)])]
+```
+
+This model was contributed by [Yih-Dar SHIEH](https://huggingface.co/ydshieh). The original code can be found [here](https://github.com/microsoft/unilm/tree/master/kosmos-2).
+
+## Kosmos2Config
+
+[[autodoc]] Kosmos2Config
+
+## Kosmos2ImageProcessor
+
+## Kosmos2Processor
+
+[[autodoc]] Kosmos2Processor
+    - __call__
+
+## Kosmos2Model
+
+[[autodoc]] Kosmos2Model
+    - forward
+
+## Kosmos2ForConditionalGeneration
+
+[[autodoc]] Kosmos2ForConditionalGeneration
+    - forward

src/transformers/__init__.py

@@ -388,6 +388,11 @@ _import_structure = {
         "JukeboxTokenizer",
         "JukeboxVQVAEConfig",
     ],
+    "models.kosmos2": [
+        "KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Kosmos2Config",
+        "Kosmos2Processor",
+    ],
     "models.layoutlm": ["LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP", "LayoutLMConfig", "LayoutLMTokenizer"],
     "models.layoutlmv2": [
         "LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP",
@@ -2051,6 +2056,14 @@ else:
             "JukeboxVQVAE",
         ]
     )
+    _import_structure["models.kosmos2"].extend(
+        [
+            "KOSMOS2_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "Kosmos2ForConditionalGeneration",
+            "Kosmos2Model",
+            "Kosmos2PreTrainedModel",
+        ]
+    )
     _import_structure["models.layoutlm"].extend(
         [
             "LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -4561,6 +4574,11 @@ if TYPE_CHECKING:
         JukeboxTokenizer,
         JukeboxVQVAEConfig,
     )
+    from .models.kosmos2 import (
+        KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        Kosmos2Config,
+        Kosmos2Processor,
+    )
     from .models.layoutlm import LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP, LayoutLMConfig, LayoutLMTokenizer
     from .models.layoutlmv2 import (
         LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -5986,6 +6004,12 @@ if TYPE_CHECKING:
             JukeboxPrior,
             JukeboxVQVAE,
         )
+        from .models.kosmos2 import (
+            KOSMOS2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Kosmos2ForConditionalGeneration,
+            Kosmos2Model,
+            Kosmos2PreTrainedModel,
+        )
         from .models.layoutlm import (
             LAYOUTLM_PRETRAINED_MODEL_ARCHIVE_LIST,
             LayoutLMForMaskedLM,

src/transformers/models/__init__.py

@@ -109,6 +109,7 @@ from . import (
     informer,
     instructblip,
     jukebox,
+    kosmos2,
     layoutlm,
     layoutlmv2,
     layoutlmv3,

src/transformers/models/auto/configuration_auto.py

@@ -117,6 +117,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("informer", "InformerConfig"),
         ("instructblip", "InstructBlipConfig"),
         ("jukebox", "JukeboxConfig"),
+        ("kosmos-2", "Kosmos2Config"),
         ("layoutlm", "LayoutLMConfig"),
         ("layoutlmv2", "LayoutLMv2Config"),
         ("layoutlmv3", "LayoutLMv3Config"),
@@ -331,6 +332,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("informer", "INFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("instructblip", "INSTRUCTBLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("jukebox", "JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("kosmos-2", "KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("layoutlm", "LAYOUTLM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("layoutlmv2", "LAYOUTLMV2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("layoutlmv3", "LAYOUTLMV3_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -546,6 +548,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("informer", "Informer"),
         ("instructblip", "InstructBLIP"),
         ("jukebox", "Jukebox"),
+        ("kosmos-2", "KOSMOS-2"),
         ("layoutlm", "LayoutLM"),
         ("layoutlmv2", "LayoutLMv2"),
         ("layoutlmv3", "LayoutLMv3"),
@@ -709,6 +712,7 @@ SPECIAL_MODEL_TYPE_TO_MODULE_NAME = OrderedDict(
         ("data2vec-text", "data2vec"),
         ("data2vec-vision", "data2vec"),
         ("donut-swin", "donut"),
+        ("kosmos-2", "kosmos2"),
         ("maskformer-swin", "maskformer"),
         ("xclip", "x_clip"),
     ]

