Add BLIP-2 (#21441)

* First draft

* More improvements

* More improvements

* Improve conversion script

* Convert all weights

* Make forward pass work

* Make logits match

* More improvements

* More improvements

* More improvements

* Use get_input_embeddings

* Improve some more

* Improve model tests

* Improve model tests

* More improvements

* Fix processor

* Update files

* Update prepare_inputs_for_generation

* More improvements

* Fix copies

* More fixes

* Make fixup

* More improvements

* Add support for seq2seq language model

* More improvements

* Fix test

* More improvements

* Improve conversion script

* Remove some todo's

* Fix README's

* Improve conversion script

* Fix generation

* Fix style and remove Blip2Model

* Fix model outputs

* More improvements

* Set eos_token_id in config

* Fix quality

* Small improvements

* Add processor tests

* More improvements

* Apply suggestions

* Apply suggestions

* Add integration test

* Update image URL

* Add integration test

* Fix model_type

* Update style

* Improve docs

* Add doc tests

* Fix copies

* Remove tests which are passing

* Improve some more

* Add tests for seq2seq language models

* Minor fix

* Convert more checkpoints

* finalize CI

* Fix blip and blip2 processors

* add `accelerate` support for `blip2`

* clean up

* make style

* Update conversion script

* Update conversion script some more

* Update organization

* revert toc file

* add blip-2 to toc file

* Some more improvements

* Fix docstring

* Improve docs

---------
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Co-authored-by: younesbelkada <younesbelkada@gmail.com>

README.md

@@ -287,6 +287,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_es.md

@@ -280,6 +280,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_hd.md

@@ -252,6 +252,7 @@ conda install -c huggingface transformers
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (फेसबुक से) साथ में कागज [एक ओपन-डोमेन चैटबॉट बनाने की विधि](https://arxiv.org /abs/2004.13637) स्टीफन रोलर, एमिली दीनन, नमन गोयल, दा जू, मैरी विलियमसन, यिनहान लियू, जिंग जू, मायल ओट, कर्ट शस्टर, एरिक एम। स्मिथ, वाई-लैन बॉरो, जेसन वेस्टन द्वारा।
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (फेसबुक से) साथ में पेपर [एक ओपन-डोमेन चैटबॉट बनाने की रेसिपी](https://arxiv .org/abs/2004.13637) स्टीफन रोलर, एमिली दीनन, नमन गोयल, दा जू, मैरी विलियमसन, यिनहान लियू, जिंग जू, मायल ओट, कर्ट शस्टर, एरिक एम स्मिथ, वाई-लैन बॉरो, जेसन वेस्टन द्वारा।
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (Salesforce से) Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi. द्वाराअनुसंधान पत्र [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) के साथ जारी किया गया
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (एलेक्सा से) कागज के साथ [बीईआरटी के लिए ऑप्टिमल सबआर्किटेक्चर एक्सट्रैक्शन](https://arxiv.org/abs/ 2010.10499) एड्रियन डी विंटर और डैनियल जे पेरी द्वारा।
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (हरबिन इंस्टिट्यूट ऑफ़ टेक्नोलॉजी/माइक्रोसॉफ्ट रिसर्च एशिया/इंटेल लैब्स से) कागज के साथ [ब्रिजटॉवर: विजन-लैंग्वेज रिप्रेजेंटेशन लर्निंग में एनकोडर्स के बीच ब्रिज बनाना](<https://arxiv.org/abs/2206.08657>) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_ja.md

@@ -314,6 +314,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (Facebook から) Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston から公開された研究論文: [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637)
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (Facebook から) Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston から公開された研究論文: [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637)
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (Salesforce から) Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi から公開された研究論文: [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086)
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (Salesforce から) Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi. から公開された研究論文 [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597)
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (BigScience workshop から) [BigScience Workshop](https://bigscience.huggingface.co/) から公開されました.
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa から) Adrian de Wynter and Daniel J. Perry から公開された研究論文: [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499)
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (Harbin Institute of Technology/Microsoft Research Asia/Intel Labs から) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_ko.md

@@ -229,6 +229,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (Salesforce 에서 제공)은 Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.의 [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597)논문과 함께 발표했습니다.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa 에서) Adrian de Wynter and Daniel J. Perry 의 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 논문과 함께 발표했습니다.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_zh-hans.md

@@ -253,6 +253,7 @@ conda install -c huggingface transformers
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (来自 Salesforce) 伴随论文 [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) 由 Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi 发布。
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (来自 Salesforce) 伴随论文 [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) 由 Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi 发布。
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (来自 Alexa) 伴随论文 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 由 Adrian de Wynter and Daniel J. Perry 发布。
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

README_zh-hant.md

@@ -265,6 +265,7 @@ conda install -c huggingface transformers
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLIP](https://huggingface.co/docs/transformers/model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](https://huggingface.co/docs/transformers/main/model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/main/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.

docs/source/en/_toctree.yml

@@ -522,6 +522,8 @@
         title: AltCLIP
       - local: model_doc/blip
         title: BLIP
+      - local: model_doc/blip-2
+        title: BLIP-2
       - local: model_doc/bridgetower
         title: BridgeTower
       - local: model_doc/chinese_clip

docs/source/en/index.mdx

@@ -66,6 +66,7 @@ The documentation is organized into five sections:
 1. **[Blenderbot](model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BLIP](model_doc/blip)** (from Salesforce) released with the paper [BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation](https://arxiv.org/abs/2201.12086) by Junnan Li, Dongxu Li, Caiming Xiong, Steven Hoi.
+1. **[BLIP-2](model_doc/blip-2)** (from Salesforce) released with the paper [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi.
 1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
@@ -253,6 +254,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          Blenderbot           |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |        BlenderbotSmall        |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             BLIP              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|            BLIP-2             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             BLOOM             |       ❌       |       ✅       |       ✅        |         ❌         |      ❌      |
 |          BridgeTower          |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |           CamemBERT           |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/blip-2.mdx

+<!--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.
+-->
+
+# BLIP-2
+
+## Overview
+
+The BLIP-2 model was proposed in [BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models](https://arxiv.org/abs/2301.12597) by
+Junnan Li, Dongxu Li, Silvio Savarese, Steven Hoi. BLIP-2 leverages frozen pre-trained image encoders and large language models (LLMs) by training a lightweight, 12-layer Transformer
+encoder in between them, achieving state-of-the-art performance on various vision-language tasks. Most notably, BLIP-2 improves upon [Flamingo](https://arxiv.org/abs/2204.14198), an 80 billion parameter model, by 8.7%
+on zero-shot VQAv2 with 54x fewer trainable parameters. 
+
+The abstract from the paper is the following:
+
+*The cost of vision-and-language pre-training has become increasingly prohibitive due to end-to-end training of large-scale models. This paper proposes BLIP-2, a generic and efficient pre-training strategy that bootstraps vision-language pre-training from off-the-shelf frozen pre-trained image encoders and frozen large language models. BLIP-2 bridges the modality gap with a lightweight Querying Transformer, which is pre-trained in two stages. The first stage bootstraps vision-language representation learning from a frozen image encoder. The second stage bootstraps vision-to-language generative learning from a frozen language model. BLIP-2 achieves state-of-the-art performance on various vision-language tasks, despite having significantly fewer trainable parameters than existing methods. For example, our model outperforms Flamingo80B by 8.7% on zero-shot VQAv2 with 54x fewer trainable parameters. We also demonstrate the model's emerging capabilities of zero-shot image-to-text generation that can follow natural language instructions.*
+
+Tips:
+
+- BLIP-2 can be used for conditional text generation given an image and an optional text prompt. At inference time, it's recommended to use the [`generate`] method.
+- One can use [`Blip2Processor`] to prepare images for the model, and decode the predicted tokens ID's back to text.
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/blip2_architecture.jpg"
+alt="drawing" width="600"/> 
+
+<small> BLIP-2 architecture. Taken from the <a href="https://arxiv.org/abs/2301.12597">original paper.</a> </small>
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/salesforce/LAVIS/tree/5ee63d688ba4cebff63acee04adaef2dee9af207).
+
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with BLIP-2.
+
+- Demo notebooks for BLIP-2 for image captioning, visual question answering (VQA) and chat-like conversations can be found [here](https://github.com/NielsRogge/Transformers-Tutorials/tree/master/BLIP-2).
+
+If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+## Blip2Config
+
+[[autodoc]] Blip2Config
+    - from_vision_qformer_text_configs
+
+## Blip2VisionConfig
+
+[[autodoc]] Blip2VisionConfig
+
+## Blip2QFormerConfig
+
+[[autodoc]] Blip2QFormerConfig
+
+## Blip2Processor
+
+[[autodoc]] Blip2Processor
+
+## Blip2VisionModel
+
+[[autodoc]] Blip2VisionModel
+    - forward
+
+## Blip2QFormerModel
+
+[[autodoc]] Blip2QFormerModel
+    - forward
+
+## Blip2ForConditionalGeneration
+
+[[autodoc]] Blip2ForConditionalGeneration
+    - forward
+    - generate
\ No newline at end of file

