Flourish commited on
Commit
f304131
·
verified ·
1 Parent(s): 69be683

Upload folder using huggingface_hub

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,202 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ Apache License
3
+ Version 2.0, January 2004
4
+ http://www.apache.org/licenses/
5
+
6
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
7
+
8
+ 1. Definitions.
9
+
10
+ "License" shall mean the terms and conditions for use, reproduction,
11
+ and distribution as defined by Sections 1 through 9 of this document.
12
+
13
+ "Licensor" shall mean the copyright owner or entity authorized by
14
+ the copyright owner that is granting the License.
15
+
16
+ "Legal Entity" shall mean the union of the acting entity and all
17
+ other entities that control, are controlled by, or are under common
18
+ control with that entity. For the purposes of this definition,
19
+ "control" means (i) the power, direct or indirect, to cause the
20
+ direction or management of such entity, whether by contract or
21
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
22
+ outstanding shares, or (iii) beneficial ownership of such entity.
23
+
24
+ "You" (or "Your") shall mean an individual or Legal Entity
25
+ exercising permissions granted by this License.
26
+
27
+ "Source" form shall mean the preferred form for making modifications,
28
+ including but not limited to software source code, documentation
29
+ source, and configuration files.
30
+
31
+ "Object" form shall mean any form resulting from mechanical
32
+ transformation or translation of a Source form, including but
33
+ not limited to compiled object code, generated documentation,
34
+ and conversions to other media types.
35
+
36
+ "Work" shall mean the work of authorship, whether in Source or
37
+ Object form, made available under the License, as indicated by a
38
+ copyright notice that is included in or attached to the work
39
+ (an example is provided in the Appendix below).
40
+
41
+ "Derivative Works" shall mean any work, whether in Source or Object
42
+ form, that is based on (or derived from) the Work and for which the
43
+ editorial revisions, annotations, elaborations, or other modifications
44
+ represent, as a whole, an original work of authorship. For the purposes
45
+ of this License, Derivative Works shall not include works that remain
46
+ separable from, or merely link (or bind by name) to the interfaces of,
47
+ the Work and Derivative Works thereof.
48
+
49
+ "Contribution" shall mean any work of authorship, including
50
+ the original version of the Work and any modifications or additions
51
+ to that Work or Derivative Works thereof, that is intentionally
52
+ submitted to Licensor for inclusion in the Work by the copyright owner
53
+ or by an individual or Legal Entity authorized to submit on behalf of
54
+ the copyright owner. For the purposes of this definition, "submitted"
55
+ means any form of electronic, verbal, or written communication sent
56
+ to the Licensor or its representatives, including but not limited to
57
+ communication on electronic mailing lists, source code control systems,
58
+ and issue tracking systems that are managed by, or on behalf of, the
59
+ Licensor for the purpose of discussing and improving the Work, but
60
+ excluding communication that is conspicuously marked or otherwise
61
+ designated in writing by the copyright owner as "Not a Contribution."
62
+
63
+ "Contributor" shall mean Licensor and any individual or Legal Entity
64
+ on behalf of whom a Contribution has been received by Licensor and
65
+ subsequently incorporated within the Work.
66
+
67
+ 2. Grant of Copyright License. Subject to the terms and conditions of
68
+ this License, each Contributor hereby grants to You a perpetual,
69
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
70
+ copyright license to reproduce, prepare Derivative Works of,
71
+ publicly display, publicly perform, sublicense, and distribute the
72
+ Work and such Derivative Works in Source or Object form.
73
+
74
+ 3. Grant of Patent License. Subject to the terms and conditions of
75
+ this License, each Contributor hereby grants to You a perpetual,
76
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
77
+ (except as stated in this section) patent license to make, have made,
78
+ use, offer to sell, sell, import, and otherwise transfer the Work,
79
+ where such license applies only to those patent claims licensable
80
+ by such Contributor that are necessarily infringed by their
81
+ Contribution(s) alone or by combination of their Contribution(s)
82
+ with the Work to which such Contribution(s) was submitted. If You
83
+ institute patent litigation against any entity (including a
84
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
85
+ or a Contribution incorporated within the Work constitutes direct
86
+ or contributory patent infringement, then any patent licenses
87
+ granted to You under this License for that Work shall terminate
88
+ as of the date such litigation is filed.
89
+
90
+ 4. Redistribution. You may reproduce and distribute copies of the
91
+ Work or Derivative Works thereof in any medium, with or without
92
+ modifications, and in Source or Object form, provided that You
93
+ meet the following conditions:
94
+
95
+ (a) You must give any other recipients of the Work or
96
+ Derivative Works a copy of this License; and
97
+
98
+ (b) You must cause any modified files to carry prominent notices
99
+ stating that You changed the files; and
100
+
101
+ (c) You must retain, in the Source form of any Derivative Works
102
+ that You distribute, all copyright, patent, trademark, and
103
+ attribution notices from the Source form of the Work,
104
+ excluding those notices that do not pertain to any part of
105
+ the Derivative Works; and
106
+
107
+ (d) If the Work includes a "NOTICE" text file as part of its
108
+ distribution, then any Derivative Works that You distribute must
109
+ include a readable copy of the attribution notices contained
110
+ within such NOTICE file, excluding those notices that do not
111
+ pertain to any part of the Derivative Works, in at least one
112
+ of the following places: within a NOTICE text file distributed
113
+ as part of the Derivative Works; within the Source form or
114
+ documentation, if provided along with the Derivative Works; or,
115
+ within a display generated by the Derivative Works, if and
116
+ wherever such third-party notices normally appear. The contents
117
+ of the NOTICE file are for informational purposes only and
118
+ do not modify the License. You may add Your own attribution
119
+ notices within Derivative Works that You distribute, alongside
120
+ or as an addendum to the NOTICE text from the Work, provided
121
+ that such additional attribution notices cannot be construed
122
+ as modifying the License.
123
+
124
+ You may add Your own copyright statement to Your modifications and
125
+ may provide additional or different license terms and conditions
126
+ for use, reproduction, or distribution of Your modifications, or
127
+ for any such Derivative Works as a whole, provided Your use,
128
+ reproduction, and distribution of the Work otherwise complies with
129
+ the conditions stated in this License.
130
+
131
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
132
+ any Contribution intentionally submitted for inclusion in the Work
133
+ by You to the Licensor shall be under the terms and conditions of
134
+ this License, without any additional terms or conditions.
135
+ Notwithstanding the above, nothing herein shall supersede or modify
136
+ the terms of any separate license agreement you may have executed
137
+ with Licensor regarding such Contributions.
138
+
139
+ 6. Trademarks. This License does not grant permission to use the trade
140
+ names, trademarks, service marks, or product names of the Licensor,
141
+ except as required for reasonable and customary use in describing the
142
+ origin of the Work and reproducing the content of the NOTICE file.
143
+
144
+ 7. Disclaimer of Warranty. Unless required by applicable law or
145
+ agreed to in writing, Licensor provides the Work (and each
146
+ Contributor provides its Contributions) on an "AS IS" BASIS,
147
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
148
+ implied, including, without limitation, any warranties or conditions
149
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
150
+ PARTICULAR PURPOSE. You are solely responsible for determining the
151
+ appropriateness of using or redistributing the Work and assume any
152
+ risks associated with Your exercise of permissions under this License.
153
+
154
+ 8. Limitation of Liability. In no event and under no legal theory,
155
+ whether in tort (including negligence), contract, or otherwise,
156
+ unless required by applicable law (such as deliberate and grossly
157
+ negligent acts) or agreed to in writing, shall any Contributor be
158
+ liable to You for damages, including any direct, indirect, special,
159
+ incidental, or consequential damages of any character arising as a
160
+ result of this License or out of the use or inability to use the
161
+ Work (including but not limited to damages for loss of goodwill,
162
+ work stoppage, computer failure or malfunction, or any and all
163
+ other commercial damages or losses), even if such Contributor
164
+ has been advised of the possibility of such damages.
165
+
166
+ 9. Accepting Warranty or Additional Liability. While redistributing
167
+ the Work or Derivative Works thereof, You may choose to offer,
168
+ and charge a fee for, acceptance of support, warranty, indemnity,
169
+ or other liability obligations and/or rights consistent with this
170
+ License. However, in accepting such obligations, You may act only
171
+ on Your own behalf and on Your sole responsibility, not on behalf
172
+ of any other Contributor, and only if You agree to indemnify,
173
+ defend, and hold each Contributor harmless for any liability
174
+ incurred by, or claims asserted against, such Contributor by reason
175
+ of your accepting any such warranty or additional liability.
176
+
177
+ END OF TERMS AND CONDITIONS
178
+
179
+ APPENDIX: How to apply the Apache License to your work.
180
+
181
+ To apply the Apache License to your work, attach the following
182
+ boilerplate notice, with the fields enclosed by brackets "[]"
183
+ replaced with your own identifying information. (Don't include
184
+ the brackets!) The text should be enclosed in the appropriate
185
+ comment syntax for the file format. We also recommend that a
186
+ file or class name and description of purpose be included on the
187
+ same "printed page" as the copyright notice for easier
188
+ identification within third-party archives.
189
+
190
+ Copyright [yyyy] [name of copyright owner]
191
+
192
+ Licensed under the Apache License, Version 2.0 (the "License");
193
+ you may not use this file except in compliance with the License.
194
+ You may obtain a copy of the License at
195
+
196
+ http://www.apache.org/licenses/LICENSE-2.0
197
+
198
+ Unless required by applicable law or agreed to in writing, software
199
+ distributed under the License is distributed on an "AS IS" BASIS,
200
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
201
+ See the License for the specific language governing permissions and
202
+ limitations under the License.
Notice ADDED
@@ -0,0 +1,79 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Copyright (C) 2025 AIDC-AI
2
+ Licensed under the Apache License, Version 2.0 (the "License").
3
+
4
+ This model was trained based on the following model:
5
+ 1. Ovis2-2B https://huggingface.co/AIDC-AI/Ovis2-2B
6
+ 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)
7
+ Apache License
8
+ Version 2.0, January 2004
9
+ http://www.apache.org/licenses/
10
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
11
+
12
+ Definitions.
13
+
14
+ "License" shall mean the terms and conditions for use, reproduction, and distribution as defined by Sections 1 through 9 of this document.
15
+
16
+ "Licensor" shall mean the copyright owner or entity authorized by the copyright owner that is granting the License.
17
+
18
+ "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.
19
+
20
+ "You" (or "Your") shall mean an individual or Legal Entity exercising permissions granted by this License.
21
+
22
+ "Source" form shall mean the preferred form for making modifications, including but not limited to software source code, documentation source, and configuration files.
23
+
24
+ "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.
25
+
26
+ "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).
27
+
28
+ "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.
29
+
30
+ "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."
31
+
32
+ "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.
33
+
34
+ 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.
35
+
36
+ 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.
37
+
38
+ 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:
39
+
40
+ (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.
41
+
42
+ 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.
43
+
44
+ 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.
45
+
46
+ 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.
47
+
48
+ 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.
49
+
50
+ 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.
51
+
52
+ END OF TERMS AND CONDITIONS
53
+
54
+ APPENDIX: How to apply the Apache License to your work.
55
+
56
+ To apply the Apache License to your work, attach the following
57
+ boilerplate notice, with the fields enclosed by brackets "[]"
58
+ replaced with your own identifying information. (Don't include
59
+ the brackets!) The text should be enclosed in the appropriate
60
+ comment syntax for the file format. We also recommend that a
61
+ file or class name and description of purpose be included on the
62
+ same "printed page" as the copyright notice for easier
63
+ identification within third-party archives.
64
+ Copyright [yyyy] [name of copyright owner]
65
+
66
+ 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
67
+
68
+ http://www.apache.org/licenses/LICENSE-2.0
69
+ 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.
70
+
71
+ 2. sdxl-vae https://huggingface.co/stabilityai/sdxl-vae
72
+ License: MIT (https://huggingface.co/datasets/choosealicense/licenses/blob/main/markdown/mit.md, SPDX-License-identifier: MIT)
73
+ MIT License
74
+ Copyright (c) [year] [fullname]
75
+ 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:
76
+ The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
77
+ 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.
78
+
79
+
README.md CHANGED
@@ -1,3 +1,179 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ language:
4
+ - en
5
+ ---
6
+
7
+ # Ovis-U1
8
+
9
+ <div align="center">
10
+ <img src=https://cdn-uploads.huggingface.co/production/uploads/637aebed7ce76c3b834cea37/3IK823BZ8w-mz_QfeYkDn.png width="30%"/>
11
+ </div>
12
+
13
+ <p align="center">
14
+ <!-- <a href="https://arxiv.org/abs/2502.12579"><img src="https://img.shields.io/badge/arXiv%20paper-2502.12579-b31b1b.svg" alt="arxiv"></a> -->
15
+ <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>
16
+ <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>
17
+ <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>
18
+ </p>
19
+
20
+
21
+ 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.
22
+
23
+ <figure>
24
+ <img src="https://cdn-uploads.huggingface.co/production/uploads/636f4c6b5d2050767e4a1491/EmEEGmot9JzaBfHP2uWld.jpeg" alt="Ovis-U1 architecture">
25
+ <figcaption style="text-align: center;">The overall architecture of Ovis-U1 (cf. Fig.2 in our report).</figcaption>
26
+ </figure>
27
+
28
+ ---
29
+
30
+ ## 🚀 News
31
+
32
+ - [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))!
33
+
34
+ ---
35
+
36
+ ## 📦 Installation
37
+
38
+ 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.
39
+
40
+ ```bash
41
+ git clone [email protected]:AIDC-AI/Ovis-U1.git
42
+ conda create -n ovis-u1 python=3.10 -y
43
+ conda activate ovis-u1
44
+ cd Ovis-U1
45
+ pip install -r requirements.txt
46
+ pip install -e .
47
+
48
+ ```
49
+
50
+ ## 📂 Model Checkpoints
51
+
52
+ We provide pretrained Ovis-U1-3B checkpoints for easy download and evaluation:
53
+
54
+ - **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)
55
+
56
+
57
+ ## 🛠️ Inference
58
+
59
+ For multimodal understanding, please run
60
+
61
+ ```bash
62
+ python ovis/eval/test_txt_generation.py
63
+ ```
64
+
65
+ For text-to-image, please run
66
+ ```bash
67
+ python ovis/eval/test_t2i.py \
68
+ --height 1024 \
69
+ --width 1024 \
70
+ --steps 50 \
71
+ --seed 42 \
72
+ --txt_cfg 5
73
+ ```
74
+
75
+ For image editing, please run
76
+ ```bash
77
+ python ovis/eval/test_img_edit.py \
78
+ --steps 50 \
79
+ --img_cfg 4 \
80
+ --txt_cfg 7.5
81
+ ```
82
+
83
+ ## 📊 Performance
84
+
85
+ #### OpenCompass Multi-modal Academic Benchmarks
86
+
87
+ | Model | MMB | MMS | MMMU | MathVista | Hallusion | AI2D | OCRBench | MMVet | Avg |
88
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
89
+ | GPT-4o | 86 | 70.2 | 72.9 | 71.6 | 57 | 86.3 | 82.2 | 76.9 | **75.4** |
90
+ | InternVL2.5-2B | 70.9 | 54.3 | 43.2 | 51.1 | 42.3 | 74.9 | 80.2 | 62.6 | **59.9** |
91
+ | SAIL-VL-2B | 73.7 | 56.5 | 44.1 | 62.8 | 45.9 | 77.4 | 83.1 | 44.2 | **61** |
92
+ | InternVL3-2B | 78 | 61.1 | 48.7 | 57.6 | 41.9 | 78.6 | 83.1 | 67 | **61.1**|
93
+ | Qwen2.5-VL-3B | 76.8 | 56.3 | 51.2 | 61.2 | 46.6 | 81.4 | 82.8 | 60 | **64.5** |
94
+ | Ovis2-2B | 76.9 | 56.7 | 45.6 | 64.1 | 50.2 | 82.7 | 87.3 | 58.3 | **65.2** |
95
+ | SAIL-VL-1.5-2B | 78.5 | 62.6 | 46.4 | 67 | 50 | 83.7 | 89.1 | 58.8 | **67** |
96
+ | Ristretto-3B | 80.2 | 62.8 | 51.3 | 67.6 | 50.2 | 84.2 | 84.7 | 60.7 | **67.7** |
97
+ | Ovis-U1 | 77.8 | 61.3 | 51.1 | 69.4 | 56.3 | 85.6 | 88.3 | 66.7 | **69.6** |
98
+
99
+ #### GenEval
100
+
101
+ | Model | Single object | Two object | Counting | Colors | Position | Attribute binding | Overall |
102
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
103
+ | GPT-4o | 0.99 | 0.92 | 0.85 | 0.92 | 0.75 | 0.61 | **0.84** |
104
+ | BAGEL | 0.99 | 0.94 | 0.81 | 0.88 | 0.64 | 0.63 | **0.82** |
105
+ | BAGEL 📝 | 0.98 | 0.95 | 0.84 | 0.95 | 0.78 | 0.77 | **0.88** |
106
+ | UniWorld-V1 | 0.99 | 0.93 | 0.79 | 0.89 | 0.49 | 0.70 | **0.80** |
107
+ | UniWorld-V1 📝 | 0.98 | 0.93 | 0.81 | 0.89 | 0.74 | 0.71 | **0.84** |
108
+ | OmniGen | 0.98 | 0.84 | 0.66 | 0.74 | 0.40 | 0.43 | **0.68** |
109
+ | OmniGen2 | 1 | 0.95 | 0.64 | 0.88 | 0.55 | 0.76 | **0.80** |
110
+ | OmniGen2 📝 | 0.99 | 0.96 | 0.74 | 0.98 | 0.71 | 0.75 | **0.86** |
111
+ | Ovis-U1 | 0.98 | 0.98 | 0.90 | 0.92 | 0.79 | 0.75 | **0.89** |
112
+
113
+ *📝 denotes using the rewritten prompts*
114
+
115
+ #### DPG-Bench
116
+
117
+ | Model | Global | Entity | Attribute | Relation | Other | Overall |
118
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|
119
+ | BAGEL | 88.94 | 90.37 | 91.29 | 90.82 | 88.67 | **85.07** |
120
+ | UniWorld-V1 | 83.64 | 88.39 | 88.44 | 89.27 | 87.22 | **81.38** |
121
+ | OmniGen | 87.90 | 88.97 | 88.47 | 87.95 | 83.56 | **81.16** |
122
+ | OmniGen2 | 88.81 | 88.83 | 90.18 | 89.37 | 90.27 | **83.57** |
123
+ | Ovis-U1 | 82.37 | 90.08 | 88.68 | 93.35 | 85.20 | **83.72** |
124
+
125
+ #### ImgEdit-Bench
126
+
127
+ | Model | Add | Adjust | Extract | Replace | Remove | Background | Style | Hybrid | Action | Overall |
128
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
129
+ | GPT-4o | 4.61 | 4.33 | 2.9 | 4.35 | 3.66 | 4.57 | 4.93 | 3.96 | 4.89 | **4.2** |
130
+ | MagicBrush | 2.84 | 1.58 | 1.51 | 1.97 | 1.58 | 1.75 | 2.38 | 1.62 | 1.22 | **1.90** |
131
+ | Instruct-P2P | 2.45 | 1.83 | 1.44 | 2.01 | 1.50 | 1.44 | 3.55 | 1.2 | 1.46 | **1.88** |
132
+ | AnyEdit | 3.18 | 2.95 | 1.88 | 2.47 | 2.23 | 2.24 | 2.85 | 1.56 | 2.65 | **2.45** |
133
+ | UltraEdit | 3.44 | 2.81 | 2.13 | 2.96 | 1.45 | 2.83 | 3.76 | 1.91 | 2.98 | **2.7** |
134
+ | OmniGen | 3.47 | 3.04 | 1.71 | 2.94 | 2.43 | 3.21 | 4.19 | 2.24 | 3.38 | **2.96** |
135
+ | Step1X-Edit | 3.88 | 3.14 | 1.76 | 3.40 | 2.41 | 3.16 | 4.63 | 2.64 | 2.52 | **3.06** |
136
+ | ICEdit | 3.58 | 3.39 | 1.73 | 3.15 | 2.93 | 3.08 | 3.84 | 2.04 | 3.68 | **3.05** |
137
+ | BAGEL | 3.56 | 3.31 | 1.7 | 3.3 | 2.62 | 3.24 | 4.49 | 2.38 | 4.17 | **3.2** |
138
+ | UniWorld-V1 | 3.82 | 3.64 | 2.27 | 3.47 | 3.24 | 2.99 | 4.21 | 2.96 | 2.74 | **3.26** |
139
+ | OmniGen2 | 3.57 | 3.06 | 1.77 | 3.74 | 3.2 | 3.57 | 4.81 | 2.52 | 4.68 | **3.44** |
140
+ | Ovis-U1 | 4.13 | 3.62 | 2.98 | 4.45 | 4.06 | 4.22 | 4.69 | 3.45 | 4.61 | **4.00** |
141
+
142
+ #### GEdit-Bench-EN
143
+
144
+ | Model | Background Change | Color Alteration | Material Modification | Motion Change | Portrait Beautification | Style Transfer | Subject Addition | Subject Removal | Subject Replacement | Text Modification | Tone Transformation | Avg |
145
+ |:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|
146
+ | 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** |
147
+ | 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** |
148
+ | 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** |
149
+ | 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** |
150
+ | 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** |
151
+ | 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** |
152
+ | 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** |
153
+ | 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** |
154
+ | 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** |
155
+ | 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** |
156
+
157
+ ## 📚 Citation
158
+
159
+ If you find Ovis-U1 useful, please cite our paper:
160
+
161
+ ```bibtex
162
+ @inproceedings{wang2025ovisu1,
163
+ title={Ovis-U1 Technical Report},
164
+ author={Ovis Team},
165
+ year={2025}
166
+ }
167
+ ```
168
+
169
+ ## 🙏 Acknowledgments
170
+
171
+ The code is built upon [Ovis](https://github.com/AIDC-AI/Ovis) and [FLUX](https://github.com/black-forest-labs/flux).
