Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

LICENSE

@@ -0,0 +1,202 @@
+
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   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.

Notice

@@ -0,0 +1,79 @@
+Copyright (C) 2025 AIDC-AI
+Licensed under the Apache License, Version 2.0 (the "License").
+
+This model was trained based on the following model:
+1. Ovis2-2B https://huggingface.co/AIDC-AI/Ovis2-2B
+License: Apache License, Version 2.0 (https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/apache-2.0.md, SPDX-License-identifier: Apache-2.0)
+Apache License
+                       Version 2.0, January 2004
+                    http://www.apache.org/licenses/
+TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+Definitions.
+
+"License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.
+
+"Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
+
+"Legal Entity" shall mean the union of the acting entity and all other entities that control, are controlled by, or are under common control with that entity. For the purposes of this definition, "control" means (i) the power, direct or indirect, to cause the direction or management of such entity, whether by contract or otherwise, or (ii) ownership of fifty percent (50%) or more of the outstanding shares, or (iii) beneficial ownership of such entity.
+
+"You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
+
+"Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
+
+"Object" form shall mean any form resulting from mechanical transformation or translation of a Source form, including but not limited to compiled object code, generated documentation, and conversions to other media types.
+
+"Work" shall mean the work of authorship, whether in Source or Object form, made available under the License, as indicated by a copyright notice that is included in or attached to the work (an example is provided in the Appendix below).
+
+"Derivative Works" shall mean any work, whether in Source or Object form, that is based on (or derived from) the Work and for which the editorial revisions, annotations, elaborations, or other modifications represent, as a whole, an original work of authorship. For the purposes of this License, Derivative Works shall not include works that remain separable from, or merely link (or bind by name) to the interfaces of, the Work and Derivative Works thereof.
+
+"Contribution" shall mean any work of authorship, including the original version of the Work and any modifications or additions to that Work or Derivative Works thereof, that is intentionally submitted to Licensor for inclusion in the Work by the copyright owner or by an individual or Legal Entity authorized to submit on behalf of the copyright owner. For the purposes of this definition, "submitted" means any form of electronic, verbal, or written communication sent to the Licensor or its representatives, including but not limited to communication on electronic mailing lists, source code control systems, and issue tracking systems that are managed by, or on behalf of, the Licensor for the purpose of discussing and improving the Work, but excluding communication that is conspicuously marked or otherwise designated in writing by the copyright owner as "Not a Contribution."
+
+"Contributor" shall mean Licensor and any individual or Legal Entity on behalf of whom a Contribution has been received by Licensor and subsequently incorporated within the Work.
+
+Grant of Copyright License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable copyright license to reproduce, prepare Derivative Works of, publicly display, publicly perform, sublicense, and distribute the Work and such Derivative Works in Source or Object form.
+
+Grant of Patent License. Subject to the terms and conditions of this License, each Contributor hereby grants to You a perpetual, worldwide, non-exclusive, no-charge, royalty-free, irrevocable (except as stated in this section) patent license to make, have made, use, offer to sell, sell, import, and otherwise transfer the Work, where such license applies only to those patent claims licensable by such Contributor that are necessarily infringed by their Contribution(s) alone or by combination of their Contribution(s) with the Work to which such Contribution(s) was submitted. If You institute patent litigation against any entity (including a cross-claim or counterclaim in a lawsuit) alleging that the Work or a Contribution incorporated within the Work constitutes direct or contributory patent infringement, then any patent licenses granted to You under this License for that Work shall terminate as of the date such litigation is filed.
+
+Redistribution. You may reproduce and distribute copies of the Work or Derivative Works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
+
+(a) You must give any other recipients of the Work or Derivative Works a copy of this License; and (b) You must cause any modified files to carry prominent notices stating that You changed the files; and (c) You must retain, in the Source form of any Derivative Works that You distribute, all copyright, patent, trademark, and attribution notices from the Source form of the Work, excluding those notices that do not pertain to any part of the Derivative Works; and (d) If the Work includes a "NOTICE" text file as part of its distribution, then any Derivative Works that You distribute must include a readable copy of the attribution notices contained within such NOTICE file, excluding those notices that do not pertain to any part of the Derivative Works, in at least one of the following places: within a NOTICE text file distributed as part of the Derivative Works; within the Source form or documentation, if provided along with the Derivative Works; or, within a display generated by the Derivative Works, if and wherever such third-party notices normally appear. The contents of the NOTICE file are for informational purposes only and do not modify the License. You may add Your own attribution notices within Derivative Works that You distribute, alongside or as an addendum to the NOTICE text from the Work, provided that such additional attribution notices cannot be construed as modifying the License. You may add Your own copyright statement to Your modifications and may provide additional or different license terms and conditions for use, reproduction, or distribution of Your modifications, or for any such Derivative Works as a whole, provided Your use, reproduction, and distribution of the Work otherwise complies with the conditions stated in this License.
+
+Submission of Contributions. Unless You explicitly state otherwise, any Contribution intentionally submitted for inclusion in the Work by You to the Licensor shall be under the terms and conditions of this License, without any additional terms or conditions. Notwithstanding the above, nothing herein shall supersede or modify the terms of any separate license agreement you may have executed with Licensor regarding such Contributions.
+
+Trademarks. This License does not grant permission to use the trade names, trademarks, service marks, or product names of the Licensor, except as required for reasonable and customary use in describing the origin of the Work and reproducing the content of the NOTICE file.
+
+Disclaimer of Warranty. Unless required by applicable law or agreed to in writing, Licensor provides the Work (and each Contributor provides its Contributions) on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied, including, without limitation, any warranties or conditions of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A PARTICULAR PURPOSE. You are solely responsible for determining the appropriateness of using or redistributing the Work and assume any risks associated with Your exercise of permissions under this License.
+
+Limitation of Liability. In no event and under no legal theory, whether in tort (including negligence), contract, or otherwise, unless required by applicable law (such as deliberate and grossly negligent acts) or agreed to in writing, shall any Contributor be liable to You for damages, including any direct, indirect, special, incidental, or consequential damages of any character arising as a result of this License or out of the use or inability to use the Work (including but not limited to damages for loss of goodwill, work stoppage, computer failure or malfunction, or any and all other commercial damages or losses), even if such Contributor has been advised of the possibility of such damages.
+
+Accepting Warranty or Additional Liability. While redistributing the Work or Derivative Works thereof, You may choose to offer, and charge a fee for, acceptance of support, warranty, indemnity, or other liability obligations and/or rights consistent with this License. However, in accepting such obligations, You may act only on Your own behalf and on Your sole responsibility, not on behalf of any other Contributor, and only if You agree to indemnify, defend, and hold each Contributor harmless for any liability incurred by, or claims asserted against, such Contributor by reason of your accepting any such warranty or additional liability.
+
+END OF TERMS AND CONDITIONS
+
+APPENDIX: How to apply the Apache License to your work.
+
+  To apply the Apache License to your work, attach the following
+  boilerplate notice, with the fields enclosed by brackets "[]"
+  replaced with your own identifying information. (Don't include
+  the brackets!)  The text should be enclosed in the appropriate
+  comment syntax for the file format. We also recommend that a
+  file or class name and description of purpose be included on the
+  same "printed page" as the copyright notice for easier
+  identification within third-party archives.
+Copyright [yyyy] [name of copyright owner]
+
+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.
+
+2. sdxl-vae https://huggingface.co/stabilityai/sdxl-vae
+License: MIT (https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md, SPDX-License-identifier: MIT)
+MIT License
+Copyright (c) [year] [fullname]
+Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
+The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
+
+

README.md

@@ -1,3 +1,179 @@
----
-license: apache-2.0
----
+---
+license: apache-2.0
+language:
+- en
+---
+
+# Ovis-U1
+
+<div align="center">
+  <img src=https://cdn-uploads.huggingface.co/production/uploads/637aebed7ce76c3b834cea37/3IK823BZ8w-mz_QfeYkDn.png width="30%"/>
+</div>
+
+<p align="center">
+  <!-- <a href="https://arxiv.org/abs/2502.12579"><img src="https://img.shields.io/badge/arXiv%20paper-2502.12579-b31b1b.svg" alt="arxiv"></a> -->
+  <a href="https://github.com/AIDC-AI/Ovis"><img src="https://img.shields.io/badge/GitHub-AIDC--AI/Ovis--U1-blue?style=flat&logo=github" alt="demo"></a>
+  <a href="https://huggingface.co/spaces/AIDC-AI/Ovis-U1-3B"><img src="https://img.shields.io/badge/🎨_HF_Spaces-AIDC--AI/Ovis--U1--3B-lightblack" alt="demo"></a>
+  <a href="https://huggingface.co/AIDC-AI/Ovis-U1-3B"><img src="https://img.shields.io/badge/🤗_Model-AIDC--AI/Ovis--U1--3B-yellow" alt="model"></a>
+</p>
+
+
+Building on the foundation of the Ovis series, Ovis-U1 is a 3-billion-parameter unified model that integrates multimodal understanding, text-to-image generation, and image editing capabilities. 
+
+<figure>
+  <img src="https://cdn-uploads.huggingface.co/production/uploads/636f4c6b5d2050767e4a1491/EmEEGmot9JzaBfHP2uWld.jpeg" alt="Ovis-U1 architecture">
+  <figcaption style="text-align: center;">The overall architecture of Ovis-U1 (cf. Fig.2 in our report).</figcaption>
+</figure>
+
+---
+
+## 🚀 News
+
+- [2025/6/28] 🔥 Announcing Ovis-U1-3B ([Model](https://huggingface.co/AIDC-AI/Ovis-U1-3B), [Demo](https://huggingface.co/spaces/AIDC-AI/Ovis-U1-3B))!
+
+---
+
+## 📦 Installation
+
+Ovis-U1 has been tested with Python 3.10, Torch 2.4.0, Transformers 4.51.3, and DeepSpeed 0.15.4. For a comprehensive list of package dependencies, please consult the requirements.txt file.
+
+```bash
+git clone [email protected]:AIDC-AI/Ovis-U1.git
+conda create -n ovis-u1 python=3.10 -y
+conda activate ovis-u1
+cd Ovis-U1
+pip install -r requirements.txt
+pip install -e .
+
+```
+
+## 📂 Model Checkpoints
+
+We provide pretrained Ovis-U1-3B checkpoints for easy download and evaluation:
+ 
+- **Model Repository**: [![Hugging Face](https://img.shields.io/badge/Hugging_Face-Ovis--U1--3B-blue?logo=Huggingface)](https://huggingface.co/AIDC-AI/Ovis-U1-3B)
+
+
+## 🛠️ Inference
+
+For multimodal understanding, please run
+
+```bash
+python ovis/eval/test_txt_generation.py
+```
+
+For text-to-image, please run
+```bash
+python ovis/eval/test_t2i.py \
+    --height 1024 \
+    --width 1024  \
+    --steps 50 \
+    --seed 42 \
+    --txt_cfg 5  
+```
+
+For image editing, please run
+```bash
+python ovis/eval/test_img_edit.py \
+    --steps 50 \
+    --img_cfg 4 \
+    --txt_cfg 7.5  
+```
+
+## 📊 Performance
+
+#### OpenCompass Multi-modal Academic Benchmarks
+
+| Model | MMB | MMS | MMMU | MathVista | Hallusion | AI2D | OCRBench | MMVet | Avg |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| GPT-4o | 86 | 70.2 | 72.9 | 71.6 | 57 | 86.3 | 82.2 | 76.9 | **75.4** |
+| InternVL2.5-2B | 70.9 | 54.3 | 43.2 | 51.1 | 42.3 | 74.9 | 80.2 | 62.6 | **59.9** |
+| SAIL-VL-2B | 73.7 | 56.5 | 44.1 | 62.8 | 45.9 | 77.4 | 83.1 | 44.2 | **61** |
+| InternVL3-2B | 78 | 61.1 | 48.7 | 57.6 | 41.9 | 78.6 | 83.1 | 67 | **61.1**|
+| Qwen2.5-VL-3B | 76.8 | 56.3 | 51.2 | 61.2 | 46.6 | 81.4 | 82.8 | 60 | **64.5** |
+| Ovis2-2B | 76.9 | 56.7 | 45.6 | 64.1 | 50.2 | 82.7 | 87.3 | 58.3 | **65.2** |
+| SAIL-VL-1.5-2B | 78.5 | 62.6 | 46.4 | 67 | 50 | 83.7 | 89.1 | 58.8 | **67** |
+| Ristretto-3B | 80.2 | 62.8 | 51.3 | 67.6 | 50.2 | 84.2 | 84.7 | 60.7 | **67.7** |
+| Ovis-U1 | 77.8 | 61.3 | 51.1 | 69.4 | 56.3 | 85.6 | 88.3 | 66.7 | **69.6** |
+
+#### GenEval
+
+| Model | Single object | Two object | Counting | Colors | Position | Attribute binding | Overall |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| GPT-4o | 0.99 | 0.92 | 0.85 | 0.92 | 0.75 | 0.61 | **0.84** |
+| BAGEL | 0.99 | 0.94 | 0.81 | 0.88 | 0.64 | 0.63 | **0.82** |
+| BAGEL 📝 | 0.98 | 0.95 | 0.84 | 0.95 | 0.78 | 0.77 | **0.88** |
+| UniWorld-V1 | 0.99 | 0.93 | 0.79 | 0.89 | 0.49 | 0.70 | **0.80** |
+| UniWorld-V1 📝 | 0.98 | 0.93 | 0.81 | 0.89 | 0.74 | 0.71 | **0.84** |
+| OmniGen | 0.98 | 0.84 | 0.66 | 0.74 | 0.40 | 0.43 | **0.68** | 
+| OmniGen2 | 1 | 0.95 | 0.64 | 0.88 | 0.55 | 0.76 | **0.80** |
+| OmniGen2 📝 | 0.99 | 0.96 | 0.74 | 0.98 | 0.71 | 0.75 | **0.86** |
+| Ovis-U1 | 0.98 | 0.98 | 0.90 | 0.92 | 0.79 | 0.75 | **0.89** |
+
+*📝 denotes using the rewritten prompts*
+
+#### DPG-Bench
+
+| Model | Global | Entity | Attribute | Relation | Other | Overall |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| BAGEL | 88.94 | 90.37 | 91.29 | 90.82 | 88.67 | **85.07** |
+| UniWorld-V1 | 83.64 | 88.39 | 88.44 | 89.27 | 87.22 | **81.38** |
+| OmniGen | 87.90 | 88.97 | 88.47 | 87.95 | 83.56 | **81.16** |
+| OmniGen2 | 88.81 | 88.83 | 90.18 | 89.37 | 90.27 | **83.57** |
+| Ovis-U1 | 82.37 | 90.08 | 88.68 | 93.35 | 85.20 | **83.72** |
+
+#### ImgEdit-Bench
+
+| Model | Add | Adjust | Extract | Replace | Remove | Background | Style | Hybrid | Action | Overall |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| GPT-4o | 4.61 | 4.33 | 2.9 | 4.35 | 3.66 | 4.57 | 4.93 | 3.96 | 4.89 | **4.2** |
+| MagicBrush | 2.84 | 1.58 | 1.51 | 1.97 | 1.58 | 1.75 | 2.38 | 1.62 | 1.22 | **1.90** |
+| Instruct-P2P | 2.45 | 1.83 | 1.44 | 2.01 | 1.50 | 1.44 | 3.55 | 1.2 | 1.46 | **1.88** |
+| AnyEdit | 3.18 | 2.95 | 1.88 | 2.47 | 2.23 | 2.24 | 2.85 | 1.56 | 2.65 | **2.45** |
+| UltraEdit | 3.44 | 2.81 | 2.13 | 2.96 | 1.45 | 2.83 | 3.76 | 1.91 | 2.98 | **2.7** |
+| OmniGen | 3.47 | 3.04 | 1.71 | 2.94 | 2.43 | 3.21 | 4.19 | 2.24 | 3.38 | **2.96** |
+| Step1X-Edit | 3.88 | 3.14 | 1.76 | 3.40 | 2.41 | 3.16 | 4.63 | 2.64 | 2.52 | **3.06** |
+| ICEdit | 3.58 | 3.39 | 1.73 | 3.15 | 2.93 | 3.08 | 3.84 | 2.04 | 3.68 | **3.05** |
+| BAGEL | 3.56 | 3.31 | 1.7 | 3.3 | 2.62 | 3.24 | 4.49 | 2.38 | 4.17 | **3.2** |
+| UniWorld-V1 | 3.82 | 3.64 | 2.27 | 3.47 | 3.24 | 2.99 | 4.21 | 2.96 | 2.74 | **3.26** |
+| OmniGen2 | 3.57 | 3.06 | 1.77 | 3.74 | 3.2 | 3.57 | 4.81 | 2.52 | 4.68 | **3.44** |
+| Ovis-U1 | 4.13 | 3.62 | 2.98 | 4.45 | 4.06 | 4.22 | 4.69 | 3.45 | 4.61 | **4.00** |
+
+#### GEdit-Bench-EN
+
+|  Model | Background Change | Color Alteration   | Material Modification  | Motion Change | Portrait Beautification  | Style Transfer  | Subject Addition  | Subject Removal  | Subject Replacement  | Text Modification  | Tone Transformation  | Avg |
+|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
+| GPT-4o | 7.205 |	6.491 |	6.607 | 8.096 |	7.768 |	6.961 |	7.622 |	8.331 |	8.067 |	7.427 |	8.301 |	**7.534** |
+| AnyEdit | 4.663	| 4.260 |	2.537 |	2.024 |	3.479	| 2.032 |	3.995 |	3.089 |	3.180 |	0.922 |	5.151 |	**3.212** |
+| Instruct-Pix2Pix | 3.825 |	5.182 |	3.688 |	3.509 |	4.339 |	4.560 |	3.461 |	2.031 |	4.237 |	0.955 |	4.733 |	**3.684** |
+| MagicBrush |	5.637 |	5.136 |	5.078 |	4.513 |	4.487 |	4.439 |	5.252 |	3.704 |	4.941 |	1.384 |	5.130 |	**4.518** |
+| OmniGen | 5.281 |	6.003 |	5.308 |	2.916 |	3.087 |	4.903 |	6.628 |	6.352 |	5.616 |	4.519 |	5.064 |	**5.062** |
+| Gemini |	6.781 |	6.369 |	6.040 |	6.938 |	5.591 |	4.676 |	7.501 |	6.447 |	7.003 |	5.765 |	6.350 |	**6.315** |
+| Step1X-Edit |	6.547 |	6.545 |	6.204 |	6.483 |	6.787 |	7.221 |	6.975 |	6.512 |	7.068 |	6.921 |	6.448 |	**6.701** |
+| Doubao |	7.430 |	7.095 |	6.339 |	6.973 |	6.972 |	6.767 |	7.674 |	6.748 |	7.447 |	3.471 |	7.383 |	**6.754** |
+| BAGEL | 7.324 |	6.909 |	6.381 |	4.753 |	4.573 |	6.150 |	7.896 |	7.164 |	7.021 |	7.320 |	6.218 |	**6.519** |
+| Ovis-U1 | 7.486 |	6.879 |	6.208 |	4.790 |	5.981 |	6.463 |	7.491 |	7.254 |	7.266 |	4.482 |	6.314 |	**6.420** |
+
+## 📚 Citation
+
+If you find Ovis-U1 useful, please cite our paper:
+
+```bibtex
+@inproceedings{wang2025ovisu1,
+title={Ovis-U1 Technical Report},
+author={Ovis Team},
+year={2025}
+}
+```
+
+## 🙏 Acknowledgments
+
+The code is built upon [Ovis](https://github.com/AIDC-AI/Ovis) and [FLUX](https://github.com/black-forest-labs/flux).
+
+## 📄 License
+
+The project is released under Apache License 2.0 (http://www.apache.org/licenses/LICENSE-2.0, SPDX-License-identifier: Apache-2.0).
+
+## 🚨 Disclaimer
+
+We used compliance checking algorithms during the training process, to ensure the compliance of the trained model to the best of our ability. Due to complex data and the diversity of language model usage scenarios, we cannot guarantee that the model is completely free of copyright issues or improper content. If you believe anything infringes on your rights or generates improper content, please contact us, and we will promptly address the matter.