src/transformers/__init__.py

@@ -180,6 +180,13 @@ _import_structure = {
         "BlipTextConfig",
         "BlipVisionConfig",
     ],
+    "models.blip_2": [
+        "BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Blip2Config",
+        "Blip2Processor",
+        "Blip2QFormerConfig",
+        "Blip2VisionConfig",
+    ],
     "models.bloom": ["BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP", "BloomConfig"],
     "models.bort": [],
     "models.bridgetower": [
@@ -1150,6 +1157,15 @@ else:
             "BlipVisionModel",
         ]
     )
+    _import_structure["models.blip_2"].extend(
+        [
+            "BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "Blip2ForConditionalGeneration",
+            "Blip2PreTrainedModel",
+            "Blip2QFormerModel",
+            "Blip2VisionModel",
+        ]
+    )
     _import_structure["models.bloom"].extend(
         [
             "BLOOM_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -3645,6 +3661,13 @@ if TYPE_CHECKING:
         BlipTextConfig,
         BlipVisionConfig,
     )
+    from .models.blip_2 import (
+        BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        Blip2Config,
+        Blip2Processor,
+        Blip2QFormerConfig,
+        Blip2VisionConfig,
+    )
     from .models.bloom import BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP, BloomConfig
     from .models.bridgetower import (
         BRIDGETOWER_PRETRAINED_CONFIG_ARCHIVE_MAP,
@@ -4498,6 +4521,13 @@ if TYPE_CHECKING:
             BlipTextModel,
             BlipVisionModel,
         )
+        from .models.blip_2 import (
+            BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Blip2ForConditionalGeneration,
+            Blip2PreTrainedModel,
+            Blip2QFormerModel,
+            Blip2VisionModel,
+        )
         from .models.bloom import (
             BLOOM_PRETRAINED_MODEL_ARCHIVE_LIST,
             BloomForCausalLM,

src/transformers/models/__init__.py

@@ -32,6 +32,7 @@ from . import (
     blenderbot,
     blenderbot_small,
     blip,
+    blip_2,
     bloom,
     bort,
     bridgetower,

src/transformers/models/auto/configuration_auto.py

@@ -43,6 +43,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("blenderbot", "BlenderbotConfig"),
         ("blenderbot-small", "BlenderbotSmallConfig"),
         ("blip", "BlipConfig"),
+        ("blip-2", "Blip2Config"),
         ("bloom", "BloomConfig"),
         ("bridgetower", "BridgeTowerConfig"),
         ("camembert", "CamembertConfig"),
@@ -212,6 +213,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("blenderbot", "BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("blenderbot-small", "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("blip", "BLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("blip-2", "BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bloom", "BLOOM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bridgetower", "BRIDGETOWER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("camembert", "CAMEMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -369,6 +371,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("blenderbot", "Blenderbot"),
         ("blenderbot-small", "BlenderbotSmall"),
         ("blip", "BLIP"),
+        ("blip-2", "BLIP-2"),
         ("bloom", "BLOOM"),
         ("bort", "BORT"),
         ("bridgetower", "BridgeTower"),

src/transformers/models/auto/image_processing_auto.py

@@ -40,6 +40,7 @@ IMAGE_PROCESSOR_MAPPING_NAMES = OrderedDict(
         ("beit", "BeitImageProcessor"),
         ("bit", "BitImageProcessor"),
         ("blip", "BlipImageProcessor"),
+        ("blip-2", "BlipImageProcessor"),
         ("bridgetower", "BridgeTowerImageProcessor"),
         ("chinese_clip", "ChineseCLIPImageProcessor"),
         ("clip", "CLIPImageProcessor"),

src/transformers/models/auto/processing_auto.py

@@ -42,7 +42,8 @@ logger = logging.get_logger(__name__)
 PROCESSOR_MAPPING_NAMES = OrderedDict(
     [
         ("altclip", "AltCLIPProcessor"),
-        ("blip", "BLIPProcessor"),
+        ("blip", "BlipProcessor"),
+        ("blip-2", "Blip2Processor"),
         ("bridgetower", "BridgeTowerProcessor"),
         ("chinese_clip", "ChineseCLIPProcessor"),
         ("clip", "CLIPProcessor"),

src/transformers/models/auto/tokenization_auto.py

@@ -78,6 +78,7 @@ else:
             ("blenderbot", ("BlenderbotTokenizer", "BlenderbotTokenizerFast")),
             ("blenderbot-small", ("BlenderbotSmallTokenizer", None)),
             ("blip", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
+            ("blip-2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
             ("bloom", (None, "BloomTokenizerFast" if is_tokenizers_available() else None)),
             ("bridgetower", ("RobertaTokenizer", "RobertaTokenizerFast" if is_tokenizers_available() else None)),
             ("byt5", ("ByT5Tokenizer", None)),

src/transformers/models/blip_2/__init__.py

+# flake8: noqa
+# There's no way to ignore "F401 '...' imported but unused" warnings in this
+# module, but to preserve other warnings. So, don't check this module at all.
+
+# Copyright 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_blip_2": [
+        "BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "Blip2Config",
+        "Blip2QFormerConfig",
+        "Blip2VisionConfig",
+    ],
+    "processing_blip_2": ["Blip2Processor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_blip_2"] = [
+        "BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "Blip2QFormerModel",
+        "Blip2PreTrainedModel",
+        "Blip2ForConditionalGeneration",
+        "Blip2VisionModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_blip_2 import (
+        BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        Blip2Config,
+        Blip2QFormerConfig,
+        Blip2VisionConfig,
+    )
+    from .processing_blip_2 import Blip2Processor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_blip_2 import (
+            BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST,
+            Blip2ForConditionalGeneration,
+            Blip2PreTrainedModel,
+            Blip2QFormerModel,
+            Blip2VisionModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/blip_2/configuration_blip_2.py