172
+
173
+ ## 📄 License
174
+
175
+ The project is released under Apache License 2.0 (http://www.apache.org/licenses/LICENSE-2.0, SPDX-License-identifier: Apache-2.0).
176
+
177
+ ## 🚨 Disclaimer
178
+
179
+ 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 ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "</think>": 151668,
3
+ "</tool_call>": 151658,
4
+ "</tool_response>": 151666,
5
+ "<think>": 151667,
6
+ "<tool_call>": 151657,
7
+ "<tool_response>": 151665,
8
+ "<|box_end|>": 151649,
9
+ "<|box_start|>": 151648,
10
+ "<|endoftext|>": 151643,
11
+ "<|file_sep|>": 151664,
12
+ "<|fim_middle|>": 151660,
13
+ "<|fim_pad|>": 151662,
14
+ "<|fim_prefix|>": 151659,
15
+ "<|fim_suffix|>": 151661,
16
+ "<|im_end|>": 151645,
17
+ "<|im_start|>": 151644,
18
+ "<|image_pad|>": 151655,
19
+ "<|object_ref_end|>": 151647,
20
+ "<|object_ref_start|>": 151646,
21
+ "<|quad_end|>": 151651,
22
+ "<|quad_start|>": 151650,
23
+ "<|repo_name|>": 151663,
24
+ "<|video_pad|>": 151656,
25
+ "<|vision_end|>": 151653,
26
+ "<|vision_pad|>": 151654,
27
+ "<|vision_start|>": 151652
28
+ }
config.json ADDED
@@ -0,0 +1,410 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "OvisU1"
4
+ ],
5
+ "auto_map": {
6
+ "AutoConfig": "configuration_ovis_u1.OvisU1Config",
7
+ "AutoModelForCausalLM": "modeling_ovis_u1.OvisU1"
8
+ },
9
+ "conversation_formatter_class": "Qwen3ConversationFormatter",
10
+ "disable_tie_weight": false,
11
+ "hidden_size": 2048,
12
+ "llm_attn_implementation": null,
13
+ "llm_config": {
14
+ "_attn_implementation_autoset": true,
15
+ "_name_or_path": "Qwen/Qwen3-1.7B",
16
+ "add_cross_attention": false,
17
+ "architectures": [
18
+ "Qwen3ForCausalLM"
19
+ ],
20
+ "attention_bias": false,
21
+ "attention_dropout": 0.0,
22
+ "bad_words_ids": null,
23
+ "begin_suppress_tokens": null,
24
+ "bos_token_id": 151643,
25
+ "chunk_size_feed_forward": 0,
26
+ "cross_attention_hidden_size": null,
27
+ "decoder_start_token_id": null,
28
+ "diversity_penalty": 0.0,
29
+ "do_sample": false,
30
+ "early_stopping": false,
31
+ "encoder_no_repeat_ngram_size": 0,
32
+ "eos_token_id": 151645,
33
+ "exponential_decay_length_penalty": null,
34
+ "finetuning_task": null,
35
+ "forced_bos_token_id": null,
36
+ "forced_eos_token_id": null,
37
+ "head_dim": 128,
38
+ "hidden_act": "silu",
39
+ "hidden_size": 2048,
40
+ "id2label": {
41
+ "0": "LABEL_0",
42
+ "1": "LABEL_1"
43
+ },
44
+ "initializer_range": 0.02,
45
+ "intermediate_size": 6144,
46
+ "is_decoder": false,
47
+ "is_encoder_decoder": false,
48
+ "label2id": {
49
+ "LABEL_0": 0,
50
+ "LABEL_1": 1
51
+ },
52
+ "length_penalty": 1.0,
53
+ "max_length": 20,
54
+ "max_position_embeddings": 40960,
55
+ "max_window_layers": 28,
56
+ "min_length": 0,
57
+ "model_type": "qwen3",
58
+ "no_repeat_ngram_size": 0,
59
+ "num_attention_heads": 16,
60
+ "num_beam_groups": 1,
61
+ "num_beams": 1,
62
+ "num_hidden_layers": 28,
63
+ "num_key_value_heads": 8,
64
+ "num_return_sequences": 1,
65
+ "output_attentions": false,
66
+ "output_hidden_states": false,
67
+ "output_scores": false,
68
+ "pad_token_id": null,
69
+ "prefix": null,
70
+ "problem_type": null,
71
+ "pruned_heads": {},
72
+ "remove_invalid_values": false,
73
+ "repetition_penalty": 1.0,
74
+ "return_dict": true,
75
+ "return_dict_in_generate": false,
76
+ "rms_norm_eps": 1e-06,
77
+ "rope_scaling": null,
78
+ "rope_theta": 1000000,
79
+ "sep_token_id": null,
80
+ "sliding_window": null,
81
+ "suppress_tokens": null,
82
+ "task_specific_params": null,
83
+ "temperature": 1.0,
84
+ "tf_legacy_loss": false,
85
+ "tie_encoder_decoder": false,
86
+ "tie_word_embeddings": true,
87
+ "tokenizer_class": null,
88
+ "top_k": 50,
89
+ "top_p": 1.0,
90
+ "torch_dtype": "bfloat16",
91
+ "torchscript": false,
92
+ "typical_p": 1.0,
93
+ "use_bfloat16": false,
94
+ "use_cache": true,
95
+ "use_sliding_window": false,
96
+ "vocab_size": 151936
97
+ },
98
+ "model_type": "ovis_u1",
99
+ "multimodal_max_length": 4496,
100
+ "torch_dtype": "bfloat16",
101
+ "transformers_version": "4.51.3",
102
+ "use_cache": false,
103
+ "visual_generator_config": {
104
+ "_attn_implementation_autoset": true,
105
+ "_name_or_path": "yak_1b/yak_1b_qwen3_trained_S8",
106
+ "add_cross_attention": false,
107
+ "architectures": [
108
+ "YakModel"
109
+ ],
110
+ "axes_dim": [
111
+ 16,
112
+ 56,
113
+ 56
114
+ ],
115
+ "bad_words_ids": null,
116
+ "base_shift": 0.5,
117
+ "begin_suppress_tokens": null,
118
+ "bos_token_id": null,
119
+ "checkpoint": false,
120
+ "chunk_size_feed_forward": 0,
121
+ "context_in_dim": 4096,
122
+ "cross_attention_hidden_size": null,
123
+ "decoder_start_token_id": null,
124
+ "depth": 6,
125
+ "depth_single_blocks": 12,
126
+ "diversity_penalty": 0.0,
127
+ "do_sample": false,
128
+ "early_stopping": false,
129
+ "encoder_no_repeat_ngram_size": 0,
130
+ "eos_token_id": null,
131
+ "exponential_decay_length_penalty": null,
132
+ "finetuning_task": null,
133
+ "forced_bos_token_id": null,
134
+ "forced_eos_token_id": null,
135
+ "guidance_embed": false,
136
+ "hidden_size": 1536,
137
+ "id2label": {
138
+ "0": "LABEL_0",
139
+ "1": "LABEL_1"
140
+ },
141
+ "in_channels": 16,
142
+ "is_decoder": false,
143
+ "is_encoder_decoder": false,
144
+ "label2id": {
145
+ "LABEL_0": 0,
146
+ "LABEL_1": 1
147
+ },
148
+ "length_penalty": 1.0,
149
+ "max_length": 20,
150
+ "max_shift": 1.15,
151
+ "min_length": 0,
152
+ "mlp_ratio": 4.0,
153
+ "model_type": "yak",
154
+ "no_repeat_ngram_size": 0,
155
+ "num_beam_groups": 1,
156
+ "num_beams": 1,
157
+ "num_heads": 12,
158
+ "num_return_sequences": 1,
159
+ "out_channels": 16,
160
+ "output_attentions": false,
161
+ "output_hidden_states": false,
162
+ "output_scores": false,
163
+ "pad_token_id": null,
164
+ "prefix": null,
165
+ "problem_type": null,
166
+ "pruned_heads": {},
167
+ "qkv_bias": true,
168
+ "remove_invalid_values": false,
169
+ "repetition_penalty": 1.0,
170
+ "return_dict": true,
171
+ "return_dict_in_generate": false,
172
+ "sep_token_id": null,
173
+ "suppress_tokens": null,
174
+ "task_specific_params": null,
175
+ "temperature": 1.0,
176
+ "tf_legacy_loss": false,
177
+ "theta": 10000,
178
+ "tie_encoder_decoder": false,
179
+ "tie_word_embeddings": true,
180
+ "timestep_shift": true,
181
+ "tokenizer_class": null,
182
+ "top_k": 50,
183
+ "top_p": 1.0,
184
+ "torch_dtype": "float32",
185
+ "torchscript": false,
186
+ "txt_type": "refiner",
187
+ "typical_p": 1.0,
188
+ "use_bfloat16": false,
189
+ "vae_config": {
190
+ "_class_name": "AutoencoderKL",
191
+ "_diffusers_version": "0.18.0.dev0",
192
+ "_name_or_path": ".",
193
+ "act_fn": "silu",
194
+ "block_out_channels": [
195
+ 128,
196
+ 256,
197
+ 512,
198
+ 512
199
+ ],
200
+ "down_block_types": [
201
+ "DownEncoderBlock2D",
202
+ "DownEncoderBlock2D",
203
+ "DownEncoderBlock2D",
204
+ "DownEncoderBlock2D"
205
+ ],
206
+ "force_upcast": false,
207
+ "in_channels": 3,
208
+ "latent_channels": 4,
209
+ "layers_per_block": 2,
210
+ "norm_num_groups": 32,
211
+ "out_channels": 3,
212
+ "sample_size": 512,
213
+ "scaling_factor": 0.13025,
214
+ "up_block_types": [
215
+ "UpDecoderBlock2D",
216
+ "UpDecoderBlock2D",
217
+ "UpDecoderBlock2D",
218
+ "UpDecoderBlock2D"
219
+ ]
220
+ },
221
+ "vec_in_dim": 1536
222
+ },
223
+ "visual_tokenizer_config": {
224
+ "_attn_implementation_autoset": true,
225
+ "_name_or_path": "",
226
+ "add_cross_attention": false,
227
+ "architectures": null,
228
+ "backbone_config": {
229
+ "_attn_implementation_autoset": false,
230
+ "_name_or_path": "aimv2/visual_tokenizer_backbone",
231
+ "add_cross_attention": false,
232
+ "architectures": [
233
+ "AIMv2Model"
234
+ ],
235
+ "attention_dropout": 0.0,
236
+ "auto_map": {
237
+ "AutoConfig": "configuration_aimv2.AIMv2Config",
238
+ "AutoModel": "modeling_aimv2.AIMv2Model",
239
+ "FlaxAutoModel": "modeling_flax_aimv2.FlaxAIMv2Model"
240
+ },
241
+ "bad_words_ids": null,
242
+ "begin_suppress_tokens": null,
243
+ "bos_token_id": null,
244
+ "chunk_size_feed_forward": 0,
245
+ "cross_attention_hidden_size": null,
246
+ "decoder_start_token_id": null,
247
+ "disable_rope": false,
248
+ "diversity_penalty": 0.0,
249
+ "do_sample": false,
250
+ "early_stopping": false,
251
+ "encoder_no_repeat_ngram_size": 0,
252
+ "eos_token_id": null,
253
+ "exponential_decay_length_penalty": null,
254
+ "finetuning_task": null,
255
+ "forced_bos_token_id": null,
256
+ "forced_eos_token_id": null,
257
+ "fullatt_block_indexes": null,
258
+ "hidden_size": 1024,
259
+ "hidden_stride": 2,
260
+ "id2label": {
261
+ "0": "LABEL_0",
262
+ "1": "LABEL_1"
263
+ },
264
+ "image_size": 448,
265
+ "intermediate_size": 2816,
266
+ "interpolate_pe_method": "two_dim",
267
+ "is_decoder": false,
268
+ "is_encoder_decoder": false,
269
+ "label2id": {
270
+ "LABEL_0": 0,
271
+ "LABEL_1": 1
272
+ },
273
+ "length_penalty": 1.0,
274
+ "max_length": 20,
275
+ "max_pixels": 2408448,
276
+ "min_length": 0,
277
+ "min_pixels": 200704,
278
+ "model_type": "aimv2",
279
+ "no_repeat_ngram_size": 0,
280
+ "num_attention_heads": 8,
281
+ "num_beam_groups": 1,
282
+ "num_beams": 1,
283
+ "num_channels": 3,
284
+ "num_hidden_layers": 24,
285
+ "num_return_sequences": 1,
286
+ "output_attentions": false,
287
+ "output_hidden_states": false,
288
+ "output_scores": false,
289
+ "pad_token_id": null,
290
+ "patch_size": 14,
291
+ "prefix": null,
292
+ "preserve_original_pe": true,
293
+ "problem_type": null,
294
+ "projection_dropout": 0.0,
295
+ "pruned_heads": {},
296
+ "qkv_bias": false,
297
+ "remove_invalid_values": false,
298
+ "repetition_penalty": 1.0,
299
+ "return_dict": true,
300
+ "return_dict_in_generate": false,
301
+ "rms_norm_eps": 1e-05,
302
+ "sep_token_id": null,
303
+ "suppress_tokens": null,
304
+ "task_specific_params": null,
305
+ "temperature": 1.0,
306
+ "temporal_patch_size": 1,
307
+ "tf_legacy_loss": false,
308
+ "tie_encoder_decoder": false,
309
+ "tie_word_embeddings": true,
310
+ "tokenizer_class": null,
311
+ "top_k": 50,
312
+ "top_p": 1.0,
313
+ "torch_dtype": "bfloat16",
314
+ "torchscript": false,
315
+ "typical_p": 1.0,
316
+ "use_bfloat16": false,
317
+ "use_bias": false,
318
+ "window_size": 112
319
+ },
320
+ "backbone_kwargs": {
321
+ "disable_rope": false,
322
+ "hidden_stride": 2,
323
+ "interpolate_pe_method": "two_dim",
324
+ "max_pixels": 2408448,
325
+ "min_pixels": 200704,
326
+ "preserve_original_pe": true,
327
+ "temporal_patch_size": 1,
328
+ "window_size": 112
329
+ },
330
+ "bad_words_ids": null,
331
+ "begin_suppress_tokens": null,
332
+ "bos_token_id": null,
333
+ "chunk_size_feed_forward": 0,
334
+ "cross_attention_hidden_size": null,
335
+ "decoder_start_token_id": null,
336
+ "depths": null,
337
+ "disable_rope": false,
338
+ "diversity_penalty": 0.0,
339
+ "do_sample": false,
340
+ "drop_cls_token": false,
341
+ "early_stopping": false,
342
+ "encoder_no_repeat_ngram_size": 0,
343
+ "eos_token_id": null,
344
+ "exponential_decay_length_penalty": null,
345
+ "finetuning_task": null,
346
+ "forced_bos_token_id": null,
347
+ "forced_eos_token_id": null,
348
+ "fullatt_block_indexes": null,
349
+ "hidden_stride": 2,
350
+ "id2label": {
351
+ "0": "LABEL_0",
352
+ "1": "LABEL_1"
353
+ },
354
+ "image_processor_new_kwargs": {
355
+ "hidden_stride": 2,
356
+ "max_pixels": 2408448,
357
+ "min_pixels": 200704,
358
+ "temporal_patch_size": 1
359
+ },
360
+ "interpolate_pe_method": "two_dim",
361
+ "is_decoder": false,
362
+ "is_encoder_decoder": false,
363
+ "label2id": {
364
+ "LABEL_0": 0,
365
+ "LABEL_1": 1
366
+ },
367
+ "length_penalty": 1.0,
368
+ "max_length": 20,
369
+ "max_pixels": 2408448,
370
+ "min_length": 0,
371
+ "min_pixels": 200704,
372
+ "model_type": "aimv2_visual_tokenizer",
373
+ "no_repeat_ngram_size": 0,
374
+ "num_beam_groups": 1,
375
+ "num_beams": 1,
376
+ "num_return_sequences": 1,
377
+ "output_attentions": false,
378
+ "output_hidden_states": false,
379
+ "output_scores": false,
380
+ "pad_token_id": null,
381
+ "prefix": null,
382
+ "preserve_original_pe": true,
383
+ "problem_type": null,
384
+ "pruned_heads": {},
385
+ "remove_invalid_values": false,
386
+ "repetition_penalty": 1.0,
387
+ "return_dict": true,
388
+ "return_dict_in_generate": false,
389
+ "sep_token_id": null,
390
+ "suppress_tokens": null,
391
+ "task_specific_params": null,
392
+ "tau": 1.0,
393
+ "temperature": 1.0,
394
+ "temporal_patch_size": 1,
395
+ "tf_legacy_loss": false,
396
+ "tie_encoder_decoder": false,
397
+ "tie_word_embeddings": true,
398
+ "tokenize_function": "softmax",
399
+ "tokenizer_class": null,
400
+ "top_k": 50,
401
+ "top_p": 1.0,
402
+ "torch_dtype": null,
403
+ "torchscript": false,
404
+ "typical_p": 1.0,
405
+ "use_bfloat16": false,
406
+ "use_indicators": false,
407
+ "vocab_size": 65536,
408
+ "window_size": 112
409
+ }
410
+ }
configuration_aimv2.py ADDED
@@ -0,0 +1,82 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # copied from https://huggingface.co/apple/aimv2-huge-patch14-448
2
+ from typing import Any
3
+
4
+ from transformers.configuration_utils import PretrainedConfig
5
+
6
+ __all__ = ["AIMv2Config"]
7
+
8
+
9
+ class AIMv2Config(PretrainedConfig):
10
+ """This is the configuration class to store the configuration of an [`AIMv2Model`].
11
+
12
+ Instantiating a configuration with the defaults will yield a similar configuration
13
+ to that of the [apple/aimv2-large-patch14-224](https://huggingface.co/apple/aimv2-large-patch14-224).
14
+
15
+ Args:
16
+ hidden_size: Dimension of the hidden representations.
17
+ intermediate_size: Dimension of the SwiGLU representations.
18
+ num_hidden_layers: Number of hidden layers in the Transformer.
19
+ num_attention_heads: Number of attention heads for each attention layer
20
+ in the Transformer.
21
+ num_channels: Number of input channels.
22
+ image_size: Image size.
23
+ patch_size: Patch size.
24
+ rms_norm_eps: Epsilon value used for the RMS normalization layer.
25
+ attention_dropout: Dropout ratio for attention probabilities.
26
+ projection_dropout: Dropout ratio for the projection layer after the attention.
27
+ qkv_bias: Whether to add a bias to the queries, keys and values.
28
+ use_bias: Whether to add a bias in the feed-forward and projection layers.
29
+ kwargs: Keyword arguments for the [`PretrainedConfig`].