added_tokens.json

@@ -0,0 +1,28 @@
+{
+  "</think>": 151668,
+  "</tool_call>": 151658,
+  "</tool_response>": 151666,
+  "<think>": 151667,
+  "<tool_call>": 151657,
+  "<tool_response>": 151665,
+  "<|box_end|>": 151649,
+  "<|box_start|>": 151648,
+  "<|endoftext|>": 151643,
+  "<|file_sep|>": 151664,
+  "<|fim_middle|>": 151660,
+  "<|fim_pad|>": 151662,
+  "<|fim_prefix|>": 151659,
+  "<|fim_suffix|>": 151661,
+  "<|im_end|>": 151645,
+  "<|im_start|>": 151644,
+  "<|image_pad|>": 151655,
+  "<|object_ref_end|>": 151647,
+  "<|object_ref_start|>": 151646,
+  "<|quad_end|>": 151651,
+  "<|quad_start|>": 151650,
+  "<|repo_name|>": 151663,
+  "<|video_pad|>": 151656,
+  "<|vision_end|>": 151653,
+  "<|vision_pad|>": 151654,
+  "<|vision_start|>": 151652
+}

config.json

@@ -0,0 +1,410 @@
+{
+  "architectures": [
+    "OvisU1"
+  ],
+  "auto_map": {
+    "AutoConfig": "configuration_ovis_u1.OvisU1Config",
+    "AutoModelForCausalLM": "modeling_ovis_u1.OvisU1"
+  },
+  "conversation_formatter_class": "Qwen3ConversationFormatter",
+  "disable_tie_weight": false,
+  "hidden_size": 2048,
+  "llm_attn_implementation": null,
+  "llm_config": {
+    "_attn_implementation_autoset": true,
+    "_name_or_path": "Qwen/Qwen3-1.7B",
+    "add_cross_attention": false,
+    "architectures": [
+      "Qwen3ForCausalLM"
+    ],
+    "attention_bias": false,
+    "attention_dropout": 0.0,
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": 151643,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": 151645,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "head_dim": 128,
+    "hidden_act": "silu",
+    "hidden_size": 2048,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "initializer_range": 0.02,
+    "intermediate_size": 6144,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_position_embeddings": 40960,
+    "max_window_layers": 28,
+    "min_length": 0,
+    "model_type": "qwen3",
+    "no_repeat_ngram_size": 0,
+    "num_attention_heads": 16,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_hidden_layers": 28,
+    "num_key_value_heads": 8,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "rms_norm_eps": 1e-06,
+    "rope_scaling": null,
+    "rope_theta": 1000000,
+    "sep_token_id": null,
+    "sliding_window": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "bfloat16",
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_cache": true,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "model_type": "ovis_u1",
+  "multimodal_max_length": 4496,
+  "torch_dtype": "bfloat16",
+  "transformers_version": "4.51.3",
+  "use_cache": false,
+  "visual_generator_config": {
+    "_attn_implementation_autoset": true,
+    "_name_or_path": "yak_1b/yak_1b_qwen3_trained_S8",
+    "add_cross_attention": false,
+    "architectures": [
+      "YakModel"
+    ],
+    "axes_dim": [
+      16,
+      56,
+      56
+    ],
+    "bad_words_ids": null,
+    "base_shift": 0.5,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "checkpoint": false,
+    "chunk_size_feed_forward": 0,
+    "context_in_dim": 4096,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "depth": 6,
+    "depth_single_blocks": 12,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "guidance_embed": false,
+    "hidden_size": 1536,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "in_channels": 16,
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_shift": 1.15,
+    "min_length": 0,
+    "mlp_ratio": 4.0,
+    "model_type": "yak",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_heads": 12,
+    "num_return_sequences": 1,
+    "out_channels": 16,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "problem_type": null,
+    "pruned_heads": {},
+    "qkv_bias": true,
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "temperature": 1.0,
+    "tf_legacy_loss": false,
+    "theta": 10000,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "timestep_shift": true,
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": "float32",
+    "torchscript": false,
+    "txt_type": "refiner",
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "vae_config": {
+      "_class_name": "AutoencoderKL",
+      "_diffusers_version": "0.18.0.dev0",
+      "_name_or_path": ".",
+      "act_fn": "silu",
+      "block_out_channels": [
+        128,
+        256,
+        512,
+        512
+      ],
+      "down_block_types": [
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D",
+        "DownEncoderBlock2D"
+      ],
+      "force_upcast": false,
+      "in_channels": 3,
+      "latent_channels": 4,
+      "layers_per_block": 2,
+      "norm_num_groups": 32,
+      "out_channels": 3,
+      "sample_size": 512,
+      "scaling_factor": 0.13025,
+      "up_block_types": [
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D",
+        "UpDecoderBlock2D"
+      ]
+    },
+    "vec_in_dim": 1536
+  },
+  "visual_tokenizer_config": {
+    "_attn_implementation_autoset": true,
+    "_name_or_path": "",
+    "add_cross_attention": false,
+    "architectures": null,
+    "backbone_config": {
+      "_attn_implementation_autoset": false,
+      "_name_or_path": "aimv2/visual_tokenizer_backbone",
+      "add_cross_attention": false,
+      "architectures": [
+        "AIMv2Model"
+      ],
+      "attention_dropout": 0.0,
+      "auto_map": {
+        "AutoConfig": "configuration_aimv2.AIMv2Config",
+        "AutoModel": "modeling_aimv2.AIMv2Model",
+        "FlaxAutoModel": "modeling_flax_aimv2.FlaxAIMv2Model"
+      },
+      "bad_words_ids": null,
+      "begin_suppress_tokens": null,
+      "bos_token_id": null,
+      "chunk_size_feed_forward": 0,
+      "cross_attention_hidden_size": null,
+      "decoder_start_token_id": null,
+      "disable_rope": false,
+      "diversity_penalty": 0.0,
+      "do_sample": false,
+      "early_stopping": false,
+      "encoder_no_repeat_ngram_size": 0,
+      "eos_token_id": null,
+      "exponential_decay_length_penalty": null,
+      "finetuning_task": null,
+      "forced_bos_token_id": null,
+      "forced_eos_token_id": null,
+      "fullatt_block_indexes": null,
+      "hidden_size": 1024,
+      "hidden_stride": 2,
+      "id2label": {
+        "0": "LABEL_0",
+        "1": "LABEL_1"
+      },
+      "image_size": 448,
+      "intermediate_size": 2816,
+      "interpolate_pe_method": "two_dim",
+      "is_decoder": false,
+      "is_encoder_decoder": false,
+      "label2id": {
+        "LABEL_0": 0,
+        "LABEL_1": 1
+      },
+      "length_penalty": 1.0,
+      "max_length": 20,
+      "max_pixels": 2408448,
+      "min_length": 0,
+      "min_pixels": 200704,
+      "model_type": "aimv2",
+      "no_repeat_ngram_size": 0,
+      "num_attention_heads": 8,
+      "num_beam_groups": 1,
+      "num_beams": 1,
+      "num_channels": 3,
+      "num_hidden_layers": 24,
+      "num_return_sequences": 1,
+      "output_attentions": false,
+      "output_hidden_states": false,
+      "output_scores": false,
+      "pad_token_id": null,
+      "patch_size": 14,
+      "prefix": null,
+      "preserve_original_pe": true,
+      "problem_type": null,
+      "projection_dropout": 0.0,
+      "pruned_heads": {},
+      "qkv_bias": false,
+      "remove_invalid_values": false,
+      "repetition_penalty": 1.0,
+      "return_dict": true,
+      "return_dict_in_generate": false,
+      "rms_norm_eps": 1e-05,
+      "sep_token_id": null,
+      "suppress_tokens": null,
+      "task_specific_params": null,
+      "temperature": 1.0,
+      "temporal_patch_size": 1,
+      "tf_legacy_loss": false,
+      "tie_encoder_decoder": false,
+      "tie_word_embeddings": true,
+      "tokenizer_class": null,
+      "top_k": 50,
+      "top_p": 1.0,
+      "torch_dtype": "bfloat16",
+      "torchscript": false,
+      "typical_p": 1.0,
+      "use_bfloat16": false,
+      "use_bias": false,
+      "window_size": 112
+    },
+    "backbone_kwargs": {
+      "disable_rope": false,
+      "hidden_stride": 2,
+      "interpolate_pe_method": "two_dim",
+      "max_pixels": 2408448,
+      "min_pixels": 200704,
+      "preserve_original_pe": true,
+      "temporal_patch_size": 1,
+      "window_size": 112
+    },
+    "bad_words_ids": null,
+    "begin_suppress_tokens": null,
+    "bos_token_id": null,
+    "chunk_size_feed_forward": 0,
+    "cross_attention_hidden_size": null,
+    "decoder_start_token_id": null,
+    "depths": null,
+    "disable_rope": false,
+    "diversity_penalty": 0.0,
+    "do_sample": false,
+    "drop_cls_token": false,
+    "early_stopping": false,
+    "encoder_no_repeat_ngram_size": 0,
+    "eos_token_id": null,
+    "exponential_decay_length_penalty": null,
+    "finetuning_task": null,
+    "forced_bos_token_id": null,
+    "forced_eos_token_id": null,
+    "fullatt_block_indexes": null,
+    "hidden_stride": 2,
+    "id2label": {
+      "0": "LABEL_0",
+      "1": "LABEL_1"
+    },
+    "image_processor_new_kwargs": {
+      "hidden_stride": 2,
+      "max_pixels": 2408448,
+      "min_pixels": 200704,
+      "temporal_patch_size": 1
+    },
+    "interpolate_pe_method": "two_dim",
+    "is_decoder": false,
+    "is_encoder_decoder": false,
+    "label2id": {
+      "LABEL_0": 0,
+      "LABEL_1": 1
+    },
+    "length_penalty": 1.0,
+    "max_length": 20,
+    "max_pixels": 2408448,
+    "min_length": 0,
+    "min_pixels": 200704,
+    "model_type": "aimv2_visual_tokenizer",
+    "no_repeat_ngram_size": 0,
+    "num_beam_groups": 1,
+    "num_beams": 1,
+    "num_return_sequences": 1,
+    "output_attentions": false,
+    "output_hidden_states": false,
+    "output_scores": false,
+    "pad_token_id": null,
+    "prefix": null,
+    "preserve_original_pe": true,
+    "problem_type": null,
+    "pruned_heads": {},
+    "remove_invalid_values": false,
+    "repetition_penalty": 1.0,
+    "return_dict": true,
+    "return_dict_in_generate": false,
+    "sep_token_id": null,
+    "suppress_tokens": null,
+    "task_specific_params": null,
+    "tau": 1.0,
+    "temperature": 1.0,
+    "temporal_patch_size": 1,
+    "tf_legacy_loss": false,
+    "tie_encoder_decoder": false,
+    "tie_word_embeddings": true,
+    "tokenize_function": "softmax",
+    "tokenizer_class": null,
+    "top_k": 50,
+    "top_p": 1.0,
+    "torch_dtype": null,
+    "torchscript": false,
+    "typical_p": 1.0,
+    "use_bfloat16": false,
+    "use_indicators": false,
+    "vocab_size": 65536,
+    "window_size": 112
+  }
+}

configuration_aimv2.py

@@ -0,0 +1,82 @@
+# copied from https://huggingface.co/apple/aimv2-huge-patch14-448
+from typing import Any
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["AIMv2Config"]
+
+
+class AIMv2Config(PretrainedConfig):
+    """This is the configuration class to store the configuration of an [`AIMv2Model`].
+
+    Instantiating a configuration with the defaults will yield a similar configuration
+    to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
+
+    Args:
+        hidden_size: Dimension of the hidden representations.
+        intermediate_size: Dimension of the SwiGLU representations.
+        num_hidden_layers: Number of hidden layers in the Transformer.
+        num_attention_heads: Number of attention heads for each attention layer
+            in the Transformer.
+        num_channels: Number of input channels.
+        image_size: Image size.
+        patch_size: Patch size.
+        rms_norm_eps: Epsilon value used for the RMS normalization layer.
+        attention_dropout: Dropout ratio for attention probabilities.
+        projection_dropout: Dropout ratio for the projection layer after the attention.
+        qkv_bias: Whether to add a bias to the queries, keys and values.
+        use_bias: Whether to add a bias in the feed-forward and projection layers.
+        kwargs: Keyword arguments for the [`PretrainedConfig`].
+    """
+
+    model_type: str = "aimv2"
+
+    def __init__(
+        self,
+        hidden_size: int = 1024,
+        intermediate_size: int = 2816,
+        num_hidden_layers: int = 24,
+        num_attention_heads: int = 8,
+        num_channels: int = 3,
+        image_size: int = 224,
+        patch_size: int = 14,
+        rms_norm_eps: float = 1e-5,
+        attention_dropout: float = 0.0,
+        projection_dropout: float = 0.0,
+        qkv_bias: bool = False,
+        use_bias: bool = False,
+        hidden_stride: int = 2,
+        window_size: int = 112,
+        fullatt_block_indexes: list = None,
+        temporal_patch_size: int = 1,
+        preserve_original_pe: bool = False,
+        interpolate_pe_method: str = 'one_dim',
+        disable_rope: bool = False,
+        min_pixels: int = 3136,
+        max_pixels: int = 1960000,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.image_size = image_size
+        self.attention_dropout = attention_dropout
+        self.rms_norm_eps = rms_norm_eps
+
+        self.projection_dropout = projection_dropout
+        self.qkv_bias = qkv_bias
+        self.use_bias = use_bias
+
+        self.hidden_stride = hidden_stride
+        self.window_size = window_size
+        self.fullatt_block_indexes = fullatt_block_indexes
+        self.temporal_patch_size = temporal_patch_size
+        self.preserve_original_pe = preserve_original_pe
+        self.interpolate_pe_method = interpolate_pe_method
+        self.disable_rope = disable_rope
+        self.min_pixels = min_pixels
+        self.max_pixels = max_pixels