+# 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.
+""" BLIP-2 model configuration"""
+
+import copy
+import os
+from typing import Union
+
+from transformers.models.auto.modeling_auto import MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import logging
+from ..auto import CONFIG_MAPPING
+
+
+logger = logging.get_logger(__name__)
+
+BLIP_2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "salesforce/blip2-opt-2.7b": "https://huggingface.co/salesforce/blip2-opt-2.7b/resolve/main/config.json",
+}
+
+
+class Blip2VisionConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Blip2VisionModel`]. It is used to instantiate a
+    BLIP-2 vision encoder according to the specified arguments, defining the model architecture. Instantiating a
+    configuration defaults will yield a similar configuration to that of the BLIP-2
+    [Salesforce/blip2-opt-2.7b](https://huggingface.co/Salesforce/blip2-opt-2.7b) 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 1408):
+            Dimensionality of the encoder layers and the pooler layer.
+        intermediate_size (`int`, *optional*, defaults to 6144):
+            Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
+        num_hidden_layers (`int`, *optional*, defaults to 39):
+            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.
+        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 `"gelu"`):
+            The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
+            `"relu"`, `"selu"` and `"gelu_new"` ``"gelu"` are supported. layer_norm_eps (`float`, *optional*, defaults
+            to 1e-5): The epsilon used by the layer normalization layers.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probabilitiy for all fully connected layers in the embeddings, encoder, and pooler.
+        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).
+        qkv_bias (`bool`, *optional*, defaults to `True`):
+            Whether to add a bias to the queries and values in the self-attention layers.
+
+    Example:
+
+    ```python
+    >>> from transformers import Blip2VisionConfig, Blip2VisionModel
+
+    >>> # Initializing a Blip2VisionConfig with Salesforce/blip2-opt-2.7b style configuration
+    >>> configuration = Blip2VisionConfig()
+
+    >>> # Initializing a Blip2VisionModel (with random weights) from the Salesforce/blip2-opt-2.7b style configuration
+    >>> model = Blip2VisionModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+
+    model_type = "blip_2_vision_model"
+
+    def __init__(
+        self,
+        hidden_size=1408,
+        intermediate_size=6144,
+        projection_dim=512,
+        num_hidden_layers=39,
+        num_attention_heads=16,
+        num_channels=3,
+        image_size=224,
+        patch_size=14,
+        hidden_act="gelu",
+        layer_norm_eps=0.00001,
+        dropout=0.0,
+        attention_dropout=0.0,
+        initializer_range=1e-10,
+        initializer_factor=1.0,
+        qkv_bias=True,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.projection_dim = projection_dim
+        self.dropout = dropout
+        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
+        self.qkv_bias = qkv_bias
+
+    @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 Blip2Config
+        if config_dict.get("model_type") == "blip-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 Blip2QFormerConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Blip2QFormerModel`]. It is used to instantiate a
+    BLIP-2 Querying Transformer (Q-Former) 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 BLIP-2
+    [Salesforce/blip2-opt-2.7b](https://huggingface.co/Salesforce/blip2-opt-2.7b) architecture. Configuration objects
+    inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the documentation from
+    [`PretrainedConfig`] for more information.
+
+    Note that [`Blip2QFormerModel`] is very similar to [`BertLMHeadModel`] with interleaved cross-attention.
+
+    Args:
+        vocab_size (`int`, *optional*, defaults to 30522):
+            Vocabulary size of the Q-Former model. Defines the number of different tokens that can be represented by
+            the `inputs_ids` passed when calling the model.
+        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).
+        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).
+        classifier_dropout (`float`, *optional*):
+            The dropout ratio for the classification head.
+        cross_attention_frequency (`int`, *optional*, defaults to 2):
+            The frequency of adding cross-attention to the Transformer layers.
+        encoder_hidden_size (`int`, *optional*, defaults to 1408):
+            The hidden size of the hidden states for cross-attention.
+
+    Examples:
+
+    ```python
+    >>> from transformers import Blip2QFormerConfig, Blip2QFormerModel
+
+    >>> # Initializing a BLIP-2 Salesforce/blip2-opt-2.7b style configuration
+    >>> configuration = Blip2QFormerConfig()
+
+    >>> # Initializing a model (with random weights) from the Salesforce/blip2-opt-2.7b style configuration
+    >>> model = Blip2QFormerModel(configuration)
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```"""
+    model_type = "blip_2_qformer"
+
+    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,
+        initializer_range=0.02,
+        layer_norm_eps=1e-12,
+        pad_token_id=0,
+        position_embedding_type="absolute",
+        classifier_dropout=None,
+        cross_attention_frequency=2,
+        encoder_hidden_size=1408,
+        **kwargs,
+    ):
+        super().__init__(pad_token_id=pad_token_id, **kwargs)
+
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.hidden_act = hidden_act
+        self.intermediate_size = intermediate_size
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.max_position_embeddings = max_position_embeddings
+        self.initializer_range = initializer_range
+        self.layer_norm_eps = layer_norm_eps
+        self.position_embedding_type = position_embedding_type
+        self.classifier_dropout = classifier_dropout
+        self.cross_attention_frequency = cross_attention_frequency
+        self.encoder_hidden_size = encoder_hidden_size
+
+    @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 qformer config dict if we are loading from Blip2Config
+        if config_dict.get("model_type") == "blip-2":
+            config_dict = config_dict["qformer_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 Blip2Config(PretrainedConfig):
+    r"""
+    [`Blip2Config`] is the configuration class to store the configuration of a [`Blip2ForConditionalGeneration`]. It is
+    used to instantiate a BLIP-2 model according to the specified arguments, defining the vision model, Q-Former model
+    and language model configs. Instantiating a configuration with the defaults will yield a similar configuration to
+    that of the BLIP-2 [Salesforce/blip2-opt-2.7b](https://huggingface.co/Salesforce/blip2-opt-2.7b) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        vision_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`Blip2VisionConfig`].
+        qformer_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize [`Blip2QFormerConfig`].
+        text_config (`dict`, *optional*):
+            Dictionary of configuration options used to initialize any [`PretrainedConfig`].
+        num_query_tokens (`int`, *optional*, defaults to 32):
+            The number of query tokens passed through the Transformer.
+
+        kwargs (*optional*):
+            Dictionary of keyword arguments.
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     Blip2VisionConfig,
+    ...     Blip2QFormerConfig,
+    ...     OPTConfig,
+    ...     Blip2Config,
+    ...     Blip2ForConditionalGeneration,
+    ... )
+
+    >>> # Initializing a Blip2Config with Salesforce/blip2-opt-2.7b style configuration
+    >>> configuration = Blip2Config()
+
+    >>> # Initializing a Blip2ForConditionalGeneration (with random weights) from the Salesforce/blip2-opt-2.7b style configuration
+    >>> model = Blip2ForConditionalGeneration(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+
+    >>> # We can also initialize a Blip2Config from a Blip2VisionConfig, Blip2QFormerConfig and any PretrainedConfig
+
+    >>> # Initializing BLIP-2 vision, BLIP-2 Q-Former and language model configurations
+    >>> vision_config = Blip2VisionConfig()
+    >>> qformer_config = Blip2QFormerConfig()
+    >>> text_config = OPTConfig()
+
+    >>> config = Blip2Config.from_text_vision_configs(vision_config, qformer_config, text_config)
+    ```"""
+
+    model_type = "blip-2"
+    is_composition = True
+
+    def __init__(self, vision_config=None, qformer_config=None, text_config=None, num_query_tokens=32, **kwargs):
+        super().__init__(**kwargs)
+
+        if vision_config is None:
+            vision_config = {}
+            logger.info("vision_config is None. initializing the Blip2VisionConfig with default values.")
+
+        if qformer_config is None:
+            qformer_config = {}
+            logger.info("qformer_config is None. Initializing the Blip2QFormerConfig with default values.")
+
+        if text_config is None:
+            text_config = {}
+            logger.info("text_config is None. Initializing the text config with default values (`OPTConfig`).")
+
+        self.vision_config = Blip2VisionConfig(**vision_config)
+        self.qformer_config = Blip2QFormerConfig(**qformer_config)
+        text_model_type = text_config["model_type"] if "model_type" in text_config else "opt"
+        self.text_config = CONFIG_MAPPING[text_model_type](**text_config)
+
+        self.num_query_tokens = num_query_tokens
+        self.qformer_config.encoder_hidden_size = self.vision_config.hidden_size
+        self.use_decoder_only_language_model = self.text_config.model_type in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES
+        self.initializer_factor = 1.0
+        self.initializer_range = 0.02
+
+    @classmethod
+    def from_vision_qformer_text_configs(
+        cls,
+        vision_config: Blip2VisionConfig,
+        qformer_config: Blip2QFormerConfig,
+        text_config: PretrainedConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`Blip2Config`] (or a derived class) from a BLIP-2 vision model, Q-Former and language model
+        configurations.
+
+        Returns:
+            [`Blip2Config`]: An instance of a configuration object
+        """
+
+        return cls(
+            vision_config=vision_config.to_dict(),
+            qformer_config=qformer_config.to_dict(),
+            text_config=text_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["vision_config"] = self.vision_config.to_dict()
+        output["qformer_config"] = self.qformer_config.to_dict()
+        output["text_config"] = self.text_config.to_dict()
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/blip_2/convert_blip_2_original_to_pytorch.py

+# 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.
+"""
+Convert BLIP-2 checkpoints from the original repository.
+
+URL: https://github.com/salesforce/LAVIS/tree/main/projects/blip2
+"""
+
+import argparse
+
+import requests
+import torch
+
+# pip3 install salesforce-lavis
+# I'm actually installing a slightly modified version: pip3 install git+https://github.com/nielsrogge/LAVIS.git@fix_lavis
+from lavis.models import load_model_and_preprocess
+from PIL import Image
+
+from transformers import (