src/transformers/models/auto/modeling_auto.py

@@ -112,6 +112,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("imagegpt", "ImageGPTModel"),
         ("informer", "InformerModel"),
         ("jukebox", "JukeboxModel"),
+        ("kosmos-2", "Kosmos2Model"),
         ("layoutlm", "LayoutLMModel"),
         ("layoutlmv2", "LayoutLMv2Model"),
         ("layoutlmv3", "LayoutLMv3Model"),
@@ -570,6 +571,7 @@ MODEL_FOR_VISION_2_SEQ_MAPPING_NAMES = OrderedDict(
         ("blip-2", "Blip2ForConditionalGeneration"),
         ("git", "GitForCausalLM"),
         ("instructblip", "InstructBlipForConditionalGeneration"),
+        ("kosmos-2", "Kosmos2ForConditionalGeneration"),
         ("pix2struct", "Pix2StructForConditionalGeneration"),
         ("vision-encoder-decoder", "VisionEncoderDecoderModel"),
     ]

src/transformers/models/auto/processing_auto.py

@@ -60,6 +60,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("hubert", "Wav2Vec2Processor"),
         ("idefics", "IdeficsProcessor"),
         ("instructblip", "InstructBlipProcessor"),
+        ("kosmos-2", "Kosmos2Processor"),
         ("layoutlmv2", "LayoutLMv2Processor"),
         ("layoutlmv3", "LayoutLMv3Processor"),
         ("markuplm", "MarkupLMProcessor"),

src/transformers/models/auto/tokenization_auto.py

@@ -181,6 +181,13 @@ else:
             ("idefics", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
             ("instructblip", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
             ("jukebox", ("JukeboxTokenizer", None)),
+            (
+                "kosmos-2",
+                (
+                    "XLMRobertaTokenizer" if is_sentencepiece_available() else None,
+                    "XLMRobertaTokenizerFast" if is_tokenizers_available() else None,
+                ),
+            ),
             ("layoutlm", ("LayoutLMTokenizer", "LayoutLMTokenizerFast" if is_tokenizers_available() else None)),
             ("layoutlmv2", ("LayoutLMv2Tokenizer", "LayoutLMv2TokenizerFast" if is_tokenizers_available() else None)),
             ("layoutlmv3", ("LayoutLMv3Tokenizer", "LayoutLMv3TokenizerFast" if is_tokenizers_available() else None)),

src/transformers/models/kosmos2/__init__.py

+# coding=utf-8
+# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+    is_vision_available,
+)
+
+
+_import_structure = {
+    "configuration_kosmos2": ["KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP", "Kosmos2Config"],
+    "processing_kosmos2": ["Kosmos2Processor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_kosmos2"] = [
+        "KOSMOS2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Kosmos2ForConditionalGeneration",
+        "Kosmos2Model",
+        "Kosmos2PreTrainedModel",
+    ]
+
+
+if TYPE_CHECKING:
+    from .configuration_kosmos2 import KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP, Kosmos2Config
+    from .processing_kosmos2 import Kosmos2Processor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_kosmos2 import (
+            KOSMOS2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Kosmos2ForConditionalGeneration,
+            Kosmos2Model,
+            Kosmos2PreTrainedModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure)