configuration_ovis_u1.py

@@ -0,0 +1,281 @@
+from typing import List, Dict, Union, Optional
+from abc import ABC, abstractmethod
+
+from transformers import PretrainedConfig, AutoConfig, AutoModel
+
+# Model Constants
+IGNORE_ID = -100
+IMAGE_TOKEN_ID = -200
+VIDEO_TOKEN_ID = -201 
+IMAGE_TOKEN = "<image>"
+VIDEO_TOKEN = "<video>"
+IMAGE_ATOM_ID = -300
+IMAGE_INDICATOR_IDS = [-301, -302, -303, -304]
+
+from .configuration_aimv2 import AIMv2Config
+from .modeling_aimv2 import AIMv2Model
+AutoConfig.register("aimv2", AIMv2Config)
+AutoModel.register(AIMv2Config, AIMv2Model)
+
+from .configuration_yak import YakConfig
+from .modeling_yak import YakModel
+AutoConfig.register("yak", YakConfig)
+AutoModel.register(YakConfig, YakModel)
+
+
+# ----------------------------------------------------------------------
+#                     Visual Tokenizer Configuration
+# ----------------------------------------------------------------------
+class BaseVisualTokenizerConfig(PretrainedConfig):
+    def __init__(self,
+        vocab_size=16384,
+        tokenize_function="softmax",
+        tau=1.0,
+        depths=None,
+        use_indicators=False,
+        drop_cls_token=False,
+        backbone_config: Optional[Union[PretrainedConfig, dict]] = None,
+        hidden_stride: int = 1,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.vocab_size = vocab_size
+        self.tokenize_function = tokenize_function
+        self.tau = tau
+        if isinstance(depths, str):
+            depths = [int(x) for x in depths.split('|')]
+        self.depths = depths
+        self.backbone_kwargs = {}
+        self.use_indicators = use_indicators
+        self.drop_cls_token = drop_cls_token
+        if backbone_config is not None:
+            assert isinstance(backbone_config, (PretrainedConfig, dict)), \
+                f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"
+            if not isinstance(backbone_config, PretrainedConfig):
+                model_type = backbone_config['model_type']
+                backbone_config.pop('model_type')
+                backbone_config = AutoConfig.for_model(model_type, **backbone_config)
+        self.backbone_config = backbone_config
+        self.hidden_stride = hidden_stride
+
+
+class Aimv2VisualTokenizerConfig(BaseVisualTokenizerConfig):
+    model_type = "aimv2_visual_tokenizer"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        if self.drop_cls_token:
+            self.drop_cls_token = False
+        if self.depths:
+            assert len(self.depths) == 1
+            self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
+
+        self.image_processor_new_kwargs = {}
+
+        if kwargs.get("min_pixels", None) is not None:
+            self.image_processor_new_kwargs['min_pixels'] = kwargs.get("min_pixels")
+            self.backbone_kwargs['min_pixels'] = self.min_pixels
+        
+        if kwargs.get("max_pixels", None) is not None:
+            self.image_processor_new_kwargs['max_pixels'] = kwargs.get("max_pixels")
+            self.backbone_kwargs['max_pixels'] = self.max_pixels
+        
+        if kwargs.get("temporal_patch_size", None) is not None:
+            self.image_processor_new_kwargs['temporal_patch_size'] = kwargs.get("temporal_patch_size")
+            self.backbone_kwargs['temporal_patch_size'] = self.temporal_patch_size
+        
+        if kwargs.get("hidden_stride", None) is not None:
+            self.image_processor_new_kwargs['hidden_stride'] = kwargs.get("hidden_stride")
+
+        if kwargs.get("patch_size", None) is not None:
+            self.image_processor_new_kwargs['patch_size'] = kwargs.get("patch_size")
+            self.backbone_kwargs['patch_size'] = self.patch_size
+
+        if kwargs.get("window_size", None) is not None:
+            self.backbone_kwargs['window_size'] = kwargs.get("window_size")
+
+        if kwargs.get("hidden_stride", None) is not None:
+            self.backbone_kwargs['hidden_stride'] = kwargs.get("hidden_stride")
+
+        if kwargs.get('fullatt_block_indexes', None) is not None:
+            self.backbone_kwargs['fullatt_block_indexes'] = [int(i) for i in kwargs.get('fullatt_block_indexes').replace(' ','').split('|')]
+        
+        if kwargs.get("preserve_original_pe", None) is not None:
+            self.backbone_kwargs['preserve_original_pe'] = kwargs.get("preserve_original_pe")
+        
+        if kwargs.get("interpolate_pe_method", None) is not None:
+            self.backbone_kwargs['interpolate_pe_method'] = kwargs.get("interpolate_pe_method")
+
+        if kwargs.get("disable_rope", None) is not None:
+            self.backbone_kwargs['disable_rope'] = kwargs.get("disable_rope")
+
+AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)
+
+
+
+# ----------------------------------------------------------------------
+#                          OvisU1 Configuration
+# ----------------------------------------------------------------------
+class OvisU1Config(PretrainedConfig):
+    model_type = "ovis_u1"
+
+    def __init__(self,
+                 llm_config: Optional[Union[PretrainedConfig, dict]] = None,
+                 visual_tokenizer_config: Optional[Union[PretrainedConfig, dict]] = None,
+                 visual_generator_config: Optional[Union[PretrainedConfig, dict]] = None,
+                 multimodal_max_length=2048,
+                 hidden_size=None,
+                 conversation_formatter_class=None,
+                 llm_attn_implementation=None,
+                 disable_tie_weight=False,
+                 **kwargs):
+        super().__init__(**kwargs)
+        if llm_config is not None:
+            assert isinstance(llm_config, (PretrainedConfig, dict)), \
+                f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"
+            if not isinstance(llm_config, PretrainedConfig):
+                model_type = llm_config['model_type']
+                llm_config.pop('model_type')
+                llm_config = AutoConfig.for_model(model_type, **llm_config)
+        self.llm_config = llm_config
+        if visual_tokenizer_config is not None:
+            assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), \
+                f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
+            if not isinstance(visual_tokenizer_config, PretrainedConfig):
+                model_type = visual_tokenizer_config['model_type']
+                visual_tokenizer_config.pop('model_type')
+                if model_type == "aimv2_native_visual_tokenizer":
+                    model_type = "aimv2_visual_tokenizer"
+                if visual_tokenizer_config['backbone_config']['model_type'] == "aimv2_native":
+                    visual_tokenizer_config['backbone_config']['model_type'] = "aimv2"
+                visual_tokenizer_config = AutoConfig.for_model(model_type, **visual_tokenizer_config)
+        self.visual_tokenizer_config = visual_tokenizer_config
+        if visual_generator_config is not None:
+            assert isinstance(visual_generator_config, (PretrainedConfig, dict)), \
+                f"expect `visual_generator_config` to be instance of PretrainedConfig or dict, but got {type(visual_generator_config)} type"
+            if not isinstance(visual_generator_config, PretrainedConfig):
+                model_type = visual_generator_config['model_type']
+                visual_generator_config.pop('model_type')
+                visual_generator_config = AutoConfig.for_model(model_type, **visual_generator_config)
+        self.visual_generator_config = visual_generator_config
+        self.multimodal_max_length = multimodal_max_length
+        self.hidden_size = hidden_size
+        self.conversation_formatter_class = conversation_formatter_class
+        self.llm_attn_implementation = llm_attn_implementation
+        self.disable_tie_weight = disable_tie_weight
+        
+
+# ----------------------------------------------------------------------
+#                         Conversation Formatter
+# ----------------------------------------------------------------------
+class ConversationFormatter(ABC):
+    support_tokenizer_types = None
+
+    def __init__(self, tokenizer):
+        tokenizer_type = type(tokenizer).__name__
+        assert tokenizer_type in self.support_tokenizer_types, \
+            f'Invalid tokenizer type, expected one from `{self.support_tokenizer_types}`, but got `{tokenizer_type}`'
+        self.tokenizer = tokenizer
+        self.image_token = IMAGE_TOKEN
+        self.image_token_id = IMAGE_TOKEN_ID
+        self.ignore_id = IGNORE_ID
+        self.im_end = None
+        self.video_token = VIDEO_TOKEN
+        self.video_token_id = VIDEO_TOKEN_ID
+
+    def _tokenize_with_image_symbol(self, text):
+        if text.find(self.video_token) != -1:
+            token = self.video_token
+            token_id = self.video_token_id
+        else:
+            token = self.image_token
+            token_id = self.image_token_id
+
+        text_chunks = [self.tokenizer(chunk, add_special_tokens=False).input_ids for chunk in
+                       text.split(token)]
+        token_ids = []
+        num_chuck = len(text_chunks)
+        for i, chunk in enumerate(text_chunks):
+            token_ids.extend(chunk)
+            if i < num_chuck - 1:
+                token_ids.append(token_id)
+        return token_ids
+
+    @abstractmethod
+    def format(self, conversations: List[Dict], generation_preface=None):
+        pass
+
+    @abstractmethod
+    def format_query(self, query, generation_preface=""):
+        pass
+
+
+class Qwen3ConversationFormatter(ConversationFormatter):
+    support_tokenizer_types = ['QWenTokenizer', 'Qwen2TokenizerFast']
+
+    def __init__(self, tokenizer):
+        super().__init__(tokenizer)
+        self.from2role = {
+            "system": "<|im_start|>system\n",
+            "human": "<|im_start|>user\n",
+            "gpt": "<|im_start|>assistant\n",
+            "ignored_gpt": "<|im_start|>assistant\n",
+        }
+        self.gpt_token_num = None
+        self.im_end = "<|im_end|>\n"
+        self.empty_think = "<think>\n\n</think>\n\n"
+
+    def format(self, conversations: List[Dict], generation_preface=None, enable_thinking=False):
+        if self.gpt_token_num is None:
+            prefilled_think = "" if enable_thinking else self.empty_think
+            self.gpt_token_num = len(
+                self.tokenizer(self.from2role["gpt"] + prefilled_think, add_special_tokens=False).input_ids
+            )
+
+        if generation_preface is not None:
+            conversations.append({
+                "from": "gpt",
+                "value": generation_preface
+            })
+
+        prompt = ""
+        input_ids = []
+        labels = []
+        num_conversation = len(conversations)
+        for i, conversation in enumerate(conversations):
+            frm = conversation["from"]
+            role = self.from2role[frm]
+            message = conversation["value"]
+            if frm == 'gpt' and not enable_thinking:
+                text = role + self.empty_think + message
+            else:
+                text = role + message
+            if i < num_conversation - 1 or generation_preface is None:
+                text += self.im_end
+            prompt += text
+            token_ids = self._tokenize_with_image_symbol(text)
+            input_ids.extend(token_ids)
+            label_ids = [self.ignore_id] * len(token_ids)
+            if frm == "gpt" and generation_preface is None:
+                # learning `\n` following `im_end` is meaningless, so the last `\n` token is ignored in label
+                label_ids[self.gpt_token_num:-1] = token_ids[self.gpt_token_num:-1]
+            labels.extend(label_ids)
+
+        assert self._tokenize_with_image_symbol(prompt) == input_ids
+        assert len(input_ids) == len(labels)
+
+        if conversations[-1]['from'] == "gpt" and generation_preface is None:
+            # remove the last `\n` following `im_end` in input_ids
+            input_ids.pop()
+            labels.pop()
+
+        return prompt, input_ids, labels
+
+    def format_query(self, query, generation_preface="", enable_thinking=False):
+        prompt, input_ids, _ = self.format([{
+            "from": "human",
+            "value": query
+        }], generation_preface=generation_preface, enable_thinking=enable_thinking)
+
+        return prompt, input_ids
+

configuration_yak.py

@@ -0,0 +1,63 @@
+from typing import Any
+from typing import Union, Optional
+
+from transformers.configuration_utils import PretrainedConfig
+
+__all__ = ["YakConfig"]
+
+
+class YakConfig(PretrainedConfig):
+    """This is the configuration class to store the configuration of an [`YakModel`].
+
+    Args:
+    """
+
+    model_type: str = "yak"
+
+    def __init__(
+        self,
+        in_channels: int = 16,
+        out_channels: int = 16,
+        vec_in_dim: int = 1536,
+        context_in_dim: int = 3072,
+        hidden_size: int = 1536,
+        mlp_ratio: int = 4,
+        num_heads: int = 12,
+        depth: int = 6,
+        depth_single_blocks: int = 12,
+        axes_dim: list = [16, 56, 56],
+        theta: int = 10_000,
+        qkv_bias: bool = True,
+        guidance_embed: bool = False,
+        checkpoint: bool = False,
+        txt_type: str = "refiner",
+        timestep_shift: bool = False,
+        base_shift: float = 0.5,
+        max_shift: float = 1.15,
+        vae_config: Optional[Union[PretrainedConfig, dict]] = None,
+        **kwargs: Any,
+    ):
+        super().__init__(**kwargs)
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.vec_in_dim = vec_in_dim
+        self.context_in_dim = context_in_dim
+        self.hidden_size = hidden_size
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.depth = depth
+        self.depth_single_blocks = depth_single_blocks
+        self.axes_dim = axes_dim
+        self.theta = theta
+        self.qkv_bias = qkv_bias
+        self.guidance_embed = guidance_embed
+        self.checkpoint = checkpoint
+        self.txt_type = txt_type
+        self.timestep_shift = timestep_shift
+        self.base_shift = base_shift
+        self.max_shift = max_shift
+
+        self.vae_config = vae_config
+
+
+    

model-00001-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:70d8644198c2531f6466b3aea7e34953fb9c5b67e04c71642f44a7b1e062b01b
+size 4061178424

model-00002-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c099dec97f53b4a18c79a206a6e2eed0893c4e8c91c701e995d85cdf53012d5f
+size 4973530020

model-00003-of-00003.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:744d9cdf3f96e946dd761454e1743992ef79b7629e2a285e086678c186db3506
+size 1212709496

modeling_aimv2.py

@@ -0,0 +1,385 @@
+# adapted from https://huggingface.co/apple/aimv2-huge-patch14-448 (modification: add gradient checkpoint support)
+from typing import Optional, Tuple, Union
+
+import torch
+from torch import nn
+from torch.nn import functional as F
+from transformers.modeling_outputs import BaseModelOutputWithNoAttention
+from transformers.modeling_utils import PreTrainedModel
+from flash_attn.layers.rotary import apply_rotary_emb
+from flash_attn import flash_attn_varlen_func
+
+from .configuration_aimv2 import AIMv2Config
+
+
+__all__ = ["AIMv2Model"]
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(dim))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
+    def extra_repr(self) -> str:
+        return f"{tuple(self.weight.shape)}, eps={self.eps}"
+
+    def _norm(self, x: torch.Tensor) -> torch.Tensor:
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+
+class AIMv2SwiGLUFFN(nn.Module):
+    def __init__(self, config: AIMv2Config):
+        super().__init__()
+        hidden_features = config.intermediate_size
+        in_features = config.hidden_size
+        bias = config.use_bias
+
+        self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
+        self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
+        self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = F.silu(self.fc1(x)) * self.fc3(x)
+        x = self.fc2(x)
+        return x
+
+
+# copied from qwen2.5-vl
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+    
+# Note: in qwen2-vl and qwen2.5-vl, 3d convolution is used.
+class AIMv2PatchEmbed(nn.Module):
+    def __init__(self, config: AIMv2Config):
+        super().__init__()
+        self.config = config
+        self.proj = nn.Conv2d(
+            config.num_channels,
+            config.hidden_size,
+            kernel_size=(config.patch_size, config.patch_size),
+            stride=(config.patch_size, config.patch_size),
+        )
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = x.view(-1, self.config.num_channels * self.config.temporal_patch_size, self.config.patch_size, self.config.patch_size)
+        x = self.proj(x).view(-1, self.config.hidden_size) #.flatten(2).transpose(1, 2) # token_len x hidden_size
+        x = self.norm(x)
+        return x
+
+class AIMv2ViTPreprocessor(nn.Module):
+    def __init__(self, config: AIMv2Config):
+        super().__init__()
+
+        num_patches = (config.image_size // config.patch_size) ** 2
+
+        self.patchifier = AIMv2PatchEmbed(config)
+
+        self.preserve_original_pe = config.preserve_original_pe
+        self.hidden_stride = config.hidden_stride
+
+        if self.preserve_original_pe:
+            self.interpolate_pe_method = config.interpolate_pe_method
+            self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.hidden_size)))
+
+    def forward(self, x: torch.Tensor, grid_thws: Optional[torch.Tensor] = None) -> torch.Tensor:
+        tokens = self.patchifier(x)
+
+        if self.preserve_original_pe:
+            assert grid_thws is not None
+            pos_embed_new = torch.zeros_like(tokens)
+            if self.interpolate_pe_method == 'one_dim':
+                pos_embed = self.pos_embed.transpose(1,2).to(tokens.device)
+            elif self.interpolate_pe_method == 'two_dim':
+                ori_h = ori_w = int(self.pos_embed.shape[1] ** 0.5)
+                pos_embed = self.pos_embed.reshape(1, ori_h, ori_w, -1).permute(0,3,1,2)
+            else:
+                raise TypeError("The interpolation method for pe should be one_dim, two_dim.")
+            cnt = 0
+            for t, h, w in grid_thws:
+                num_patches = h * w
+                thw = t * h * w
+                if self.interpolate_pe_method == 'one_dim':
+                    pe = F.interpolate(pos_embed, size=num_patches, mode='linear', align_corners=False).transpose(1,2)
+                elif self.interpolate_pe_method == 'two_dim':
+                    # 1, 1024, 32, 32
+                    pe = F.interpolate(pos_embed, size=(h,w), mode='bicubic', align_corners=False)
+                    # 1, 1024, 1024
+                    pe = pe.permute(0,2,3,1).reshape(1, h*w, -1)
+                # 1024, 1024
+                pe = pe[0].repeat(t,1)
+                # 1, 16, 2, 16, 2, 1024
+                pe = pe.reshape(t, h//self.hidden_stride, self.hidden_stride, w//self.hidden_stride, self.hidden_stride, -1)
+                # 1024, 1024
+                pe = pe.permute(0,1,3,2,4,5).reshape(thw,-1)
+                pos_embed_new[cnt:cnt+thw] = pe
+
+                cnt += thw
+
+            tokens = tokens + pos_embed_new
+        return tokens
+
+# copied from qwen2.5-vl
+def apply_rotary_pos_emb_flashatt(
+    q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    cos = cos.chunk(2, dim=-1)[0].contiguous()
+    sin = sin.chunk(2, dim=-1)[0].contiguous()
+    q_embed = apply_rotary_emb(q.float(), cos.float(), sin.float()).type_as(q)
+    k_embed = apply_rotary_emb(k.float(), cos.float(), sin.float()).type_as(k)
+    return q_embed, k_embed
+
+class AIMv2FlashAttention2(nn.Module):
+    def __init__(self, config: AIMv2Config) -> None:
+        super().__init__()
+        dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
+        self.proj = nn.Linear(dim, dim, bias=config.use_bias)
+
+        self.use_rope = not config.disable_rope
+        
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+
+        seq_length = hidden_states.shape[0]
+        q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        if self.use_rope:
+            cos, sin = position_embeddings
+            q, k = apply_rotary_pos_emb_flashatt(q.unsqueeze(0), k.unsqueeze(0), cos, sin)
+            q = q.squeeze(0)
+            k = k.squeeze(0)
+
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+            seq_length, -1
+        )
+        attn_output = self.proj(attn_output)
+        return attn_output
+
+class AIMv2Block(nn.Module):
+    def __init__(self, config: AIMv2Config):
+        super().__init__()
+        self.attn = AIMv2FlashAttention2(config)
+        self.norm_1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.mlp = AIMv2SwiGLUFFN(config)
+        self.norm_2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self, x: torch.Tensor, cu_seqlens: torch.Tensor, position_embeddings: torch.Tensor
+    ) -> torch.Tensor:
+        x = x + self.attn(self.norm_1(x), cu_seqlens=cu_seqlens, position_embeddings=position_embeddings)
+        x = x + self.mlp(self.norm_2(x))
+        return x
+
+
+class AIMv2Transformer(nn.Module):
+    def __init__(self, config: AIMv2Config):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [AIMv2Block(config) for _ in range(config.num_hidden_layers)]
+        )
+        self.post_trunk_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.gradient_checkpointing = False
+
+        self.rotary_pos_emb = VisionRotaryEmbedding(config.hidden_size // config.num_attention_heads // 2)
+        
+        self.hidden_stride = config.hidden_stride
+        self.patch_size = config.patch_size
+        self.window_size = config.window_size
+        self.spatial_merge_unit = config.hidden_stride * config.hidden_stride
+        
+        self.fullatt_block_indexes = config.fullatt_block_indexes
+
+    # copied from qwen2.5_vl
+    def rot_pos_emb(self, grid_thw):
+        pos_ids = []
+        for t, h, w in grid_thw:
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // self.hidden_stride,
+                self.hidden_stride,
+                w // self.hidden_stride,
+                self.hidden_stride,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // self.hidden_stride,
+                self.hidden_stride,
+                w // self.hidden_stride,
+                self.hidden_stride,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_thw[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+        return rotary_pos_emb
+
+    def get_window_index(self, grid_thw):
+        window_index: list = []
+        cu_window_seqlens: list = [0]
+        window_index_id = 0
+        vit_merger_window_size = self.window_size // self.hidden_stride // self.patch_size # patch (after merge) number in each window
+
+        for grid_t, grid_h, grid_w in grid_thw:
+            llm_grid_h, llm_grid_w = (
+                grid_h // self.hidden_stride, # number of patch after merge
+                grid_w // self.hidden_stride,
+            )
+            index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
+            pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
+            pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
+            num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
+            num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
+            index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
+            index_padded = index_padded.reshape(
+                grid_t,
+                num_windows_h,
+                vit_merger_window_size,
+                num_windows_w,
+                vit_merger_window_size,
+            )
+            index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
+                grid_t,
+                num_windows_h * num_windows_w,
+                vit_merger_window_size,
+                vit_merger_window_size,
+            )
+            seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
+            index_padded = index_padded.reshape(-1)
+            index_new = index_padded[index_padded != -100]
+            window_index.append(index_new + window_index_id)
+            cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
+            cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
+            window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
+        window_index = torch.cat(window_index, dim=0)
+
+        return window_index, cu_window_seqlens
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        grid_thws: torch.Tensor,
+        output_hidden_states: bool = False,
+    ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, ...]]]:
+        # RoPE, modified from qwen2.5_vl
+        rotary_pos_emb = self.rot_pos_emb(grid_thws)
+        window_index, cu_window_seqlens = self.get_window_index(grid_thws)
+        cu_window_seqlens = torch.tensor(
+            cu_window_seqlens,
+            device=tokens.device,
+            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
+
+        seq_len, _ = tokens.size()
+        tokens = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        tokens = tokens[window_index, :, :]
+        tokens = tokens.reshape(seq_len, -1)
+        rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        rotary_pos_emb = rotary_pos_emb[window_index, :, :]
+        rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+
+        cu_seqlens = torch.repeat_interleave(grid_thws[:, 1] * grid_thws[:, 2], grid_thws[:, 0]).cumsum(
+            dim=0,
+            # Select dtype based on the following factors:
+            #  - FA2 requires that cu_seqlens_q must have dtype int32
+            #  - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
+            # See https://github.com/huggingface/transformers/pull/34852 for more information
+            dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
+        )
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        reverse_indices = torch.argsort(window_index)
+        
+        hidden_states = () if output_hidden_states else None
+        for index, block in enumerate(self.blocks):
+            if self.fullatt_block_indexes is None or index in self.fullatt_block_indexes:
+                cu_seqlens_tmp = cu_seqlens
+            else:
+                cu_seqlens_tmp = cu_window_seqlens
+            if self.gradient_checkpointing and self.training:
+                tokens = self._gradient_checkpointing_func(block.__call__, tokens, cu_seqlens_tmp, position_embeddings)
+            else:
+                tokens = block(tokens, cu_seqlens_tmp, position_embeddings)
+            if output_hidden_states:
+                tokens_ = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+                hidden_states += (tokens_[reverse_indices,:].reshape(seq_len, -1),)
+        tokens = self.post_trunk_norm(tokens)
+        tokens = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
+        tokens = tokens[reverse_indices,:].reshape(seq_len, -1)
+        
+        return tokens, hidden_states
+
+
+class AIMv2PretrainedModel(PreTrainedModel):
+    config_class = AIMv2Config
+    base_model_prefix = "aimv2"
+    supports_gradient_checkpointing = True
+    main_input_name = "pixel_values"
+    _no_split_modules = ["AIMv2ViTPreprocessor", "AIMv2Block"]
+    _supports_sdpa = True
+
+
+class AIMv2Model(AIMv2PretrainedModel):
+    def __init__(self, config: AIMv2Config):
+        super().__init__(config)
+        self.preprocessor = AIMv2ViTPreprocessor(config)
+        self.trunk = AIMv2Transformer(config)
+
+    def forward(
+        self,
+        pixel_values: torch.Tensor,
+        grid_thws: torch.Tensor,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+    ) -> Union[
+        Tuple[torch.Tensor],
+        Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
+        BaseModelOutputWithNoAttention,
+    ]:
+        if output_hidden_states is None:
+            output_hidden_states = self.config.output_hidden_states
+        if return_dict is None:
+            return_dict = self.config.use_return_dict
+
+        x = self.preprocessor(pixel_values, grid_thws=grid_thws)
+        
+        x, hidden_states = self.trunk(
+            x, grid_thws=grid_thws, output_hidden_states=output_hidden_states
+        )
+
+        if not return_dict:
+            res = (x,)
+            res += (hidden_states,) if output_hidden_states else ()
+            return res
+
+        return BaseModelOutputWithNoAttention(
+            last_hidden_state=x,
+            hidden_states=hidden_states,
+        )