+    AutoTokenizer,
+    Blip2Config,
+    Blip2ForConditionalGeneration,
+    Blip2Processor,
+    Blip2VisionConfig,
+    BlipImageProcessor,
+    OPTConfig,
+    T5Config,
+)
+from transformers.utils.constants import OPENAI_CLIP_MEAN, OPENAI_CLIP_STD
+
+
+def load_demo_image():
+    url = "https://storage.googleapis.com/sfr-vision-language-research/LAVIS/assets/merlion.png"
+    image = Image.open(requests.get(url, stream=True).raw).convert("RGB")
+
+    return image
+
+
+# here we list all keys to be renamed (original name on the left, our name on the right)
+def create_rename_keys(config):
+    rename_keys = []
+    # fmt: off
+
+    # vision encoder
+    rename_keys.append(("visual_encoder.cls_token", "vision_model.embeddings.class_embedding"))
+    rename_keys.append(("visual_encoder.pos_embed", "vision_model.embeddings.position_embedding"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.weight", "vision_model.embeddings.patch_embedding.weight"))
+    rename_keys.append(("visual_encoder.patch_embed.proj.bias", "vision_model.embeddings.patch_embedding.bias"))
+    rename_keys.append(("ln_vision.weight", "vision_model.post_layernorm.weight"))
+    rename_keys.append(("ln_vision.bias", "vision_model.post_layernorm.bias"))
+
+    for i in range(config.vision_config.num_hidden_layers):
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.weight", f"vision_model.encoder.layers.{i}.layer_norm1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm1.bias", f"vision_model.encoder.layers.{i}.layer_norm1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.weight", f"vision_model.encoder.layers.{i}.layer_norm2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.norm2.bias", f"vision_model.encoder.layers.{i}.layer_norm2.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.qkv.weight", f"vision_model.encoder.layers.{i}.self_attn.qkv.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.weight", f"vision_model.encoder.layers.{i}.self_attn.projection.weight",))
+        rename_keys.append((f"visual_encoder.blocks.{i}.attn.proj.bias", f"vision_model.encoder.layers.{i}.self_attn.projection.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.weight", f"vision_model.encoder.layers.{i}.mlp.fc1.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc1.bias", f"vision_model.encoder.layers.{i}.mlp.fc1.bias"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.weight", f"vision_model.encoder.layers.{i}.mlp.fc2.weight"))
+        rename_keys.append((f"visual_encoder.blocks.{i}.mlp.fc2.bias", f"vision_model.encoder.layers.{i}.mlp.fc2.bias"))
+
+    # QFormer
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.weight", "qformer.layernorm.weight"))
+    rename_keys.append(("Qformer.bert.embeddings.LayerNorm.bias", "qformer.layernorm.bias"))
+
+    # fmt: on
+    return rename_keys
+
+
+def rename_key(dct, old, new):
+    val = dct.pop(old)
+    dct[new] = val
+
+
+def read_in_q_v_bias(state_dict, config):
+    for i in range(config.vision_config.num_hidden_layers):
+        # read in original q and v biases
+        q_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.q_bias")
+        v_bias = state_dict.pop(f"visual_encoder.blocks.{i}.attn.v_bias")
+
+        # next, set bias in the state dict
+        qkv_bias = torch.cat((q_bias, torch.zeros_like(v_bias, requires_grad=False), v_bias))
+        state_dict[f"vision_model.encoder.layers.{i}.self_attn.qkv.bias"] = qkv_bias
+
+
+def get_blip2_config(model_name, eos_token_id):
+    image_size = 364 if "coco" in model_name else 224
+    vision_config = Blip2VisionConfig(image_size=image_size).to_dict()
+
+    # make sure the models have proper bos_token_id and eos_token_id set (important for generation)
+    # seems like flan-T5 models don't have bos_token_id properly set?
+    if "opt-2.7b" in model_name:
+        text_config = OPTConfig.from_pretrained("facebook/opt-2.7b", eos_token_id=eos_token_id).to_dict()
+    elif "opt-6.7b" in model_name:
+        text_config = OPTConfig.from_pretrained("facebook/opt-6.7b", eos_token_id=eos_token_id).to_dict()
+    elif "t5-xl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+    elif "t5-xxl" in model_name:
+        text_config = T5Config.from_pretrained("google/flan-t5-xxl", dense_act_fn="gelu", bos_token_id=1).to_dict()
+
+    config = Blip2Config(vision_config=vision_config, text_config=text_config)
+
+    return config, image_size
+
+
+@torch.no_grad()
+def convert_blip2_checkpoint(model_name, pytorch_dump_folder_path=None, push_to_hub=False):
+    """
+    Copy/paste/tweak model's weights to Transformers design.
+    """
+    tokenizer = (
+        AutoTokenizer.from_pretrained("facebook/opt-2.7b")
+        if "opt" in model_name
+        else AutoTokenizer.from_pretrained("google/flan-t5-xl")
+    )
+    eos_token_id = tokenizer("\n", add_special_tokens=False).input_ids[0]
+    config, image_size = get_blip2_config(model_name, eos_token_id=eos_token_id)
+
+    hf_model = Blip2ForConditionalGeneration(config).eval()
+
+    model_name_to_original = {
+        "blip2-opt-2.7b": ("blip2_opt", "pretrain_opt2.7b"),
+        "blip2-opt-6.7b": ("blip2_opt", "pretrain_opt6.7b"),
+        "blip2-opt-2.7b-coco": ("blip2_opt", "caption_coco_opt2.7b"),
+        "blip2-opt-6.7b-coco": ("blip2_opt", "caption_coco_opt6.7b"),
+        "blip2-flan-t5-xl": ("blip2_t5", "pretrain_flant5xl"),
+        "blip2-flan-t5-xl-coco": ("blip2_t5", "caption_coco_flant5xl"),
+        "blip2-flan-t5-xxl": ("blip2_t5", "pretrain_flant5xxl"),
+    }
+
+    name, type = model_name_to_original[model_name]
+
+    # load original model
+    print("Loading original model...")
+    device = "cuda" if torch.cuda.is_available() else "cpu"
+    original_model, vis_processors, _ = load_model_and_preprocess(
+        name=name, model_type=type, is_eval=True, device=device
+    )
+    original_model.eval()
+    print("Done!")
+
+    # update state dict keys
+    state_dict = original_model.state_dict()
+    rename_keys = create_rename_keys(config)
+    for src, dest in rename_keys:
+        rename_key(state_dict, src, dest)
+
+    # some keys can be renamed efficiently
+    for key, val in state_dict.copy().items():
+        val = state_dict.pop(key)
+        if key.startswith("Qformer.bert"):
+            key = key.replace("Qformer.bert", "qformer")
+        if "attention.self" in key:
+            key = key.replace("self", "attention")
+        if "opt_proj" in key:
+            key = key.replace("opt_proj", "language_projection")
+        if "t5_proj" in key:
+            key = key.replace("t5_proj", "language_projection")
+        if key.startswith("opt"):
+            key = key.replace("opt", "language")
+        if key.startswith("t5"):
+            key = key.replace("t5", "language")
+        state_dict[key] = val
+
+    # read in qv biases
+    read_in_q_v_bias(state_dict, config)
+
+    missing_keys, unexpected_keys = hf_model.load_state_dict(state_dict, strict=False)
+    assert len(missing_keys) == 0
+    assert unexpected_keys == ["qformer.embeddings.position_ids"]
+
+    image = load_demo_image()
+    original_pixel_values = vis_processors["eval"](image).unsqueeze(0).to(device)
+    input_ids = tokenizer(["\n"], return_tensors="pt").input_ids.to(device)
+
+    # create processor
+    image_processor = BlipImageProcessor(
+        size={"height": image_size, "width": image_size}, image_mean=OPENAI_CLIP_MEAN, image_std=OPENAI_CLIP_STD
+    )
+    processor = Blip2Processor(image_processor=image_processor, tokenizer=tokenizer)
+    pixel_values = processor(images=image, return_tensors="pt").pixel_values.to(device)
+
+    # make sure processor creates exact same pixel values
+    assert torch.allclose(pixel_values, original_pixel_values)
+
+    original_model.to(device)
+    hf_model.to(device)
+    with torch.no_grad():
+        if "opt" in model_name:
+            original_logits = original_model({"image": original_pixel_values, "text_input": [""]}).logits
+            logits = hf_model(original_pixel_values, input_ids).logits
+        else:
+            original_logits = original_model(
+                {"image": original_pixel_values, "text_input": ["\n"], "text_output": ["\n"]}
+            ).logits
+            labels = input_ids.masked_fill(input_ids == tokenizer.pad_token_id, -100)
+            logits = hf_model(original_pixel_values, input_ids, labels=labels).logits
+
+    assert original_logits.shape == logits.shape
+    print("First values of original logits:", original_logits[0, :3, :3])
+    print("First values of HF logits:", logits[0, :3, :3])
+
+    # assert values
+    if model_name == "blip2-flan-t5-xl":
+        expected_slice_logits = torch.tensor(
+            [[-41.5850, -4.4440, -8.9922], [-47.4322, -5.9143, -1.7340]], device=device
+        )
+        assert torch.allclose(logits[0, :3, :3], expected_slice_logits, atol=1e-4)
+    elif model_name == "blip2-flan-t5-xl-coco":
+        expected_slice_logits = torch.tensor(
+            [[-57.0109, -9.8967, -12.6280], [-68.6578, -12.7191, -10.5065]], device=device
+        )
+    else:
+        # cast to same type
+        target_dtype = logits.dtype
+        assert torch.allclose(original_logits.to(target_dtype), logits, atol=1e-2)
+    print("Looks ok!")
+
+    print("Generating a caption...")
+    prompt = ""
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(device)
+
+    original_outputs = original_model.generate({"image": original_pixel_values})
+    outputs = hf_model.generate(
+        original_pixel_values,
+        input_ids,
+        do_sample=False,
+        num_beams=5,
+        max_length=30,
+        min_length=1,
+        top_p=0.9,
+        repetition_penalty=1.0,
+        length_penalty=1.0,
+        temperature=1,
+    )
+    print("Original generation:", original_outputs)
+    prompt_length = input_ids.shape[1]
+    output_text = processor.batch_decode(outputs[:, prompt_length:], skip_special_tokens=True)
+    output_text = [text.strip() for text in output_text]
+    print("HF generation:", output_text)
+
+    if pytorch_dump_folder_path is not None:
+        processor.save_pretrained(pytorch_dump_folder_path)
+        hf_model.save_pretrained(pytorch_dump_folder_path)
+
+    if push_to_hub:
+        processor.push_to_hub(f"nielsr/{model_name}")
+        hf_model.push_to_hub(f"nielsr/{model_name}")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    choices = [
+        "blip2-opt-2.7b",
+        "blip2-opt-6.7b",
+        "blip2-opt-2.7b-coco",
+        "blip2-opt-6.7b-coco",
+        "blip2-flan-t5-xl",
+        "blip2-flan-t5-xl-coco",
+        "blip2-flan-t5-xxl",
+    ]
+    parser.add_argument(
+        "--model_name",
+        default="blip2-opt-2.7b",
+        choices=choices,
+        type=str,
+        help="Path to hf config.json of model to convert",
+    )
+    parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument(
+        "--push_to_hub",
+        action="store_true",
+        help="Whether to push the model and processor to the hub after converting",
+    )
+
+    args = parser.parse_args()
+
+    convert_blip2_checkpoint(args.model_name, args.pytorch_dump_folder_path, args.push_to_hub)

src/transformers/models/blip_2/modeling_blip_2.py

+# coding=utf-8
+# Copyright 2023 The Salesforce Authors and The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" PyTorch BLIP-2 model."""
+
+import math
+from dataclasses import dataclass
+from typing import Any, Optional, Tuple, Union