30
+ """
31
+
32
+ model_type: str = "aimv2"
33
+
34
+ def __init__(
35
+ self,
36
+ hidden_size: int = 1024,
37
+ intermediate_size: int = 2816,
38
+ num_hidden_layers: int = 24,
39
+ num_attention_heads: int = 8,
40
+ num_channels: int = 3,
41
+ image_size: int = 224,
42
+ patch_size: int = 14,
43
+ rms_norm_eps: float = 1e-5,
44
+ attention_dropout: float = 0.0,
45
+ projection_dropout: float = 0.0,
46
+ qkv_bias: bool = False,
47
+ use_bias: bool = False,
48
+ hidden_stride: int = 2,
49
+ window_size: int = 112,
50
+ fullatt_block_indexes: list = None,
51
+ temporal_patch_size: int = 1,
52
+ preserve_original_pe: bool = False,
53
+ interpolate_pe_method: str = 'one_dim',
54
+ disable_rope: bool = False,
55
+ min_pixels: int = 3136,
56
+ max_pixels: int = 1960000,
57
+ **kwargs: Any,
58
+ ):
59
+ super().__init__(**kwargs)
60
+ self.hidden_size = hidden_size
61
+ self.intermediate_size = intermediate_size
62
+ self.num_hidden_layers = num_hidden_layers
63
+ self.num_attention_heads = num_attention_heads
64
+ self.num_channels = num_channels
65
+ self.patch_size = patch_size
66
+ self.image_size = image_size
67
+ self.attention_dropout = attention_dropout
68
+ self.rms_norm_eps = rms_norm_eps
69
+
70
+ self.projection_dropout = projection_dropout
71
+ self.qkv_bias = qkv_bias
72
+ self.use_bias = use_bias
73
+
74
+ self.hidden_stride = hidden_stride
75
+ self.window_size = window_size
76
+ self.fullatt_block_indexes = fullatt_block_indexes
77
+ self.temporal_patch_size = temporal_patch_size
78
+ self.preserve_original_pe = preserve_original_pe
79
+ self.interpolate_pe_method = interpolate_pe_method
80
+ self.disable_rope = disable_rope
81
+ self.min_pixels = min_pixels
82
+ self.max_pixels = max_pixels
configuration_ovis_u1.py ADDED
@@ -0,0 +1,281 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import List, Dict, Union, Optional
2
+ from abc import ABC, abstractmethod
3
+
4
+ from transformers import PretrainedConfig, AutoConfig, AutoModel
5
+
6
+ # Model Constants
7
+ IGNORE_ID = -100
8
+ IMAGE_TOKEN_ID = -200
9
+ VIDEO_TOKEN_ID = -201
10
+ IMAGE_TOKEN = "<image>"
11
+ VIDEO_TOKEN = "<video>"
12
+ IMAGE_ATOM_ID = -300
13
+ IMAGE_INDICATOR_IDS = [-301, -302, -303, -304]
14
+
15
+ from .configuration_aimv2 import AIMv2Config
16
+ from .modeling_aimv2 import AIMv2Model
17
+ AutoConfig.register("aimv2", AIMv2Config)
18
+ AutoModel.register(AIMv2Config, AIMv2Model)
19
+
20
+ from .configuration_yak import YakConfig
21
+ from .modeling_yak import YakModel
22
+ AutoConfig.register("yak", YakConfig)
23
+ AutoModel.register(YakConfig, YakModel)
24
+
25
+
26
+ # ----------------------------------------------------------------------
27
+ # Visual Tokenizer Configuration
28
+ # ----------------------------------------------------------------------
29
+ class BaseVisualTokenizerConfig(PretrainedConfig):
30
+ def __init__(self,
31
+ vocab_size=16384,
32
+ tokenize_function="softmax",
33
+ tau=1.0,
34
+ depths=None,
35
+ use_indicators=False,
36
+ drop_cls_token=False,
37
+ backbone_config: Optional[Union[PretrainedConfig, dict]] = None,
38
+ hidden_stride: int = 1,
39
+ **kwargs
40
+ ):
41
+ super().__init__(**kwargs)
42
+ self.vocab_size = vocab_size
43
+ self.tokenize_function = tokenize_function
44
+ self.tau = tau
45
+ if isinstance(depths, str):
46
+ depths = [int(x) for x in depths.split('|')]
47
+ self.depths = depths
48
+ self.backbone_kwargs = {}
49
+ self.use_indicators = use_indicators
50
+ self.drop_cls_token = drop_cls_token
51
+ if backbone_config is not None:
52
+ assert isinstance(backbone_config, (PretrainedConfig, dict)), \
53
+ f"expect `backbone_config` to be instance of PretrainedConfig or dict, but got {type(backbone_config)} type"
54
+ if not isinstance(backbone_config, PretrainedConfig):
55
+ model_type = backbone_config['model_type']
56
+ backbone_config.pop('model_type')
57
+ backbone_config = AutoConfig.for_model(model_type, **backbone_config)
58
+ self.backbone_config = backbone_config
59
+ self.hidden_stride = hidden_stride
60
+
61
+
62
+ class Aimv2VisualTokenizerConfig(BaseVisualTokenizerConfig):
63
+ model_type = "aimv2_visual_tokenizer"
64
+
65
+ def __init__(self, **kwargs):
66
+ super().__init__(**kwargs)
67
+ if self.drop_cls_token:
68
+ self.drop_cls_token = False
69
+ if self.depths:
70
+ assert len(self.depths) == 1
71
+ self.backbone_kwargs['num_hidden_layers'] = self.depths[0]
72
+
73
+ self.image_processor_new_kwargs = {}
74
+
75
+ if kwargs.get("min_pixels", None) is not None:
76
+ self.image_processor_new_kwargs['min_pixels'] = kwargs.get("min_pixels")
77
+ self.backbone_kwargs['min_pixels'] = self.min_pixels
78
+
79
+ if kwargs.get("max_pixels", None) is not None:
80
+ self.image_processor_new_kwargs['max_pixels'] = kwargs.get("max_pixels")
81
+ self.backbone_kwargs['max_pixels'] = self.max_pixels
82
+
83
+ if kwargs.get("temporal_patch_size", None) is not None:
84
+ self.image_processor_new_kwargs['temporal_patch_size'] = kwargs.get("temporal_patch_size")
85
+ self.backbone_kwargs['temporal_patch_size'] = self.temporal_patch_size
86
+
87
+ if kwargs.get("hidden_stride", None) is not None:
88
+ self.image_processor_new_kwargs['hidden_stride'] = kwargs.get("hidden_stride")
89
+
90
+ if kwargs.get("patch_size", None) is not None:
91
+ self.image_processor_new_kwargs['patch_size'] = kwargs.get("patch_size")
92
+ self.backbone_kwargs['patch_size'] = self.patch_size
93
+
94
+ if kwargs.get("window_size", None) is not None:
95
+ self.backbone_kwargs['window_size'] = kwargs.get("window_size")
96
+
97
+ if kwargs.get("hidden_stride", None) is not None:
98
+ self.backbone_kwargs['hidden_stride'] = kwargs.get("hidden_stride")
99
+
100
+ if kwargs.get('fullatt_block_indexes', None) is not None:
101
+ self.backbone_kwargs['fullatt_block_indexes'] = [int(i) for i in kwargs.get('fullatt_block_indexes').replace(' ','').split('|')]
102
+
103
+ if kwargs.get("preserve_original_pe", None) is not None:
104
+ self.backbone_kwargs['preserve_original_pe'] = kwargs.get("preserve_original_pe")
105
+
106
+ if kwargs.get("interpolate_pe_method", None) is not None:
107
+ self.backbone_kwargs['interpolate_pe_method'] = kwargs.get("interpolate_pe_method")
108
+
109
+ if kwargs.get("disable_rope", None) is not None:
110
+ self.backbone_kwargs['disable_rope'] = kwargs.get("disable_rope")
111
+
112
+ AutoConfig.register("aimv2_visual_tokenizer", Aimv2VisualTokenizerConfig)
113
+
114
+
115
+
116
+ # ----------------------------------------------------------------------
117
+ # OvisU1 Configuration
118
+ # ----------------------------------------------------------------------
119
+ class OvisU1Config(PretrainedConfig):
120
+ model_type = "ovis_u1"
121
+
122
+ def __init__(self,
123
+ llm_config: Optional[Union[PretrainedConfig, dict]] = None,
124
+ visual_tokenizer_config: Optional[Union[PretrainedConfig, dict]] = None,
125
+ visual_generator_config: Optional[Union[PretrainedConfig, dict]] = None,
126
+ multimodal_max_length=2048,
127
+ hidden_size=None,
128
+ conversation_formatter_class=None,
129
+ llm_attn_implementation=None,
130
+ disable_tie_weight=False,
131
+ **kwargs):
132
+ super().__init__(**kwargs)
133
+ if llm_config is not None:
134
+ assert isinstance(llm_config, (PretrainedConfig, dict)), \
135
+ f"expect `llm_config` to be instance of PretrainedConfig or dict, but got {type(llm_config)} type"
136
+ if not isinstance(llm_config, PretrainedConfig):
137
+ model_type = llm_config['model_type']
138
+ llm_config.pop('model_type')
139
+ llm_config = AutoConfig.for_model(model_type, **llm_config)
140
+ self.llm_config = llm_config
141
+ if visual_tokenizer_config is not None:
142
+ assert isinstance(visual_tokenizer_config, (PretrainedConfig, dict)), \
143
+ f"expect `visual_tokenizer_config` to be instance of PretrainedConfig or dict, but got {type(visual_tokenizer_config)} type"
144
+ if not isinstance(visual_tokenizer_config, PretrainedConfig):
145
+ model_type = visual_tokenizer_config['model_type']
146
+ visual_tokenizer_config.pop('model_type')
147
+ if model_type == "aimv2_native_visual_tokenizer":
148
+ model_type = "aimv2_visual_tokenizer"
149
+ if visual_tokenizer_config['backbone_config']['model_type'] == "aimv2_native":
150
+ visual_tokenizer_config['backbone_config']['model_type'] = "aimv2"
151
+ visual_tokenizer_config = AutoConfig.for_model(model_type, **visual_tokenizer_config)
152
+ self.visual_tokenizer_config = visual_tokenizer_config
153
+ if visual_generator_config is not None:
154
+ assert isinstance(visual_generator_config, (PretrainedConfig, dict)), \
155
+ f"expect `visual_generator_config` to be instance of PretrainedConfig or dict, but got {type(visual_generator_config)} type"
156
+ if not isinstance(visual_generator_config, PretrainedConfig):
157
+ model_type = visual_generator_config['model_type']
158
+ visual_generator_config.pop('model_type')
159
+ visual_generator_config = AutoConfig.for_model(model_type, **visual_generator_config)
160
+ self.visual_generator_config = visual_generator_config
161
+ self.multimodal_max_length = multimodal_max_length
162
+ self.hidden_size = hidden_size
163
+ self.conversation_formatter_class = conversation_formatter_class
164
+ self.llm_attn_implementation = llm_attn_implementation
165
+ self.disable_tie_weight = disable_tie_weight
166
+
167
+
168
+ # ----------------------------------------------------------------------
169
+ # Conversation Formatter
170
+ # ----------------------------------------------------------------------
171
+ class ConversationFormatter(ABC):
172
+ support_tokenizer_types = None
173
+
174
+ def __init__(self, tokenizer):
175
+ tokenizer_type = type(tokenizer).__name__
176
+ assert tokenizer_type in self.support_tokenizer_types, \
177
+ f'Invalid tokenizer type, expected one from `{self.support_tokenizer_types}`, but got `{tokenizer_type}`'
178
+ self.tokenizer = tokenizer
179
+ self.image_token = IMAGE_TOKEN
180
+ self.image_token_id = IMAGE_TOKEN_ID
181
+ self.ignore_id = IGNORE_ID
182
+ self.im_end = None
183
+ self.video_token = VIDEO_TOKEN
184
+ self.video_token_id = VIDEO_TOKEN_ID
185
+
186
+ def _tokenize_with_image_symbol(self, text):
187
+ if text.find(self.video_token) != -1:
188
+ token = self.video_token
189
+ token_id = self.video_token_id
190
+ else:
191
+ token = self.image_token
192
+ token_id = self.image_token_id
193
+
194
+ text_chunks = [self.tokenizer(chunk, add_special_tokens=False).input_ids for chunk in
195
+ text.split(token)]
196
+ token_ids = []
197
+ num_chuck = len(text_chunks)
198
+ for i, chunk in enumerate(text_chunks):
199
+ token_ids.extend(chunk)
200
+ if i < num_chuck - 1:
201
+ token_ids.append(token_id)
202
+ return token_ids
203
+
204
+ @abstractmethod
205
+ def format(self, conversations: List[Dict], generation_preface=None):
206
+ pass
207
+
208
+ @abstractmethod
209
+ def format_query(self, query, generation_preface=""):
210
+ pass
211
+
212
+
213
+ class Qwen3ConversationFormatter(ConversationFormatter):
214
+ support_tokenizer_types = ['QWenTokenizer', 'Qwen2TokenizerFast']
215
+
216
+ def __init__(self, tokenizer):
217
+ super().__init__(tokenizer)
218
+ self.from2role = {
219
+ "system": "<|im_start|>system\n",
220
+ "human": "<|im_start|>user\n",
221
+ "gpt": "<|im_start|>assistant\n",
222
+ "ignored_gpt": "<|im_start|>assistant\n",
223
+ }
224
+ self.gpt_token_num = None
225
+ self.im_end = "<|im_end|>\n"
226
+ self.empty_think = "<think>\n\n</think>\n\n"
227
+
228
+ def format(self, conversations: List[Dict], generation_preface=None, enable_thinking=False):
229
+ if self.gpt_token_num is None:
230
+ prefilled_think = "" if enable_thinking else self.empty_think
231
+ self.gpt_token_num = len(
232
+ self.tokenizer(self.from2role["gpt"] + prefilled_think, add_special_tokens=False).input_ids
233
+ )
234
+
235
+ if generation_preface is not None:
236
+ conversations.append({
237
+ "from": "gpt",
238
+ "value": generation_preface
239
+ })
240
+
241
+ prompt = ""
242
+ input_ids = []
243
+ labels = []
244
+ num_conversation = len(conversations)
245
+ for i, conversation in enumerate(conversations):
246
+ frm = conversation["from"]
247
+ role = self.from2role[frm]
248
+ message = conversation["value"]
249
+ if frm == 'gpt' and not enable_thinking:
250
+ text = role + self.empty_think + message
251
+ else:
252
+ text = role + message
253
+ if i < num_conversation - 1 or generation_preface is None:
254
+ text += self.im_end
255
+ prompt += text
256
+ token_ids = self._tokenize_with_image_symbol(text)
257
+ input_ids.extend(token_ids)
258
+ label_ids = [self.ignore_id] * len(token_ids)
259
+ if frm == "gpt" and generation_preface is None:
260
+ # learning `\n` following `im_end` is meaningless, so the last `\n` token is ignored in label
261
+ label_ids[self.gpt_token_num:-1] = token_ids[self.gpt_token_num:-1]
262
+ labels.extend(label_ids)
263
+
264
+ assert self._tokenize_with_image_symbol(prompt) == input_ids
265
+ assert len(input_ids) == len(labels)
266
+
267
+ if conversations[-1]['from'] == "gpt" and generation_preface is None:
268
+ # remove the last `\n` following `im_end` in input_ids
269
+ input_ids.pop()
270
+ labels.pop()
271
+
272
+ return prompt, input_ids, labels
273
+
274
+ def format_query(self, query, generation_preface="", enable_thinking=False):
275
+ prompt, input_ids, _ = self.format([{
276
+ "from": "human",
277
+ "value": query
278
+ }], generation_preface=generation_preface, enable_thinking=enable_thinking)
279
+
280
+ return prompt, input_ids
281
+
configuration_yak.py ADDED
@@ -0,0 +1,63 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Any
2
+ from typing import Union, Optional
3
+
4
+ from transformers.configuration_utils import PretrainedConfig
5
+
6
+ __all__ = ["YakConfig"]
7
+
8
+
9
+ class YakConfig(PretrainedConfig):
10
+ """This is the configuration class to store the configuration of an [`YakModel`].
11
+
12
+ Args:
13
+ """
14
+
15
+ model_type: str = "yak"
16
+
17
+ def __init__(
18
+ self,
19
+ in_channels: int = 16,
20
+ out_channels: int = 16,
21
+ vec_in_dim: int = 1536,
22
+ context_in_dim: int = 3072,
23
+ hidden_size: int = 1536,
24
+ mlp_ratio: int = 4,
25
+ num_heads: int = 12,
26
+ depth: int = 6,
27
+ depth_single_blocks: int = 12,
28
+ axes_dim: list = [16, 56, 56],
29
+ theta: int = 10_000,
30
+ qkv_bias: bool = True,
31
+ guidance_embed: bool = False,
32
+ checkpoint: bool = False,
33
+ txt_type: str = "refiner",
34
+ timestep_shift: bool = False,
35
+ base_shift: float = 0.5,
36
+ max_shift: float = 1.15,
37
+ vae_config: Optional[Union[PretrainedConfig, dict]] = None,
38
+ **kwargs: Any,
39
+ ):
40
+ super().__init__(**kwargs)
41
+ self.in_channels = in_channels
42
+ self.out_channels = out_channels
43
+ self.vec_in_dim = vec_in_dim
44
+ self.context_in_dim = context_in_dim
45
+ self.hidden_size = hidden_size
46
+ self.mlp_ratio = mlp_ratio
47
+ self.num_heads = num_heads
48
+ self.depth = depth
49
+ self.depth_single_blocks = depth_single_blocks
50
+ self.axes_dim = axes_dim
51
+ self.theta = theta
52
+ self.qkv_bias = qkv_bias
53
+ self.guidance_embed = guidance_embed
54
+ self.checkpoint = checkpoint
55
+ self.txt_type = txt_type
56
+ self.timestep_shift = timestep_shift
57
+ self.base_shift = base_shift
58
+ self.max_shift = max_shift
59
+
60
+ self.vae_config = vae_config
61
+
62
+
63
+
merges.txt ADDED
The diff for this file is too large to render. See raw diff
 
model-00001-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:70d8644198c2531f6466b3aea7e34953fb9c5b67e04c71642f44a7b1e062b01b
3
+ size 4061178424
model-00002-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c099dec97f53b4a18c79a206a6e2eed0893c4e8c91c701e995d85cdf53012d5f
3
+ size 4973530020
model-00003-of-00003.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:744d9cdf3f96e946dd761454e1743992ef79b7629e2a285e086678c186db3506
3
+ size 1212709496
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
modeling_aimv2.py ADDED
@@ -0,0 +1,385 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # adapted from https://huggingface.co/apple/aimv2-huge-patch14-448 (modification: add gradient checkpoint support)
2
+ from typing import Optional, Tuple, Union
3
+
4
+ import torch
5
+ from torch import nn
6
+ from torch.nn import functional as F
7
+ from transformers.modeling_outputs import BaseModelOutputWithNoAttention
8
+ from transformers.modeling_utils import PreTrainedModel
9
+ from flash_attn.layers.rotary import apply_rotary_emb
10
+ from flash_attn import flash_attn_varlen_func
11
+
12
+ from .configuration_aimv2 import AIMv2Config
13
+
14
+
15
+ __all__ = ["AIMv2Model"]
16
+
17
+
18
+ class RMSNorm(nn.Module):
19
+ def __init__(self, dim: int, eps: float = 1e-6):
20
+ super().__init__()
21
+ self.weight = nn.Parameter(torch.ones(dim))
22
+ self.eps = eps
23
+
24
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
25
+ output = self._norm(x.float()).type_as(x)
26
+ return output * self.weight
27
+
28
+ def extra_repr(self) -> str:
29
+ return f"{tuple(self.weight.shape)}, eps={self.eps}"
30
+
31
+ def _norm(self, x: torch.Tensor) -> torch.Tensor:
32
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
33
+
34
+
35
+ class AIMv2SwiGLUFFN(nn.Module):
36
+ def __init__(self, config: AIMv2Config):
37
+ super().__init__()
38
+ hidden_features = config.intermediate_size
39
+ in_features = config.hidden_size
40
+ bias = config.use_bias
41
+
42
+ self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
43
+ self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
44
+ self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)
45
+
46
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
47
+ x = F.silu(self.fc1(x)) * self.fc3(x)
48
+ x = self.fc2(x)
49
+ return x
50
+
51
+
52
+ # copied from qwen2.5-vl
53
+ class VisionRotaryEmbedding(nn.Module):
54
+ def __init__(self, dim: int, theta: float = 10000.0) -> None:
55
+ super().__init__()
56
+ inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
57
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
58
+
59
+ def forward(self, seqlen: int) -> torch.Tensor:
60
+ seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
61
+ freqs = torch.outer(seq, self.inv_freq)
62
+ return freqs
63
+
64
+ # Note: in qwen2-vl and qwen2.5-vl, 3d convolution is used.
65
+ class AIMv2PatchEmbed(nn.Module):
66
+ def __init__(self, config: AIMv2Config):
67
+ super().__init__()
68
+ self.config = config
69
+ self.proj = nn.Conv2d(
70
+ config.num_channels,
71
+ config.hidden_size,
72
+ kernel_size=(config.patch_size, config.patch_size),
73
+ stride=(config.patch_size, config.patch_size),
74
+ )
75
+ self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
76
+
77
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
78
+ x = x.view(-1, self.config.num_channels * self.config.temporal_patch_size, self.config.patch_size, self.config.patch_size)
79
+ x = self.proj(x).view(-1, self.config.hidden_size) #.flatten(2).transpose(1, 2) # token_len x hidden_size
80
+ x = self.norm(x)
81
+ return x
82
+
83
+ class AIMv2ViTPreprocessor(nn.Module):
84
+ def __init__(self, config: AIMv2Config):
85
+ super().__init__()
86
+
87
+ num_patches = (config.image_size // config.patch_size) ** 2
88
+
89
+ self.patchifier = AIMv2PatchEmbed(config)
90
+
91
+ self.preserve_original_pe = config.preserve_original_pe
92
+ self.hidden_stride = config.hidden_stride
93
+
94
+ if self.preserve_original_pe:
95
+ self.interpolate_pe_method = config.interpolate_pe_method
96
+ self.pos_embed = nn.Parameter(torch.zeros((1, num_patches, config.hidden_size)))
97
+
98
+ def forward(self, x: torch.Tensor, grid_thws: Optional[torch.Tensor] = None) -> torch.Tensor:
99
+ tokens = self.patchifier(x)
100
+
101
+ if self.preserve_original_pe:
102
+ assert grid_thws is not None
103
+ pos_embed_new = torch.zeros_like(tokens)
104
+ if self.interpolate_pe_method == 'one_dim':
105
+ pos_embed = self.pos_embed.transpose(1,2).to(tokens.device)
106
+ elif self.interpolate_pe_method == 'two_dim':
107
+ ori_h = ori_w = int(self.pos_embed.shape[1] ** 0.5)
108
+ pos_embed = self.pos_embed.reshape(1, ori_h, ori_w, -1).permute(0,3,1,2)
109
+ else:
110
+ raise TypeError("The interpolation method for pe should be one_dim, two_dim.")