modeling_ovis_u1.py

@@ -0,0 +1,921 @@
+import logging
+import math
+from datetime import datetime
+from importlib import import_module
+from typing import List, Union, Optional, Dict
+
+import numpy as np
+import PIL.Image
+import torch
+from torch import Tensor
+from torch.nn import init
+from torch.nn.functional import softmax, gumbel_softmax, pad
+from torchvision import transforms
+import transformers
+from transformers import AutoImageProcessor
+from transformers import PreTrainedModel, AutoConfig, AutoModel, AutoTokenizer, AutoModelForCausalLM
+from transformers.generation.utils import GenerateOutput
+from transformers import CLIPImageProcessor
+
+from .modeling_aimv2 import AIMv2Model
+from .configuration_ovis_u1 import BaseVisualTokenizerConfig, Aimv2VisualTokenizerConfig
+from .configuration_ovis_u1 import OvisU1Config, ConversationFormatter
+from .configuration_ovis_u1 import IGNORE_ID, IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS, IMAGE_TOKEN_ID, VIDEO_TOKEN_ID
+
+# ----------------------------------------------------------------------
+#                            Visual Tokenizer
+# ----------------------------------------------------------------------
+class BaseVisualTokenizer(PreTrainedModel):
+    base_model_prefix = "backbone"
+    main_input_name = None
+    _image_processor_class = None
+    _image_processor_kwargs = {}
+    _backbone_class = None
+
+    def __init__(self, config: BaseVisualTokenizerConfig, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        if kwargs.get('train_from_scratch'):
+            # for key in self._image_processor_kwargs.keys():
+            #     self._image_processor_kwargs[key] = getattr(self.config, key, self._image_processor_kwargs[key])
+            image_processor = self._image_processor_class.from_pretrained(kwargs['backbone_name_or_path'],
+                                                                               **self._image_processor_kwargs)
+
+            self.backbone = self._backbone_class.from_pretrained(kwargs['backbone_name_or_path'], **self.config.backbone_kwargs)
+            self.config.backbone_config = self.backbone.config
+
+            config = image_processor.to_dict()
+            if getattr(self.config, 'image_processor_new_kwargs', None) is not None:
+                for key in self.config.image_processor_new_kwargs.keys():
+                    config[key] = self.config.image_processor_new_kwargs[key]
+            if 'patch_size' not in config:
+                assert getattr(self.backbone.config, 'patch_size'), "Patch size must be set."
+                config['patch_size'] = self.backbone.config.patch_size
+            self.image_processor = self._image_processor_class.from_dict(config)
+
+        else:
+            self.image_processor = AutoImageProcessor.from_pretrained(kwargs['image_processor_name_or_path'])
+            self.backbone = AutoModel.from_config(self.config.backbone_config)
+        head_dim = self.config.vocab_size - len(IMAGE_INDICATOR_IDS)  # reserved tokens for IMAGE_INDICATORS
+        self.head = torch.nn.Sequential(
+            torch.nn.Linear(
+                self.backbone.config.hidden_size * self.config.hidden_stride * self.config.hidden_stride, head_dim,
+                bias=False
+            ),
+            torch.nn.LayerNorm(head_dim)
+        )
+        assert all((self.image_processor.do_resize,
+                    not getattr(self.image_processor, 'do_center_crop', False),
+                    self.image_processor.do_rescale,
+                    self.image_processor.do_normalize
+                    )), f"image_processor `{self.image_processor}` is not supported currently"
+
+    def get_backbone(self):
+        return self.backbone
+
+    def get_monitor_tensors(self):
+        raise NotImplementedError
+
+    def get_image_processor(self):
+        return self.image_processor
+
+    def mock_input(self):
+        height, width = self.get_image_size()
+        return torch.zeros(1, 3, height, width), self.construct_image_placeholders((1, 1))
+
+    def get_head(self):
+        return self.head
+
+    def get_image_size(self):
+        raise NotImplementedError
+
+    @staticmethod
+    def construct_image_placeholders(grid, data_type='image'):
+        if data_type == 'image':
+            image_placeholders = [IMAGE_INDICATOR_IDS[0], IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS[1]]
+        elif data_type == 'video':
+            image_placeholders = [IMAGE_INDICATOR_IDS[2], IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS[2]]
+        else:
+            raise TypeError
+        
+        return image_placeholders
+
+    @staticmethod
+    def _partition(img_size, grid):
+        w, h = img_size
+        row_height = h // grid[0]
+        col_width = w // grid[1]
+
+        partition = []
+        for row in range(grid[0]):
+            for col in range(grid[1]):
+                left = col * col_width
+                upper = row * row_height
+                right = w if col == grid[1] - 1 else (col + 1) * col_width
+                lower = h if row == grid[0] - 1 else (row + 1) * row_height
+                partition.append((left, upper, right, lower))
+
+        return partition
+
+    @staticmethod
+    def get_best_grid(img_size, side, max_partition, covering_threshold):
+
+        def _covering_area(left, upper, right, lower, side):
+            w = right - left
+            h = lower - upper
+            w, h = max(w, h), min(w, h)
+            if w > side:
+                h = h / w * side
+                w = side
+            return w * h
+
+        img_area = img_size[0] * img_size[1]
+
+        candidate_grids = []
+        for i in range(1, max_partition + 1):
+            for j in range(1, max_partition + 1):
+                if i * j <= max_partition:
+                    candidate_grids.append((i, j))
+
+        all_grids = []
+        good_grids = []
+        for grid in candidate_grids:
+            partition = BaseVisualTokenizer._partition(img_size, grid)
+            covering_ratio = sum([_covering_area(*p, side) for p in partition]) / img_area
+            assert covering_ratio <= 1.0
+            all_grids.append((grid, covering_ratio))
+            if covering_ratio > covering_threshold:
+                good_grids.append((grid, covering_ratio))
+
+        if len(good_grids) > 0:
+            # pick the good partition with minimum #sub_images and break the tie using covering_ratio
+            return sorted(good_grids, key=lambda x: (x[0][0] * x[0][1], -x[1]))[0][0]
+        else:
+            # pick the partition with maximum covering_ratio and break the tie using #sub_images
+            return sorted(all_grids, key=lambda x: (-x[1], x[0][0] * x[0][1]))[0][0]
+
+    def preprocess_image(self, image: PIL.Image.Image, max_partition=4, covering_threshold=0.9, convert_to_rgb=True):
+        def _preprocess(img: PIL.Image.Image, side):
+            # first resize and preprocess
+            w, h = img.size
+            if w == h:
+                new_width = new_height = side
+            elif w > h:
+                new_width = side
+                new_height = int(h / w * new_width)
+            else:
+                new_height = side
+                new_width = int(w / h * new_height)
+            new_size = dict(height=new_height, width=new_width)
+            pixel_values = self.image_processor.preprocess(img, size=new_size, return_tensors='pt')['pixel_values']
+
+            # then pad to square
+            square_values = torch.zeros([1, 3, side, side], dtype=pixel_values.dtype, device=pixel_values.device)
+            new_height, new_width = pixel_values.shape[2:]
+            if new_height == new_width:
+                square_values[:, :, :, :] = pixel_values
+            elif new_height > new_width:
+                from_index = (side - new_width) // 2
+                square_values[:, :, :, from_index:from_index + new_width] = pixel_values
+            else:
+                from_index = (side - new_height) // 2
+                square_values[:, :, from_index:from_index + new_height, :] = pixel_values
+
+            return square_values
+
+        if convert_to_rgb and image.mode != 'RGB':
+            image = image.convert('RGB')
+
+        sides = self.get_image_size()
+        if sides[0] != sides[1]:
+            raise ValueError('get_image_size() returns non-square size')
+        side = sides[0]
+        grid = self.get_best_grid(image.size, side, max_partition, covering_threshold)
+        partition = self._partition(image.size, grid)
+        crops = [image.crop(p) for p in partition]
+        if len(crops) > 1:
+            crops.insert(0, image)
+        pixel_values = torch.cat([_preprocess(crop, side) for crop in crops], dim=0)
+        image_placeholders = self.construct_image_placeholders(grid)
+        return pixel_values, image_placeholders
+
+    def get_backbone_layer(self, index):
+        if 'aimv2' in self.config.model_type:
+            return self.backbone.trunk.blocks[index]
+        else:
+            return self.backbone.vision_model.encoder.layers[index]
+
+    def tokenize(self, logits):
+        def st_argmax(y_soft, dim):  # straight-through softmax
+            index = y_soft.max(dim, keepdim=True)[1]
+            y_hard = torch.zeros_like(y_soft, memory_format=torch.legacy_contiguous_format).scatter_(dim, index, 1.0)
+            ret = y_hard - y_soft.detach() + y_soft
+            return ret
+
+        if self.config.tokenize_function == 'softmax':
+            tokens = softmax(logits, dim=-1, dtype=torch.float32).to(logits.dtype)
+        elif self.config.tokenize_function == 'gumbel_argmax':
+            tokens = gumbel_softmax(logits, tau=self.config.tau, hard=True)
+        elif self.config.tokenize_function == 'st_argmax':
+            tokens = st_argmax(logits, dim=-1)
+        else:
+            raise ValueError(
+                f'Invalid `max_type`, expected softmax or gumbel_argmax or st_argmax, but got {self.config.tokenize_function}')
+        return tokens
+
+    def encode(self, pixel_values):
+        output = self.backbone(pixel_values, output_hidden_states=True, return_dict=True)
+        features = output.hidden_states[-1]
+        if self.config.drop_cls_token:
+            features = features[:, 1:, :]
+
+        # merge number of `hidden_stride * hidden_stride` hidden states together to reduce token sequence length
+        # e.g., for hidden_stride=3, this leads to a token length reduction: 729 -> 81 for siglip
+        if self.config.hidden_stride > 1:
+            n, l, d = features.shape  # this `d` maybe different from the above `d
+            sqrt_l = int(l ** 0.5)
+            assert sqrt_l ** 2 == l, "The token sequence length should be a perfect square."
+            features = features.reshape(n, sqrt_l, sqrt_l, d)
+            pl = (self.config.hidden_stride - (sqrt_l % self.config.hidden_stride)) % self.config.hidden_stride
+            features = pad(features, (0, 0, 0, pl, 0, pl), "constant", 0)
+            sqrt_l += pl
+            features = features.reshape(n, sqrt_l // self.config.hidden_stride, self.config.hidden_stride,
+                                        sqrt_l // self.config.hidden_stride, self.config.hidden_stride, d)
+            features = features.permute(0, 1, 3, 2, 4, 5)  # [n, sqrt_l/hs, sqrt_l/hs, hs, hs, d]
+            features = features.flatten(3)  # [n, sqrt_l/hs, sqrt_l/hs, hs*hs*d]
+            features = features.reshape(
+                n, -1, self.config.hidden_stride * self.config.hidden_stride * d)
+
+        return features
+
+    def forward(self, pixel_values) -> torch.Tensor:  # [BatchSize, ImageShape] -> [BatchSize, #Token, VocabSize]
+        features = self.encode(pixel_values)
+        logits = self.head(features)
+        tokens = self.tokenize(logits)
+        # tokens' shape is [BatchSize, #Token, VocabSize-5], so padding with [BatchSize, #Token, 5], after
+        # which, tokens' shape should become [BatchSize, #Token, VocabSize]
+        batch_size, token_len, _ = tokens.shape
+        padding_tensor = torch.zeros(size=(batch_size, token_len, len(IMAGE_INDICATOR_IDS)),
+                                     dtype=tokens.dtype,
+                                     device=tokens.device,
+                                     layout=tokens.layout,
+                                     requires_grad=False)
+        tokens = torch.cat((tokens, padding_tensor), dim=2)
+        return tokens
+
+class Aimv2VisualTokenizer(BaseVisualTokenizer):
+    config_class = Aimv2VisualTokenizerConfig
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["AIMv2ViTPreprocessor", "AIMv2Block"]
+    _image_processor_class = CLIPImageProcessor
+    _image_processor_kwargs = dict(do_center_crop=False, crop_size={'height': -1, 'width': -1}, size={'shortest_edge':-1})
+    _backbone_class = AIMv2Model
+    
+    # Copied from qwen2_vl
+    def smart_resize(self, 
+        height: int, width: int, factor: int = 28, min_pixels: int = 56 * 56, max_pixels: int = 14 * 14 * 4 * 1280
+    ):
+        """Rescales the image so that the following conditions are met:
+
+        1. Both dimensions (height and width) are divisible by 'factor'.
+
+        2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
+
+        3. The aspect ratio of the image is maintained as closely as possible.
+
+        """
+        
+        if height < factor or width < factor:
+            print(f"height:{height} or width:{width} must be larger than factor:{factor}")
+            if height < width:
+                width = round(factor/height*width)
+                height = factor
+            else:
+                height = round(factor/width*height)
+                width = factor
+
+        elif max(height, width) / min(height, width) > 200:
+            print(
+                f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}"
+            )
+            if height > width:
+                height = 200 * width
+            else:
+                width = 200 * height
+
+        h_bar = round(height / factor) * factor
+        w_bar = round(width / factor) * factor
+        if h_bar * w_bar > max_pixels:
+            beta = math.sqrt((height * width) / max_pixels)
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+        elif h_bar * w_bar < min_pixels:
+            beta = math.sqrt(min_pixels / (height * width))
+            h_bar = math.ceil(height * beta / factor) * factor
+            w_bar = math.ceil(width * beta / factor) * factor
+        return h_bar, w_bar
+
+    def get_monitor_tensors(self):
+        return dict(
+            backbone_bottom=self.backbone.trunk.blocks[0].attn.qkv.weight,
+            backbone_top=self.backbone.trunk.blocks[-1].attn.qkv.weight,
+            head=self.head[0].weight
+        )
+
+    def get_min_image_size(self):
+        min_pixels = self.image_processor.min_pixels
+        max_pixels = self.image_processor.max_pixels
+        height = int(min_pixels**0.5)
+        width = int(min_pixels**0.5)
+        patch_size = self.image_processor.patch_size
+        hidden_stride = self.image_processor.hidden_stride
+        height, width = self.smart_resize(height, width, patch_size * hidden_stride, min_pixels, max_pixels)
+        return height, width
+    
+    def get_image_size(self):
+        min_pixels = self.image_processor.min_pixels
+        max_pixels = self.image_processor.max_pixels
+        num_pixels = (min_pixels+max_pixels) / 2
+        height = int(num_pixels**0.5)
+        width = int(num_pixels**0.5)
+        patch_size = self.image_processor.patch_size
+        hidden_stride = self.image_processor.hidden_stride
+        height, width = self.smart_resize(height, width, patch_size * hidden_stride, min_pixels, max_pixels)
+        return height, width
+
+    def get_token_length(self, width: int,
+                            height: int, 
+                            n_frames: int = 1,
+                            num_images: int = 1):
+        patch_size = self.image_processor.patch_size
+        temporal_patch_size = self.image_processor.temporal_patch_size
+        hidden_stride = self.image_processor.hidden_stride
+        min_pixels = self.image_processor.min_pixels
+        max_pixels = self.image_processor.max_pixels
+        
+        max_pixels = max_pixels // num_images
+        min_pixels = min(max_pixels, min_pixels)
+        
+        resized_height, resized_width = height, width
+        resized_height, resized_width = self.smart_resize(
+                    height,
+                    width,
+                    factor=patch_size * hidden_stride,
+                    min_pixels=min_pixels,
+                    max_pixels=max_pixels,
+                )
+       
+        if n_frames % temporal_patch_size != 0:
+            n_frames = n_frames + temporal_patch_size - 1
+        grid_t = n_frames // temporal_patch_size
+        grid_h, grid_w = resized_height // patch_size // hidden_stride, resized_width // patch_size // hidden_stride
+
+        return grid_t * grid_w * grid_h
+
+    def mock_input(self):
+        height, width = self.get_min_image_size()
+        return torch.zeros(1, 3, height, width), self.construct_image_placeholders((1, 1))
+
+    def preprocess_image(self, images: Union[PIL.Image.Image, List[PIL.Image.Image]], 
+                            convert_to_rgb: Optional[bool] = True, 
+                            num_images: Optional[int] = 1,
+                            min_pixels: Optional[int] = None, 
+                            max_pixels: Optional[int] = None,
+                            multimodal_type: Optional[str] = 'single_image'):
+
+
+        patch_size = self.image_processor.patch_size # 14
+        temporal_patch_size = self.image_processor.temporal_patch_size # 1
+        hidden_stride = self.image_processor.hidden_stride # 2
+        min_pixels = min_pixels or self.image_processor.min_pixels # 200704
+        max_pixels = max_pixels or self.image_processor.max_pixels # 3211264
+        
+        max_pixels = max_pixels // num_images
+        min_pixels = min(max_pixels, min_pixels)
+
+        if not isinstance(images, list):
+            images = [images]
+        if multimodal_type == 'video':
+            assert len(images) >= 1
+        else:
+            assert len(images) == 1
+        images = [image.convert("RGB") if convert_to_rgb and image.mode != 'RGB' else image for image in images ]
+        # images = [np.array(image) for image in images]
+        
+        width, height = images[0].size
+        resized_height, resized_width = height, width
+        processed_images = []
+        for image in images:
+            resized_height, resized_width = self.smart_resize(
+                height,
+                width,
+                factor=patch_size * hidden_stride,
+                min_pixels=min_pixels,
+                max_pixels=max_pixels,
+            )
+            new_size = dict(height=resized_height, width=resized_width)
+            image_pt = self.image_processor.preprocess(image, size=new_size, return_tensors="np")['pixel_values'][0]
+            
+            processed_images.append(image_pt)
+
+        patches = np.array(processed_images)
+        # if data_format == ChannelDimension.LAST:
+        #     patches = patches.transpose(0, 3, 1, 2)
+        if patches.shape[0] % temporal_patch_size != 0:
+            repeats = np.repeat(patches[-1][np.newaxis], temporal_patch_size - 1, axis=0)
+            patches = np.concatenate([patches, repeats], axis=0)
+        channel = patches.shape[1]
+        grid_t = patches.shape[0] // temporal_patch_size # 1
+        grid_h, grid_w = resized_height // patch_size, resized_width // patch_size # 32, 32
+        
+        patches = patches.reshape(
+            grid_t,
+            temporal_patch_size,
+            channel,
+            grid_h // hidden_stride,
+            hidden_stride,
+            patch_size,
+            grid_w // hidden_stride,
+            hidden_stride,
+            patch_size,
+        )
+        patches = patches.transpose(0, 3, 6, 4, 7, 2, 1, 5, 8)
+        flatten_patches = patches.reshape(
+            grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size
+        )
+        # 1024, 588
+
+        image_placeholders = self.construct_image_placeholders((1, 1), data_type='video' if multimodal_type=='video' else 'image') # [-301, -300, -302, -305]
+        
+        # print(flatten_patches.shape, len(images))
+        return torch.tensor(flatten_patches), torch.tensor([[grid_t, grid_h, grid_w]]), image_placeholders
+    
+    def encode(self, pixel_values, grid_thws):
+        output = self.backbone(pixel_values, grid_thws, output_hidden_states=True, return_dict=True)
+        features = output.hidden_states[-1]
+        # default: false
+        # if self.config.drop_cls_token:
+        #     features = features[:, 1:, :]
+        
+        # refer to qwen2.5-vl patchmerger
+        seq_len, _ = features.shape
+        features = features.reshape(seq_len//(self.config.hidden_stride ** 2), -1)
+        
+        return features
+
+    def forward(self, pixel_values, grid_thws) -> torch.Tensor:  # [BatchSize, ImageShape] -> [BatchSize, #Token, VocabSize]
+        features = self.encode(pixel_values, grid_thws)
+        logits = self.head(features)
+        tokens = self.tokenize(logits)
+        # tokens' shape is [#Token, VocabSize-5], so padding with [#Token, 5], after
+        # which, tokens' shape should become [#Token, VocabSize];
+        # this is different from original aimv2 which has [BatchSize, #Token, VocabSize-5]
+        token_len, _ = tokens.shape
+        padding_tensor = torch.zeros(size=(token_len, len(IMAGE_INDICATOR_IDS)),
+                                     dtype=tokens.dtype,
+                                     device=tokens.device,