+
+import torch
+import torch.utils.checkpoint
+from torch import nn
+from torch.nn import CrossEntropyLoss
+
+from ...activations import ACT2FN
+from ...modeling_outputs import (
+    BaseModelOutput,
+    BaseModelOutputWithPastAndCrossAttentions,
+    BaseModelOutputWithPooling,
+    BaseModelOutputWithPoolingAndCrossAttentions,
+)
+from ...modeling_utils import PreTrainedModel
+from ...pytorch_utils import apply_chunking_to_forward, find_pruneable_heads_and_indices, prune_linear_layer
+from ...utils import (
+    ModelOutput,
+    add_start_docstrings,
+    add_start_docstrings_to_model_forward,
+    logging,
+    replace_return_docstrings,
+)
+from ..auto import AutoModelForCausalLM, AutoModelForSeq2SeqLM
+from .configuration_blip_2 import Blip2Config, Blip2QFormerConfig, Blip2VisionConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CHECKPOINT_FOR_DOC = "Salesforce/blip2-opt-2.7b"
+
+BLIP_2_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "Salesforce/blip2-opt-2.7b",
+    # See all BLIP-2 models at https://huggingface.co/models?filter=blip
+]
+
+
+@dataclass
+class Blip2ForConditionalGenerationModelOutput(ModelOutput):
+    """
+    Class defining the outputs of [`Blip2ForConditionalGeneration`].
+
+    Args:
+        loss (`torch.FloatTensor`, *optional*, returned when `labels` is provided, `torch.FloatTensor` of shape `(1,)`):
+            Language modeling loss from the language model.
+        logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+            Prediction scores of the language modeling head of the language model.
+        vision_outputs (`BaseModelOutputWithPooling`):
+            Outputs of the vision encoder.
+        qformer_outputs (`BaseModelOutputWithPoolingAndCrossAttentions`):
+            Outputs of the Q-Former (Querying Transformer).
+        language_model_outputs (`CausalLMOutputWithPast` or `Seq2SeqLMOutput`):
+            Outputs of the language model.
+    """
+
+    loss: Optional[Tuple[torch.FloatTensor]] = None
+    logits: Optional[Tuple[torch.FloatTensor]] = None
+    vision_outputs: Optional[torch.FloatTensor] = None
+    qformer_outputs: Optional[Tuple[torch.FloatTensor]] = None
+    language_model_outputs: Optional[Tuple[torch.FloatTensor]] = None
+
+    def to_tuple(self) -> Tuple[Any]:
+        return tuple(
+            self[k]
+            if k not in ["vision_outputs", "qformer_outputs", "language_model_outputs"]
+            else getattr(self, k).to_tuple()
+            for k in self.keys()
+        )
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipVisionEmbeddings with Blip->Blip2
+class Blip2VisionEmbeddings(nn.Module):
+    def __init__(self, config: Blip2VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = nn.Parameter(
+            torch.randn(1, 1, self.embed_dim),
+        )
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=3, out_channels=self.embed_dim, kernel_size=self.patch_size, stride=self.patch_size
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+
+        self.position_embedding = nn.Parameter(torch.randn(1, self.num_positions, self.embed_dim))
+
+    def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
+        batch_size = pixel_values.shape[0]
+        target_dtype = self.patch_embedding.weight.dtype
+        patch_embeds = self.patch_embedding(pixel_values)  # shape = [*, width, grid, grid]
+        patch_embeds = patch_embeds.flatten(2).transpose(1, 2)
+
+        class_embeds = self.class_embedding.expand(batch_size, 1, -1).to(target_dtype)
+        embeddings = torch.cat([class_embeds, patch_embeds], dim=1)
+        embeddings = embeddings + self.position_embedding[:, : embeddings.size(1), :].to(target_dtype)
+        return embeddings
+
+
+class Blip2Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = nn.Dropout(config.attention_dropout)
+
+        # small tweak here compared to CLIP, no bias here
+        self.qkv = nn.Linear(self.embed_dim, 3 * self.embed_dim, bias=False)
+
+        if config.qkv_bias:
+            q_bias = nn.Parameter(torch.zeros(self.embed_dim))
+            v_bias = nn.Parameter(torch.zeros(self.embed_dim))
+        else:
+            q_bias = None
+            v_bias = None
+
+        if q_bias is not None:
+            qkv_bias = torch.cat((q_bias, torch.zeros_like(v_bias, requires_grad=False), v_bias))
+            self.qkv.bias = nn.Parameter(qkv_bias)
+
+        self.projection = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
+        return tensor.view(bsz, seq_len, self.num_heads, self.head_dim).transpose(1, 2).contiguous()
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        head_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        """Input shape: Batch x Time x Channel"""
+
+        bsz, tgt_len, embed_dim = hidden_states.size()
+
+        mixed_qkv = self.qkv(hidden_states)
+
+        mixed_qkv = mixed_qkv.reshape(bsz, tgt_len, 3, self.num_heads, embed_dim // self.num_heads).permute(
+            2, 0, 3, 1, 4
+        )
+        query_states, key_states, value_states = (
+            mixed_qkv[0],
+            mixed_qkv[1],
+            mixed_qkv[2],
+        )
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_states, key_states.transpose(-1, -2))
+
+        attention_scores = attention_scores * self.scale
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.functional.softmax(attention_scores, dim=-1)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs = attention_probs * head_mask
+
+        context_layer = torch.matmul(attention_probs, value_states).permute(0, 2, 1, 3)
+
+        new_context_layer_shape = context_layer.size()[:-2] + (self.embed_dim,)
+        context_layer = context_layer.reshape(new_context_layer_shape)
+
+        output = self.projection(context_layer)
+
+        outputs = (output, attention_probs) if output_attentions else (output, None)
+
+        return outputs
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipMLP
+class Blip2MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipEncoderLayer with Blip->Blip2
+class Blip2EncoderLayer(nn.Module):
+    def __init__(self, config: Blip2Config):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = Blip2Attention(config)
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = Blip2MLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.FloatTensor]:
+        """
+        Args:
+            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+            attention_mask (`torch.FloatTensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+                `(config.encoder_attention_heads,)`.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+        """
+        residual = hidden_states
+
+        hidden_states = self.layer_norm1(hidden_states)
+        hidden_states, attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            head_mask=attention_mask,
+            output_attentions=output_attentions,
+        )
+        hidden_states = hidden_states + residual
+        residual = hidden_states
+        hidden_states = self.layer_norm2(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+
+        hidden_states = hidden_states + residual
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (attn_weights,)
+
+        return outputs
+
+
+class Blip2PreTrainedModel(PreTrainedModel):
+    """
+    An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
+    models.
+    """
+
+    config_class = Blip2Config
+    base_model_prefix = "blip"
+    supports_gradient_checkpointing = True
+    _keys_to_ignore_on_load_missing = [
+        r"position_ids",
+        r"language_model.encoder.embed_tokens.weight",
+        r"language_model.decoder.embed_tokens.weight",
+    ]
+    _no_split_modules = ["Blip2Attention", "T5Block", "OPTDecoderLayer"]
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        factor = self.config.initializer_range
+        if isinstance(module, nn.Conv2d) or isinstance(module, nn.Embedding) or isinstance(module, nn.Linear):
+            module.weight.data.normal_(mean=0.0, std=factor)
+            if hasattr(module, "bias") and module.bias is not None:
+                module.bias.data.zero_()
+
+        if isinstance(module, Blip2VisionEmbeddings):
+            if hasattr(self.config, "vision_config"):
+                factor = self.config.vision_config.initializer_range
+            nn.init.trunc_normal_(module.position_embedding, mean=0.0, std=factor)
+            nn.init.trunc_normal_(module.class_embedding, mean=0.0, std=factor)
+
+        elif isinstance(module, nn.LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
+        elif isinstance(module, nn.Linear) and module.bias is not None:
+            module.bias.data.zero_()
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if isinstance(module, Blip2Encoder):
+            module.gradient_checkpointing = value
+
+
+BLIP_2_START_DOCSTRING = r"""
+    This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+    Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+    and behavior.
+
+    Parameters:
+        config ([`Blip2Config`]): Model configuration class with all the parameters of the model.
+            Initializing with a config file does not load the weights associated with the model, only the
+            configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+BLIP_2_VISION_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`Blip2Processor`]. See [`Blip2Processor.__call__`] for
+            details.
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+BLIP_2_INPUTS_DOCSTRING = r"""
+    Args:
+        pixel_values (`torch.FloatTensor` of shape `(batch_size, num_channels, height, width)`):
+            Pixel values. Pixel values can be obtained using [`Blip2Processor`]. See [`Blip2Processor.__call__`] for
+            details.
+        
+        input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Indices of input sequence tokens in the vocabulary of the language model. Input tokens can optionally be
+            provided to serve as text prompt, which the language model can continue.
+
+            Indices can be obtained using [`Blip2Processor`]. See [`Blip2Processor.__call__`] for details.
+
+            [What are input IDs?](../glossary#input-ids)
+        attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+
+            [What are attention masks?](../glossary#attention-mask)
+
+        decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary of the language model. Only relevant in case an
+            encoder-decoder language model (like T5) is used.
+            
+            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids)
+        
+        decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            Only relevant in case an encoder-decoder language model (like T5) is used.
+
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipEncoder with Blip->Blip2
+class Blip2Encoder(nn.Module):
+    """
+    Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
+    [`Blip2EncoderLayer`].
+
+    Args:
+        config (`Blip2Config`):
+            The corresponding vision configuration for the `Blip2Encoder`.
+    """
+
+    def __init__(self, config: Blip2Config):
+        super().__init__()
+        self.config = config
+        self.layers = nn.ModuleList([Blip2EncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        inputs_embeds,
+        attention_mask: Optional[torch.Tensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutput]:
+        r"""