111
+ cnt = 0
112
+ for t, h, w in grid_thws:
113
+ num_patches = h * w
114
+ thw = t * h * w
115
+ if self.interpolate_pe_method == 'one_dim':
116
+ pe = F.interpolate(pos_embed, size=num_patches, mode='linear', align_corners=False).transpose(1,2)
117
+ elif self.interpolate_pe_method == 'two_dim':
118
+ # 1, 1024, 32, 32
119
+ pe = F.interpolate(pos_embed, size=(h,w), mode='bicubic', align_corners=False)
120
+ # 1, 1024, 1024
121
+ pe = pe.permute(0,2,3,1).reshape(1, h*w, -1)
122
+ # 1024, 1024
123
+ pe = pe[0].repeat(t,1)
124
+ # 1, 16, 2, 16, 2, 1024
125
+ pe = pe.reshape(t, h//self.hidden_stride, self.hidden_stride, w//self.hidden_stride, self.hidden_stride, -1)
126
+ # 1024, 1024
127
+ pe = pe.permute(0,1,3,2,4,5).reshape(thw,-1)
128
+ pos_embed_new[cnt:cnt+thw] = pe
129
+
130
+ cnt += thw
131
+
132
+ tokens = tokens + pos_embed_new
133
+ return tokens
134
+
135
+ # copied from qwen2.5-vl
136
+ def apply_rotary_pos_emb_flashatt(
137
+ q: torch.Tensor, k: torch.Tensor, cos: torch.Tensor, sin: torch.Tensor
138
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
139
+ cos = cos.chunk(2, dim=-1)[0].contiguous()
140
+ sin = sin.chunk(2, dim=-1)[0].contiguous()
141
+ q_embed = apply_rotary_emb(q.float(), cos.float(), sin.float()).type_as(q)
142
+ k_embed = apply_rotary_emb(k.float(), cos.float(), sin.float()).type_as(k)
143
+ return q_embed, k_embed
144
+
145
+ class AIMv2FlashAttention2(nn.Module):
146
+ def __init__(self, config: AIMv2Config) -> None:
147
+ super().__init__()
148
+ dim = config.hidden_size
149
+ self.num_heads = config.num_attention_heads
150
+ self.qkv = nn.Linear(dim, dim * 3, bias=config.qkv_bias)
151
+ self.proj = nn.Linear(dim, dim, bias=config.use_bias)
152
+
153
+ self.use_rope = not config.disable_rope
154
+
155
+ def forward(
156
+ self,
157
+ hidden_states: torch.Tensor,
158
+ cu_seqlens: torch.Tensor,
159
+ position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,
160
+ ) -> torch.Tensor:
161
+
162
+ seq_length = hidden_states.shape[0]
163
+ q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
164
+ if self.use_rope:
165
+ cos, sin = position_embeddings
166
+ q, k = apply_rotary_pos_emb_flashatt(q.unsqueeze(0), k.unsqueeze(0), cos, sin)
167
+ q = q.squeeze(0)
168
+ k = k.squeeze(0)
169
+
170
+ max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
171
+ attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
172
+ seq_length, -1
173
+ )
174
+ attn_output = self.proj(attn_output)
175
+ return attn_output
176
+
177
+ class AIMv2Block(nn.Module):
178
+ def __init__(self, config: AIMv2Config):
179
+ super().__init__()
180
+ self.attn = AIMv2FlashAttention2(config)
181
+ self.norm_1 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
182
+ self.mlp = AIMv2SwiGLUFFN(config)
183
+ self.norm_2 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
184
+
185
+ def forward(
186
+ self, x: torch.Tensor, cu_seqlens: torch.Tensor, position_embeddings: torch.Tensor
187
+ ) -> torch.Tensor:
188
+ x = x + self.attn(self.norm_1(x), cu_seqlens=cu_seqlens, position_embeddings=position_embeddings)
189
+ x = x + self.mlp(self.norm_2(x))
190
+ return x
191
+
192
+
193
+ class AIMv2Transformer(nn.Module):
194
+ def __init__(self, config: AIMv2Config):
195
+ super().__init__()
196
+ self.blocks = nn.ModuleList(
197
+ [AIMv2Block(config) for _ in range(config.num_hidden_layers)]
198
+ )
199
+ self.post_trunk_norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
200
+ self.gradient_checkpointing = False
201
+
202
+ self.rotary_pos_emb = VisionRotaryEmbedding(config.hidden_size // config.num_attention_heads // 2)
203
+
204
+ self.hidden_stride = config.hidden_stride
205
+ self.patch_size = config.patch_size
206
+ self.window_size = config.window_size
207
+ self.spatial_merge_unit = config.hidden_stride * config.hidden_stride
208
+
209
+ self.fullatt_block_indexes = config.fullatt_block_indexes
210
+
211
+ # copied from qwen2.5_vl
212
+ def rot_pos_emb(self, grid_thw):
213
+ pos_ids = []
214
+ for t, h, w in grid_thw:
215
+ hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
216
+ hpos_ids = hpos_ids.reshape(
217
+ h // self.hidden_stride,
218
+ self.hidden_stride,
219
+ w // self.hidden_stride,
220
+ self.hidden_stride,
221
+ )
222
+ hpos_ids = hpos_ids.permute(0, 2, 1, 3)
223
+ hpos_ids = hpos_ids.flatten()
224
+
225
+ wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
226
+ wpos_ids = wpos_ids.reshape(
227
+ h // self.hidden_stride,
228
+ self.hidden_stride,
229
+ w // self.hidden_stride,
230
+ self.hidden_stride,
231
+ )
232
+ wpos_ids = wpos_ids.permute(0, 2, 1, 3)
233
+ wpos_ids = wpos_ids.flatten()
234
+ pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
235
+ pos_ids = torch.cat(pos_ids, dim=0)
236
+ max_grid_size = grid_thw[:, 1:].max()
237
+ rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
238
+ rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
239
+ return rotary_pos_emb
240
+
241
+ def get_window_index(self, grid_thw):
242
+ window_index: list = []
243
+ cu_window_seqlens: list = [0]
244
+ window_index_id = 0
245
+ vit_merger_window_size = self.window_size // self.hidden_stride // self.patch_size # patch (after merge) number in each window
246
+
247
+ for grid_t, grid_h, grid_w in grid_thw:
248
+ llm_grid_h, llm_grid_w = (
249
+ grid_h // self.hidden_stride, # number of patch after merge
250
+ grid_w // self.hidden_stride,
251
+ )
252
+ index = torch.arange(grid_t * llm_grid_h * llm_grid_w).reshape(grid_t, llm_grid_h, llm_grid_w)
253
+ pad_h = vit_merger_window_size - llm_grid_h % vit_merger_window_size
254
+ pad_w = vit_merger_window_size - llm_grid_w % vit_merger_window_size
255
+ num_windows_h = (llm_grid_h + pad_h) // vit_merger_window_size
256
+ num_windows_w = (llm_grid_w + pad_w) // vit_merger_window_size
257
+ index_padded = F.pad(index, (0, pad_w, 0, pad_h), "constant", -100)
258
+ index_padded = index_padded.reshape(
259
+ grid_t,
260
+ num_windows_h,
261
+ vit_merger_window_size,
262
+ num_windows_w,
263
+ vit_merger_window_size,
264
+ )
265
+ index_padded = index_padded.permute(0, 1, 3, 2, 4).reshape(
266
+ grid_t,
267
+ num_windows_h * num_windows_w,
268
+ vit_merger_window_size,
269
+ vit_merger_window_size,
270
+ )
271
+ seqlens = (index_padded != -100).sum([2, 3]).reshape(-1)
272
+ index_padded = index_padded.reshape(-1)
273
+ index_new = index_padded[index_padded != -100]
274
+ window_index.append(index_new + window_index_id)
275
+ cu_seqlens_tmp = seqlens.cumsum(0) * self.spatial_merge_unit + cu_window_seqlens[-1]
276
+ cu_window_seqlens.extend(cu_seqlens_tmp.tolist())
277
+ window_index_id += (grid_t * llm_grid_h * llm_grid_w).item()
278
+ window_index = torch.cat(window_index, dim=0)
279
+
280
+ return window_index, cu_window_seqlens
281
+
282
+ def forward(
283
+ self,
284
+ tokens: torch.Tensor,
285
+ grid_thws: torch.Tensor,
286
+ output_hidden_states: bool = False,
287
+ ) -> Tuple[torch.Tensor, Optional[Tuple[torch.Tensor, ...]]]:
288
+ # RoPE, modified from qwen2.5_vl
289
+ rotary_pos_emb = self.rot_pos_emb(grid_thws)
290
+ window_index, cu_window_seqlens = self.get_window_index(grid_thws)
291
+ cu_window_seqlens = torch.tensor(
292
+ cu_window_seqlens,
293
+ device=tokens.device,
294
+ dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
295
+ )
296
+ cu_window_seqlens = torch.unique_consecutive(cu_window_seqlens)
297
+
298
+ seq_len, _ = tokens.size()
299
+ tokens = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
300
+ tokens = tokens[window_index, :, :]
301
+ tokens = tokens.reshape(seq_len, -1)
302
+ rotary_pos_emb = rotary_pos_emb.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
303
+ rotary_pos_emb = rotary_pos_emb[window_index, :, :]
304
+ rotary_pos_emb = rotary_pos_emb.reshape(seq_len, -1)
305
+ emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
306
+ position_embeddings = (emb.cos(), emb.sin())
307
+
308
+ cu_seqlens = torch.repeat_interleave(grid_thws[:, 1] * grid_thws[:, 2], grid_thws[:, 0]).cumsum(
309
+ dim=0,
310
+ # Select dtype based on the following factors:
311
+ # - FA2 requires that cu_seqlens_q must have dtype int32
312
+ # - torch.onnx.export requires that cu_seqlens_q must have same dtype as grid_thw
313
+ # See https://github.com/huggingface/transformers/pull/34852 for more information
314
+ dtype=grid_thws.dtype if torch.jit.is_tracing() else torch.int32,
315
+ )
316
+ cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
317
+
318
+ reverse_indices = torch.argsort(window_index)
319
+
320
+ hidden_states = () if output_hidden_states else None
321
+ for index, block in enumerate(self.blocks):
322
+ if self.fullatt_block_indexes is None or index in self.fullatt_block_indexes:
323
+ cu_seqlens_tmp = cu_seqlens
324
+ else:
325
+ cu_seqlens_tmp = cu_window_seqlens
326
+ if self.gradient_checkpointing and self.training:
327
+ tokens = self._gradient_checkpointing_func(block.__call__, tokens, cu_seqlens_tmp, position_embeddings)
328
+ else:
329
+ tokens = block(tokens, cu_seqlens_tmp, position_embeddings)
330
+ if output_hidden_states:
331
+ tokens_ = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
332
+ hidden_states += (tokens_[reverse_indices,:].reshape(seq_len, -1),)
333
+ tokens = self.post_trunk_norm(tokens)
334
+ tokens = tokens.reshape(seq_len // self.spatial_merge_unit, self.spatial_merge_unit, -1)
335
+ tokens = tokens[reverse_indices,:].reshape(seq_len, -1)
336
+
337
+ return tokens, hidden_states
338
+
339
+
340
+ class AIMv2PretrainedModel(PreTrainedModel):
341
+ config_class = AIMv2Config
342
+ base_model_prefix = "aimv2"
343
+ supports_gradient_checkpointing = True
344
+ main_input_name = "pixel_values"
345
+ _no_split_modules = ["AIMv2ViTPreprocessor", "AIMv2Block"]
346
+ _supports_sdpa = True
347
+
348
+
349
+ class AIMv2Model(AIMv2PretrainedModel):
350
+ def __init__(self, config: AIMv2Config):
351
+ super().__init__(config)
352
+ self.preprocessor = AIMv2ViTPreprocessor(config)
353
+ self.trunk = AIMv2Transformer(config)
354
+
355
+ def forward(
356
+ self,
357
+ pixel_values: torch.Tensor,
358
+ grid_thws: torch.Tensor,
359
+ output_hidden_states: Optional[bool] = None,
360
+ return_dict: Optional[bool] = None,
361
+ ) -> Union[
362
+ Tuple[torch.Tensor],
363
+ Tuple[torch.Tensor, Tuple[torch.Tensor, ...]],
364
+ BaseModelOutputWithNoAttention,
365
+ ]:
366
+ if output_hidden_states is None:
367
+ output_hidden_states = self.config.output_hidden_states
368
+ if return_dict is None:
369
+ return_dict = self.config.use_return_dict
370
+
371
+ x = self.preprocessor(pixel_values, grid_thws=grid_thws)
372
+
373
+ x, hidden_states = self.trunk(
374
+ x, grid_thws=grid_thws, output_hidden_states=output_hidden_states
375
+ )
376
+
377
+ if not return_dict:
378
+ res = (x,)
379
+ res += (hidden_states,) if output_hidden_states else ()
380
+ return res
381
+
382
+ return BaseModelOutputWithNoAttention(
383
+ last_hidden_state=x,
384
+ hidden_states=hidden_states,
385
+ )
modeling_ovis_u1.py ADDED
@@ -0,0 +1,921 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import logging
2
+ import math
3
+ from datetime import datetime
4
+ from importlib import import_module
5
+ from typing import List, Union, Optional, Dict
6
+
7
+ import numpy as np
8
+ import PIL.Image
9
+ import torch
10
+ from torch import Tensor
11
+ from torch.nn import init
12
+ from torch.nn.functional import softmax, gumbel_softmax, pad
13
+ from torchvision import transforms
14
+ import transformers
15
+ from transformers import AutoImageProcessor
16
+ from transformers import PreTrainedModel, AutoConfig, AutoModel, AutoTokenizer, AutoModelForCausalLM
17
+ from transformers.generation.utils import GenerateOutput
18
+ from transformers import CLIPImageProcessor
19
+
20
+ from .modeling_aimv2 import AIMv2Model
21
+ from .configuration_ovis_u1 import BaseVisualTokenizerConfig, Aimv2VisualTokenizerConfig
22
+ from .configuration_ovis_u1 import OvisU1Config, ConversationFormatter
23
+ from .configuration_ovis_u1 import IGNORE_ID, IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS, IMAGE_TOKEN_ID, VIDEO_TOKEN_ID
24
+
25
+ # ----------------------------------------------------------------------
26
+ # Visual Tokenizer
27
+ # ----------------------------------------------------------------------
28
+ class BaseVisualTokenizer(PreTrainedModel):
29
+ base_model_prefix = "backbone"
30
+ main_input_name = None
31
+ _image_processor_class = None
32
+ _image_processor_kwargs = {}
33
+ _backbone_class = None
34
+
35
+ def __init__(self, config: BaseVisualTokenizerConfig, *inputs, **kwargs):
36
+ super().__init__(config, *inputs, **kwargs)
37
+ if kwargs.get('train_from_scratch'):
38
+ # for key in self._image_processor_kwargs.keys():
39
+ # self._image_processor_kwargs[key] = getattr(self.config, key, self._image_processor_kwargs[key])
40
+ image_processor = self._image_processor_class.from_pretrained(kwargs['backbone_name_or_path'],
41
+ **self._image_processor_kwargs)
42
+
43
+ self.backbone = self._backbone_class.from_pretrained(kwargs['backbone_name_or_path'], **self.config.backbone_kwargs)
44
+ self.config.backbone_config = self.backbone.config
45
+
46
+ config = image_processor.to_dict()
47
+ if getattr(self.config, 'image_processor_new_kwargs', None) is not None:
48
+ for key in self.config.image_processor_new_kwargs.keys():
49
+ config[key] = self.config.image_processor_new_kwargs[key]
50
+ if 'patch_size' not in config:
51
+ assert getattr(self.backbone.config, 'patch_size'), "Patch size must be set."
52
+ config['patch_size'] = self.backbone.config.patch_size
53
+ self.image_processor = self._image_processor_class.from_dict(config)
54
+
55
+ else:
56
+ self.image_processor = AutoImageProcessor.from_pretrained(kwargs['image_processor_name_or_path'])
57
+ self.backbone = AutoModel.from_config(self.config.backbone_config)
58
+ head_dim = self.config.vocab_size - len(IMAGE_INDICATOR_IDS) # reserved tokens for IMAGE_INDICATORS
59
+ self.head = torch.nn.Sequential(
60
+ torch.nn.Linear(
61
+ self.backbone.config.hidden_size * self.config.hidden_stride * self.config.hidden_stride, head_dim,
62
+ bias=False
63
+ ),
64
+ torch.nn.LayerNorm(head_dim)
65
+ )
66
+ assert all((self.image_processor.do_resize,
67
+ not getattr(self.image_processor, 'do_center_crop', False),
68
+ self.image_processor.do_rescale,
69
+ self.image_processor.do_normalize
70
+ )), f"image_processor `{self.image_processor}` is not supported currently"
71
+
72
+ def get_backbone(self):
73
+ return self.backbone
74
+
75
+ def get_monitor_tensors(self):
76
+ raise NotImplementedError
77
+
78
+ def get_image_processor(self):
79
+ return self.image_processor
80
+
81
+ def mock_input(self):
82
+ height, width = self.get_image_size()
83
+ return torch.zeros(1, 3, height, width), self.construct_image_placeholders((1, 1))
84
+
85
+ def get_head(self):
86
+ return self.head
87
+
88
+ def get_image_size(self):
89
+ raise NotImplementedError
90
+
91
+ @staticmethod
92
+ def construct_image_placeholders(grid, data_type='image'):
93
+ if data_type == 'image':
94
+ image_placeholders = [IMAGE_INDICATOR_IDS[0], IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS[1]]
95
+ elif data_type == 'video':
96
+ image_placeholders = [IMAGE_INDICATOR_IDS[2], IMAGE_ATOM_ID, IMAGE_INDICATOR_IDS[2]]
97
+ else:
98
+ raise TypeError
99
+
100
+ return image_placeholders
101
+
102
+ @staticmethod
103
+ def _partition(img_size, grid):
104
+ w, h = img_size
105
+ row_height = h // grid[0]
106
+ col_width = w // grid[1]
107
+
108
+ partition = []
109
+ for row in range(grid[0]):
110
+ for col in range(grid[1]):
111
+ left = col * col_width
112
+ upper = row * row_height
113
+ right = w if col == grid[1] - 1 else (col + 1) * col_width
114
+ lower = h if row == grid[0] - 1 else (row + 1) * row_height
115
+ partition.append((left, upper, right, lower))
116
+
117
+ return partition
118
+
119
+ @staticmethod
120
+ def get_best_grid(img_size, side, max_partition, covering_threshold):
121
+
122
+ def _covering_area(left, upper, right, lower, side):
123
+ w = right - left
124
+ h = lower - upper
125
+ w, h = max(w, h), min(w, h)
126
+ if w > side:
127
+ h = h / w * side
128
+ w = side
129
+ return w * h
130
+
131
+ img_area = img_size[0] * img_size[1]
132
+
133
+ candidate_grids = []
134
+ for i in range(1, max_partition + 1):
135
+ for j in range(1, max_partition + 1):
136
+ if i * j <= max_partition:
137
+ candidate_grids.append((i, j))
138
+
139
+ all_grids = []
140
+ good_grids = []
141
+ for grid in candidate_grids:
142
+ partition = BaseVisualTokenizer._partition(img_size, grid)
143
+ covering_ratio = sum([_covering_area(*p, side) for p in partition]) / img_area
144
+ assert covering_ratio <= 1.0
145
+ all_grids.append((grid, covering_ratio))
146
+ if covering_ratio > covering_threshold:
147
+ good_grids.append((grid, covering_ratio))
148
+
149
+ if len(good_grids) > 0:
150
+ # pick the good partition with minimum #sub_images and break the tie using covering_ratio
151
+ return sorted(good_grids, key=lambda x: (x[0][0] * x[0][1], -x[1]))[0][0]
152
+ else:
153
+ # pick the partition with maximum covering_ratio and break the tie using #sub_images
154
+ return sorted(all_grids, key=lambda x: (-x[1], x[0][0] * x[0][1]))[0][0]
155
+
156
+ def preprocess_image(self, image: PIL.Image.Image, max_partition=4, covering_threshold=0.9, convert_to_rgb=True):
157
+ def _preprocess(img: PIL.Image.Image, side):
158
+ # first resize and preprocess
159
+ w, h = img.size
160
+ if w == h:
161
+ new_width = new_height = side
162
+ elif w > h:
163
+ new_width = side
164
+ new_height = int(h / w * new_width)
165
+ else:
166
+ new_height = side
167
+ new_width = int(w / h * new_height)
168
+ new_size = dict(height=new_height, width=new_width)
169
+ pixel_values = self.image_processor.preprocess(img, size=new_size, return_tensors='pt')['pixel_values']
170
+
171
+ # then pad to square
172
+ square_values = torch.zeros([1, 3, side, side], dtype=pixel_values.dtype, device=pixel_values.device)
173
+ new_height, new_width = pixel_values.shape[2:]
174
+ if new_height == new_width:
175
+ square_values[:, :, :, :] = pixel_values
176
+ elif new_height > new_width:
177
+ from_index = (side - new_width) // 2
178
+ square_values[:, :, :, from_index:from_index + new_width] = pixel_values
179
+ else:
180
+ from_index = (side - new_height) // 2
181
+ square_values[:, :, from_index:from_index + new_height, :] = pixel_values
182
+
183
+ return square_values
184
+
185
+ if convert_to_rgb and image.mode != 'RGB':
186
+ image = image.convert('RGB')
187
+
188
+ sides = self.get_image_size()
189
+ if sides[0] != sides[1]:
190
+ raise ValueError('get_image_size() returns non-square size')
191
+ side = sides[0]
192
+ grid = self.get_best_grid(image.size, side, max_partition, covering_threshold)
193
+ partition = self._partition(image.size, grid)
194
+ crops = [image.crop(p) for p in partition]
195
+ if len(crops) > 1:
196
+ crops.insert(0, image)
197
+ pixel_values = torch.cat([_preprocess(crop, side) for crop in crops], dim=0)
198
+ image_placeholders = self.construct_image_placeholders(grid)
199
+ return pixel_values, image_placeholders
200
+
201
+ def get_backbone_layer(self, index):
202
+ if 'aimv2' in self.config.model_type:
203
+ return self.backbone.trunk.blocks[index]
204
+ else:
205
+ return self.backbone.vision_model.encoder.layers[index]
206
+
207
+ def tokenize(self, logits):
208
+ def st_argmax(y_soft, dim): # straight-through softmax
209
+ index = y_soft.max(dim, keepdim=True)[1]
210
+ y_hard = torch.zeros_like(y_soft, memory_format=torch.legacy_contiguous_format).scatter_(dim, index, 1.0)
211
+ ret = y_hard - y_soft.detach() + y_soft
212
+ return ret
213
+
214
+ if self.config.tokenize_function == 'softmax':
215
+ tokens = softmax(logits, dim=-1, dtype=torch.float32).to(logits.dtype)
216
+ elif self.config.tokenize_function == 'gumbel_argmax':
217
+ tokens = gumbel_softmax(logits, tau=self.config.tau, hard=True)
218
+ elif self.config.tokenize_function == 'st_argmax':
219
+ tokens = st_argmax(logits, dim=-1)
220
+ else:
221
+ raise ValueError(
222
+ f'Invalid `max_type`, expected softmax or gumbel_argmax or st_argmax, but got {self.config.tokenize_function}')
223
+ return tokens
224
+
225
+ def encode(self, pixel_values):
226
+ output = self.backbone(pixel_values, output_hidden_states=True, return_dict=True)
227
+ features = output.hidden_states[-1]
228
+ if self.config.drop_cls_token:
229
+ features = features[:, 1:, :]
230
+
231
+ # merge number of `hidden_stride * hidden_stride` hidden states together to reduce token sequence length
232
+ # e.g., for hidden_stride=3, this leads to a token length reduction: 729 -> 81 for siglip
233
+ if self.config.hidden_stride > 1:
234
+ n, l, d = features.shape # this `d` maybe different from the above `d
235
+ sqrt_l = int(l ** 0.5)
236
+ assert sqrt_l ** 2 == l, "The token sequence length should be a perfect square."