+                                     layout=tokens.layout,
+                                     requires_grad=False)
+        tokens = torch.cat((tokens, padding_tensor), dim=1)
+        return tokens
+
+AutoModel.register(Aimv2VisualTokenizerConfig, Aimv2VisualTokenizer)
+
+
+
+
+# ----------------------------------------------------------------------
+#                           Visual Generator
+# ----------------------------------------------------------------------
+from .configuration_yak import YakConfig
+from .modeling_yak import YakModel
+AutoConfig.register("yak", YakConfig)
+AutoModel.register(YakConfig, YakModel)
+
+
+
+# ----------------------------------------------------------------------
+#                               OvisU1
+# ----------------------------------------------------------------------
+class VisualEmbedding(torch.nn.Embedding):
+    def forward(self, visual_tokens: Tensor) -> Tensor:
+        if visual_tokens.dtype in [torch.int8, torch.int16, torch.int32, torch.int64, torch.long]:
+            return super().forward(visual_tokens)
+        return torch.matmul(visual_tokens, self.weight)
+
+    def reset_parameters(self, mean=0., std=1.) -> None:
+        init.normal_(self.weight, mean=mean, std=std)
+        self._fill_padding_idx_with_zero()
+
+
+class OvisU1PreTrainedModel(PreTrainedModel):
+    config_class = OvisU1Config
+    base_model_prefix = "ovis_u1"
+
+
+class OvisU1(OvisU1PreTrainedModel):
+    
+    def __init__(self, config: OvisU1Config, *inputs, **kwargs):
+        super().__init__(config, *inputs, **kwargs)
+        attn_kwargs = dict()
+        if self.config.llm_attn_implementation:
+            attn_kwargs['attn_implementation'] = self.config.llm_attn_implementation
+        self.llm = AutoModelForCausalLM.from_config(self.config.llm_config, **attn_kwargs)
+        assert self.config.hidden_size == self.llm.config.hidden_size, "hidden size mismatch"
+        self.text_tokenizer = AutoTokenizer.from_pretrained(self.config.name_or_path)
+        self.visual_tokenizer = AutoModel.from_config(self.config.visual_tokenizer_config,
+                                                    image_processor_name_or_path=self.config.name_or_path)
+        self.visual_generator = AutoModel.from_config(self.config.visual_generator_config)
+        self.vte = VisualEmbedding(self.config.visual_tokenizer_config.vocab_size, self.config.hidden_size,
+                                    device=self.visual_tokenizer.device, dtype=self.visual_tokenizer.dtype)
+
+        def _merge_modules(modules_list: tuple):
+            merged_modules = []
+            for modules in modules_list:
+                merged_modules.extend(modules if modules else [])
+            return merged_modules
+
+        self._no_split_modules = _merge_modules((self.llm._no_split_modules, self.visual_tokenizer._no_split_modules))
+        self._skip_keys_device_placement = self.llm._skip_keys_device_placement
+        self._keep_in_fp32_modules = _merge_modules(
+            (self.llm._keep_in_fp32_modules, self.visual_tokenizer._keep_in_fp32_modules))
+        self._supports_flash_attn_2 = True
+        self.is_parallelizable = all((self.llm.is_parallelizable, self.visual_tokenizer.is_parallelizable, self.visual_generator.is_parallelizable))
+        self.supports_gradient_checkpointing = all(
+            (self.llm.supports_gradient_checkpointing, self.visual_tokenizer.supports_gradient_checkpointing, self.visual_generator.supports_gradient_checkpointing))
+        self._supports_sdpa = all((self.llm._supports_sdpa, self.visual_tokenizer._supports_sdpa, self.visual_generator._supports_sdpa))
+
+    def get_text_tokenizer(self):
+        return self.text_tokenizer
+
+    def get_visual_tokenizer(self):
+        return self.visual_tokenizer
+    
+    def get_visual_generator(self):
+        return self.visual_generator
+
+    def tie_weights(self):
+        if not self.config.disable_tie_weight:
+            self.get_llm().tie_weights()
+
+    def get_lm_head(self):
+        return self.get_llm().get_output_embeddings()
+
+    def get_llm(self):
+        return self.llm
+
+    def get_vte(self):
+        return self.vte
+
+    def get_wte(self):
+        return self.llm.get_input_embeddings()
+
+    def get_conversation_formatter(self) -> ConversationFormatter:
+        if getattr(self, 'conversation_formatter', None) is None:
+            self.conversation_formatter = getattr(import_module(".configuration_ovis_u1", __package__),
+                                                  self.config.conversation_formatter_class)(self.text_tokenizer)
+        return self.conversation_formatter
+
+    def merge_multimodal(
+            self,
+            text_input_ids: torch.Tensor,
+            text_attention_masks: torch.Tensor,
+            text_labels: Optional[torch.Tensor],
+            pixel_values: Optional[torch.Tensor],
+            grid_thws: Optional[torch.Tensor],
+            left_padding: bool = False
+    ):
+        input_device = text_input_ids.device
+        visual_vocab_szie = self.get_visual_tokenizer().config.vocab_size
+        visual_indicator_embeds = self.get_vte()(
+            torch.tensor(
+                list(range(visual_vocab_szie - 5, visual_vocab_szie)),
+                dtype=torch.long,
+                device=self.get_visual_tokenizer().device
+            )
+        ).to(device=input_device)
+
+        if self.training:
+            # When training, to be compatible with deepspeed zero, each sample has to include pixel_value tensor.
+            # For text-only sample, one can simply use a full zero tensor as pixel_value, which will be ignored
+            # (see below in this function); so, the gradient will not be affected.
+            num_images = [x.prod() // (self.visual_tokenizer.config.hidden_stride**2) for x in grid_thws]
+            
+            visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
+
+            visual_embeds_ = torch.split(self.get_vte()(visual_tokens).to(dtype=self.dtype, device=input_device),
+                                        split_size_or_sections=num_images, dim=0)
+            
+
+
+            visual_input_ids_ = torch.split(torch.argmax(visual_tokens, dim=-1).to(device=input_device),
+                                           split_size_or_sections=num_images, dim=0)
+
+
+            visual_labels_ = [torch.full(x.shape, IGNORE_ID, dtype=torch.long, device=input_device) for x in
+                             visual_input_ids_]
+
+            
+            visual_embeds = []
+            visual_input_ids = []
+            visual_labels = []
+            ind = 0
+            for text_input_id in text_input_ids:
+                image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
+                n = len(image_atom_positions)
+                if n > 0:
+                    visual_embeds.append(visual_embeds_[ind:ind+n])
+                    visual_input_ids.append(visual_input_ids_[ind:ind+n])
+                    visual_labels.append(visual_labels_[ind:ind+n])
+                    ind += n
+                else:
+                    visual_embeds.append(visual_embeds_[ind:ind+1])
+                    visual_input_ids.append(visual_input_ids_[ind:ind+1])
+                    visual_labels.append(visual_labels_[ind:ind+1])
+                    ind += 1
+                
+
+        else:
+            # TODO: Not modified yet
+            # When inference, sample can include only text with `None` pixel_value
+            # num_images = [x.shape[0] if x is not None else 0 for x in pixel_values]
+            num_images = [x.prod() // (self.visual_tokenizer.config.hidden_stride**2) if x is not None else 0 for x in grid_thws]
+            if sum(num_images) > 0:
+                visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
+                try:
+                    visual_embeds_ = torch.split(self.get_vte()(visual_tokens).to(dtype=self.dtype, device=input_device),
+                                        split_size_or_sections=num_images, dim=0)
+                except Exception as e:
+                    print(e)
+                    print(pixel_values.shape, grid_thws.shape, visual_tokens.shape, num_images)
+            
+
+                visual_input_ids_ = torch.split(torch.argmax(visual_tokens, dim=-1).to(device=input_device),
+                                            split_size_or_sections=num_images, dim=0)
+
+
+                visual_labels_ = [torch.full(x.shape, IGNORE_ID, dtype=torch.long, device=input_device) for x in
+                                visual_input_ids_]
+                
+                visual_embeds = []
+                visual_input_ids = []
+                visual_labels = []
+                ind = 0
+                for text_input_id in text_input_ids:
+                    image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
+                    n = len(image_atom_positions)
+                    if n > 0:
+                        visual_embeds.append(visual_embeds_[ind:ind+n])
+                        visual_input_ids.append(visual_input_ids_[ind:ind+n])
+                        visual_labels.append(visual_labels_[ind:ind+n])
+                        ind += n
+                    else:
+                        visual_embeds.append(visual_embeds_[ind:ind+1])
+                        visual_input_ids.append(visual_input_ids_[ind:ind+1])
+                        visual_labels.append(visual_labels_[ind:ind+1])
+                        ind += 1
+                        
+            else:
+                # just placeholders
+                visual_embeds = [None] * len(num_images)
+                visual_input_ids = [None] * len(num_images)
+                visual_labels = [None] * len(num_images)
+            
+        # just placeholders
+        if text_labels is None:
+            text_labels = torch.full(text_input_ids.shape, IGNORE_ID, dtype=torch.long, device=input_device)
+
+        input_embeds = []
+        attention_masks = []
+        labels = []
+        input_img_poss = []
+        for text_input_id, text_label, text_attention_mask, visual_embed, visual_input_id, visual_label in zip(
+            text_input_ids, text_labels, text_attention_masks, visual_embeds, visual_input_ids, visual_labels
+        ):
+            placeholder_token_mask = torch.lt(text_input_id, 0)
+            text_embed = self.get_wte()(torch.masked_fill(text_input_id, placeholder_token_mask, 0))
+            for i, indicator_id in enumerate(IMAGE_INDICATOR_IDS):
+                text_embed[text_input_id == indicator_id] = visual_indicator_embeds[i]
+            image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
+            if len(image_atom_positions) > 0:
+                input_embed_parts = []
+                attention_mask_parts = []
+                label_parts = []
+                input_img_pos_parts = []
+                prev_image_atom_position = -1
+                for index, image_atom_position in enumerate(image_atom_positions):
+                    input_embed_parts.append(
+                        text_embed[prev_image_atom_position + 1:image_atom_position, :])
+                    label_parts.append(
+                        text_label[prev_image_atom_position + 1:image_atom_position])
+                    input_img_pos_parts.append(
+                        torch.zeros_like(text_label[prev_image_atom_position + 1:image_atom_position])
+                    )
+                    attention_mask_parts.append(
+                        text_attention_mask[prev_image_atom_position + 1:image_atom_position])
+                    input_embed_parts.append(visual_embed[index])
+                    attention_mask_parts.append(
+                        torch.ones_like(visual_label[index], dtype=torch.bool))
+                    label_parts.append(visual_label[index])
+                    input_img_pos_parts.append(
+                        torch.ones_like(visual_label[index])
+                    )
+                    prev_image_atom_position = image_atom_position
+                if prev_image_atom_position + 1 < text_input_id.shape[0]:
+                    input_embed_parts.append(
+                        text_embed[prev_image_atom_position + 1:, :])
+                    attention_mask_parts.append(
+                        text_attention_mask[prev_image_atom_position + 1:])
+                    label_parts.append(
+                        text_label[prev_image_atom_position + 1:])
+                    input_img_pos_parts.append(
+                        torch.zeros_like(text_label[prev_image_atom_position + 1:])
+                    )
+                input_embed = torch.cat(input_embed_parts, dim=0)
+                attention_mask = torch.cat(attention_mask_parts, dim=0)
+                label = torch.cat(label_parts, dim=0)
+                input_img_pos = torch.cat(input_img_pos_parts, dim=0)
+            else:
+                input_embed = text_embed
+                attention_mask = text_attention_mask
+                label = text_label
+                input_img_pos = torch.zeros_like(text_label)
+                if self.training:
+                    # Make visual_embed & visual_indicator_embeds involved in the backward graph,
+                    # to be compatible with deepspeed zero and ddp.
+                    input_embed += torch.sum(visual_embed[0] * 0.0) + torch.sum(visual_indicator_embeds * 0.0)
+            input_embeds.append(input_embed)
+            attention_masks.append(attention_mask)
+            labels.append(label)
+            input_img_poss.append(input_img_pos)
+
+        batch_input_embeds = self.pad_truncate_sequence(input_embeds, batch_first=True, padding_value=0.0, left_padding=left_padding)
+        batch_attention_mask = self.pad_truncate_sequence(attention_masks, batch_first=True, padding_value=False, left_padding=left_padding)
+        batch_labels = self.pad_truncate_sequence(labels, batch_first=True, padding_value=IGNORE_ID, left_padding=left_padding)
+        batch_input_img_labels = self.pad_truncate_sequence(input_img_poss, batch_first=True, padding_value=0.0, left_padding=left_padding)
+
+        return visual_input_ids, batch_input_embeds, batch_labels, batch_attention_mask, batch_input_img_labels
+
+    def pad_truncate_sequence(self, sequences: List[torch.Tensor], batch_first: bool = True, padding_value: float = 0.0, left_padding: bool = False) -> torch.Tensor:
+        if left_padding == False:
+            pad_sequence = torch.nn.utils.rnn.pad_sequence(sequences, batch_first=batch_first, padding_value=padding_value)
+            return pad_sequence[:,:self.config.multimodal_max_length]
+        else:
+            pad_sequence = torch.nn.utils.rnn.pad_sequence([i.flip(dims=[0]) for i in sequences],batch_first=True, padding_value=padding_value).flip(dims=[1])
+            return pad_sequence[:,-self.config.multimodal_max_length:]
+
+    def preprocess_inputs(
+        self,
+        text_or_conversations: Union[List[Dict], str],
+        images: Optional[Union[List[PIL.Image.Image], List[List[PIL.Image.Image]]]],
+        generation_preface='',
+        return_labels=False,
+        propagate_exception=True,
+        frame_selector=None,
+        multimodal_type="single_image",
+        fix_sample_overall_length_navit=False,
+        min_pixels=None,
+        max_pixels=None,
+        enable_thinking=False
+    ):
+        # convert text to conversations
+        if isinstance(text_or_conversations, str):
+            conversations = [{
+                "from": "human",
+                "value": text_or_conversations
+            }]
+        elif isinstance(text_or_conversations, list):
+            conversations = text_or_conversations
+        else:
+            raise ValueError(f'[{datetime.now()}] Invalid type of `text_or_conversations`, expected `List[Dict]` or `str`,'
+                             f' but got {type(text_or_conversations)}')
+
+        if frame_selector is not None:
+            conversations, images = frame_selector(conversations=conversations,frames=images,clear_prompt=True)
+
+        # format conversations
+        prompt, raw_input_ids, raw_labels = self.get_conversation_formatter().format(
+            conversations, generation_preface=generation_preface, enable_thinking=enable_thinking)
+
+        # place image placeholders
+        input_ids = []
+        labels = []
+        pixel_values = []
+        grid_thws = []
+        invalidate_label = False
+        image_token_indices = [i for i, v in enumerate(raw_input_ids) if v == IMAGE_TOKEN_ID or v == VIDEO_TOKEN_ID]
+        last_image_token_index = -1
+        for i in range(len(image_token_indices)):
+            head = 0 if i == 0 else image_token_indices[i - 1] + 1
+            tail = image_token_indices[i]
+            last_image_token_index = tail
+            input_ids.extend(raw_input_ids[head:tail])
+            labels.extend(raw_labels[head:tail])
+            try:
+                # currently, do not support multiple videos
+                if multimodal_type == "video":
+                    image = images
+                else:
+                    image = images[i]
+                raw_pixel_values, image_grid_thws, image_placeholders = self.visual_tokenizer.preprocess_image(
+                    image, num_images=len(images) if fix_sample_overall_length_navit else 1, min_pixels=min_pixels, max_pixels=max_pixels,
+                    multimodal_type=multimodal_type)
+            except Exception as e:
+                if propagate_exception:
+                    raise e
+                logging.exception(e)
+                invalidate_label = True
+                # raw_pixel_values, image_placeholders = self.visual_tokenizer.mock_input() # TODO
+                raw_pixel_values, _ = self.visual_tokenizer.mock_input()
+                mock_image = transforms.ToPILImage()(raw_pixel_values[0])
+                raw_pixel_values, image_grid_thws, image_placeholders = self.visual_tokenizer.preprocess_image(
+                            mock_image, min_pixels=min_pixels, max_pixels=max_pixels)
+                
+            input_ids.extend(image_placeholders)
+            labels.extend([IGNORE_ID] * len(image_placeholders))
+            pixel_values.append(raw_pixel_values)
+            grid_thws.append(image_grid_thws)
+        input_ids.extend(raw_input_ids[last_image_token_index + 1:])
+        labels.extend(raw_labels[last_image_token_index + 1:])
+
+        # return tensors
+        input_ids = torch.tensor(input_ids, dtype=torch.long)
+        labels = torch.tensor([IGNORE_ID] * len(labels) if invalidate_label else labels, dtype=torch.long)
+        pixel_values = torch.cat(pixel_values, dim=0) if len(pixel_values) > 0 else None
+        grid_thws = torch.cat(grid_thws, dim=0) if len(grid_thws) > 0 else None
+
+        if return_labels:
+            return prompt, input_ids, pixel_values, grid_thws, labels
+        else:
+            return prompt, input_ids, pixel_values, grid_thws
+
+    def generate(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        # assert inputs.shape[0] == 1, 'Currently, only support `batch_size=1`'
+        _, inputs_embeds, labels, attention_mask, input_img_labels = self.merge_multimodal(
+            text_input_ids=inputs,
+            text_attention_masks=kwargs.pop('attention_mask'),
+            text_labels=None,
+            pixel_values=kwargs.pop('pixel_values'),
+            grid_thws=kwargs.pop('grid_thws'),
+            left_padding=True
+        )
+        inputs_embeds = inputs_embeds.detach()
+        torch.cuda.empty_cache()
+        return self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
+
+    def generate_condition(
+            self,
+            inputs: Optional[torch.Tensor] = None,
+            **kwargs,
+    ):
+        # assert inputs.shape[0] == 1, 'Currently, only support `batch_size=1`'
+        _, inputs_embeds, labels, attention_mask, input_img_labels = self.merge_multimodal(
+            text_input_ids=inputs,
+            text_attention_masks=kwargs.pop('attention_mask'),
+            text_labels=None,
+            pixel_values=kwargs.pop('pixel_values'),
+            grid_thws=kwargs.pop('grid_thws'),
+            left_padding=True
+        )
+        inputs_embeds = inputs_embeds.detach()
+        torch.cuda.empty_cache()
+        device = self.llm.device
+        outputs = self.llm(inputs_embeds=inputs_embeds.to(device), 
+                            labels=labels.to(device), 
+                            attention_mask=attention_mask.to(device), 
+                            output_hidden_states=True, 
+                            **kwargs)
+        semantic_cond_0 = outputs.hidden_states[-1]
+        semantic_cond_1 = outputs.hidden_states[-2]
+        semantic_cond = torch.cat([semantic_cond_0, semantic_cond_1], dim=-1)
+        return dict(
+            txt=semantic_cond
+        )
+    
+    def generate_img(
+        self,
+        inputs: Optional[torch.Tensor] = None,
+        cond = None,
+        no_both_cond = None,
+        no_txt_cond = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        if cond is None:
+            cond = self.generate_condition(inputs, **kwargs)
+        
+        height = kwargs.get('height', 1024)
+        width = kwargs.get('width', 1024)
+        num_steps = kwargs.get('num_steps', 50)
+        seed = kwargs.get('seed', 42)
+        img_cfg = kwargs.pop('img_cfg', 1.5)
+        txt_cfg = kwargs.pop('txt_cfg', 5)
+        yak_output = self.visual_generator.generate_image(
+            cond=cond, no_txt_cond=no_txt_cond, no_both_cond=no_both_cond,
+            height=height, width=width, 
+            num_steps=num_steps, seed=seed, 
+            img_cfg=img_cfg, txt_cfg=txt_cfg,
+            output_type="pil")
+        return yak_output