+        Args:
+            inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
+                Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
+                This is useful if you want more control over how to convert `input_ids` indices into associated vectors
+                than the model's internal embedding lookup matrix.
+            attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        hidden_states = inputs_embeds
+        for idx, encoder_layer in enumerate(self.layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            if self.gradient_checkpointing and self.training:
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, output_attentions)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(encoder_layer),
+                    hidden_states,
+                    attention_mask,
+                )
+            else:
+                layer_outputs = encoder_layer(
+                    hidden_states,
+                    attention_mask,
+                    output_attentions=output_attentions,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return BaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+# Copied from transformers.models.blip.modeling_blip.BlipVisionModel with Blip->Blip2, BLIP->BLIP_2
+class Blip2VisionModel(Blip2PreTrainedModel):
+    main_input_name = "pixel_values"
+    config_class = Blip2VisionConfig
+
+    def __init__(self, config: Blip2VisionConfig):
+        super().__init__(config)
+        self.config = config
+        embed_dim = config.hidden_size
+
+        self.embeddings = Blip2VisionEmbeddings(config)
+        self.encoder = Blip2Encoder(config)
+        self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.post_init()
+
+    @add_start_docstrings_to_model_forward(BLIP_2_VISION_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=BaseModelOutputWithPooling, config_class=Blip2VisionConfig)
+    def forward(
+        self,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, BaseModelOutputWithPooling]:
+        r"""
+        Returns:
+
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if pixel_values is None:
+            raise ValueError("You have to specify pixel_values")
+
+        hidden_states = self.embeddings(pixel_values)
+
+        encoder_outputs = self.encoder(
+            inputs_embeds=hidden_states,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+
+        last_hidden_state = encoder_outputs[0]
+        last_hidden_state = self.post_layernorm(last_hidden_state)
+
+        pooled_output = last_hidden_state[:, 0, :]
+        pooled_output = self.post_layernorm(pooled_output)
+
+        if not return_dict:
+            return (last_hidden_state, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPooling(
+            last_hidden_state=last_hidden_state,
+            pooler_output=pooled_output,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+        )
+
+    def get_input_embeddings(self):
+        return self.embeddings
+
+
+class Blip2QFormerMultiHeadAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.config = config
+        if config.hidden_size % config.num_attention_heads != 0 and not hasattr(config, "embedding_size"):
+            raise ValueError(
+                "The hidden size (%d) is not a multiple of the number of attention heads (%d)"
+                % (config.hidden_size, config.num_attention_heads)
+            )
+
+        self.num_attention_heads = config.num_attention_heads
+        self.attention_head_size = int(config.hidden_size / config.num_attention_heads)
+        self.all_head_size = self.num_attention_heads * self.attention_head_size
+
+        self.query = nn.Linear(config.hidden_size, self.all_head_size)
+        if is_cross_attention:
+            self.key = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.encoder_hidden_size, self.all_head_size)
+        else:
+            self.key = nn.Linear(config.hidden_size, self.all_head_size)
+            self.value = nn.Linear(config.hidden_size, self.all_head_size)
+
+        self.dropout = nn.Dropout(config.attention_probs_dropout_prob)
+        self.position_embedding_type = getattr(config, "position_embedding_type", "absolute")
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            self.max_position_embeddings = config.max_position_embeddings
+            self.distance_embedding = nn.Embedding(2 * config.max_position_embeddings - 1, self.attention_head_size)
+        self.save_attention = False
+
+    def save_attn_gradients(self, attn_gradients):
+        self.attn_gradients = attn_gradients
+
+    def get_attn_gradients(self):
+        return self.attn_gradients
+
+    def save_attention_map(self, attention_map):
+        self.attention_map = attention_map
+
+    def get_attention_map(self):
+        return self.attention_map
+
+    def transpose_for_scores(self, x):
+        new_x_shape = x.size()[:-1] + (self.num_attention_heads, self.attention_head_size)
+        x = x.view(*new_x_shape)
+        return x.permute(0, 2, 1, 3)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+    ):
+        # If this is instantiated as a cross-attention module, the keys
+        # and values come from an encoder; the attention mask needs to be
+        # such that the encoder's padding tokens are not attended to.
+        is_cross_attention = encoder_hidden_states is not None
+
+        if is_cross_attention:
+            key_layer = self.transpose_for_scores(self.key(encoder_hidden_states))
+            value_layer = self.transpose_for_scores(self.value(encoder_hidden_states))
+            attention_mask = encoder_attention_mask
+        elif past_key_value is not None:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+            key_layer = torch.cat([past_key_value[0], key_layer], dim=2)
+            value_layer = torch.cat([past_key_value[1], value_layer], dim=2)
+        else:
+            key_layer = self.transpose_for_scores(self.key(hidden_states))
+            value_layer = self.transpose_for_scores(self.value(hidden_states))
+
+        mixed_query_layer = self.query(hidden_states)
+
+        query_layer = self.transpose_for_scores(mixed_query_layer)
+
+        past_key_value = (key_layer, value_layer)
+
+        # Take the dot product between "query" and "key" to get the raw attention scores.
+        attention_scores = torch.matmul(query_layer, key_layer.transpose(-1, -2))
+
+        if self.position_embedding_type == "relative_key" or self.position_embedding_type == "relative_key_query":
+            seq_length = hidden_states.size()[1]
+            position_ids_l = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(-1, 1)
+            position_ids_r = torch.arange(seq_length, dtype=torch.long, device=hidden_states.device).view(1, -1)
+            distance = position_ids_l - position_ids_r
+            positional_embedding = self.distance_embedding(distance + self.max_position_embeddings - 1)
+            positional_embedding = positional_embedding.to(dtype=query_layer.dtype)  # fp16 compatibility
+
+            if self.position_embedding_type == "relative_key":
+                relative_position_scores = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores
+            elif self.position_embedding_type == "relative_key_query":
+                relative_position_scores_query = torch.einsum("bhld,lrd->bhlr", query_layer, positional_embedding)
+                relative_position_scores_key = torch.einsum("bhrd,lrd->bhlr", key_layer, positional_embedding)
+                attention_scores = attention_scores + relative_position_scores_query + relative_position_scores_key
+
+        attention_scores = attention_scores / math.sqrt(self.attention_head_size)
+
+        if attention_mask is not None:
+            # Apply the attention mask is (precomputed for all layers in BertModel forward() function)
+            attention_scores = attention_scores + attention_mask
+
+        # Normalize the attention scores to probabilities.
+        attention_probs = nn.Softmax(dim=-1)(attention_scores)
+
+        if is_cross_attention and self.save_attention:
+            self.save_attention_map(attention_probs)
+            attention_probs.register_hook(self.save_attn_gradients)
+
+        # This is actually dropping out entire tokens to attend to, which might
+        # seem a bit unusual, but is taken from the original Transformer paper.
+        attention_probs_dropped = self.dropout(attention_probs)
+
+        # Mask heads if we want to
+        if head_mask is not None:
+            attention_probs_dropped = attention_probs_dropped * head_mask
+
+        context_layer = torch.matmul(attention_probs_dropped, value_layer)
+
+        context_layer = context_layer.permute(0, 2, 1, 3).contiguous()
+        new_context_layer_shape = context_layer.size()[:-2] + (self.all_head_size,)
+        context_layer = context_layer.view(*new_context_layer_shape)
+
+        outputs = (context_layer, attention_probs) if output_attentions else (context_layer,)
+
+        outputs = outputs + (past_key_value,)
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput with Bert->Blip2QFormer
+class Blip2QFormerSelfOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class Blip2QFormerAttention(nn.Module):
+    def __init__(self, config, is_cross_attention=False):
+        super().__init__()
+        self.attention = Blip2QFormerMultiHeadAttention(config, is_cross_attention)
+        self.output = Blip2QFormerSelfOutput(config)
+        self.pruned_heads = set()
+
+    def prune_heads(self, heads):
+        if len(heads) == 0:
+            return
+        heads, index = find_pruneable_heads_and_indices(
+            heads, self.attention.num_attention_heads, self.attention.attention_head_size, self.pruned_heads
+        )
+
+        # Prune linear layers
+        self.attention.query = prune_linear_layer(self.attention.query, index)
+        self.attention.key = prune_linear_layer(self.attention.key, index)
+        self.attention.value = prune_linear_layer(self.attention.value, index)
+        self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
+
+        # Update hyper params and store pruned heads
+        self.attention.num_attention_heads = self.attention.num_attention_heads - len(heads)
+        self.attention.all_head_size = self.attention.attention_head_size * self.attention.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.FloatTensor] = None,
+        head_mask: Optional[torch.FloatTensor] = None,
+        encoder_hidden_states: Optional[torch.FloatTensor] = None,
+        encoder_attention_mask: Optional[torch.FloatTensor] = None,
+        past_key_value: Optional[Tuple[Tuple[torch.FloatTensor]]] = None,
+        output_attentions: Optional[bool] = False,
+    ) -> Tuple[torch.Tensor]:
+        self_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            encoder_hidden_states,
+            encoder_attention_mask,
+            past_key_value,
+            output_attentions,
+        )
+        attention_output = self.output(self_outputs[0], hidden_states)
+        outputs = (attention_output,) + self_outputs[1:]  # add attentions if we output them
+        return outputs
+
+
+# Copied from transformers.models.bert.modeling_bert.BertIntermediate with Bert->Blip2QFormer
+class Blip2QFormerIntermediate(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.intermediate_size)
+        if isinstance(config.hidden_act, str):