237
+ features = features.reshape(n, sqrt_l, sqrt_l, d)
238
+ pl = (self.config.hidden_stride - (sqrt_l % self.config.hidden_stride)) % self.config.hidden_stride
239
+ features = pad(features, (0, 0, 0, pl, 0, pl), "constant", 0)
240
+ sqrt_l += pl
241
+ features = features.reshape(n, sqrt_l // self.config.hidden_stride, self.config.hidden_stride,
242
+ sqrt_l // self.config.hidden_stride, self.config.hidden_stride, d)
243
+ features = features.permute(0, 1, 3, 2, 4, 5) # [n, sqrt_l/hs, sqrt_l/hs, hs, hs, d]
244
+ features = features.flatten(3) # [n, sqrt_l/hs, sqrt_l/hs, hs*hs*d]
245
+ features = features.reshape(
246
+ n, -1, self.config.hidden_stride * self.config.hidden_stride * d)
247
+
248
+ return features
249
+
250
+ def forward(self, pixel_values) -> torch.Tensor: # [BatchSize, ImageShape] -> [BatchSize, #Token, VocabSize]
251
+ features = self.encode(pixel_values)
252
+ logits = self.head(features)
253
+ tokens = self.tokenize(logits)
254
+ # tokens' shape is [BatchSize, #Token, VocabSize-5], so padding with [BatchSize, #Token, 5], after
255
+ # which, tokens' shape should become [BatchSize, #Token, VocabSize]
256
+ batch_size, token_len, _ = tokens.shape
257
+ padding_tensor = torch.zeros(size=(batch_size, token_len, len(IMAGE_INDICATOR_IDS)),
258
+ dtype=tokens.dtype,
259
+ device=tokens.device,
260
+ layout=tokens.layout,
261
+ requires_grad=False)
262
+ tokens = torch.cat((tokens, padding_tensor), dim=2)
263
+ return tokens
264
+
265
+ class Aimv2VisualTokenizer(BaseVisualTokenizer):
266
+ config_class = Aimv2VisualTokenizerConfig
267
+ supports_gradient_checkpointing = True
268
+ _no_split_modules = ["AIMv2ViTPreprocessor", "AIMv2Block"]
269
+ _image_processor_class = CLIPImageProcessor
270
+ _image_processor_kwargs = dict(do_center_crop=False, crop_size={'height': -1, 'width': -1}, size={'shortest_edge':-1})
271
+ _backbone_class = AIMv2Model
272
+
273
+ # Copied from qwen2_vl
274
+ def smart_resize(self,
275
+ height: int, width: int, factor: int = 28, min_pixels: int = 56 * 56, max_pixels: int = 14 * 14 * 4 * 1280
276
+ ):
277
+ """Rescales the image so that the following conditions are met:
278
+
279
+ 1. Both dimensions (height and width) are divisible by 'factor'.
280
+
281
+ 2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
282
+
283
+ 3. The aspect ratio of the image is maintained as closely as possible.
284
+
285
+ """
286
+
287
+ if height < factor or width < factor:
288
+ print(f"height:{height} or width:{width} must be larger than factor:{factor}")
289
+ if height < width:
290
+ width = round(factor/height*width)
291
+ height = factor
292
+ else:
293
+ height = round(factor/width*height)
294
+ width = factor
295
+
296
+ elif max(height, width) / min(height, width) > 200:
297
+ print(
298
+ f"absolute aspect ratio must be smaller than 200, got {max(height, width) / min(height, width)}"
299
+ )
300
+ if height > width:
301
+ height = 200 * width
302
+ else:
303
+ width = 200 * height
304
+
305
+ h_bar = round(height / factor) * factor
306
+ w_bar = round(width / factor) * factor
307
+ if h_bar * w_bar > max_pixels:
308
+ beta = math.sqrt((height * width) / max_pixels)
309
+ h_bar = math.floor(height / beta / factor) * factor
310
+ w_bar = math.floor(width / beta / factor) * factor
311
+ elif h_bar * w_bar < min_pixels:
312
+ beta = math.sqrt(min_pixels / (height * width))
313
+ h_bar = math.ceil(height * beta / factor) * factor
314
+ w_bar = math.ceil(width * beta / factor) * factor
315
+ return h_bar, w_bar
316
+
317
+ def get_monitor_tensors(self):
318
+ return dict(
319
+ backbone_bottom=self.backbone.trunk.blocks[0].attn.qkv.weight,
320
+ backbone_top=self.backbone.trunk.blocks[-1].attn.qkv.weight,
321
+ head=self.head[0].weight
322
+ )
323
+
324
+ def get_min_image_size(self):
325
+ min_pixels = self.image_processor.min_pixels
326
+ max_pixels = self.image_processor.max_pixels
327
+ height = int(min_pixels**0.5)
328
+ width = int(min_pixels**0.5)
329
+ patch_size = self.image_processor.patch_size
330
+ hidden_stride = self.image_processor.hidden_stride
331
+ height, width = self.smart_resize(height, width, patch_size * hidden_stride, min_pixels, max_pixels)
332
+ return height, width
333
+
334
+ def get_image_size(self):
335
+ min_pixels = self.image_processor.min_pixels
336
+ max_pixels = self.image_processor.max_pixels
337
+ num_pixels = (min_pixels+max_pixels) / 2
338
+ height = int(num_pixels**0.5)
339
+ width = int(num_pixels**0.5)
340
+ patch_size = self.image_processor.patch_size
341
+ hidden_stride = self.image_processor.hidden_stride
342
+ height, width = self.smart_resize(height, width, patch_size * hidden_stride, min_pixels, max_pixels)
343
+ return height, width
344
+
345
+ def get_token_length(self, width: int,
346
+ height: int,
347
+ n_frames: int = 1,
348
+ num_images: int = 1):
349
+ patch_size = self.image_processor.patch_size
350
+ temporal_patch_size = self.image_processor.temporal_patch_size
351
+ hidden_stride = self.image_processor.hidden_stride
352
+ min_pixels = self.image_processor.min_pixels
353
+ max_pixels = self.image_processor.max_pixels
354
+
355
+ max_pixels = max_pixels // num_images
356
+ min_pixels = min(max_pixels, min_pixels)
357
+
358
+ resized_height, resized_width = height, width
359
+ resized_height, resized_width = self.smart_resize(
360
+ height,
361
+ width,
362
+ factor=patch_size * hidden_stride,
363
+ min_pixels=min_pixels,
364
+ max_pixels=max_pixels,
365
+ )
366
+
367
+ if n_frames % temporal_patch_size != 0:
368
+ n_frames = n_frames + temporal_patch_size - 1
369
+ grid_t = n_frames // temporal_patch_size
370
+ grid_h, grid_w = resized_height // patch_size // hidden_stride, resized_width // patch_size // hidden_stride
371
+
372
+ return grid_t * grid_w * grid_h
373
+
374
+ def mock_input(self):
375
+ height, width = self.get_min_image_size()
376
+ return torch.zeros(1, 3, height, width), self.construct_image_placeholders((1, 1))
377
+
378
+ def preprocess_image(self, images: Union[PIL.Image.Image, List[PIL.Image.Image]],
379
+ convert_to_rgb: Optional[bool] = True,
380
+ num_images: Optional[int] = 1,
381
+ min_pixels: Optional[int] = None,
382
+ max_pixels: Optional[int] = None,
383
+ multimodal_type: Optional[str] = 'single_image'):
384
+
385
+
386
+ patch_size = self.image_processor.patch_size # 14
387
+ temporal_patch_size = self.image_processor.temporal_patch_size # 1
388
+ hidden_stride = self.image_processor.hidden_stride # 2
389
+ min_pixels = min_pixels or self.image_processor.min_pixels # 200704
390
+ max_pixels = max_pixels or self.image_processor.max_pixels # 3211264
391
+
392
+ max_pixels = max_pixels // num_images
393
+ min_pixels = min(max_pixels, min_pixels)
394
+
395
+ if not isinstance(images, list):
396
+ images = [images]
397
+ if multimodal_type == 'video':
398
+ assert len(images) >= 1
399
+ else:
400
+ assert len(images) == 1
401
+ images = [image.convert("RGB") if convert_to_rgb and image.mode != 'RGB' else image for image in images ]
402
+ # images = [np.array(image) for image in images]
403
+
404
+ width, height = images[0].size
405
+ resized_height, resized_width = height, width
406
+ processed_images = []
407
+ for image in images:
408
+ resized_height, resized_width = self.smart_resize(
409
+ height,
410
+ width,
411
+ factor=patch_size * hidden_stride,
412
+ min_pixels=min_pixels,
413
+ max_pixels=max_pixels,
414
+ )
415
+ new_size = dict(height=resized_height, width=resized_width)
416
+ image_pt = self.image_processor.preprocess(image, size=new_size, return_tensors="np")['pixel_values'][0]
417
+
418
+ processed_images.append(image_pt)
419
+
420
+ patches = np.array(processed_images)
421
+ # if data_format == ChannelDimension.LAST:
422
+ # patches = patches.transpose(0, 3, 1, 2)
423
+ if patches.shape[0] % temporal_patch_size != 0:
424
+ repeats = np.repeat(patches[-1][np.newaxis], temporal_patch_size - 1, axis=0)
425
+ patches = np.concatenate([patches, repeats], axis=0)
426
+ channel = patches.shape[1]
427
+ grid_t = patches.shape[0] // temporal_patch_size # 1
428
+ grid_h, grid_w = resized_height // patch_size, resized_width // patch_size # 32, 32
429
+
430
+ patches = patches.reshape(
431
+ grid_t,
432
+ temporal_patch_size,
433
+ channel,
434
+ grid_h // hidden_stride,
435
+ hidden_stride,
436
+ patch_size,
437
+ grid_w // hidden_stride,
438
+ hidden_stride,
439
+ patch_size,
440
+ )
441
+ patches = patches.transpose(0, 3, 6, 4, 7, 2, 1, 5, 8)
442
+ flatten_patches = patches.reshape(
443
+ grid_t * grid_h * grid_w, channel * temporal_patch_size * patch_size * patch_size
444
+ )
445
+ # 1024, 588
446
+
447
+ image_placeholders = self.construct_image_placeholders((1, 1), data_type='video' if multimodal_type=='video' else 'image') # [-301, -300, -302, -305]
448
+
449
+ # print(flatten_patches.shape, len(images))
450
+ return torch.tensor(flatten_patches), torch.tensor([[grid_t, grid_h, grid_w]]), image_placeholders
451
+
452
+ def encode(self, pixel_values, grid_thws):
453
+ output = self.backbone(pixel_values, grid_thws, output_hidden_states=True, return_dict=True)
454
+ features = output.hidden_states[-1]
455
+ # default: false
456
+ # if self.config.drop_cls_token:
457
+ # features = features[:, 1:, :]
458
+
459
+ # refer to qwen2.5-vl patchmerger
460
+ seq_len, _ = features.shape
461
+ features = features.reshape(seq_len//(self.config.hidden_stride ** 2), -1)
462
+
463
+ return features
464
+
465
+ def forward(self, pixel_values, grid_thws) -> torch.Tensor: # [BatchSize, ImageShape] -> [BatchSize, #Token, VocabSize]
466
+ features = self.encode(pixel_values, grid_thws)
467
+ logits = self.head(features)
468
+ tokens = self.tokenize(logits)
469
+ # tokens' shape is [#Token, VocabSize-5], so padding with [#Token, 5], after
470
+ # which, tokens' shape should become [#Token, VocabSize];
471
+ # this is different from original aimv2 which has [BatchSize, #Token, VocabSize-5]
472
+ token_len, _ = tokens.shape
473
+ padding_tensor = torch.zeros(size=(token_len, len(IMAGE_INDICATOR_IDS)),
474
+ dtype=tokens.dtype,
475
+ device=tokens.device,
476
+ layout=tokens.layout,
477
+ requires_grad=False)
478
+ tokens = torch.cat((tokens, padding_tensor), dim=1)
479
+ return tokens
480
+
481
+ AutoModel.register(Aimv2VisualTokenizerConfig, Aimv2VisualTokenizer)
482
+
483
+
484
+
485
+
486
+ # ----------------------------------------------------------------------
487
+ # Visual Generator
488
+ # ----------------------------------------------------------------------
489
+ from .configuration_yak import YakConfig
490
+ from .modeling_yak import YakModel
491
+ AutoConfig.register("yak", YakConfig)
492
+ AutoModel.register(YakConfig, YakModel)
493
+
494
+
495
+
496
+ # ----------------------------------------------------------------------
497
+ # OvisU1
498
+ # ----------------------------------------------------------------------
499
+ class VisualEmbedding(torch.nn.Embedding):
500
+ def forward(self, visual_tokens: Tensor) -> Tensor:
501
+ if visual_tokens.dtype in [torch.int8, torch.int16, torch.int32, torch.int64, torch.long]:
502
+ return super().forward(visual_tokens)
503
+ return torch.matmul(visual_tokens, self.weight)
504
+
505
+ def reset_parameters(self, mean=0., std=1.) -> None:
506
+ init.normal_(self.weight, mean=mean, std=std)
507
+ self._fill_padding_idx_with_zero()
508
+
509
+
510
+ class OvisU1PreTrainedModel(PreTrainedModel):
511
+ config_class = OvisU1Config
512
+ base_model_prefix = "ovis_u1"
513
+
514
+
515
+ class OvisU1(OvisU1PreTrainedModel):
516
+
517
+ def __init__(self, config: OvisU1Config, *inputs, **kwargs):
518
+ super().__init__(config, *inputs, **kwargs)
519
+ attn_kwargs = dict()
520
+ if self.config.llm_attn_implementation:
521
+ attn_kwargs['attn_implementation'] = self.config.llm_attn_implementation
522
+ self.llm = AutoModelForCausalLM.from_config(self.config.llm_config, **attn_kwargs)
523
+ assert self.config.hidden_size == self.llm.config.hidden_size, "hidden size mismatch"
524
+ self.text_tokenizer = AutoTokenizer.from_pretrained(self.config.name_or_path)
525
+ self.visual_tokenizer = AutoModel.from_config(self.config.visual_tokenizer_config,
526
+ image_processor_name_or_path=self.config.name_or_path)
527
+ self.visual_generator = AutoModel.from_config(self.config.visual_generator_config)
528
+ self.vte = VisualEmbedding(self.config.visual_tokenizer_config.vocab_size, self.config.hidden_size,
529
+ device=self.visual_tokenizer.device, dtype=self.visual_tokenizer.dtype)
530
+
531
+ def _merge_modules(modules_list: tuple):
532
+ merged_modules = []
533
+ for modules in modules_list:
534
+ merged_modules.extend(modules if modules else [])
535
+ return merged_modules
536
+
537
+ self._no_split_modules = _merge_modules((self.llm._no_split_modules, self.visual_tokenizer._no_split_modules))
538
+ self._skip_keys_device_placement = self.llm._skip_keys_device_placement
539
+ self._keep_in_fp32_modules = _merge_modules(
540
+ (self.llm._keep_in_fp32_modules, self.visual_tokenizer._keep_in_fp32_modules))
541
+ self._supports_flash_attn_2 = True
542
+ self.is_parallelizable = all((self.llm.is_parallelizable, self.visual_tokenizer.is_parallelizable, self.visual_generator.is_parallelizable))
543
+ self.supports_gradient_checkpointing = all(
544
+ (self.llm.supports_gradient_checkpointing, self.visual_tokenizer.supports_gradient_checkpointing, self.visual_generator.supports_gradient_checkpointing))
545
+ self._supports_sdpa = all((self.llm._supports_sdpa, self.visual_tokenizer._supports_sdpa, self.visual_generator._supports_sdpa))
546
+
547
+ def get_text_tokenizer(self):
548
+ return self.text_tokenizer
549
+
550
+ def get_visual_tokenizer(self):
551
+ return self.visual_tokenizer
552
+
553
+ def get_visual_generator(self):
554
+ return self.visual_generator
555
+
556
+ def tie_weights(self):
557
+ if not self.config.disable_tie_weight:
558
+ self.get_llm().tie_weights()
559
+
560
+ def get_lm_head(self):
561
+ return self.get_llm().get_output_embeddings()
562
+
563
+ def get_llm(self):
564
+ return self.llm
565
+
566
+ def get_vte(self):
567
+ return self.vte
568
+
569
+ def get_wte(self):
570
+ return self.llm.get_input_embeddings()
571
+
572
+ def get_conversation_formatter(self) -> ConversationFormatter:
573
+ if getattr(self, 'conversation_formatter', None) is None:
574
+ self.conversation_formatter = getattr(import_module(".configuration_ovis_u1", __package__),
575
+ self.config.conversation_formatter_class)(self.text_tokenizer)
576
+ return self.conversation_formatter
577
+
578
+ def merge_multimodal(
579
+ self,
580
+ text_input_ids: torch.Tensor,
581
+ text_attention_masks: torch.Tensor,
582
+ text_labels: Optional[torch.Tensor],
583
+ pixel_values: Optional[torch.Tensor],
584
+ grid_thws: Optional[torch.Tensor],
585
+ left_padding: bool = False
586
+ ):
587
+ input_device = text_input_ids.device
588
+ visual_vocab_szie = self.get_visual_tokenizer().config.vocab_size
589
+ visual_indicator_embeds = self.get_vte()(
590
+ torch.tensor(
591
+ list(range(visual_vocab_szie - 5, visual_vocab_szie)),
592
+ dtype=torch.long,
593
+ device=self.get_visual_tokenizer().device
594
+ )
595
+ ).to(device=input_device)
596
+
597
+ if self.training:
598
+ # When training, to be compatible with deepspeed zero, each sample has to include pixel_value tensor.
599
+ # For text-only sample, one can simply use a full zero tensor as pixel_value, which will be ignored
600
+ # (see below in this function); so, the gradient will not be affected.