modeling_yak.py

@@ -0,0 +1,1461 @@
+from typing import Optional, Callable
+import math
+from dataclasses import dataclass
+import collections.abc
+from itertools import repeat as iter_repeat
+
+import numpy as np
+import torch
+from torch import Tensor, nn
+import torchvision
+from torchvision import transforms
+from diffusers import AutoencoderKL
+from PIL import Image
+from PIL.ImageOps import exif_transpose
+from torch.nn import functional as F
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import ModelOutput
+from einops import rearrange, repeat
+
+from .configuration_yak import YakConfig
+
+
+def _ntuple(n):
+    def parse(x):
+        if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+            x = tuple(x)
+            if len(x) == 1:
+                x = tuple(iter_repeat(x[0], n))
+            return x
+        return tuple(iter_repeat(x, n))
+    return parse
+
+
+to_1tuple = _ntuple(1)
+to_2tuple = _ntuple(2)
+to_3tuple = _ntuple(3)
+to_4tuple = _ntuple(4)
+
+
+def as_tuple(x):
+    if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
+        return tuple(x)
+    if x is None or isinstance(x, (int, float, str)):
+        return (x,)
+    else:
+        raise ValueError(f"Unknown type {type(x)}")
+
+
+def as_list_of_2tuple(x):
+    x = as_tuple(x)
+    if len(x) == 1:
+        x = (x[0], x[0])
+    assert len(x) % 2 == 0, f"Expect even length, got {len(x)}."
+    lst = []
+    for i in range(0, len(x), 2):
+        lst.append((x[i], x[i + 1]))
+    return lst
+
+def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor=None, attn_mask=None) -> Tensor:
+    if pe is None:
+        if attn_mask is not None and attn_mask.dtype != torch.bool:
+            attn_mask = attn_mask.to(q.dtype)
+        x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
+        x = rearrange(x, "B H L D -> B L (H D)")
+    else:
+        q, k = apply_rope(q, k, pe)
+        x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "B H L D -> B L (H D)")
+    return x
+
+
+def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
+    assert dim % 2 == 0
+    scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+    omega = 1.0 / (theta**scale)
+    out = torch.einsum("...n,d->...nd", pos, omega)
+    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
+    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
+    return out.float()
+
+
+def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:
+    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
+    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
+    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
+    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
+    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
+
+
+class EmbedND(nn.Module):
+    def __init__(self, dim: int, theta: int, axes_dim: list[int]):
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+        self.axes_dim = axes_dim
+
+    def forward(self, ids: Tensor) -> Tensor:
+        n_axes = ids.shape[-1]
+        emb = torch.cat(
+            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
+            dim=-3,
+        )
+
+        return emb.unsqueeze(1)
+
+
+def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
+    """
+    Create sinusoidal timestep embeddings.
+    :param t: a 1-D Tensor of N indices, one per batch element.
+                      These may be fractional.
+    :param dim: the dimension of the output.
+    :param max_period: controls the minimum frequency of the embeddings.
+    :return: an (N, D) Tensor of positional embeddings.
+    """
+    t = time_factor * t
+    half = dim // 2
+    freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
+        t.device
+    )
+
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    if torch.is_floating_point(t):
+        embedding = embedding.to(t)
+    return embedding
+
+
+class MLPEmbedder(nn.Module):
+    def __init__(self, in_dim: int, hidden_dim: int):
+        super().__init__()
+        self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
+        self.silu = nn.SiLU()
+        self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
+
+    def forward(self, x: Tensor) -> Tensor:
+        return self.out_layer(self.silu(self.in_layer(x)))
+
+
+class RMSNorm(torch.nn.Module):
+    def __init__(self, dim: int, scale_factor=1.0, eps:float=1e-6):
+        super().__init__()
+        self.scale = nn.Parameter(torch.ones(dim) * scale_factor)
+        self.eps = eps
+
+    def forward(self, x: Tensor):
+        x_dtype = x.dtype
+        x = x.float()
+        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + self.eps)
+        return (x * rrms).to(dtype=x_dtype) * self.scale
+
+
+class QKNorm(torch.nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.query_norm = RMSNorm(dim)
+        self.key_norm = RMSNorm(dim)
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
+        q = self.query_norm(q)
+        k = self.key_norm(k)
+        return q.to(v), k.to(v)
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.norm = QKNorm(head_dim)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(self, x: Tensor, pe: Tensor) -> Tensor:
+        qkv = self.qkv(x)
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
+        q, k = self.norm(q, k, v)
+        x = attention(q, k, v, pe=pe)
+        x = self.proj(x)
+        return x
+
+
+@dataclass
+class ModulationOut:
+    shift: Tensor
+    scale: Tensor
+    gate: Tensor
+
+
+class Modulation(nn.Module):
+    def __init__(self, dim: int, double: bool):
+        super().__init__()
+        self.is_double = double
+        self.multiplier = 6 if double else 3
+        self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
+
+    def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
+        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
+
+        return (
+            ModulationOut(*out[:3]),
+            ModulationOut(*out[3:]) if self.is_double else None,
+        )
+
+class TriModulation(nn.Module):
+    def __init__(self, dim: int):
+        super().__init__()
+        self.multiplier = 9
+        self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
+
+    def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
+        out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
+
+        return (
+            ModulationOut(*out[:3]),
+            ModulationOut(*out[3:6]),
+            ModulationOut(*out[6:]),
+        )
+
+
+# from https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
+class DoubleStreamXBlockProcessor:
+    def __call__(self, attn, img, txt, vec, pe, **attention_kwargs):
+        img_mod1, img_mod2, img_mod3 = attn.img_mod(vec)
+        txt_mod1, txt_mod2 = attn.txt_mod(vec)
+
+        # prepare image for attention
+        img_modulated = attn.img_norm1(img)
+        img_cos_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
+        img_qkv = attn.img_attn.qkv(img_cos_modulated)
+        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
+        img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)
+
+        # prepare image for self-attention
+        img_self_modulated = (1 + img_mod3.scale) * img_modulated + img_mod3.shift
+        img_self_qkv = attn.img_self_attn.qkv(img_self_modulated)
+        img_self_q, img_self_k, img_self_v = rearrange(img_self_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
+        img_self_q, img_self_k = attn.img_self_attn.norm(img_self_q, img_self_k, img_self_v)
+        txt_pe, img_pe = torch.split(pe, [txt.shape[1], img.shape[1]], dim=2)
+        img_self_attn = attention(img_self_q, img_self_k, img_self_v, pe=img_pe)
+
+        # prepare txt for attention
+        txt_modulated = attn.txt_norm1(txt)
+        txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
+        txt_qkv = attn.txt_attn.qkv(txt_modulated)
+        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
+        txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)
+
+        # run actual attention
+        q = torch.cat((txt_q, img_q), dim=2)
+        k = torch.cat((txt_k, img_k), dim=2)
+        v = torch.cat((txt_v, img_v), dim=2)
+
+        attn1 = attention(q, k, v, pe=pe)
+        txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]
+
+        # calculate the img bloks
+        img = img + img_mod1.gate * attn.img_attn.proj(img_attn)
+        img = img + img_mod3.gate * attn.img_self_attn.proj(img_self_attn)
+        img = img + img_mod2.gate * attn.img_mlp((1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift)
+
+        # calculate the txt bloks
+        txt = txt + txt_mod1.gate * attn.txt_attn.proj(txt_attn)
+        txt = txt + txt_mod2.gate * attn.txt_mlp((1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift)
+        return img, txt
+    
+class DoubleStreamXBlock(nn.Module):
+    def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
+        super().__init__()
+
+        mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.img_mod = TriModulation(hidden_size)
+        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+        self.img_self_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.img_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+
+        self.txt_mod = Modulation(hidden_size, double=True)
+        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
+
+        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.txt_mlp = nn.Sequential(
+            nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
+        )
+        processor = DoubleStreamXBlockProcessor()
+        self.set_processor(processor)
+    
+    def set_processor(self, processor) -> None:
+        self.processor = processor
+
+    def get_processor(self):
+        return self.processor
+
+    def forward(
+        self,
+        img: Tensor,
+        txt: Tensor,
+        vec: Tensor,
+        pe: Tensor,
+        image_proj: Tensor = None,
+        ip_scale: float =1.0,
+    ) -> tuple[Tensor, Tensor]:
+        if image_proj is None:
+            return self.processor(self, img, txt, vec, pe)
+        else:
+            return self.processor(self, img, txt, vec, pe, image_proj, ip_scale)
+
+class SingleStreamBlockProcessor:
+    def __call__(self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
+        mod, _ = attn.modulation(vec)
+        x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
+        qkv, mlp = torch.split(attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1)
+
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
+        q, k = attn.norm(q, k, v)
+
+        # compute attention
+        attn_1 = attention(q, k, v, pe=pe)
+
+        # compute activation in mlp stream, cat again and run second linear layer
+        output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
+        output = x + mod.gate * output
+        return output
+
+
+class SingleStreamBlock(nn.Module):
+    """
+    A DiT block with parallel linear layers as described in
+    https://arxiv.org/abs/2302.05442 and adapted modulation interface.
+    """
+
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        qk_scale: float | None = None,
+    ):
+        super().__init__()
+        self.hidden_dim = hidden_size
+        self.num_heads = num_heads
+        head_dim = hidden_size // num_heads
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
+        # qkv and mlp_in
+        self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
+        # proj and mlp_out
+        self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
+
+        self.norm = QKNorm(head_dim)
+
+        self.hidden_size = hidden_size
+        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+
+        self.mlp_act = nn.GELU(approximate="tanh")
+        self.modulation = Modulation(hidden_size, double=False)
+
+        processor = SingleStreamBlockProcessor()
+        self.set_processor(processor)
+
+
+    def set_processor(self, processor) -> None:
+        self.processor = processor
+
+    def get_processor(self):
+        return self.processor
+
+    def forward(
+        self,
+        x: Tensor,
+        vec: Tensor,
+        pe: Tensor,
+        image_proj: Tensor | None = None,
+        ip_scale: float = 1.0
+    ) -> Tensor:
+        if image_proj is None:
+            return self.processor(self, x, vec, pe)
+        else:
+            return self.processor(self, x, vec, pe, image_proj, ip_scale)
+
+
+class LastLayer(nn.Module):
+    def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+    def forward(self, x: Tensor, vec: Tensor) -> Tensor:
+        shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
+        x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
+        x = self.linear(x)
+        return x
+
+    
+
+def get_norm_layer(norm_layer):
+    """
+    Get the normalization layer.
+
+    Args:
+        norm_layer (str): The type of normalization layer.
+
+    Returns:
+        norm_layer (nn.Module): The normalization layer.
+    """
+    if norm_layer == "layer":
+        return nn.LayerNorm
+    elif norm_layer == "rms":
+        return RMSNorm
+    else:
+        raise NotImplementedError(f"Norm layer {norm_layer} is not implemented")   
+  
+def get_activation_layer(act_type):
+    """get activation layer
+
+    Args:
+        act_type (str): the activation type
+
+    Returns:
+        torch.nn.functional: the activation layer
+    """
+    if act_type == "gelu":
+        return lambda: nn.GELU()
+    elif act_type == "gelu_tanh":
+        # Approximate `tanh` requires torch >= 1.13
+        return lambda: nn.GELU(approximate="tanh")
+    elif act_type == "relu":
+        return nn.ReLU
+    elif act_type == "silu":
+        return nn.SiLU
+    else:
+        raise ValueError(f"Unknown activation type: {act_type}")
+
+def modulate(x, shift=None, scale=None):
+    """modulate by shift and scale
+
+    Args:
+        x (torch.Tensor): input tensor.
+        shift (torch.Tensor, optional): shift tensor. Defaults to None.
+        scale (torch.Tensor, optional): scale tensor. Defaults to None.
+
+    Returns:
+        torch.Tensor: the output tensor after modulate.
+    """
+    if scale is None and shift is None:
+        return x
+    elif shift is None:
+        return x * (1 + scale.unsqueeze(1))
+    elif scale is None:
+        return x + shift.unsqueeze(1)
+    else:
+        return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+def apply_gate(x, gate=None, tanh=False):
+    """AI is creating summary for apply_gate
+
+    Args:
+        x (torch.Tensor): input tensor.
+        gate (torch.Tensor, optional): gate tensor. Defaults to None.
+        tanh (bool, optional): whether to use tanh function. Defaults to False.
+
+    Returns:
+        torch.Tensor: the output tensor after apply gate.
+    """
+    if gate is None:
+        return x
+    if tanh:
+        return x * gate.unsqueeze(1).tanh()
+    else:
+        return x * gate.unsqueeze(1)
+
+class MLP(nn.Module):
+    """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
+
+    def __init__(
+        self,
+        in_channels,
+        hidden_channels=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        norm_layer=None,
+        bias=True,
+        drop=0.0,
+        use_conv=False,
+        device=None,
+        dtype=None,
+    ):
+        factory_kwargs = {"device": device, "dtype": dtype}
+        super().__init__()
+        out_features = out_features or in_channels
+        hidden_channels = hidden_channels or in_channels
+        bias = to_2tuple(bias)
+        drop_probs = to_2tuple(drop)
+        linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
+
+        self.fc1 = linear_layer(
+            in_channels, hidden_channels, bias=bias[0], **factory_kwargs
+        )
+        self.act = act_layer()
+        self.drop1 = nn.Dropout(drop_probs[0])
+        self.norm = (
+            norm_layer(hidden_channels, **factory_kwargs)
+            if norm_layer is not None
+            else nn.Identity()
+        )
+        self.fc2 = linear_layer(
+            hidden_channels, out_features, bias=bias[1], **factory_kwargs
+        )
+        self.drop2 = nn.Dropout(drop_probs[1])
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop1(x)
+        x = self.norm(x)
+        x = self.fc2(x)
+        x = self.drop2(x)
+        return x
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+class TextProjection(nn.Module):
+    """
+    Projects text embeddings. Also handles dropout for classifier-free guidance.
+
+    Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
+    """
+
+    def __init__(self, in_channels, hidden_size, act_layer):
+        super().__init__()
+        self.linear_1 = nn.Linear(
+            in_features=in_channels,
+            out_features=hidden_size,
+            bias=True,    
+        )
+        self.act_1 = act_layer()
+        self.linear_2 = nn.Linear(
+            in_features=hidden_size,
+            out_features=hidden_size,
+            bias=True,
+        )
+
+    def forward(self, caption):
+        hidden_states = self.linear_1(caption)
+        hidden_states = self.act_1(hidden_states)
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+
+def timestep_embedding_refiner(t, dim, max_period=10000):
+    """
+    Create sinusoidal timestep embeddings.
+
+    Args:
+        t (torch.Tensor): a 1-D Tensor of N indices, one per batch element. These may be fractional.
+        dim (int): the dimension of the output.
+        max_period (int): controls the minimum frequency of the embeddings.
+
+    Returns:
+        embedding (torch.Tensor): An (N, D) Tensor of positional embeddings.
+
+    .. ref_link: https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+    """
+    half = dim // 2
+    freqs = torch.exp(
+        -math.log(max_period)
+        * torch.arange(start=0, end=half, dtype=torch.float32)
+        / half
+    ).to(device=t.device)
+    args = t[:, None].float() * freqs[None]
+    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+    if dim % 2:
+        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+    return embedding
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+
+    def __init__(
+        self,
+        hidden_size,
+        act_layer,
+        frequency_embedding_size=256,
+        max_period=10000,
+        out_size=None,
+    ):
+        super().__init__()
+        self.frequency_embedding_size = frequency_embedding_size
+        self.max_period = max_period
+        if out_size is None:
+            out_size = hidden_size
+
+        self.mlp = nn.Sequential(
+            nn.Linear(
+                frequency_embedding_size, hidden_size, bias=True, 
+            ),
+            act_layer(),
+            nn.Linear(hidden_size, out_size, bias=True, ),
+        )
+        nn.init.normal_(self.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.mlp[2].weight, std=0.02)
+
+    def forward(self, t):
+        t_freq = timestep_embedding_refiner(
+            t, self.frequency_embedding_size, self.max_period
+        ).type(self.mlp[0].weight.dtype)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class IndividualTokenRefinerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads_num,
+        mlp_width_ratio: str = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+    ):
+        super().__init__()
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
+
+        self.norm1 = nn.LayerNorm(
+            hidden_size, elementwise_affine=True, eps=1e-6, 
+        )
+        self.self_attn_qkv = nn.Linear(
+            hidden_size, hidden_size * 3, bias=qkv_bias, 
+        )
+        qk_norm_layer = get_norm_layer(qk_norm_type)
+        self.self_attn_q_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
+            if qk_norm
+            else nn.Identity()
+        )
+        self.self_attn_k_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
+            if qk_norm
+            else nn.Identity()
+        )
+        self.self_attn_proj = nn.Linear(
+            hidden_size, hidden_size, bias=qkv_bias, 
+        )
+
+        self.norm2 = nn.LayerNorm(
+            hidden_size, elementwise_affine=True, eps=1e-6, 
+        )
+        act_layer = get_activation_layer(act_type)
+        self.mlp = MLP(
+            in_channels=hidden_size,
+            hidden_channels=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=mlp_drop_rate,
+        )
+
+        self.adaLN_modulation = nn.Sequential(
+            act_layer(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True, ),
+        )
+        # Zero-initialize the modulation
+        nn.init.zeros_(self.adaLN_modulation[1].weight)
+        nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        c: torch.Tensor,  # timestep_aware_representations + context_aware_representations
+        attn_mask: torch.Tensor = None,
+    ):
+        gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1)
+
+        norm_x = self.norm1(x)
+        qkv = self.self_attn_qkv(norm_x)
+        q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num)
+        # Apply QK-Norm if needed
+        q = self.self_attn_q_norm(q).to(v)
+        k = self.self_attn_k_norm(k).to(v)
+
+        # Self-Attention
+        q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
+        attn = attention(q, k, v, attn_mask=attn_mask)
+        x = x + apply_gate(self.self_attn_proj(attn), gate_msa)
+
+        # FFN Layer
+        x = x + apply_gate(self.mlp(self.norm2(x)), gate_mlp)
+
+        return x
+
+
+class CrossTokenRefinerBlock(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads_num,
+        mlp_width_ratio: str = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+    ):
+        super().__init__()
+        self.heads_num = heads_num
+        head_dim = hidden_size // heads_num
+        mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
+
+        self.norm1 = nn.LayerNorm(
+            hidden_size, elementwise_affine=True, eps=1e-6, 
+        )
+        self.self_attn_q = nn.Linear(
+            hidden_size, hidden_size, bias=qkv_bias, 
+        )
+        self.norm_y = nn.LayerNorm(
+            hidden_size, elementwise_affine=True, eps=1e-6, 
+        )
+        self.self_attn_kv = nn.Linear(
+            hidden_size, hidden_size*2, bias=qkv_bias, 
+        )
+        qk_norm_layer = get_norm_layer(qk_norm_type)
+        self.self_attn_q_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
+            if qk_norm
+            else nn.Identity()
+        )
+        self.self_attn_k_norm = (
+            qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
+            if qk_norm
+            else nn.Identity()
+        )
+        self.self_attn_proj = nn.Linear(
+            hidden_size, hidden_size, bias=qkv_bias, 
+        )
+
+        self.norm2 = nn.LayerNorm(
+            hidden_size, elementwise_affine=True, eps=1e-6, 
+        )
+        act_layer = get_activation_layer(act_type)
+        self.mlp = MLP(
+            in_channels=hidden_size,
+            hidden_channels=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=mlp_drop_rate,
+        )
+
+        self.adaLN_modulation = nn.Sequential(
+            act_layer(),
+            nn.Linear(hidden_size, 2 * hidden_size, bias=True, ),
+        )
+        # Zero-initialize the modulation
+        nn.init.zeros_(self.adaLN_modulation[1].weight)
+        nn.init.zeros_(self.adaLN_modulation[1].bias)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        y: torch.Tensor,
+        c: torch.Tensor,  # timestep_aware_representations + context_aware_representations
+        attn_mask: torch.Tensor = None,
+    ):
+        gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1)
+
+        norm_x = self.norm1(x)
+        q = self.self_attn_q(norm_x)