+            self.intermediate_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.intermediate_act_fn = config.hidden_act
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.intermediate_act_fn(hidden_states)
+        return hidden_states
+
+
+# Copied from transformers.models.bert.modeling_bert.BertOutput with Bert->Blip2QFormer
+class Blip2QFormerOutput(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.intermediate_size, config.hidden_size)
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+    def forward(self, hidden_states: torch.Tensor, input_tensor: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.dropout(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states + input_tensor)
+        return hidden_states
+
+
+class Blip2QFormerLayer(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.chunk_size_feed_forward = config.chunk_size_feed_forward
+        self.seq_len_dim = 1
+        self.attention = Blip2QFormerAttention(config)
+
+        self.layer_idx = layer_idx
+
+        if layer_idx % config.cross_attention_frequency == 0:
+            self.crossattention = Blip2QFormerAttention(config, is_cross_attention=True)
+            self.has_cross_attention = True
+        else:
+            self.has_cross_attention = False
+
+        self.intermediate_query = Blip2QFormerIntermediate(config)
+        self.output_query = Blip2QFormerOutput(config)
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_value=None,
+        output_attentions=False,
+        query_length=0,
+    ):
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        self_attention_outputs = self.attention(
+            hidden_states,
+            attention_mask,
+            head_mask,
+            output_attentions=output_attentions,
+            past_key_value=self_attn_past_key_value,
+        )
+        attention_output = self_attention_outputs[0]
+        outputs = self_attention_outputs[1:-1]
+
+        present_key_value = self_attention_outputs[-1]
+
+        if query_length > 0:
+            query_attention_output = attention_output[:, :query_length, :]
+
+            if self.has_cross_attention:
+                if encoder_hidden_states is None:
+                    raise ValueError("encoder_hidden_states must be given for cross-attention layers")
+                cross_attention_outputs = self.crossattention(
+                    query_attention_output,
+                    attention_mask,
+                    head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    output_attentions=output_attentions,
+                )
+                query_attention_output = cross_attention_outputs[0]
+                # add cross attentions if we output attention weights
+                outputs = outputs + cross_attention_outputs[1:-1]
+
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk_query,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                query_attention_output,
+            )
+
+            if attention_output.shape[1] > query_length:
+                layer_output_text = apply_chunking_to_forward(
+                    self.feed_forward_chunk,
+                    self.chunk_size_feed_forward,
+                    self.seq_len_dim,
+                    attention_output[:, query_length:, :],
+                )
+                layer_output = torch.cat([layer_output, layer_output_text], dim=1)
+        else:
+            layer_output = apply_chunking_to_forward(
+                self.feed_forward_chunk,
+                self.chunk_size_feed_forward,
+                self.seq_len_dim,
+                attention_output,
+            )
+        outputs = (layer_output,) + outputs
+
+        outputs = outputs + (present_key_value,)
+
+        return outputs
+
+    def feed_forward_chunk(self, attention_output):
+        intermediate_output = self.intermediate(attention_output)
+        layer_output = self.output(intermediate_output, attention_output)
+        return layer_output
+
+    def feed_forward_chunk_query(self, attention_output):
+        intermediate_output = self.intermediate_query(attention_output)
+        layer_output = self.output_query(intermediate_output, attention_output)
+        return layer_output
+
+
+class Blip2QFormerEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.layer = nn.ModuleList(
+            [Blip2QFormerLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+        )
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=False,
+        output_hidden_states=False,
+        return_dict=True,
+        query_length=0,
+    ):
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attentions = () if output_attentions else None
+        all_cross_attentions = () if output_attentions else None
+
+        next_decoder_cache = () if use_cache else None
+
+        for i in range(self.config.num_hidden_layers):
+            layer_module = self.layer[i]
+            if output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_head_mask = head_mask[i] if head_mask is not None else None
+            past_key_value = past_key_values[i] if past_key_values is not None else None
+
+            if getattr(self.config, "gradient_checkpointing", False) and self.training:
+                if use_cache:
+                    logger.warn(
+                        "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                    )
+                    use_cache = False
+
+                def create_custom_forward(module):
+                    def custom_forward(*inputs):
+                        return module(*inputs, past_key_value, output_attentions, query_length)
+
+                    return custom_forward
+
+                layer_outputs = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(layer_module),
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                )
+            else:
+                layer_outputs = layer_module(
+                    hidden_states,
+                    attention_mask,
+                    layer_head_mask,
+                    encoder_hidden_states,
+                    encoder_attention_mask,
+                    past_key_value,
+                    output_attentions,
+                    query_length,
+                )
+
+            hidden_states = layer_outputs[0]
+            if use_cache:
+                next_decoder_cache += (layer_outputs[-1],)
+            if output_attentions:
+                all_self_attentions = all_self_attentions + (layer_outputs[1],)
+                if layer_module.has_cross_attention:
+                    all_cross_attentions = all_cross_attentions + (layer_outputs[2],)
+
+        if output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(
+                v
+                for v in [
+                    hidden_states,
+                    next_decoder_cache,
+                    all_hidden_states,
+                    all_self_attentions,
+                    all_cross_attentions,
+                ]
+                if v is not None
+            )
+        return BaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_decoder_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attentions,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+class Blip2QFormerModel(Blip2PreTrainedModel):
+    """
+    Querying Transformer (Q-Former), used in BLIP-2.
+    """
+
+    def __init__(self, config: Blip2QFormerConfig):
+        super().__init__(config)
+        self.config = config
+
+        self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+
+        self.encoder = Blip2QFormerEncoder(config)
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """
+        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
+        class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    def get_extended_attention_mask(
+        self,
+        attention_mask: torch.Tensor,
+        input_shape: Tuple[int],
+        device: torch.device,
+        has_query: bool = False,
+    ) -> torch.Tensor:
+        """
+        Makes broadcastable attention and causal masks so that future and masked tokens are ignored.
+
+        Arguments:
+            attention_mask (`torch.Tensor`):
+                Mask with ones indicating tokens to attend to, zeros for tokens to ignore.
+            input_shape (`Tuple[int]`):
+                The shape of the input to the model.
+            device: (`torch.device`):
+                The device of the input to the model.
+
+        Returns:
+            `torch.Tensor` The extended attention mask, with a the same dtype as `attention_mask.dtype`.
+        """
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        if attention_mask.dim() == 3:
+            extended_attention_mask = attention_mask[:, None, :, :]
+        elif attention_mask.dim() == 2:
+            # Provided a padding mask of dimensions [batch_size, seq_length]
+            # - the model is an encoder, so make the mask broadcastable to [batch_size, num_heads, seq_length, seq_length]
+            extended_attention_mask = attention_mask[:, None, None, :]
+        else:
+            raise ValueError(
+                "Wrong shape for input_ids (shape {}) or attention_mask (shape {})".format(
+                    input_shape, attention_mask.shape
+                )
+            )
+
+        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
+        # masked positions, this operation will create a tensor which is 0.0 for
+        # positions we want to attend and -10000.0 for masked positions.
+        # Since we are adding it to the raw scores before the softmax, this is
+        # effectively the same as removing these entirely.
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        return extended_attention_mask
+
+    def forward(
+        self,
+        query_embeds,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+        past_key_values=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+    ):
+        r"""
+        encoder_hidden_states  (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, `optional`):
+            Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if
+            the model is configured as a decoder.
+        encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, `optional`):
+            Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in
+            the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`:
+            - 1 for tokens that are **not masked**,
+            - 0 for tokens that are **masked**.
+        past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of:
+            shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and
+            value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are
+            used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key
+            value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape
+            `(batch_size, sequence_length)`.
+        use_cache (`bool`, `optional`):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # past_key_values_length
+        past_key_values_length = (
+            past_key_values[0][0].shape[2] - self.config.query_length if past_key_values is not None else 0
+        )
+
+        query_length = query_embeds.shape[1] if query_embeds is not None else 0
+
+        embedding_output = self.layernorm(query_embeds)
+        embedding_output = self.dropout(embedding_output)
+
+        input_shape = embedding_output.size()[:-1]
+        batch_size, seq_length = input_shape