601
+ num_images = [x.prod() // (self.visual_tokenizer.config.hidden_stride**2) for x in grid_thws]
602
+
603
+ visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
604
+
605
+ visual_embeds_ = torch.split(self.get_vte()(visual_tokens).to(dtype=self.dtype, device=input_device),
606
+ split_size_or_sections=num_images, dim=0)
607
+
608
+
609
+
610
+ visual_input_ids_ = torch.split(torch.argmax(visual_tokens, dim=-1).to(device=input_device),
611
+ split_size_or_sections=num_images, dim=0)
612
+
613
+
614
+ visual_labels_ = [torch.full(x.shape, IGNORE_ID, dtype=torch.long, device=input_device) for x in
615
+ visual_input_ids_]
616
+
617
+
618
+ visual_embeds = []
619
+ visual_input_ids = []
620
+ visual_labels = []
621
+ ind = 0
622
+ for text_input_id in text_input_ids:
623
+ image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
624
+ n = len(image_atom_positions)
625
+ if n > 0:
626
+ visual_embeds.append(visual_embeds_[ind:ind+n])
627
+ visual_input_ids.append(visual_input_ids_[ind:ind+n])
628
+ visual_labels.append(visual_labels_[ind:ind+n])
629
+ ind += n
630
+ else:
631
+ visual_embeds.append(visual_embeds_[ind:ind+1])
632
+ visual_input_ids.append(visual_input_ids_[ind:ind+1])
633
+ visual_labels.append(visual_labels_[ind:ind+1])
634
+ ind += 1
635
+
636
+
637
+ else:
638
+ # TODO: Not modified yet
639
+ # When inference, sample can include only text with `None` pixel_value
640
+ # num_images = [x.shape[0] if x is not None else 0 for x in pixel_values]
641
+ num_images = [x.prod() // (self.visual_tokenizer.config.hidden_stride**2) if x is not None else 0 for x in grid_thws]
642
+ if sum(num_images) > 0:
643
+ visual_tokens = self.visual_tokenizer(pixel_values, grid_thws)
644
+ try:
645
+ visual_embeds_ = torch.split(self.get_vte()(visual_tokens).to(dtype=self.dtype, device=input_device),
646
+ split_size_or_sections=num_images, dim=0)
647
+ except Exception as e:
648
+ print(e)
649
+ print(pixel_values.shape, grid_thws.shape, visual_tokens.shape, num_images)
650
+
651
+
652
+ visual_input_ids_ = torch.split(torch.argmax(visual_tokens, dim=-1).to(device=input_device),
653
+ split_size_or_sections=num_images, dim=0)
654
+
655
+
656
+ visual_labels_ = [torch.full(x.shape, IGNORE_ID, dtype=torch.long, device=input_device) for x in
657
+ visual_input_ids_]
658
+
659
+ visual_embeds = []
660
+ visual_input_ids = []
661
+ visual_labels = []
662
+ ind = 0
663
+ for text_input_id in text_input_ids:
664
+ image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
665
+ n = len(image_atom_positions)
666
+ if n > 0:
667
+ visual_embeds.append(visual_embeds_[ind:ind+n])
668
+ visual_input_ids.append(visual_input_ids_[ind:ind+n])
669
+ visual_labels.append(visual_labels_[ind:ind+n])
670
+ ind += n
671
+ else:
672
+ visual_embeds.append(visual_embeds_[ind:ind+1])
673
+ visual_input_ids.append(visual_input_ids_[ind:ind+1])
674
+ visual_labels.append(visual_labels_[ind:ind+1])
675
+ ind += 1
676
+
677
+ else:
678
+ # just placeholders
679
+ visual_embeds = [None] * len(num_images)
680
+ visual_input_ids = [None] * len(num_images)
681
+ visual_labels = [None] * len(num_images)
682
+
683
+ # just placeholders
684
+ if text_labels is None:
685
+ text_labels = torch.full(text_input_ids.shape, IGNORE_ID, dtype=torch.long, device=input_device)
686
+
687
+ input_embeds = []
688
+ attention_masks = []
689
+ labels = []
690
+ input_img_poss = []
691
+ for text_input_id, text_label, text_attention_mask, visual_embed, visual_input_id, visual_label in zip(
692
+ text_input_ids, text_labels, text_attention_masks, visual_embeds, visual_input_ids, visual_labels
693
+ ):
694
+ placeholder_token_mask = torch.lt(text_input_id, 0)
695
+ text_embed = self.get_wte()(torch.masked_fill(text_input_id, placeholder_token_mask, 0))
696
+ for i, indicator_id in enumerate(IMAGE_INDICATOR_IDS):
697
+ text_embed[text_input_id == indicator_id] = visual_indicator_embeds[i]
698
+ image_atom_positions = torch.where(torch.eq(text_input_id, IMAGE_ATOM_ID))[0].tolist()
699
+ if len(image_atom_positions) > 0:
700
+ input_embed_parts = []
701
+ attention_mask_parts = []
702
+ label_parts = []
703
+ input_img_pos_parts = []
704
+ prev_image_atom_position = -1
705
+ for index, image_atom_position in enumerate(image_atom_positions):
706
+ input_embed_parts.append(
707
+ text_embed[prev_image_atom_position + 1:image_atom_position, :])
708
+ label_parts.append(
709
+ text_label[prev_image_atom_position + 1:image_atom_position])
710
+ input_img_pos_parts.append(
711
+ torch.zeros_like(text_label[prev_image_atom_position + 1:image_atom_position])
712
+ )
713
+ attention_mask_parts.append(
714
+ text_attention_mask[prev_image_atom_position + 1:image_atom_position])
715
+ input_embed_parts.append(visual_embed[index])
716
+ attention_mask_parts.append(
717
+ torch.ones_like(visual_label[index], dtype=torch.bool))
718
+ label_parts.append(visual_label[index])
719
+ input_img_pos_parts.append(
720
+ torch.ones_like(visual_label[index])
721
+ )
722
+ prev_image_atom_position = image_atom_position
723
+ if prev_image_atom_position + 1 < text_input_id.shape[0]:
724
+ input_embed_parts.append(
725
+ text_embed[prev_image_atom_position + 1:, :])
726
+ attention_mask_parts.append(
727
+ text_attention_mask[prev_image_atom_position + 1:])
728
+ label_parts.append(
729
+ text_label[prev_image_atom_position + 1:])
730
+ input_img_pos_parts.append(
731
+ torch.zeros_like(text_label[prev_image_atom_position + 1:])
732
+ )
733
+ input_embed = torch.cat(input_embed_parts, dim=0)
734
+ attention_mask = torch.cat(attention_mask_parts, dim=0)
735
+ label = torch.cat(label_parts, dim=0)
736
+ input_img_pos = torch.cat(input_img_pos_parts, dim=0)
737
+ else:
738
+ input_embed = text_embed
739
+ attention_mask = text_attention_mask
740
+ label = text_label
741
+ input_img_pos = torch.zeros_like(text_label)
742
+ if self.training:
743
+ # Make visual_embed & visual_indicator_embeds involved in the backward graph,
744
+ # to be compatible with deepspeed zero and ddp.
745
+ input_embed += torch.sum(visual_embed[0] * 0.0) + torch.sum(visual_indicator_embeds * 0.0)
746
+ input_embeds.append(input_embed)
747
+ attention_masks.append(attention_mask)
748
+ labels.append(label)
749
+ input_img_poss.append(input_img_pos)
750
+
751
+ batch_input_embeds = self.pad_truncate_sequence(input_embeds, batch_first=True, padding_value=0.0, left_padding=left_padding)
752
+ batch_attention_mask = self.pad_truncate_sequence(attention_masks, batch_first=True, padding_value=False, left_padding=left_padding)
753
+ batch_labels = self.pad_truncate_sequence(labels, batch_first=True, padding_value=IGNORE_ID, left_padding=left_padding)
754
+ batch_input_img_labels = self.pad_truncate_sequence(input_img_poss, batch_first=True, padding_value=0.0, left_padding=left_padding)
755
+
756
+ return visual_input_ids, batch_input_embeds, batch_labels, batch_attention_mask, batch_input_img_labels
757
+
758
+ def pad_truncate_sequence(self, sequences: List[torch.Tensor], batch_first: bool = True, padding_value: float = 0.0, left_padding: bool = False) -> torch.Tensor:
759
+ if left_padding == False:
760
+ pad_sequence = torch.nn.utils.rnn.pad_sequence(sequences, batch_first=batch_first, padding_value=padding_value)
761
+ return pad_sequence[:,:self.config.multimodal_max_length]
762
+ else:
763
+ 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])
764
+ return pad_sequence[:,-self.config.multimodal_max_length:]
765
+
766
+ def preprocess_inputs(
767
+ self,
768
+ text_or_conversations: Union[List[Dict], str],
769
+ images: Optional[Union[List[PIL.Image.Image], List[List[PIL.Image.Image]]]],
770
+ generation_preface='',
771
+ return_labels=False,
772
+ propagate_exception=True,
773
+ frame_selector=None,
774
+ multimodal_type="single_image",
775
+ fix_sample_overall_length_navit=False,
776
+ min_pixels=None,
777
+ max_pixels=None,
778
+ enable_thinking=False
779
+ ):
780
+ # convert text to conversations
781
+ if isinstance(text_or_conversations, str):
782
+ conversations = [{
783
+ "from": "human",
784
+ "value": text_or_conversations
785
+ }]
786
+ elif isinstance(text_or_conversations, list):
787
+ conversations = text_or_conversations
788
+ else:
789
+ raise ValueError(f'[{datetime.now()}] Invalid type of `text_or_conversations`, expected `List[Dict]` or `str`,'
790
+ f' but got {type(text_or_conversations)}')
791
+
792
+ if frame_selector is not None:
793
+ conversations, images = frame_selector(conversations=conversations,frames=images,clear_prompt=True)
794
+
795
+ # format conversations
796
+ prompt, raw_input_ids, raw_labels = self.get_conversation_formatter().format(
797
+ conversations, generation_preface=generation_preface, enable_thinking=enable_thinking)
798
+
799
+ # place image placeholders
800
+ input_ids = []
801
+ labels = []
802
+ pixel_values = []
803
+ grid_thws = []
804
+ invalidate_label = False
805
+ image_token_indices = [i for i, v in enumerate(raw_input_ids) if v == IMAGE_TOKEN_ID or v == VIDEO_TOKEN_ID]
806
+ last_image_token_index = -1
807
+ for i in range(len(image_token_indices)):
808
+ head = 0 if i == 0 else image_token_indices[i - 1] + 1
809
+ tail = image_token_indices[i]
810
+ last_image_token_index = tail
811
+ input_ids.extend(raw_input_ids[head:tail])
812
+ labels.extend(raw_labels[head:tail])
813
+ try:
814
+ # currently, do not support multiple videos
815
+ if multimodal_type == "video":
816
+ image = images
817
+ else:
818
+ image = images[i]
819
+ raw_pixel_values, image_grid_thws, image_placeholders = self.visual_tokenizer.preprocess_image(
820
+ image, num_images=len(images) if fix_sample_overall_length_navit else 1, min_pixels=min_pixels, max_pixels=max_pixels,
821
+ multimodal_type=multimodal_type)
822
+ except Exception as e:
823
+ if propagate_exception:
824
+ raise e
825
+ logging.exception(e)
826
+ invalidate_label = True
827
+ # raw_pixel_values, image_placeholders = self.visual_tokenizer.mock_input() # TODO
828
+ raw_pixel_values, _ = self.visual_tokenizer.mock_input()
829
+ mock_image = transforms.ToPILImage()(raw_pixel_values[0])
830
+ raw_pixel_values, image_grid_thws, image_placeholders = self.visual_tokenizer.preprocess_image(
831
+ mock_image, min_pixels=min_pixels, max_pixels=max_pixels)
832
+
833
+ input_ids.extend(image_placeholders)
834
+ labels.extend([IGNORE_ID] * len(image_placeholders))
835
+ pixel_values.append(raw_pixel_values)
836
+ grid_thws.append(image_grid_thws)
837
+ input_ids.extend(raw_input_ids[last_image_token_index + 1:])
838
+ labels.extend(raw_labels[last_image_token_index + 1:])
839
+
840
+ # return tensors
841
+ input_ids = torch.tensor(input_ids, dtype=torch.long)
842
+ labels = torch.tensor([IGNORE_ID] * len(labels) if invalidate_label else labels, dtype=torch.long)
843
+ pixel_values = torch.cat(pixel_values, dim=0) if len(pixel_values) > 0 else None
844
+ grid_thws = torch.cat(grid_thws, dim=0) if len(grid_thws) > 0 else None
845
+
846
+ if return_labels:
847
+ return prompt, input_ids, pixel_values, grid_thws, labels
848
+ else:
849
+ return prompt, input_ids, pixel_values, grid_thws
850
+
851
+ def generate(
852
+ self,
853
+ inputs: Optional[torch.Tensor] = None,
854
+ **kwargs,
855
+ ) -> Union[GenerateOutput, torch.LongTensor]:
856
+ # assert inputs.shape[0] == 1, 'Currently, only support `batch_size=1`'
857
+ _, inputs_embeds, labels, attention_mask, input_img_labels = self.merge_multimodal(
858
+ text_input_ids=inputs,
859
+ text_attention_masks=kwargs.pop('attention_mask'),
860
+ text_labels=None,
861
+ pixel_values=kwargs.pop('pixel_values'),
862
+ grid_thws=kwargs.pop('grid_thws'),
863
+ left_padding=True
864
+ )
865
+ inputs_embeds = inputs_embeds.detach()
866
+ torch.cuda.empty_cache()
867
+ return self.llm.generate(inputs=None, inputs_embeds=inputs_embeds, attention_mask=attention_mask, **kwargs)
868
+
869
+ def generate_condition(
870
+ self,
871
+ inputs: Optional[torch.Tensor] = None,
872
+ **kwargs,
873
+ ):
874
+ # assert inputs.shape[0] == 1, 'Currently, only support `batch_size=1`'
875
+ _, inputs_embeds, labels, attention_mask, input_img_labels = self.merge_multimodal(
876
+ text_input_ids=inputs,
877
+ text_attention_masks=kwargs.pop('attention_mask'),
878
+ text_labels=None,
879
+ pixel_values=kwargs.pop('pixel_values'),
880
+ grid_thws=kwargs.pop('grid_thws'),
881
+ left_padding=True
882
+ )
883
+ inputs_embeds = inputs_embeds.detach()
884
+ torch.cuda.empty_cache()
885
+ device = self.llm.device
886
+ outputs = self.llm(inputs_embeds=inputs_embeds.to(device),
887
+ labels=labels.to(device),
888
+ attention_mask=attention_mask.to(device),
889
+ output_hidden_states=True,
890
+ **kwargs)
891
+ semantic_cond_0 = outputs.hidden_states[-1]
892
+ semantic_cond_1 = outputs.hidden_states[-2]
893
+ semantic_cond = torch.cat([semantic_cond_0, semantic_cond_1], dim=-1)
894
+ return dict(
895
+ txt=semantic_cond
896
+ )
897
+
898
+ def generate_img(
899
+ self,
900
+ inputs: Optional[torch.Tensor] = None,
901
+ cond = None,
902
+ no_both_cond = None,
903
+ no_txt_cond = None,
904
+ **kwargs,
905
+ ) -> Union[GenerateOutput, torch.LongTensor]:
906
+ if cond is None:
907
+ cond = self.generate_condition(inputs, **kwargs)
908
+
909
+ height = kwargs.get('height', 1024)
910
+ width = kwargs.get('width', 1024)
911
+ num_steps = kwargs.get('num_steps', 50)
912
+ seed = kwargs.get('seed', 42)
913
+ img_cfg = kwargs.pop('img_cfg', 1.5)
914
+ txt_cfg = kwargs.pop('txt_cfg', 5)
915
+ yak_output = self.visual_generator.generate_image(
916
+ cond=cond, no_txt_cond=no_txt_cond, no_both_cond=no_both_cond,
917
+ height=height, width=width,
918
+ num_steps=num_steps, seed=seed,
919
+ img_cfg=img_cfg, txt_cfg=txt_cfg,
920
+ output_type="pil")
921
+ return yak_output
modeling_yak.py ADDED
@@ -0,0 +1,1461 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Callable
2
+ import math
3
+ from dataclasses import dataclass
4
+ import collections.abc
5
+ from itertools import repeat as iter_repeat
6
+
7
+ import numpy as np
8
+ import torch
9
+ from torch import Tensor, nn
10
+ import torchvision
11
+ from torchvision import transforms
12
+ from diffusers import AutoencoderKL
13
+ from PIL import Image
14
+ from PIL.ImageOps import exif_transpose
15
+ from torch.nn import functional as F
16
+ from transformers.modeling_utils import PreTrainedModel
17
+ from transformers.utils import ModelOutput
18
+ from einops import rearrange, repeat
19
+
20
+ from .configuration_yak import YakConfig
21
+
22
+
23
+ def _ntuple(n):
24
+ def parse(x):
25
+ if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
26
+ x = tuple(x)
27
+ if len(x) == 1:
28
+ x = tuple(iter_repeat(x[0], n))
29
+ return x
30
+ return tuple(iter_repeat(x, n))
31
+ return parse
32
+
33
+
34
+ to_1tuple = _ntuple(1)
35
+ to_2tuple = _ntuple(2)
36
+ to_3tuple = _ntuple(3)
37
+ to_4tuple = _ntuple(4)
38
+
39
+
40
+ def as_tuple(x):
41
+ if isinstance(x, collections.abc.Iterable) and not isinstance(x, str):
42
+ return tuple(x)
43
+ if x is None or isinstance(x, (int, float, str)):
44
+ return (x,)
45
+ else:
46
+ raise ValueError(f"Unknown type {type(x)}")
47
+
48
+
49
+ def as_list_of_2tuple(x):
50
+ x = as_tuple(x)
51
+ if len(x) == 1:
52
+ x = (x[0], x[0])
53
+ assert len(x) % 2 == 0, f"Expect even length, got {len(x)}."
54
+ lst = []
55
+ for i in range(0, len(x), 2):
56
+ lst.append((x[i], x[i + 1]))
57
+ return lst
58
+
59
+ def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor=None, attn_mask=None) -> Tensor:
60
+ if pe is None:
61
+ if attn_mask is not None and attn_mask.dtype != torch.bool:
62
+ attn_mask = attn_mask.to(q.dtype)
63
+ x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
64
+ x = rearrange(x, "B H L D -> B L (H D)")
65
+ else:
66
+ q, k = apply_rope(q, k, pe)
67
+ x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
68
+ x = rearrange(x, "B H L D -> B L (H D)")
69
+ return x
70
+
71
+
72
+ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
73
+ assert dim % 2 == 0
74
+ scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
75
+ omega = 1.0 / (theta**scale)
76
+ out = torch.einsum("...n,d->...nd", pos, omega)
77
+ out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
78
+ out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
79
+ return out.float()
80
+
81
+
82
+ def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:
83
+ xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
84
+ xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
85
+ xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
86
+ xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
87
+ return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
88
+
89
+
90
+ class EmbedND(nn.Module):
91
+ def __init__(self, dim: int, theta: int, axes_dim: list[int]):
92
+ super().__init__()
93
+ self.dim = dim
94
+ self.theta = theta
95
+ self.axes_dim = axes_dim
96
+
97
+ def forward(self, ids: Tensor) -> Tensor:
98
+ n_axes = ids.shape[-1]
99
+ emb = torch.cat(
100
+ [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)],
101
+ dim=-3,
102
+ )
103
+
104
+ return emb.unsqueeze(1)
105
+
106
+
107
+ def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
108
+ """
109
+ Create sinusoidal timestep embeddings.
110
+ :param t: a 1-D Tensor of N indices, one per batch element.
111
+ These may be fractional.
112
+ :param dim: the dimension of the output.
113
+ :param max_period: controls the minimum frequency of the embeddings.
114
+ :return: an (N, D) Tensor of positional embeddings.
115
+ """
116
+ t = time_factor * t
117
+ half = dim // 2
118
+ freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
119
+ t.device
120
+ )
121
+
122
+ args = t[:, None].float() * freqs[None]
123
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
124
+ if dim % 2:
125
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
126
+ if torch.is_floating_point(t):
127
+ embedding = embedding.to(t)
128
+ return embedding
129
+
130
+
131
+ class MLPEmbedder(nn.Module):
132
+ def __init__(self, in_dim: int, hidden_dim: int):
133
+ super().__init__()
134
+ self.in_layer = nn.Linear(in_dim, hidden_dim, bias=True)
135
+ self.silu = nn.SiLU()
136
+ self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=True)
137
+
138
+ def forward(self, x: Tensor) -> Tensor:
139
+ return self.out_layer(self.silu(self.in_layer(x)))
140
+
141
+
142
+ class RMSNorm(torch.nn.Module):
143
+ def __init__(self, dim: int, scale_factor=1.0, eps:float=1e-6):
144
+ super().__init__()
145
+ self.scale = nn.Parameter(torch.ones(dim) * scale_factor)
146
+ self.eps = eps
147
+
148
+ def forward(self, x: Tensor):
149
+ x_dtype = x.dtype
150
+ x = x.float()
151
+ rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + self.eps)
152
+ return (x * rrms).to(dtype=x_dtype) * self.scale
153
+
154
+
155
+ class QKNorm(torch.nn.Module):
156
+ def __init__(self, dim: int):
157
+ super().__init__()
158
+ self.query_norm = RMSNorm(dim)
159
+ self.key_norm = RMSNorm(dim)
160
+
161
+ def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
162
+ q = self.query_norm(q)
163
+ k = self.key_norm(k)
164
+ return q.to(v), k.to(v)
165
+
166
+
167
+ class SelfAttention(nn.Module):
168
+ def __init__(self, dim: int, num_heads: int = 8, qkv_bias: bool = False):
169
+ super().__init__()
170
+ self.num_heads = num_heads
171
+ head_dim = dim // num_heads
172
+
173
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
174
+ self.norm = QKNorm(head_dim)
175
+ self.proj = nn.Linear(dim, dim)
176
+
177
+ def forward(self, x: Tensor, pe: Tensor) -> Tensor:
178
+ qkv = self.qkv(x)
179
+ q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
180
+ q, k = self.norm(q, k, v)
181
+ x = attention(q, k, v, pe=pe)
182
+ x = self.proj(x)
183
+ return x
184
+
185
+
186
+ @dataclass
187
+ class ModulationOut:
188
+ shift: Tensor
189
+ scale: Tensor
190
+ gate: Tensor
191
+
192
+
193
+ class Modulation(nn.Module):
194
+ def __init__(self, dim: int, double: bool):
195
+ super().__init__()
196
+ self.is_double = double
197
+ self.multiplier = 6 if double else 3
198
+ self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
199
+
200
+ def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
201
+ out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
202
+
203
+ return (
204
+ ModulationOut(*out[:3]),
205
+ ModulationOut(*out[3:]) if self.is_double else None,
206
+ )
207
+
208
+ class TriModulation(nn.Module):
209
+ def __init__(self, dim: int):
210
+ super().__init__()
211
+ self.multiplier = 9
212
+ self.lin = nn.Linear(dim, self.multiplier * dim, bias=True)
213
+
214
+ def forward(self, vec: Tensor) -> tuple[ModulationOut, ModulationOut | None]:
215
+ out = self.lin(nn.functional.silu(vec))[:, None, :].chunk(self.multiplier, dim=-1)
216
+
217
+ return (
218
+ ModulationOut(*out[:3]),
219
+ ModulationOut(*out[3:6]),
220
+ ModulationOut(*out[6:]),
221
+ )
222
+
223
+
224
+ # from https://huggingface.co/stabilityai/stable-diffusion-3.5-medium
225
+ class DoubleStreamXBlockProcessor:
226
+ def __call__(self, attn, img, txt, vec, pe, **attention_kwargs):
227
+ img_mod1, img_mod2, img_mod3 = attn.img_mod(vec)
228
+ txt_mod1, txt_mod2 = attn.txt_mod(vec)
229
+
230
+ # prepare image for attention
231
+ img_modulated = attn.img_norm1(img)
232
+ img_cos_modulated = (1 + img_mod1.scale) * img_modulated + img_mod1.shift
233
+ img_qkv = attn.img_attn.qkv(img_cos_modulated)
234
+ 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)
235
+ img_q, img_k = attn.img_attn.norm(img_q, img_k, img_v)
236
+
237
+ # prepare image for self-attention
238
+ img_self_modulated = (1 + img_mod3.scale) * img_modulated + img_mod3.shift
239
+ img_self_qkv = attn.img_self_attn.qkv(img_self_modulated)
240
+ 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)
241
+ img_self_q, img_self_k = attn.img_self_attn.norm(img_self_q, img_self_k, img_self_v)
242
+ txt_pe, img_pe = torch.split(pe, [txt.shape[1], img.shape[1]], dim=2)
243
+ img_self_attn = attention(img_self_q, img_self_k, img_self_v, pe=img_pe)
244
+
245
+ # prepare txt for attention
246
+ txt_modulated = attn.txt_norm1(txt)
247
+ txt_modulated = (1 + txt_mod1.scale) * txt_modulated + txt_mod1.shift
248
+ txt_qkv = attn.txt_attn.qkv(txt_modulated)
249
+ 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)
250
+ txt_q, txt_k = attn.txt_attn.norm(txt_q, txt_k, txt_v)
251
+
252
+ # run actual attention
253
+ q = torch.cat((txt_q, img_q), dim=2)
254
+ k = torch.cat((txt_k, img_k), dim=2)
255
+ v = torch.cat((txt_v, img_v), dim=2)
256
+
257
+ attn1 = attention(q, k, v, pe=pe)
258
+ txt_attn, img_attn = attn1[:, : txt.shape[1]], attn1[:, txt.shape[1] :]
259
+
260
+ # calculate the img bloks
261
+ img = img + img_mod1.gate * attn.img_attn.proj(img_attn)
262
+ img = img + img_mod3.gate * attn.img_self_attn.proj(img_self_attn)
263
+ img = img + img_mod2.gate * attn.img_mlp((1 + img_mod2.scale) * attn.img_norm2(img) + img_mod2.shift)
264
+
265
+ # calculate the txt bloks
266
+ txt = txt + txt_mod1.gate * attn.txt_attn.proj(txt_attn)
267
+ txt = txt + txt_mod2.gate * attn.txt_mlp((1 + txt_mod2.scale) * attn.txt_norm2(txt) + txt_mod2.shift)
268
+ return img, txt
269
+
270
+ class DoubleStreamXBlock(nn.Module):
271
+ def __init__(self, hidden_size: int, num_heads: int, mlp_ratio: float, qkv_bias: bool = False):
272
+ super().__init__()
273
+
274
+ mlp_hidden_dim = int(hidden_size * mlp_ratio)
275
+ self.num_heads = num_heads
276
+ self.hidden_size = hidden_size
277
+ self.img_mod = TriModulation(hidden_size)
278
+ self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
279
+ self.img_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
280
+ self.img_self_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
281
+
282
+ self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
283
+ self.img_mlp = nn.Sequential(
284
+ nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
285
+ nn.GELU(approximate="tanh"),
286
+ nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
287
+ )
288
+
289
+ self.txt_mod = Modulation(hidden_size, double=True)
290
+ self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
291
+ self.txt_attn = SelfAttention(dim=hidden_size, num_heads=num_heads, qkv_bias=qkv_bias)
292
+
293
+ self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
294
+ self.txt_mlp = nn.Sequential(
295
+ nn.Linear(hidden_size, mlp_hidden_dim, bias=True),
296
+ nn.GELU(approximate="tanh"),
297
+ nn.Linear(mlp_hidden_dim, hidden_size, bias=True),
298
+ )
299
+ processor = DoubleStreamXBlockProcessor()
300
+ self.set_processor(processor)
301
+
302
+ def set_processor(self, processor) -> None:
303
+ self.processor = processor
304
+
305
+ def get_processor(self):
306
+ return self.processor
307
+
308
+ def forward(
309
+ self,
310
+ img: Tensor,
311
+ txt: Tensor,
312
+ vec: Tensor,
313
+ pe: Tensor,
314
+ image_proj: Tensor = None,
315
+ ip_scale: float =1.0,
316
+ ) -> tuple[Tensor, Tensor]:
317
+ if image_proj is None:
318
+ return self.processor(self, img, txt, vec, pe)
319
+ else:
320
+ return self.processor(self, img, txt, vec, pe, image_proj, ip_scale)
321
+
322
+ class SingleStreamBlockProcessor:
323
+ def __call__(self, attn: nn.Module, x: Tensor, vec: Tensor, pe: Tensor) -> Tensor:
324
+ mod, _ = attn.modulation(vec)
325
+ x_mod = (1 + mod.scale) * attn.pre_norm(x) + mod.shift
326
+ qkv, mlp = torch.split(attn.linear1(x_mod), [3 * attn.hidden_size, attn.mlp_hidden_dim], dim=-1)
327
+
328
+ q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=attn.num_heads)
329
+ q, k = attn.norm(q, k, v)
330
+
331
+ # compute attention
332
+ attn_1 = attention(q, k, v, pe=pe)
333
+
334
+ # compute activation in mlp stream, cat again and run second linear layer
335
+ output = attn.linear2(torch.cat((attn_1, attn.mlp_act(mlp)), 2))
336
+ output = x + mod.gate * output
337
+ return output
338
+
339
+
340
+ class SingleStreamBlock(nn.Module):
341
+ """
342
+ A DiT block with parallel linear layers as described in
343
+ https://arxiv.org/abs/2302.05442 and adapted modulation interface.