src/transformers/models/kosmos2/configuration_kosmos2.py

+# coding=utf-8
+# Copyright 2023 Microsoft Research and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" KOSMOS-2 model configuration"""
+
+import os
+from typing import Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+KOSMOS2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "microsoft/kosmos-2-patch14-224": (
+        "https://huggingface.co/microsoft/kosmos-2-patch14-224/resolve/main/config.json"
+    ),
+    # See all KOSMOS-2 models at https://huggingface.co/models?filter=kosmos-2
+}
+
+
+class Kosmos2TextConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Kosmos2TextModel`]. It is used to instantiate a
+    KOSMOS-2 text decoder 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 decoder of the KOSMOS-2
+    [microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 65037):
+            Vocabulary size of the Kosmos2 model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`Kosmos2Model`].
+        max_position_embeddings (`int`, *optional*, defaults to 2048):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        embed_dim (`int`, *optional*, defaults to 2048):
+            Dimensionality of the layers and the pooler layer.
+        layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        ffn_dim (`int`, *optional*, defaults to 8192):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the Transformer encoder.
+        attention_heads (`int`, *optional*, defaults to 32):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        activation_function (`str` or `function`, *optional*, defaults to `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"silu"` and `"gelu_new"` are supported.
+        dropout (`float`, *optional*, defaults to 0.1):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        attention_dropout (`float`, *optional*, defaults to 0.1):
+            The dropout ratio for the attention probabilities.
+        activation_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for activations inside the fully connected layer.
+        layerdrop (`float`, *optional*, defaults to 0.0):
+            The LayerDrop probability for the decoder. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556)
+            for more details.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        init_std (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        scale_embedding (`bool`, *optional*, defaults to `True`):
+            Scale embeddings by diving by sqrt(embed_dim).
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+    ```"""
+    model_type = "kosmos_2_text_model"
+    keys_to_ignore_at_inference = ["past_key_values"]
+    attribute_map = {
+        "num_attention_heads": "attention_heads",
+        "hidden_size": "embed_dim",
+        "num_hidden_layers": "layers",
+    }
+
+    def __init__(
+        self,
+        vocab_size=65037,
+        max_position_embeddings=2048,
+        embed_dim=2048,
+        layers=24,
+        ffn_dim=8192,
+        attention_heads=32,
+        activation_function="gelu",
+        dropout=0.1,
+        attention_dropout=0.1,
+        activation_dropout=0.0,
+        layerdrop=0.0,
+        layer_norm_eps=1e-5,
+        init_std=0.02,
+        scale_embedding=True,
+        use_cache=True,
+        pad_token_id=1,
+        bos_token_id=0,
+        eos_token_id=2,
+        **kwargs,
+    ):
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )
+
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.embed_dim = embed_dim
+        self.layers = layers
+        self.ffn_dim = ffn_dim
+        self.attention_heads = attention_heads
+        self.activation_function = activation_function
+        self.dropout = dropout
+        self.attention_dropout = attention_dropout
+        self.activation_dropout = activation_dropout
+        self.layerdrop = layerdrop
+        self.layer_norm_eps = layer_norm_eps
+        self.init_std = init_std
+        self.scale_embedding = scale_embedding
+        self.use_cache = use_cache
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the text config dict if we are loading from Kosmos2Config
+        if config_dict.get("model_type") == "kosmos-2":
+            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 Kosmos2VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Kosmos2VisionModel`]. It is used to instantiate a
+    KOSMOS-2 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 KOSMOS-2
+    [microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        hidden_size (`int`, *optional*, defaults to 1024):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 4096):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 24):
+            Number of hidden layers in the Transformer encoder.
+        num_attention_heads (`int`, *optional*, defaults to 16):
+            Number of attention heads for each attention layer in the Transformer encoder.
+        num_channels (`int`, *optional*, defaults to 3):
+            The number of input channels.
+        image_size (`int`, *optional*, defaults to 224):
+            The size (resolution) of each image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The size (resolution) of each patch.
+        hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
+        layer_norm_eps (`float`, *optional*, defaults to 1e-5):
+            The epsilon used by the layer normalization layers.
+        attention_dropout (`float`, *optional*, defaults to 0.0):
+            The dropout ratio for the attention probabilities.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        initializer_factor (`float`, *optional*, defaults to 1):
+            A factor for initializing all weight matrices (should be kept to 1, used internally for initialization