+        device = embedding_output.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(((batch_size, seq_length + past_key_values_length)), device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D or 3D attention mask is provided for the cross-attention
+        # we need to make broadcastable to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_hidden_states is not None:
+            if type(encoder_hidden_states) == list:
+                encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states[0].size()
+            else:
+                (
+                    encoder_batch_size,
+                    encoder_sequence_length,
+                    _,
+                ) = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+
+            if type(encoder_attention_mask) == list:
+                encoder_extended_attention_mask = [self.invert_attention_mask(mask) for mask in encoder_attention_mask]
+            elif encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+            else:
+                encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+            past_key_values=past_key_values,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            query_length=query_length,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = sequence_output[:, 0, :]
+
+        if not return_dict:
+            return (sequence_output, pooled_output) + encoder_outputs[1:]
+
+        return BaseModelOutputWithPoolingAndCrossAttentions(
+            last_hidden_state=sequence_output,
+            pooler_output=pooled_output,
+            past_key_values=encoder_outputs.past_key_values,
+            hidden_states=encoder_outputs.hidden_states,
+            attentions=encoder_outputs.attentions,
+            cross_attentions=encoder_outputs.cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    """
+    BLIP-2 Model for generating text given an image and an optional text prompt. The model consists of a vision
+    encoder, Querying Transformer (Q-Former) and a language model.
+
+    One can optionally pass `input_ids` to the model, which serve as a text prompt, to make the language model continue
+    the prompt. Otherwise, the language model starts generating text from the [BOS] (beginning-of-sequence) token.
+    """,
+    BLIP_2_START_DOCSTRING,
+)
+class Blip2ForConditionalGeneration(Blip2PreTrainedModel):
+    config_class = Blip2Config
+    main_input_name = "pixel_values"
+
+    def __init__(self, config: Blip2Config):
+        super().__init__(config)
+
+        self.vision_model = Blip2VisionModel(config.vision_config)
+
+        self.query_tokens = nn.Parameter(torch.zeros(1, config.num_query_tokens, config.qformer_config.hidden_size))
+        self.qformer = Blip2QFormerModel(config.qformer_config)
+
+        self.language_projection = nn.Linear(config.qformer_config.hidden_size, config.text_config.hidden_size)
+        if config.use_decoder_only_language_model:
+            language_model = AutoModelForCausalLM.from_config(config.text_config)
+        else:
+            language_model = AutoModelForSeq2SeqLM.from_config(config.text_config)
+        self.language_model = language_model
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self) -> nn.Module:
+        return self.vision_model.embeddings.patch_embedding
+
+    @add_start_docstrings_to_model_forward(BLIP_2_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=Blip2ForConditionalGenerationModelOutput, config_class=Blip2VisionConfig)
+    def forward(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: torch.FloatTensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        decoder_input_ids: Optional[torch.LongTensor] = None,
+        decoder_attention_mask: Optional[torch.LongTensor] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        labels: Optional[torch.LongTensor] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[Tuple, Blip2ForConditionalGenerationModelOutput]:
+        r"""
+        Returns:
+
+        Examples:
+
+        Image captioning (without providing a text prompt):
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        >>> import torch
+
+        >>> device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        >>> processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        >>> model = Blip2ForConditionalGeneration.from_pretrained(
+        ...     "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
+        ... )
+        >>> model.to(device)  # doctest: +IGNORE_RESULT
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> inputs = processor(images=image, return_tensors="pt").to(device, torch.float16)
+
+        >>> generated_ids = model.generate(**inputs)
+        >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+        >>> print(generated_text)
+        two cats laying on a couch
+        ```
+
+        Visual question answering (prompt = question):
+
+        ```python
+        >>> from PIL import Image
+        >>> import requests
+        >>> from transformers import Blip2Processor, Blip2ForConditionalGeneration
+        >>> import torch
+
+        >>> device = "cuda" if torch.cuda.is_available() else "cpu"
+
+        >>> processor = Blip2Processor.from_pretrained("Salesforce/blip2-opt-2.7b")
+        >>> model = Blip2ForConditionalGeneration.from_pretrained(
+        ...     "Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16
+        ... )
+        >>> model.to(device)  # doctest: +IGNORE_RESULT
+
+        >>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
+        >>> image = Image.open(requests.get(url, stream=True).raw)
+
+        >>> prompt = "Question: how many cats are there? Answer:"
+        >>> inputs = processor(images=image, text=prompt, return_tensors="pt").to(device, torch.float16)
+
+        >>> generated_ids = model.generate(**inputs)
+        >>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+        >>> print(generated_text)
+        two
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # step 1: forward the images through the vision encoder,
+        # to get image embeddings of shape (batch_size, seq_len, hidden_size)
+        vision_outputs = self.vision_model(
+            pixel_values=pixel_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        image_embeds = vision_outputs[0]
+
+        # step 2: forward the query tokens through the QFormer, using the image embeddings for cross-attention
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_outputs = self.qformer(
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+        )
+        query_output = query_outputs[0]
+
+        # step 3: use the language model, conditioned on the query outputs and the prompt
+        language_model_inputs = self.language_projection(query_output)
+        language_model_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds], dim=1)
+
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        expected_device = language_model_attention_mask.device
+        attention_mask = torch.cat([language_model_attention_mask, attention_mask.to(expected_device)], dim=1)
+
+        if self.config.use_decoder_only_language_model:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+            )
+            logits = outputs.logits if return_dict else outputs[0]
+            loss = None
+            # we compute the loss here since we need to take into account the sequence length of the query embeds
+            if labels is not None:
+                logits = logits[:, -labels.size(1) :, :]
+                # Shift so that tokens < n predict n
+                shift_logits = logits[..., :-1, :].contiguous()
+                shift_labels = labels[..., 1:].contiguous().to(logits.device)
+
+                # Flatten the tokens
+                loss_fct = CrossEntropyLoss(reduction="mean")
+
+                loss = loss_fct(shift_logits.view(-1, self.config.text_config.vocab_size), shift_labels.view(-1))
+        else:
+            outputs = self.language_model(
+                inputs_embeds=inputs_embeds,
+                attention_mask=attention_mask,
+                decoder_input_ids=decoder_input_ids,
+                decoder_attention_mask=decoder_attention_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                labels=labels,
+            )
+            loss = outputs.loss if return_dict else outputs[0]
+            logits = outputs.logits if return_dict else outputs[1]
+
+        if not return_dict:
+            output = (logits, vision_outputs, query_outputs, outputs)
+            return ((loss,) + output) if loss is not None else output
+
+        return Blip2ForConditionalGenerationModelOutput(
+            loss=loss,
+            logits=logits,
+            vision_outputs=vision_outputs,
+            qformer_outputs=query_outputs,
+            language_model_outputs=outputs,
+        )
+
+    @torch.no_grad()
+    def generate(
+        self,
+        pixel_values: torch.FloatTensor,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        **generate_kwargs,
+    ) -> torch.LongTensor:
+        """
+        Overrides `generate` function to be able to use the model as a conditional generator.
+
+        Args:
+            pixel_values (`torch.FloatTensor` of shape (batch_size, num_channels, height, width)):
+                Input images to be processed.
+            input_ids (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                The sequence used as a prompt for the generation.
+            attention_mask (`torch.LongTensor` of shape (batch_size, sequence_length), *optional*):
+                Mask to avoid performing attention on padding token indices
+
+        Returns:
+            captions (list): A list of strings of length batch_size * num_captions.
+        """
+        batch_size = pixel_values.shape[0]
+        image_embeds = self.vision_model(pixel_values, return_dict=True).last_hidden_state
+        image_attention_mask = torch.ones(image_embeds.size()[:-1], dtype=torch.long, device=image_embeds.device)
+
+        query_tokens = self.query_tokens.expand(image_embeds.shape[0], -1, -1)
+        query_outputs = self.qformer(
+            query_embeds=query_tokens,
+            encoder_hidden_states=image_embeds,
+            encoder_attention_mask=image_attention_mask,
+            return_dict=True,
+        )
+        query_output = query_outputs.last_hidden_state
+
+        language_model_inputs = self.language_projection(query_output)
+        language_attention_mask = torch.ones(
+            language_model_inputs.size()[:-1], dtype=torch.long, device=language_model_inputs.device
+        )
+        if input_ids is None:
+            input_ids = (
+                torch.LongTensor([[self.config.text_config.bos_token_id]])
+                .repeat(batch_size, 1)
+                .to(image_embeds.device)
+            )
+        if attention_mask is None:
+            attention_mask = torch.ones_like(input_ids)
+        attention_mask = torch.cat([language_attention_mask, attention_mask], dim=1)
+
+        # concatenate query embeddings with prompt embeddings
+        inputs_embeds = self.language_model.get_input_embeddings()(input_ids)
+        inputs_embeds = torch.cat([language_model_inputs, inputs_embeds], dim=1)
+
+        outputs = self.language_model.generate(
+            inputs_embeds=inputs_embeds,
+            attention_mask=attention_mask,
+            **generate_kwargs,
+        )
+
+        return outputs