344
+ """
345
+
346
+ def __init__(
347
+ self,
348
+ hidden_size: int,
349
+ num_heads: int,
350
+ mlp_ratio: float = 4.0,
351
+ qk_scale: float | None = None,
352
+ ):
353
+ super().__init__()
354
+ self.hidden_dim = hidden_size
355
+ self.num_heads = num_heads
356
+ head_dim = hidden_size // num_heads
357
+ self.scale = qk_scale or head_dim**-0.5
358
+
359
+ self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
360
+ # qkv and mlp_in
361
+ self.linear1 = nn.Linear(hidden_size, hidden_size * 3 + self.mlp_hidden_dim)
362
+ # proj and mlp_out
363
+ self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size)
364
+
365
+ self.norm = QKNorm(head_dim)
366
+
367
+ self.hidden_size = hidden_size
368
+ self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
369
+
370
+ self.mlp_act = nn.GELU(approximate="tanh")
371
+ self.modulation = Modulation(hidden_size, double=False)
372
+
373
+ processor = SingleStreamBlockProcessor()
374
+ self.set_processor(processor)
375
+
376
+
377
+ def set_processor(self, processor) -> None:
378
+ self.processor = processor
379
+
380
+ def get_processor(self):
381
+ return self.processor
382
+
383
+ def forward(
384
+ self,
385
+ x: Tensor,
386
+ vec: Tensor,
387
+ pe: Tensor,
388
+ image_proj: Tensor | None = None,
389
+ ip_scale: float = 1.0
390
+ ) -> Tensor:
391
+ if image_proj is None:
392
+ return self.processor(self, x, vec, pe)
393
+ else:
394
+ return self.processor(self, x, vec, pe, image_proj, ip_scale)
395
+
396
+
397
+ class LastLayer(nn.Module):
398
+ def __init__(self, hidden_size: int, patch_size: int, out_channels: int):
399
+ super().__init__()
400
+ self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
401
+ self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
402
+ self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
403
+
404
+ def forward(self, x: Tensor, vec: Tensor) -> Tensor:
405
+ shift, scale = self.adaLN_modulation(vec).chunk(2, dim=1)
406
+ x = (1 + scale[:, None, :]) * self.norm_final(x) + shift[:, None, :]
407
+ x = self.linear(x)
408
+ return x
409
+
410
+
411
+
412
+ def get_norm_layer(norm_layer):
413
+ """
414
+ Get the normalization layer.
415
+
416
+ Args:
417
+ norm_layer (str): The type of normalization layer.
418
+
419
+ Returns:
420
+ norm_layer (nn.Module): The normalization layer.
421
+ """
422
+ if norm_layer == "layer":
423
+ return nn.LayerNorm
424
+ elif norm_layer == "rms":
425
+ return RMSNorm
426
+ else:
427
+ raise NotImplementedError(f"Norm layer {norm_layer} is not implemented")
428
+
429
+ def get_activation_layer(act_type):
430
+ """get activation layer
431
+
432
+ Args:
433
+ act_type (str): the activation type
434
+
435
+ Returns:
436
+ torch.nn.functional: the activation layer
437
+ """
438
+ if act_type == "gelu":
439
+ return lambda: nn.GELU()
440
+ elif act_type == "gelu_tanh":
441
+ # Approximate `tanh` requires torch >= 1.13
442
+ return lambda: nn.GELU(approximate="tanh")
443
+ elif act_type == "relu":
444
+ return nn.ReLU
445
+ elif act_type == "silu":
446
+ return nn.SiLU
447
+ else:
448
+ raise ValueError(f"Unknown activation type: {act_type}")
449
+
450
+ def modulate(x, shift=None, scale=None):
451
+ """modulate by shift and scale
452
+
453
+ Args:
454
+ x (torch.Tensor): input tensor.
455
+ shift (torch.Tensor, optional): shift tensor. Defaults to None.
456
+ scale (torch.Tensor, optional): scale tensor. Defaults to None.
457
+
458
+ Returns:
459
+ torch.Tensor: the output tensor after modulate.
460
+ """
461
+ if scale is None and shift is None:
462
+ return x
463
+ elif shift is None:
464
+ return x * (1 + scale.unsqueeze(1))
465
+ elif scale is None:
466
+ return x + shift.unsqueeze(1)
467
+ else:
468
+ return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
469
+
470
+ def apply_gate(x, gate=None, tanh=False):
471
+ """AI is creating summary for apply_gate
472
+
473
+ Args:
474
+ x (torch.Tensor): input tensor.
475
+ gate (torch.Tensor, optional): gate tensor. Defaults to None.
476
+ tanh (bool, optional): whether to use tanh function. Defaults to False.
477
+
478
+ Returns:
479
+ torch.Tensor: the output tensor after apply gate.
480
+ """
481
+ if gate is None:
482
+ return x
483
+ if tanh:
484
+ return x * gate.unsqueeze(1).tanh()
485
+ else:
486
+ return x * gate.unsqueeze(1)
487
+
488
+ class MLP(nn.Module):
489
+ """MLP as used in Vision Transformer, MLP-Mixer and related networks"""
490
+
491
+ def __init__(
492
+ self,
493
+ in_channels,
494
+ hidden_channels=None,
495
+ out_features=None,
496
+ act_layer=nn.GELU,
497
+ norm_layer=None,
498
+ bias=True,
499
+ drop=0.0,
500
+ use_conv=False,
501
+ device=None,
502
+ dtype=None,
503
+ ):
504
+ factory_kwargs = {"device": device, "dtype": dtype}
505
+ super().__init__()
506
+ out_features = out_features or in_channels
507
+ hidden_channels = hidden_channels or in_channels
508
+ bias = to_2tuple(bias)
509
+ drop_probs = to_2tuple(drop)
510
+ linear_layer = partial(nn.Conv2d, kernel_size=1) if use_conv else nn.Linear
511
+
512
+ self.fc1 = linear_layer(
513
+ in_channels, hidden_channels, bias=bias[0], **factory_kwargs
514
+ )
515
+ self.act = act_layer()
516
+ self.drop1 = nn.Dropout(drop_probs[0])
517
+ self.norm = (
518
+ norm_layer(hidden_channels, **factory_kwargs)
519
+ if norm_layer is not None
520
+ else nn.Identity()
521
+ )
522
+ self.fc2 = linear_layer(
523
+ hidden_channels, out_features, bias=bias[1], **factory_kwargs
524
+ )
525
+ self.drop2 = nn.Dropout(drop_probs[1])
526
+
527
+ def forward(self, x):
528
+ x = self.fc1(x)
529
+ x = self.act(x)
530
+ x = self.drop1(x)
531
+ x = self.norm(x)
532
+ x = self.fc2(x)
533
+ x = self.drop2(x)
534
+ return x
535
+
536
+
537
+
538
+
539
+
540
+
541
+
542
+
543
+
544
+
545
+
546
+
547
+
548
+
549
+
550
+
551
+
552
+
553
+ class TextProjection(nn.Module):
554
+ """
555
+ Projects text embeddings. Also handles dropout for classifier-free guidance.
556
+
557
+ Adapted from https://github.com/PixArt-alpha/PixArt-alpha/blob/master/diffusion/model/nets/PixArt_blocks.py
558
+ """
559
+
560
+ def __init__(self, in_channels, hidden_size, act_layer):
561
+ super().__init__()
562
+ self.linear_1 = nn.Linear(
563
+ in_features=in_channels,
564
+ out_features=hidden_size,
565
+ bias=True,
566
+ )
567
+ self.act_1 = act_layer()
568
+ self.linear_2 = nn.Linear(
569
+ in_features=hidden_size,
570
+ out_features=hidden_size,
571
+ bias=True,
572
+ )
573
+
574
+ def forward(self, caption):
575
+ hidden_states = self.linear_1(caption)
576
+ hidden_states = self.act_1(hidden_states)
577
+ hidden_states = self.linear_2(hidden_states)
578
+ return hidden_states
579
+
580
+
581
+ def timestep_embedding_refiner(t, dim, max_period=10000):
582
+ """
583
+ Create sinusoidal timestep embeddings.
584
+
585
+ Args:
586
+ t (torch.Tensor): a 1-D Tensor of N indices, one per batch element. These may be fractional.
587
+ dim (int): the dimension of the output.
588
+ max_period (int): controls the minimum frequency of the embeddings.
589
+
590
+ Returns:
591
+ embedding (torch.Tensor): An (N, D) Tensor of positional embeddings.
592
+
593
+ .. ref_link: https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
594
+ """
595
+ half = dim // 2
596
+ freqs = torch.exp(
597
+ -math.log(max_period)
598
+ * torch.arange(start=0, end=half, dtype=torch.float32)
599
+ / half
600
+ ).to(device=t.device)
601
+ args = t[:, None].float() * freqs[None]
602
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
603
+ if dim % 2:
604
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
605
+ return embedding
606
+
607
+
608
+ class TimestepEmbedder(nn.Module):
609
+ """
610
+ Embeds scalar timesteps into vector representations.
611
+ """
612
+
613
+ def __init__(
614
+ self,
615
+ hidden_size,
616
+ act_layer,
617
+ frequency_embedding_size=256,
618
+ max_period=10000,
619
+ out_size=None,
620
+ ):
621
+ super().__init__()
622
+ self.frequency_embedding_size = frequency_embedding_size
623
+ self.max_period = max_period
624
+ if out_size is None:
625
+ out_size = hidden_size
626
+
627
+ self.mlp = nn.Sequential(
628
+ nn.Linear(
629
+ frequency_embedding_size, hidden_size, bias=True,
630
+ ),
631
+ act_layer(),
632
+ nn.Linear(hidden_size, out_size, bias=True, ),
633
+ )
634
+ nn.init.normal_(self.mlp[0].weight, std=0.02)
635
+ nn.init.normal_(self.mlp[2].weight, std=0.02)
636
+
637
+ def forward(self, t):
638
+ t_freq = timestep_embedding_refiner(
639
+ t, self.frequency_embedding_size, self.max_period
640
+ ).type(self.mlp[0].weight.dtype)
641
+ t_emb = self.mlp(t_freq)
642
+ return t_emb
643
+
644
+
645
+ class IndividualTokenRefinerBlock(nn.Module):
646
+ def __init__(
647
+ self,
648
+ hidden_size,
649
+ heads_num,
650
+ mlp_width_ratio: str = 4.0,
651
+ mlp_drop_rate: float = 0.0,
652
+ act_type: str = "silu",
653
+ qk_norm: bool = False,
654
+ qk_norm_type: str = "layer",
655
+ qkv_bias: bool = True,
656
+ ):
657
+ super().__init__()
658
+ self.heads_num = heads_num
659
+ head_dim = hidden_size // heads_num
660
+ mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
661
+
662
+ self.norm1 = nn.LayerNorm(
663
+ hidden_size, elementwise_affine=True, eps=1e-6,
664
+ )
665
+ self.self_attn_qkv = nn.Linear(
666
+ hidden_size, hidden_size * 3, bias=qkv_bias,
667
+ )
668
+ qk_norm_layer = get_norm_layer(qk_norm_type)
669
+ self.self_attn_q_norm = (
670
+ qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
671
+ if qk_norm
672
+ else nn.Identity()
673
+ )
674
+ self.self_attn_k_norm = (
675
+ qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
676
+ if qk_norm
677
+ else nn.Identity()
678
+ )
679
+ self.self_attn_proj = nn.Linear(
680
+ hidden_size, hidden_size, bias=qkv_bias,
681
+ )
682
+
683
+ self.norm2 = nn.LayerNorm(
684
+ hidden_size, elementwise_affine=True, eps=1e-6,
685
+ )
686
+ act_layer = get_activation_layer(act_type)
687
+ self.mlp = MLP(
688
+ in_channels=hidden_size,
689
+ hidden_channels=mlp_hidden_dim,
690
+ act_layer=act_layer,
691
+ drop=mlp_drop_rate,
692
+ )
693
+
694
+ self.adaLN_modulation = nn.Sequential(
695
+ act_layer(),
696
+ nn.Linear(hidden_size, 2 * hidden_size, bias=True, ),
697
+ )
698
+ # Zero-initialize the modulation
699
+ nn.init.zeros_(self.adaLN_modulation[1].weight)
700
+ nn.init.zeros_(self.adaLN_modulation[1].bias)
701
+
702
+ def forward(
703
+ self,
704
+ x: torch.Tensor,
705
+ c: torch.Tensor, # timestep_aware_representations + context_aware_representations
706
+ attn_mask: torch.Tensor = None,
707
+ ):
708
+ gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1)
709
+
710
+ norm_x = self.norm1(x)
711
+ qkv = self.self_attn_qkv(norm_x)
712
+ q, k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=3, H=self.heads_num)
713
+ # Apply QK-Norm if needed
714
+ q = self.self_attn_q_norm(q).to(v)
715
+ k = self.self_attn_k_norm(k).to(v)
716
+
717
+ # Self-Attention
718
+ q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
719
+ attn = attention(q, k, v, attn_mask=attn_mask)
720
+ x = x + apply_gate(self.self_attn_proj(attn), gate_msa)
721
+
722
+ # FFN Layer
723
+ x = x + apply_gate(self.mlp(self.norm2(x)), gate_mlp)
724
+
725
+ return x
726
+
727
+
728
+ class CrossTokenRefinerBlock(nn.Module):
729
+ def __init__(
730
+ self,
731
+ hidden_size,
732
+ heads_num,
733
+ mlp_width_ratio: str = 4.0,
734
+ mlp_drop_rate: float = 0.0,
735
+ act_type: str = "silu",
736
+ qk_norm: bool = False,
737
+ qk_norm_type: str = "layer",
738
+ qkv_bias: bool = True,
739
+ ):
740
+ super().__init__()
741
+ self.heads_num = heads_num
742
+ head_dim = hidden_size // heads_num
743
+ mlp_hidden_dim = int(hidden_size * mlp_width_ratio)
744
+
745
+ self.norm1 = nn.LayerNorm(
746
+ hidden_size, elementwise_affine=True, eps=1e-6,
747
+ )
748
+ self.self_attn_q = nn.Linear(
749
+ hidden_size, hidden_size, bias=qkv_bias,
750
+ )
751
+ self.norm_y = nn.LayerNorm(
752
+ hidden_size, elementwise_affine=True, eps=1e-6,
753
+ )
754
+ self.self_attn_kv = nn.Linear(
755
+ hidden_size, hidden_size*2, bias=qkv_bias,
756
+ )
757
+ qk_norm_layer = get_norm_layer(qk_norm_type)
758
+ self.self_attn_q_norm = (
759
+ qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
760
+ if qk_norm
761
+ else nn.Identity()
762
+ )
763
+ self.self_attn_k_norm = (
764
+ qk_norm_layer(head_dim, elementwise_affine=True, eps=1e-6, )
765
+ if qk_norm
766
+ else nn.Identity()
767
+ )
768
+ self.self_attn_proj = nn.Linear(
769
+ hidden_size, hidden_size, bias=qkv_bias,
770
+ )
771
+
772
+ self.norm2 = nn.LayerNorm(
773
+ hidden_size, elementwise_affine=True, eps=1e-6,
774
+ )
775
+ act_layer = get_activation_layer(act_type)
776
+ self.mlp = MLP(
777
+ in_channels=hidden_size,
778
+ hidden_channels=mlp_hidden_dim,
779
+ act_layer=act_layer,
780
+ drop=mlp_drop_rate,
781
+ )
782
+
783
+ self.adaLN_modulation = nn.Sequential(
784
+ act_layer(),
785
+ nn.Linear(hidden_size, 2 * hidden_size, bias=True, ),
786
+ )
787
+ # Zero-initialize the modulation
788
+ nn.init.zeros_(self.adaLN_modulation[1].weight)
789
+ nn.init.zeros_(self.adaLN_modulation[1].bias)
790
+
791
+ def forward(
792
+ self,
793
+ x: torch.Tensor,
794
+ y: torch.Tensor,
795
+ c: torch.Tensor, # timestep_aware_representations + context_aware_representations
796
+ attn_mask: torch.Tensor = None,
797
+ ):
798
+ gate_msa, gate_mlp = self.adaLN_modulation(c).chunk(2, dim=1)
799
+
800
+ norm_x = self.norm1(x)
801
+ q = self.self_attn_q(norm_x)
802
+ q = rearrange(qkv, "B L (H D) -> B L H D", H=self.heads_num)
803
+ norm_y = self.norm_y(y)
804
+ kv = self.self_attn_kv(norm_y)
805
+ k, v = rearrange(qkv, "B L (K H D) -> K B L H D", K=2, H=self.heads_num)
806
+ # Apply QK-Norm if needed
807
+ q = self.self_attn_q_norm(q).to(v)
808
+ k = self.self_attn_k_norm(k).to(v)
809
+
810
+ # Self-Attention
811
+ attn = attention(q, k, v, attn_mask=attn_mask)
812
+ x = x + apply_gate(self.self_attn_proj(attn), gate_msa)
813
+
814
+ # FFN Layer
815
+ x = x + apply_gate(self.mlp(self.norm2(x)), gate_mlp)
816
+
817
+ return x
818
+
819
+ class IndividualTokenRefiner(nn.Module):
820
+ def __init__(
821
+ self,
822
+ hidden_size,
823
+ heads_num,
824
+ depth,
825
+ mlp_width_ratio: float = 4.0,
826
+ mlp_drop_rate: float = 0.0,
827
+ act_type: str = "silu",
828
+ qk_norm: bool = False,
829
+ qk_norm_type: str = "layer",
830
+ qkv_bias: bool = True,
831
+ ):
832
+ super().__init__()
833
+ self.blocks = nn.ModuleList(
834
+ [
835
+ IndividualTokenRefinerBlock(
836
+ hidden_size=hidden_size,
837
+ heads_num=heads_num,
838
+ mlp_width_ratio=mlp_width_ratio,
839
+ mlp_drop_rate=mlp_drop_rate,
840
+ act_type=act_type,
841
+ qk_norm=qk_norm,
842
+ qk_norm_type=qk_norm_type,
843
+ qkv_bias=qkv_bias,
844
+ )
845
+ for _ in range(depth)
846
+ ]
847
+ )
848
+
849
+ def forward(
850
+ self,
851
+ x: torch.Tensor,
852
+ c: torch.LongTensor,
853
+ mask: Optional[torch.Tensor] = None,
854
+ ):
855
+ self_attn_mask = None
856
+ if mask is not None:
857
+ batch_size = mask.shape[0]
858
+ seq_len = mask.shape[1]
859
+ mask = mask.to(x.device)
860
+ # batch_size x 1 x seq_len x seq_len
861
+ self_attn_mask_1 = mask.view(batch_size, 1, 1, seq_len).repeat(
862
+ 1, 1, seq_len, 1
863
+ )
864
+ # batch_size x 1 x seq_len x seq_len
865
+ self_attn_mask_2 = self_attn_mask_1.transpose(2, 3)
866
+ # batch_size x 1 x seq_len x seq_len, 1 for broadcasting of heads_num
867
+ self_attn_mask = (self_attn_mask_1 & self_attn_mask_2).bool()
868
+ # avoids self-attention weight being NaN for padding tokens
869
+ self_attn_mask[:, :, :, 0] = True
870
+
871
+ for block in self.blocks:
872
+ x = block(x, c, self_attn_mask)
873
+ return x
874
+
875
+
876
+ class SingleTokenRefiner(nn.Module):
877
+ """
878
+ A single token refiner block for llm text embedding refine.