+            testing).
+    ```"""
+
+    model_type = "kosmos_2_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=1024,
+        intermediate_size=4096,
+        num_hidden_layers=24,
+        num_attention_heads=16,
+        num_channels=3,
+        image_size=224,
+        patch_size=14,
+        hidden_act="quick_gelu",
+        layer_norm_eps=1e-5,
+        attention_dropout=0.0,
+        initializer_range=0.02,
+        initializer_factor=1.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.initializer_range = initializer_range
+        self.initializer_factor = initializer_factor
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    @classmethod
+    def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+        cls._set_token_in_kwargs(kwargs)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the vision config dict if we are loading from Kosmos2Config
+        if config_dict.get("model_type") == "kosmos-2":
+            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 Kosmos2Config(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Kosmos2Model`]. It is used to instantiate a
+    KOSMOS-2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
+    with the defaults will yield a similar configuration to that of the KOSMOS-2
+    [microsoft/kosmos-2-patch14-224](https://huggingface.co/microsoft/kosmos-2-patch14-224) architecture.
+
+    Args:
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`Kosmos2TextConfig`].
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`Kosmos2VisionConfig`].
+        latent_query_num (`int`, *optional*, defaults to 64):
+            The number of latent query tokens that represent the image features used in the text decoder component.
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import Kosmos2Config, Kosmos2Model
+
+    >>> # Initializing a Kosmos-2 kosmos-2-patch14-224 style configuration
+    >>> configuration = Kosmos2Config()
+
+    >>> # Initializing a model (with random weights) from the kosmos-2-patch14-224 style configuration
+    >>> model = Kosmos2Model(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "kosmos-2"
+    is_composition = True
+
+    def __init__(
+        self,
+        text_config=None,
+        vision_config=None,
+        latent_query_num=64,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        if text_config is None:
+            text_config = {}
+            logger.info("`text_config` is `None`. Initializing the `Kosmos2TextConfig` with default values.")
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("`vision_config` is `None`. Initializing the `Kosmos2VisionConfig` with default values.")
+
+        self.text_config = Kosmos2TextConfig(**text_config)
+        self.vision_config = Kosmos2VisionConfig(**vision_config)
+
+        self.latent_query_num = latent_query_num

src/transformers/models/kosmos2/convert_kosmos2_original_pytorch_checkpoint_to_pytorch.py

+import argparse
+
+from fairseq.checkpoint_utils import load_checkpoint_to_cpu
+
+from transformers import Kosmos2Config, Kosmos2ForConditionalGeneration
+
+
+KEYS_TO_MODIFY_MAPPING = {
+    "gpt_model.decoder.output_projection": "text_model.lm_head",
+    "gpt_model.decoder": "text_model.model",
+    "img_connector": "image_to_text_projection",
+    "img_model.visual.class_embedding": "vision_model.model.embeddings.class_embedding",
+    "img_model.visual.positional_embedding": "vision_model.model.embeddings.position_embedding.weight",
+    "img_model.visual.conv1": "vision_model.model.embeddings.patch_embedding",
+    "img_model.visual": "vision_model.model",
+    "ln_pre": "pre_layrnorm",
+    "ln_post": "post_layernorm",
+    "transformer.resblocks": "encoder.layers",
+    "ts_attn": "self_attn",
+    "ln_1": "layer_norm1",
+    "ln_2": "layer_norm2",
+    "c_fc": "fc1",
+    "c_proj": "fc2",
+}
+
+
+KEYS_TO_IGNORE = [
+    # this buffer in the original code is only used to send weights to the desired device
+    "gpt_model.decoder.embed_positions._float_tensor",
+    # this weight is never used in the forward in the original KOSMOS-2)
+    "gpt_model.decoder.self_attn_sope.scale",
+]
+
+
+def rename_key(key):
+    for key_to_modify, new_key in KEYS_TO_MODIFY_MAPPING.items():
+        if key_to_modify in key:
+            key = key.replace(key_to_modify, new_key)
+
+    return key
+
+
+def convert_kosmos2_checkpoint_to_pytorch(checkpoint_path, pytorch_dump_folder_path):
+    state = load_checkpoint_to_cpu(checkpoint_path)
+    state_dict = state["model"]
+    state_dict_keys = list(state_dict.keys())
+
+    config = Kosmos2Config()
+    # This is necessary to match the results given by the original demo
+    config.text_config.no_repeat_ngram_size = 3
+    model = Kosmos2ForConditionalGeneration(config)
+
+    # convert (by renaming keys)
+    converted_state_dict = {}
+    for key in state_dict_keys:
+        if key in KEYS_TO_IGNORE:
+            continue
+        renamed_key = rename_key(key)
+        converted_state_dict[renamed_key] = state_dict[key]
+
+    # check weight loading
+    model.load_state_dict(converted_state_dict, strict=True)
+    # save the result
+    model.save_pretrained(pytorch_dump_folder_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+    parser.add_argument(
+        "--kosmos2_checkpoint_path", default=None, type=str, required=True, help="Path the official PyTorch dump."
+    )
+    parser.add_argument(
+        "--pytorch_dump_folder_path", default=None, type=str, required=True, help="Path to the output PyTorch model."
+    )
+    args = parser.parse_args()
+    convert_kosmos2_checkpoint_to_pytorch(args.kosmos2_checkpoint_path, args.pytorch_dump_folder_path)