+        q = rearrange(qkv, "B L (H D) -> B L H D", H=self.heads_num)
+        norm_y = self.norm_y(y)
+        kv = self.self_attn_kv(norm_y)
+        k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=2, H=self.heads_num)
+        # Apply QK-Norm if needed
+        q = self.self_attn_q_norm(q).to(v)
+        k = self.self_attn_k_norm(k).to(v)
+
+        # Self-Attention
+        attn = attention(q, k, v, attn_mask=attn_mask)
+        x = x + apply_gate(self.self_attn_proj(attn), gate_msa)
+
+        # FFN Layer
+        x = x + apply_gate(self.mlp(self.norm2(x)), gate_mlp)
+
+        return x
+
+class IndividualTokenRefiner(nn.Module):
+    def __init__(
+        self,
+        hidden_size,
+        heads_num,
+        depth,
+        mlp_width_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+    ):
+        super().__init__()
+        self.blocks = nn.ModuleList(
+            [
+                IndividualTokenRefinerBlock(
+                    hidden_size=hidden_size,
+                    heads_num=heads_num,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_drop_rate=mlp_drop_rate,
+                    act_type=act_type,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    qkv_bias=qkv_bias,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        c: torch.LongTensor,
+        mask: Optional[torch.Tensor] = None,
+    ):
+        self_attn_mask = None
+        if mask is not None:
+            batch_size = mask.shape[0]
+            seq_len = mask.shape[1]
+            mask = mask.to(x.device)
+            # batch_size x 1 x seq_len x seq_len
+            self_attn_mask_1 = mask.view(batch_size, 1, 1, seq_len).repeat(
+                1, 1, seq_len, 1
+            )
+            # batch_size x 1 x seq_len x seq_len
+            self_attn_mask_2 = self_attn_mask_1.transpose(2, 3)
+            # batch_size x 1 x seq_len x seq_len, 1 for broadcasting of heads_num
+            self_attn_mask = (self_attn_mask_1 & self_attn_mask_2).bool()
+            # avoids self-attention weight being NaN for padding tokens
+            self_attn_mask[:, :, :, 0] = True
+
+        for block in self.blocks:
+            x = block(x, c, self_attn_mask)
+        return x
+
+
+class SingleTokenRefiner(nn.Module):
+    """
+    A single token refiner block for llm text embedding refine.
+    """
+    def __init__(
+        self,
+        in_channels,
+        hidden_size,
+        heads_num,
+        depth,
+        mlp_width_ratio: float = 4.0,
+        mlp_drop_rate: float = 0.0,
+        act_type: str = "silu",
+        qk_norm: bool = False,
+        qk_norm_type: str = "layer",
+        qkv_bias: bool = True,
+        attn_mode: str = "torch",
+        enable_cls_token: bool = False,
+        enable_cross_attn: bool = False,
+        length: int = 29,
+    ):
+        super().__init__()
+        self.attn_mode = attn_mode
+        assert self.attn_mode == "torch", "Only support 'torch' mode for token refiner."
+        self.in_channels = in_channels
+        self.enable_cross_attn = enable_cross_attn
+        if self.enable_cross_attn:
+            self.length = length
+            self.input_embedder = nn.Linear(
+                in_channels//length, hidden_size, bias=True, 
+            )
+            self.kv_embedder = nn.Linear(
+                in_channels//length*(length-1), hidden_size, bias=True, 
+            )
+            self.fusion = CrossTokenRefinerBlock(
+                    hidden_size=hidden_size,
+                    heads_num=heads_num,
+                    mlp_width_ratio=mlp_width_ratio,
+                    mlp_drop_rate=mlp_drop_rate,
+                    act_type=act_type,
+                    qk_norm=qk_norm,
+                    qk_norm_type=qk_norm_type,
+                    qkv_bias=qkv_bias,
+                )
+        else:
+            self.input_embedder = nn.Linear(
+                in_channels, hidden_size, bias=True, 
+            )
+
+        act_layer = get_activation_layer(act_type)
+        # Build timestep embedding layer
+        # self.t_embedder = TimestepEmbedder(hidden_size, act_layer,)
+        # Build context embedding layer
+        self.c_embedder = TextProjection(
+            in_channels, hidden_size, act_layer, 
+        )
+
+        self.individual_token_refiner = IndividualTokenRefiner(
+            hidden_size=hidden_size,
+            heads_num=heads_num,
+            depth=depth,
+            mlp_width_ratio=mlp_width_ratio,
+            mlp_drop_rate=mlp_drop_rate,
+            act_type=act_type,
+            qk_norm=qk_norm,
+            qk_norm_type=qk_norm_type,
+            qkv_bias=qkv_bias,
+        )
+
+        self.enable_cls_token = enable_cls_token
+        if self.enable_cls_token:
+            self.cls_token = nn.Parameter(torch.zeros(1, 1, hidden_size))
+            nn.init.normal_(self.cls_token, std=1e-6)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        mask: Optional[torch.LongTensor] = None,
+    ):
+        if mask is None:
+            context_aware_representations = x.mean(dim=1)
+        else:
+            mask_float = mask.float().unsqueeze(-1)  # [b, s1, 1]
+            context_aware_representations = (x * mask_float).sum(
+                dim=1
+            ) / mask_float.sum(dim=1)
+        c = self.c_embedder(context_aware_representations)
+        if self.enable_cross_attn:
+            single_channels = self.in_channels // self.length
+            x, y = torch.split(x, [single_channels, single_channels*(self.length-1)], dim=-1)
+            x = self.input_embedder(x)
+            y = self.kv_embedder(y)
+        else:
+            x = self.input_embedder(x)
+        if self.enable_cls_token:
+            B, L, C = x.shape
+            x = torch.cat([self.cls_token.expand(B, -1, -1), x], dim=1)
+        
+        if self.enable_cross_attn:
+            x = self.fusion(x, y, c)
+        x = self.individual_token_refiner(x, c, mask)
+        if self.enable_cls_token:
+            x_global = x[:, 0]
+            x = x[:, 1:]
+        else:
+            x_global = x.mean(dim=1)
+        return dict(
+            txt_fea=x,
+            txt_fea_avg=x_global
+        )
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+__all__ = ["YakModel"]
+
+@dataclass
+class VisualGeneratorOutput(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+
+
+class YakTransformer(nn.Module):
+    def __init__(self, config: YakConfig):
+        super().__init__()
+        self.config = config
+        self.in_channels = config.in_channels
+        self.out_channels = config.out_channels
+        if config.hidden_size % config.num_heads != 0:
+            raise ValueError(
+                f"Hidden size {config.hidden_size} must be divisible by num_heads {config.num_heads}"
+            )
+        pe_dim = config.hidden_size // config.num_heads
+        if sum(config.axes_dim) != pe_dim:
+            raise ValueError(f"Got {config.axes_dim} but expected positional dim {pe_dim}")
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_heads
+        self.pe_embedder = EmbedND(dim=pe_dim, theta=config.theta, axes_dim=config.axes_dim)
+        self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
+        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
+        self.vector_in = MLPEmbedder(config.vec_in_dim, self.hidden_size)
+        self.guidance_in = (
+            MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if config.guidance_embed else nn.Identity()
+        )
+        self.txt_type = config.txt_type
+        self.txt_in = SingleTokenRefiner(
+            config.context_in_dim, 
+            self.hidden_size, 
+            heads_num=config.num_heads * 2, 
+            depth=2, 
+            enable_cls_token=True
+        )
+
+        self.double_blocks = nn.ModuleList(
+            [
+                DoubleStreamXBlock(
+                    self.hidden_size,
+                    self.num_heads,
+                    mlp_ratio=config.mlp_ratio,
+                    qkv_bias=config.qkv_bias,
+                )
+                for _ in range(config.depth)
+            ]
+        )
+
+        self.single_blocks = nn.ModuleList(
+            [
+                SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=config.mlp_ratio)
+                for _ in range(config.depth_single_blocks)
+            ]
+        )
+
+        self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
+        self.gradient_checkpointing = False
+
+    def forward(
+        self,
+        img: Tensor,
+        img_ids: Tensor,
+        txt: Tensor,
+        txt_ids: Tensor,
+        timesteps: Tensor,
+        guidance: Tensor | None = None,
+        cond_img: Tensor = None,
+        cond_img_ids: Tensor = None,
+    ):
+        if img.ndim != 3 or txt.ndim != 3:
+            raise ValueError("Input img and txt tensors must have 3 dimensions.")
+
+        # running on sequences img
+        img_tokens = img.shape[1]
+        if cond_img is not None:
+            img = torch.cat([img, cond_img], dim=1)
+            img_ids = torch.cat([img_ids, cond_img_ids], dim=1)
+        img = self.img_in(img)
+
+        vec = self.time_in(timestep_embedding(timesteps, 256))
+        if self.config.guidance_embed:
+            if guidance is None:
+                raise ValueError("Didn't get guidance strength for guidance distilled model.")
+            vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
+        txt_dict = self.txt_in(txt)
+        txt = txt_dict["txt_fea"]
+        y = txt_dict["txt_fea_avg"]
+        vec = vec + self.vector_in(y)
+
+        ids = torch.cat((txt_ids, img_ids), dim=1)
+        pe = self.pe_embedder(ids)
+
+        for block in self.double_blocks:
+            if self.training and self.gradient_checkpointing:
+                img, txt = self._gradient_checkpointing_func(
+                    block.__call__,
+                    img,
+                    txt,
+                    vec,
+                    pe,
+                )
+            else:
+                img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
+
+        img = torch.cat((txt, img), 1)
+        for block in self.single_blocks:
+            if self.training and self.gradient_checkpointing:
+                img = self._gradient_checkpointing_func(
+                    block.__call__,
+                    img,
+                    vec,
+                    pe,
+                )
+            else:
+                img = block(img, vec=vec, pe=pe)
+        img = img[:, txt.shape[1] :, ...]
+
+        img = self.final_layer(img, vec)  # (N, T, patch_size ** 2 * out_channels)
+        if cond_img is not None:
+            img = torch.split(img, img_tokens, dim=1)[0]
+        return img
+
+def time_shift(mu: float, sigma: float, t: Tensor):
+    return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
+
+
+def get_lin_function(
+    x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15
+) -> Callable[[float], float]:
+    m = (y2 - y1) / (x2 - x1)
+    b = y1 - m * x1
+    return lambda x: m * x + b
+
+def get_noise(
+    num_samples: int,
+    channel: int,
+    height: int,
+    width: int,
+    device: torch.device,
+    dtype: torch.dtype,
+    seed: int,
+):
+    return torch.randn(
+        num_samples,
+        channel,
+        # allow for packing
+        2 * math.ceil(height / 16),
+        2 * math.ceil(width / 16),
+        device=device,
+        dtype=dtype,
+        generator=torch.Generator(device=device).manual_seed(seed),
+    )
+
+def unpack(x: Tensor, height: int, width: int) -> Tensor:
+    return rearrange(
+        x,
+        "b (h w) (c ph pw) -> b c (h ph) (w pw)",
+        h=math.ceil(height / 16),
+        w=math.ceil(width / 16),
+        ph=2,
+        pw=2,
+    )
+
+class YakPretrainedModel(PreTrainedModel):
+    config_class = YakConfig
+    base_model_prefix = "yak"
+    supports_gradient_checkpointing = True
+    main_input_name = "pixel_values"
+    _supports_sdpa = True
+
+
+class YakModel(YakPretrainedModel):
+    def __init__(self, config: YakConfig):
+        super().__init__(config)
+        self.vae = AutoencoderKL.from_config(config.vae_config)
+        self.backbone = YakTransformer(config)
+
+    def get_refiner(self):
+        return self.backbone.txt_in
+    
+    def get_cls_refiner(self):
+        return self.backbone.vector_in
+
+    def get_backbone(self):
+        return self.backbone
+
+    def get_vae(self):
+        return self.vae
+
+    def preprocess_image(self, image: Image.Image, size, convert_to_rgb=True, Norm=True, output_type="tensor"):
+        image = exif_transpose(image)
+        if not image.mode == "RGB" and convert_to_rgb:
+            image = image.convert("RGB")
+
+        image = torchvision.transforms.functional.resize(
+            image, size, interpolation=transforms.InterpolationMode.BICUBIC
+        )
+
+        arr = np.array(image)
+        h = arr.shape[0]
+        w = arr.shape[1]
+        crop_y = (h - size) // 2
+        crop_x = (w - size) // 2
+        pil_image = image.crop([crop_x, crop_y, crop_x+size, crop_y+size])
+        if output_type == "pil_image":
+            return pil_image
+        
+        image_np = arr[crop_y : crop_y + size, crop_x : crop_x + size]
+        hidden_h = h // 16
+        hidden_w = w // 16
+        hidden_size = size // 16
+        img_ids = torch.zeros(hidden_h, hidden_w, 3)
+        
+        img_ids[..., 1] = img_ids[..., 1] + torch.arange(hidden_h)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.arange(hidden_w)[None, :]
+        crop_y = (hidden_h - hidden_size) // 2
+        crop_x = (hidden_w - hidden_size) // 2
+        img_ids = img_ids[crop_y : crop_y + hidden_size, crop_x : crop_x + hidden_size]
+        img_ids = rearrange(img_ids, "h w c -> (h w) c")
+
+        image_tensor = torchvision.transforms.functional.to_tensor(image_np)
+        if Norm:
+            image_tensor = torchvision.transforms.functional.normalize(image_tensor, 
+            mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+        return pil_image, image_tensor, img_ids
+
+    def process_image_aspectratio(self, image, size):
+        w, h = image.size
+        t_w, t_h = size
+        resize_r = max(float(t_w)/w, float(t_h)/h)
+        resize_size = (int(resize_r * h), int(resize_r * w))
+        image = torchvision.transforms.functional.resize(
+            image, resize_size, interpolation=transforms.InterpolationMode.BICUBIC
+        )
+        pil_image = torchvision.transforms.functional.center_crop(
+            image, (t_h, t_w)
+        )
+        hidden_h = t_h // 16
+        hidden_w = t_w // 16
+        img_ids = torch.zeros(hidden_h, hidden_w, 3)
+        
+        img_ids[..., 1] = img_ids[..., 1] + torch.arange(hidden_h)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.arange(hidden_w)[None, :]
+        img_ids = rearrange(img_ids, "h w c -> (h w) c")
+        image_tensor = torchvision.transforms.functional.to_tensor(pil_image)
+        image_tensor = torchvision.transforms.functional.normalize(image_tensor, 
+            mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
+        return pil_image, image_tensor, img_ids
+    
+    def compute_vae_encodings(self, pixel_values, with_ids=True, time=0):
+        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+        pixel_values = pixel_values.to(self.vae.device, dtype=self.vae.dtype)
+        with torch.no_grad():
+            model_input = self.vae.encode(pixel_values).latent_dist.sample()
+            if hasattr(self.vae.config, 'shift_factor') and self.vae.config.shift_factor is not None:
+                model_input = model_input - self.vae.config.shift_factor
+            if hasattr(self.vae.config, 'scaling_factor') and self.vae.config.scaling_factor is not None:
+                model_input = model_input * self.vae.config.scaling_factor
+        # patch for transformer
+        bs, c, h, w = model_input.shape
+        model_input = rearrange(model_input, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+        if with_ids:
+            img_ids = torch.zeros(h // 2, w // 2, 3)
+            img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
+            img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
+            img_ids[..., 0] = time
+            img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
+            return model_input, img_ids
+        else:
+            return model_input
+
+    def generate_image(
+            self, 
+            cond,
+            height, 
+            width, 
+            num_steps, 
+            seed, 
+            no_both_cond=None, 
+            no_txt_cond=None,
+            img_cfg=1.0,
+            txt_cfg=1.0,
+            output_type="pil"
+        ):
+        txt = cond["txt"]
+        bs = len(txt)
+        channel = self.vae.config.latent_channels
+        height = 16 * (height // 16)
+        width = 16 * (width // 16)
+        torch_device = next(self.backbone.parameters()).device
+        x = get_noise(
+            bs,
+            channel,
+            height,
+            width,
+            device=torch_device,
+            dtype=torch.bfloat16,
+            seed=seed,
+        )
+        # prepare inputs
+        img = x
+        bs, c, h, w = img.shape
+
+        img = rearrange(img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
+        if img.shape[0] == 1 and bs > 1:
+            img = repeat(img, "1 ... -> bs ...", bs=bs)
+
+        img_ids = torch.zeros(h // 2, w // 2, 3)
+        img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
+        img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
+        img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs).to(img.device)
+
+        if "vae_pixel_values" in cond:
+            img_vae_cond, cond_ids = self.compute_vae_encodings(
+                pixel_values=cond["vae_pixel_values"], with_ids=True, time=1.0)
+            cond_ids = cond_ids.to(img.device)
+
+        if txt.shape[0] == 1 and bs > 1:
+            txt = repeat(txt, "1 ... -> bs ...", bs=bs)
+        txt_ids = torch.zeros(bs, txt.shape[1], 3).to(img.device)
+
+        timesteps = self.get_schedule(
+            num_steps, img.shape[1], shift=self.config.timestep_shift,
+            base_shift=self.config.base_shift, max_shift=self.config.max_shift)
+        no_both_txt = no_both_cond["txt"]
+        if no_txt_cond is not None:
+            no_txt_txt = no_txt_cond["txt"]
+            x = self.edit_denoise(img, img_ids, 
+                                  txt, txt_ids, 
+                                  no_txt_txt, 
+                                  no_both_txt,
+                                  img_vae_cond, cond_ids.to(img.device),
+                                  timesteps=timesteps, 
+                                  img_cfg=img_cfg, txt_cfg=txt_cfg)
+        else:
+            x = self.denoise(img, img_ids, txt, txt_ids, 
+                             timesteps=timesteps, cfg=txt_cfg, 
+                             neg_txt=no_both_txt)
+        x = unpack(x.float(), height, width)
+
+        with torch.autocast(device_type=torch_device.type, dtype=torch.float32):
+            if hasattr(self.vae.config, 'scaling_factor') and self.vae.config.scaling_factor is not None:
+                x = x / self.vae.config.scaling_factor
+            if hasattr(self.vae.config, 'shift_factor') and self.vae.config.shift_factor is not None:
+                x = x + self.vae.config.shift_factor
+            x = self.vae.decode(x, return_dict=False)[0]
+        # bring into PIL format and save
+        x = x.clamp(-1, 1)
+        x = rearrange(x, "b c h w -> b h w c")
+        x = (127.5 * (x + 1.0)).cpu().byte().numpy()
+        if output_type == "np":
+            return x
+        images = []
+        for i in range(bs):
+            img = Image.fromarray(x[i])
+            images.append(img)
+        return images
+
+
+    def get_schedule(self,
+        num_steps: int,
+        image_seq_len: int,
+        base_shift: float = 0.5,
+        max_shift: float = 1.15,
+        shift: bool = True,
+    ) -> list[float]:
+        # extra step for zero
+        timesteps = torch.linspace(1, 0, num_steps + 1)
+        # shifting the schedule to favor high timesteps for higher signal images
+        if shift:
+            # eastimate mu based on linear estimation between two points
+            mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
+            timesteps = time_shift(mu, 1.0, timesteps)
+
+        return timesteps.tolist()
+
+    def denoise(self, 
+                input_img: Tensor,
+                img_ids: Tensor,
+                txt: Tensor,
+                txt_ids: Tensor,
+                # sampling parameters
+                timesteps: list[float],
+                cfg: float = 1.0,
+                neg_txt = None):
+        bs = input_img.shape[0]
+        for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:]):
+            t_vec = torch.full((bs,), t_curr, dtype=input_img.dtype, device=input_img.device)
+            txt_ids = torch.zeros(bs, txt.shape[1], 3).to(txt.device)
+            cond_eps = self.backbone(
+                img=input_img,
+                img_ids=img_ids,
+                txt=txt,
+                txt_ids=txt_ids,
+                timesteps=t_vec,
+            )
+            txt_ids = torch.zeros(bs, neg_txt.shape[1], 3).to(neg_txt.device)
+            uncond_eps = self.backbone(
+                img=input_img,
+                img_ids=img_ids,
+                txt=neg_txt,
+                txt_ids=txt_ids,
+                timesteps=t_vec,
+            )
+            pred = uncond_eps + cfg * (cond_eps - uncond_eps)
+            input_img = input_img + (t_prev - t_curr) * pred
+        return input_img
+    
+    def edit_denoise(self, 
+                input_img: Tensor,
+                img_ids: Tensor,
+                txt: Tensor,
+                txt_ids: Tensor,
+                no_txt_txt: Tensor,
+                no_both_txt: Tensor,
+                img_cond,
+                cond_img_ids,
+                # sampling parameters
+                timesteps: list[float],
+                img_cfg: float = 1.0,
+                txt_cfg: float = 1.0,):
+        bs = input_img.shape[0]
+        for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:]):
+            t_vec = torch.full((bs * 1,), t_curr, dtype=input_img.dtype, device=input_img.device)
+            txt_ids = torch.zeros(bs, txt.shape[1], 3).to(txt.device)
+            cond_eps = self.backbone(
+                img=input_img,
+                img_ids=img_ids,
+                txt=txt,
+                txt_ids=txt_ids,
+                timesteps=t_vec,
+                cond_img=img_cond,
+                cond_img_ids=cond_img_ids,
+            )
+            txt_ids = torch.zeros(bs, no_both_txt.shape[1], 3).to(no_both_txt.device)
+            no_both_eps = self.backbone(
+                img=input_img,
+                img_ids=img_ids,
+                txt=no_both_txt,
+                txt_ids=txt_ids,
+                timesteps=t_vec,
+            )
+            txt_ids = torch.zeros(bs, no_txt_txt.shape[1], 3).to(no_txt_txt.device)
+            no_txt_eps = self.backbone(
+                img=input_img,
+                img_ids=img_ids,
+                txt=no_txt_txt,
+                txt_ids=txt_ids,
+                timesteps=t_vec,
+                cond_img=img_cond,
+                cond_img_ids=cond_img_ids,
+            )
+            pred = no_both_eps 
+            pred += img_cfg * (no_txt_eps - no_both_eps) 
+            pred += txt_cfg * (cond_eps - no_txt_eps)
+            input_img = input_img + (t_prev - t_curr) * pred
+        return input_img
+

preprocessor_config.json

@@ -0,0 +1,32 @@
+{
+  "crop_size": {
+    "height": -1,
+    "width": -1
+  },
+  "do_center_crop": false,
+  "do_convert_rgb": true,
+  "do_normalize": true,
+  "do_rescale": true,
+  "do_resize": true,
+  "hidden_stride": 2,
+  "image_mean": [
+    0.48145466,
+    0.4578275,
+    0.40821073
+  ],
+  "image_processor_type": "CLIPImageProcessor",
+  "image_std": [
+    0.26862954,
+    0.26130258,
+    0.27577711
+  ],
+  "max_pixels": 2408448,
+  "min_pixels": 200704,
+  "patch_size": 14,
+  "resample": 3,
+  "rescale_factor": 0.00392156862745098,
+  "size": {
+    "shortest_edge": -1
+  },
+  "temporal_patch_size": 1
+}

special_tokens_map.json

@@ -0,0 +1,31 @@
+{
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "eos_token": {
+    "content": "<|im_end|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  },
+  "pad_token": {
+    "content": "<|endoftext|>",
+    "lstrip": false,
+    "normalized": false,
+    "rstrip": false,
+    "single_word": false
+  }
+}

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
+size 11422654

tokenizer_config.json

@@ -0,0 +1,240 @@
+{
+  "add_bos_token": false,
+  "add_prefix_space": false,
+  "added_tokens_decoder": {
+    "151643": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151644": {
+      "content": "<|im_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151645": {
+      "content": "<|im_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151646": {
+      "content": "<|object_ref_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151647": {
+      "content": "<|object_ref_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151648": {
+      "content": "<|box_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151649": {
+      "content": "<|box_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151650": {
+      "content": "<|quad_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151651": {
+      "content": "<|quad_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151652": {
+      "content": "<|vision_start|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151653": {
+      "content": "<|vision_end|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151654": {
+      "content": "<|vision_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151655": {
+      "content": "<|image_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151656": {
+      "content": "<|video_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151657": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151658": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151659": {
+      "content": "<|fim_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151660": {
+      "content": "<|fim_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151661": {
+      "content": "<|fim_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151662": {
+      "content": "<|fim_pad|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151663": {
+      "content": "<|repo_name|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151664": {
+      "content": "<|file_sep|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151665": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151666": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151667": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151668": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|im_start|>",
+    "<|im_end|>",
+    "<|object_ref_start|>",
+    "<|object_ref_end|>",
+    "<|box_start|>",
+    "<|box_end|>",
+    "<|quad_start|>",
+    "<|quad_end|>",
+    "<|vision_start|>",
+    "<|vision_end|>",
+    "<|vision_pad|>",
+    "<|image_pad|>",
+    "<|video_pad|>"
+  ],
+  "bos_token": null,
+  "chat_template": "{%- if tools %}\n    {{- '<|im_start|>system\\n' }}\n    {%- if messages[0].role == 'system' %}\n        {{- messages[0].content + '\\n\\n' }}\n    {%- endif %}\n    {{- \"# Tools\\n\\nYou may call one or more functions to assist with the user query.\\n\\nYou are provided with function signatures within <tools></tools> XML tags:\\n<tools>\" }}\n    {%- for tool in tools %}\n        {{- \"\\n\" }}\n        {{- tool | tojson }}\n    {%- endfor %}\n    {{- \"\\n</tools>\\n\\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\\n<tool_call>\\n{\\\"name\\\": <function-name>, \\\"arguments\\\": <args-json-object>}\\n</tool_call><|im_end|>\\n\" }}\n{%- else %}\n    {%- if messages[0].role == 'system' %}\n        {{- '<|im_start|>system\\n' + messages[0].content + '<|im_end|>\\n' }}\n    {%- endif %}\n{%- endif %}\n{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}\n{%- for message in messages[::-1] %}\n    {%- set index = (messages|length - 1) - loop.index0 %}\n    {%- if ns.multi_step_tool and message.role == \"user\" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}\n        {%- set ns.multi_step_tool = false %}\n        {%- set ns.last_query_index = index %}\n    {%- endif %}\n{%- endfor %}\n{%- for message in messages %}\n    {%- if (message.role == \"user\") or (message.role == \"system\" and not loop.first) %}\n        {{- '<|im_start|>' + message.role + '\\n' + message.content + '<|im_end|>' + '\\n' }}\n    {%- elif message.role == \"assistant\" %}\n        {%- set content = message.content %}\n        {%- set reasoning_content = '' %}\n        {%- if message.reasoning_content is defined and message.reasoning_content is not none %}\n            {%- set reasoning_content = message.reasoning_content %}\n        {%- else %}\n            {%- if '</think>' in message.content %}\n                {%- set content = message.content.split('</think>')[-1].lstrip('\\n') %}\n                {%- set reasoning_content = message.content.split('</think>')[0].rstrip('\\n').split('<think>')[-1].lstrip('\\n') %}\n            {%- endif %}\n        {%- endif %}\n        {%- if loop.index0 > ns.last_query_index %}\n            {%- if loop.last or (not loop.last and reasoning_content) %}\n                {{- '<|im_start|>' + message.role + '\\n<think>\\n' + reasoning_content.strip('\\n') + '\\n</think>\\n\\n' + content.lstrip('\\n') }}\n            {%- else %}\n                {{- '<|im_start|>' + message.role + '\\n' + content }}\n            {%- endif %}\n        {%- else %}\n            {{- '<|im_start|>' + message.role + '\\n' + content }}\n        {%- endif %}\n        {%- if message.tool_calls %}\n            {%- for tool_call in message.tool_calls %}\n                {%- if (loop.first and content) or (not loop.first) %}\n                    {{- '\\n' }}\n                {%- endif %}\n                {%- if tool_call.function %}\n                    {%- set tool_call = tool_call.function %}\n                {%- endif %}\n                {{- '<tool_call>\\n{\"name\": \"' }}\n                {{- tool_call.name }}\n                {{- '\", \"arguments\": ' }}\n                {%- if tool_call.arguments is string %}\n                    {{- tool_call.arguments }}\n                {%- else %}\n                    {{- tool_call.arguments | tojson }}\n                {%- endif %}\n                {{- '}\\n</tool_call>' }}\n            {%- endfor %}\n        {%- endif %}\n        {{- '<|im_end|>\\n' }}\n    {%- elif message.role == \"tool\" %}\n        {%- if loop.first or (messages[loop.index0 - 1].role != \"tool\") %}\n            {{- '<|im_start|>user' }}\n        {%- endif %}\n        {{- '\\n<tool_response>\\n' }}\n        {{- message.content }}\n        {{- '\\n</tool_response>' }}\n        {%- if loop.last or (messages[loop.index0 + 1].role != \"tool\") %}\n            {{- '<|im_end|>\\n' }}\n        {%- endif %}\n    {%- endif %}\n{%- endfor %}\n{%- if add_generation_prompt %}\n    {{- '<|im_start|>assistant\\n' }}\n    {%- if enable_thinking is defined and enable_thinking is false %}\n        {{- '<think>\\n\\n</think>\\n\\n' }}\n    {%- endif %}\n{%- endif %}",
+  "clean_up_tokenization_spaces": false,
+  "eos_token": "<|im_end|>",
+  "errors": "replace",
+  "extra_special_tokens": {},
+  "model_max_length": 131072,
+  "pad_token": "<|endoftext|>",
+  "split_special_tokens": false,
+  "tokenizer_class": "Qwen2Tokenizer",
+  "unk_token": null
+}