879
+ """
880
+ def __init__(
881
+ self,
882
+ in_channels,
883
+ hidden_size,
884
+ heads_num,
885
+ depth,
886
+ mlp_width_ratio: float = 4.0,
887
+ mlp_drop_rate: float = 0.0,
888
+ act_type: str = "silu",
889
+ qk_norm: bool = False,
890
+ qk_norm_type: str = "layer",
891
+ qkv_bias: bool = True,
892
+ attn_mode: str = "torch",
893
+ enable_cls_token: bool = False,
894
+ enable_cross_attn: bool = False,
895
+ length: int = 29,
896
+ ):
897
+ super().__init__()
898
+ self.attn_mode = attn_mode
899
+ assert self.attn_mode == "torch", "Only support 'torch' mode for token refiner."
900
+ self.in_channels = in_channels
901
+ self.enable_cross_attn = enable_cross_attn
902
+ if self.enable_cross_attn:
903
+ self.length = length
904
+ self.input_embedder = nn.Linear(
905
+ in_channels//length, hidden_size, bias=True,
906
+ )
907
+ self.kv_embedder = nn.Linear(
908
+ in_channels//length*(length-1), hidden_size, bias=True,
909
+ )
910
+ self.fusion = CrossTokenRefinerBlock(
911
+ hidden_size=hidden_size,
912
+ heads_num=heads_num,
913
+ mlp_width_ratio=mlp_width_ratio,
914
+ mlp_drop_rate=mlp_drop_rate,
915
+ act_type=act_type,
916
+ qk_norm=qk_norm,
917
+ qk_norm_type=qk_norm_type,
918
+ qkv_bias=qkv_bias,
919
+ )
920
+ else:
921
+ self.input_embedder = nn.Linear(
922
+ in_channels, hidden_size, bias=True,
923
+ )
924
+
925
+ act_layer = get_activation_layer(act_type)
926
+ # Build timestep embedding layer
927
+ # self.t_embedder = TimestepEmbedder(hidden_size, act_layer,)
928
+ # Build context embedding layer
929
+ self.c_embedder = TextProjection(
930
+ in_channels, hidden_size, act_layer,
931
+ )
932
+
933
+ self.individual_token_refiner = IndividualTokenRefiner(
934
+ hidden_size=hidden_size,
935
+ heads_num=heads_num,
936
+ depth=depth,
937
+ mlp_width_ratio=mlp_width_ratio,
938
+ mlp_drop_rate=mlp_drop_rate,
939
+ act_type=act_type,
940
+ qk_norm=qk_norm,
941
+ qk_norm_type=qk_norm_type,
942
+ qkv_bias=qkv_bias,
943
+ )
944
+
945
+ self.enable_cls_token = enable_cls_token
946
+ if self.enable_cls_token:
947
+ self.cls_token = nn.Parameter(torch.zeros(1, 1, hidden_size))
948
+ nn.init.normal_(self.cls_token, std=1e-6)
949
+
950
+ def forward(
951
+ self,
952
+ x: torch.Tensor,
953
+ mask: Optional[torch.LongTensor] = None,
954
+ ):
955
+ if mask is None:
956
+ context_aware_representations = x.mean(dim=1)
957
+ else:
958
+ mask_float = mask.float().unsqueeze(-1) # [b, s1, 1]
959
+ context_aware_representations = (x * mask_float).sum(
960
+ dim=1
961
+ ) / mask_float.sum(dim=1)
962
+ c = self.c_embedder(context_aware_representations)
963
+ if self.enable_cross_attn:
964
+ single_channels = self.in_channels // self.length
965
+ x, y = torch.split(x, [single_channels, single_channels*(self.length-1)], dim=-1)
966
+ x = self.input_embedder(x)
967
+ y = self.kv_embedder(y)
968
+ else:
969
+ x = self.input_embedder(x)
970
+ if self.enable_cls_token:
971
+ B, L, C = x.shape
972
+ x = torch.cat([self.cls_token.expand(B, -1, -1), x], dim=1)
973
+
974
+ if self.enable_cross_attn:
975
+ x = self.fusion(x, y, c)
976
+ x = self.individual_token_refiner(x, c, mask)
977
+ if self.enable_cls_token:
978
+ x_global = x[:, 0]
979
+ x = x[:, 1:]
980
+ else:
981
+ x_global = x.mean(dim=1)
982
+ return dict(
983
+ txt_fea=x,
984
+ txt_fea_avg=x_global
985
+ )
986
+
987
+
988
+
989
+
990
+
991
+
992
+
993
+
994
+
995
+
996
+
997
+
998
+
999
+
1000
+
1001
+
1002
+ __all__ = ["YakModel"]
1003
+
1004
+ @dataclass
1005
+ class VisualGeneratorOutput(ModelOutput):
1006
+ loss: Optional[torch.FloatTensor] = None
1007
+
1008
+
1009
+ class YakTransformer(nn.Module):
1010
+ def __init__(self, config: YakConfig):
1011
+ super().__init__()
1012
+ self.config = config
1013
+ self.in_channels = config.in_channels
1014
+ self.out_channels = config.out_channels
1015
+ if config.hidden_size % config.num_heads != 0:
1016
+ raise ValueError(
1017
+ f"Hidden size {config.hidden_size} must be divisible by num_heads {config.num_heads}"
1018
+ )
1019
+ pe_dim = config.hidden_size // config.num_heads
1020
+ if sum(config.axes_dim) != pe_dim:
1021
+ raise ValueError(f"Got {config.axes_dim} but expected positional dim {pe_dim}")
1022
+ self.hidden_size = config.hidden_size
1023
+ self.num_heads = config.num_heads
1024
+ self.pe_embedder = EmbedND(dim=pe_dim, theta=config.theta, axes_dim=config.axes_dim)
1025
+ self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=True)
1026
+ self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size)
1027
+ self.vector_in = MLPEmbedder(config.vec_in_dim, self.hidden_size)
1028
+ self.guidance_in = (
1029
+ MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size) if config.guidance_embed else nn.Identity()
1030
+ )
1031
+ self.txt_type = config.txt_type
1032
+ self.txt_in = SingleTokenRefiner(
1033
+ config.context_in_dim,
1034
+ self.hidden_size,
1035
+ heads_num=config.num_heads * 2,
1036
+ depth=2,
1037
+ enable_cls_token=True
1038
+ )
1039
+
1040
+ self.double_blocks = nn.ModuleList(
1041
+ [
1042
+ DoubleStreamXBlock(
1043
+ self.hidden_size,
1044
+ self.num_heads,
1045
+ mlp_ratio=config.mlp_ratio,
1046
+ qkv_bias=config.qkv_bias,
1047
+ )
1048
+ for _ in range(config.depth)
1049
+ ]
1050
+ )
1051
+
1052
+ self.single_blocks = nn.ModuleList(
1053
+ [
1054
+ SingleStreamBlock(self.hidden_size, self.num_heads, mlp_ratio=config.mlp_ratio)
1055
+ for _ in range(config.depth_single_blocks)
1056
+ ]
1057
+ )
1058
+
1059
+ self.final_layer = LastLayer(self.hidden_size, 1, self.out_channels)
1060
+ self.gradient_checkpointing = False
1061
+
1062
+ def forward(
1063
+ self,
1064
+ img: Tensor,
1065
+ img_ids: Tensor,
1066
+ txt: Tensor,
1067
+ txt_ids: Tensor,
1068
+ timesteps: Tensor,
1069
+ guidance: Tensor | None = None,
1070
+ cond_img: Tensor = None,
1071
+ cond_img_ids: Tensor = None,
1072
+ ):
1073
+ if img.ndim != 3 or txt.ndim != 3:
1074
+ raise ValueError("Input img and txt tensors must have 3 dimensions.")
1075
+
1076
+ # running on sequences img
1077
+ img_tokens = img.shape[1]
1078
+ if cond_img is not None:
1079
+ img = torch.cat([img, cond_img], dim=1)
1080
+ img_ids = torch.cat([img_ids, cond_img_ids], dim=1)
1081
+ img = self.img_in(img)
1082
+
1083
+ vec = self.time_in(timestep_embedding(timesteps, 256))
1084
+ if self.config.guidance_embed:
1085
+ if guidance is None:
1086
+ raise ValueError("Didn't get guidance strength for guidance distilled model.")
1087
+ vec = vec + self.guidance_in(timestep_embedding(guidance, 256))
1088
+ txt_dict = self.txt_in(txt)
1089
+ txt = txt_dict["txt_fea"]
1090
+ y = txt_dict["txt_fea_avg"]
1091
+ vec = vec + self.vector_in(y)
1092
+
1093
+ ids = torch.cat((txt_ids, img_ids), dim=1)
1094
+ pe = self.pe_embedder(ids)
1095
+
1096
+ for block in self.double_blocks:
1097
+ if self.training and self.gradient_checkpointing:
1098
+ img, txt = self._gradient_checkpointing_func(
1099
+ block.__call__,
1100
+ img,
1101
+ txt,
1102
+ vec,
1103
+ pe,
1104
+ )
1105
+ else:
1106
+ img, txt = block(img=img, txt=txt, vec=vec, pe=pe)
1107
+
1108
+ img = torch.cat((txt, img), 1)
1109
+ for block in self.single_blocks:
1110
+ if self.training and self.gradient_checkpointing:
1111
+ img = self._gradient_checkpointing_func(
1112
+ block.__call__,
1113
+ img,
1114
+ vec,
1115
+ pe,
1116
+ )
1117
+ else:
1118
+ img = block(img, vec=vec, pe=pe)
1119
+ img = img[:, txt.shape[1] :, ...]
1120
+
1121
+ img = self.final_layer(img, vec) # (N, T, patch_size ** 2 * out_channels)
1122
+ if cond_img is not None:
1123
+ img = torch.split(img, img_tokens, dim=1)[0]
1124
+ return img
1125
+
1126
+ def time_shift(mu: float, sigma: float, t: Tensor):
1127
+ return math.exp(mu) / (math.exp(mu) + (1 / t - 1) ** sigma)
1128
+
1129
+
1130
+ def get_lin_function(
1131
+ x1: float = 256, y1: float = 0.5, x2: float = 4096, y2: float = 1.15
1132
+ ) -> Callable[[float], float]:
1133
+ m = (y2 - y1) / (x2 - x1)
1134
+ b = y1 - m * x1
1135
+ return lambda x: m * x + b
1136
+
1137
+ def get_noise(
1138
+ num_samples: int,
1139
+ channel: int,
1140
+ height: int,
1141
+ width: int,
1142
+ device: torch.device,
1143
+ dtype: torch.dtype,
1144
+ seed: int,
1145
+ ):
1146
+ return torch.randn(
1147
+ num_samples,
1148
+ channel,
1149
+ # allow for packing
1150
+ 2 * math.ceil(height / 16),
1151
+ 2 * math.ceil(width / 16),
1152
+ device=device,
1153
+ dtype=dtype,
1154
+ generator=torch.Generator(device=device).manual_seed(seed),
1155
+ )
1156
+
1157
+ def unpack(x: Tensor, height: int, width: int) -> Tensor:
1158
+ return rearrange(
1159
+ x,
1160
+ "b (h w) (c ph pw) -> b c (h ph) (w pw)",
1161
+ h=math.ceil(height / 16),
1162
+ w=math.ceil(width / 16),
1163
+ ph=2,
1164
+ pw=2,
1165
+ )
1166
+
1167
+ class YakPretrainedModel(PreTrainedModel):
1168
+ config_class = YakConfig
1169
+ base_model_prefix = "yak"
1170
+ supports_gradient_checkpointing = True
1171
+ main_input_name = "pixel_values"
1172
+ _supports_sdpa = True
1173
+
1174
+
1175
+ class YakModel(YakPretrainedModel):
1176
+ def __init__(self, config: YakConfig):
1177
+ super().__init__(config)
1178
+ self.vae = AutoencoderKL.from_config(config.vae_config)
1179
+ self.backbone = YakTransformer(config)
1180
+
1181
+ def get_refiner(self):
1182
+ return self.backbone.txt_in
1183
+
1184
+ def get_cls_refiner(self):
1185
+ return self.backbone.vector_in
1186
+
1187
+ def get_backbone(self):
1188
+ return self.backbone
1189
+
1190
+ def get_vae(self):
1191
+ return self.vae
1192
+
1193
+ def preprocess_image(self, image: Image.Image, size, convert_to_rgb=True, Norm=True, output_type="tensor"):
1194
+ image = exif_transpose(image)
1195
+ if not image.mode == "RGB" and convert_to_rgb:
1196
+ image = image.convert("RGB")
1197
+
1198
+ image = torchvision.transforms.functional.resize(
1199
+ image, size, interpolation=transforms.InterpolationMode.BICUBIC
1200
+ )
1201
+
1202
+ arr = np.array(image)
1203
+ h = arr.shape[0]
1204
+ w = arr.shape[1]
1205
+ crop_y = (h - size) // 2
1206
+ crop_x = (w - size) // 2
1207
+ pil_image = image.crop([crop_x, crop_y, crop_x+size, crop_y+size])
1208
+ if output_type == "pil_image":
1209
+ return pil_image
1210
+
1211
+ image_np = arr[crop_y : crop_y + size, crop_x : crop_x + size]
1212
+ hidden_h = h // 16
1213
+ hidden_w = w // 16
1214
+ hidden_size = size // 16
1215
+ img_ids = torch.zeros(hidden_h, hidden_w, 3)
1216
+
1217
+ img_ids[..., 1] = img_ids[..., 1] + torch.arange(hidden_h)[:, None]
1218
+ img_ids[..., 2] = img_ids[..., 2] + torch.arange(hidden_w)[None, :]
1219
+ crop_y = (hidden_h - hidden_size) // 2
1220
+ crop_x = (hidden_w - hidden_size) // 2
1221
+ img_ids = img_ids[crop_y : crop_y + hidden_size, crop_x : crop_x + hidden_size]
1222
+ img_ids = rearrange(img_ids, "h w c -> (h w) c")
1223
+
1224
+ image_tensor = torchvision.transforms.functional.to_tensor(image_np)
1225
+ if Norm:
1226
+ image_tensor = torchvision.transforms.functional.normalize(image_tensor,
1227
+ mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
1228
+ return pil_image, image_tensor, img_ids
1229
+
1230
+ def process_image_aspectratio(self, image, size):
1231
+ w, h = image.size
1232
+ t_w, t_h = size
1233
+ resize_r = max(float(t_w)/w, float(t_h)/h)
1234
+ resize_size = (int(resize_r * h), int(resize_r * w))
1235
+ image = torchvision.transforms.functional.resize(
1236
+ image, resize_size, interpolation=transforms.InterpolationMode.BICUBIC
1237
+ )
1238
+ pil_image = torchvision.transforms.functional.center_crop(
1239
+ image, (t_h, t_w)
1240
+ )
1241
+ hidden_h = t_h // 16
1242
+ hidden_w = t_w // 16
1243
+ img_ids = torch.zeros(hidden_h, hidden_w, 3)
1244
+
1245
+ img_ids[..., 1] = img_ids[..., 1] + torch.arange(hidden_h)[:, None]
1246
+ img_ids[..., 2] = img_ids[..., 2] + torch.arange(hidden_w)[None, :]
1247
+ img_ids = rearrange(img_ids, "h w c -> (h w) c")
1248
+ image_tensor = torchvision.transforms.functional.to_tensor(pil_image)
1249
+ image_tensor = torchvision.transforms.functional.normalize(image_tensor,
1250
+ mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
1251
+ return pil_image, image_tensor, img_ids
1252
+
1253
+ def compute_vae_encodings(self, pixel_values, with_ids=True, time=0):
1254
+ pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
1255
+ pixel_values = pixel_values.to(self.vae.device, dtype=self.vae.dtype)
1256
+ with torch.no_grad():
1257
+ model_input = self.vae.encode(pixel_values).latent_dist.sample()
1258
+ if hasattr(self.vae.config, 'shift_factor') and self.vae.config.shift_factor is not None:
1259
+ model_input = model_input - self.vae.config.shift_factor
1260
+ if hasattr(self.vae.config, 'scaling_factor') and self.vae.config.scaling_factor is not None:
1261
+ model_input = model_input * self.vae.config.scaling_factor
1262
+ # patch for transformer
1263
+ bs, c, h, w = model_input.shape
1264
+ model_input = rearrange(model_input, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
1265
+ if with_ids:
1266
+ img_ids = torch.zeros(h // 2, w // 2, 3)
1267
+ img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
1268
+ img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
1269
+ img_ids[..., 0] = time
1270
+ img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs)
1271
+ return model_input, img_ids
1272
+ else:
1273
+ return model_input
1274
+
1275
+ def generate_image(
1276
+ self,
1277
+ cond,
1278
+ height,
1279
+ width,
1280
+ num_steps,
1281
+ seed,
1282
+ no_both_cond=None,
1283
+ no_txt_cond=None,
1284
+ img_cfg=1.0,
1285
+ txt_cfg=1.0,
1286
+ output_type="pil"
1287
+ ):
1288
+ txt = cond["txt"]
1289
+ bs = len(txt)
1290
+ channel = self.vae.config.latent_channels
1291
+ height = 16 * (height // 16)
1292
+ width = 16 * (width // 16)
1293
+ torch_device = next(self.backbone.parameters()).device
1294
+ x = get_noise(
1295
+ bs,
1296
+ channel,
1297
+ height,
1298
+ width,
1299
+ device=torch_device,
1300
+ dtype=torch.bfloat16,
1301
+ seed=seed,
1302
+ )
1303
+ # prepare inputs
1304
+ img = x
1305
+ bs, c, h, w = img.shape
1306
+
1307
+ img = rearrange(img, "b c (h ph) (w pw) -> b (h w) (c ph pw)", ph=2, pw=2)
1308
+ if img.shape[0] == 1 and bs > 1:
1309
+ img = repeat(img, "1 ... -> bs ...", bs=bs)
1310
+
1311
+ img_ids = torch.zeros(h // 2, w // 2, 3)
1312
+ img_ids[..., 1] = img_ids[..., 1] + torch.arange(h // 2)[:, None]
1313
+ img_ids[..., 2] = img_ids[..., 2] + torch.arange(w // 2)[None, :]
1314
+ img_ids = repeat(img_ids, "h w c -> b (h w) c", b=bs).to(img.device)
1315
+
1316
+ if "vae_pixel_values" in cond:
1317
+ img_vae_cond, cond_ids = self.compute_vae_encodings(
1318
+ pixel_values=cond["vae_pixel_values"], with_ids=True, time=1.0)
1319
+ cond_ids = cond_ids.to(img.device)
1320
+
1321
+ if txt.shape[0] == 1 and bs > 1:
1322
+ txt = repeat(txt, "1 ... -> bs ...", bs=bs)
1323
+ txt_ids = torch.zeros(bs, txt.shape[1], 3).to(img.device)
1324
+
1325
+ timesteps = self.get_schedule(
1326
+ num_steps, img.shape[1], shift=self.config.timestep_shift,
1327
+ base_shift=self.config.base_shift, max_shift=self.config.max_shift)
1328
+ no_both_txt = no_both_cond["txt"]
1329
+ if no_txt_cond is not None:
1330
+ no_txt_txt = no_txt_cond["txt"]
1331
+ x = self.edit_denoise(img, img_ids,
1332
+ txt, txt_ids,
1333
+ no_txt_txt,
1334
+ no_both_txt,
1335
+ img_vae_cond, cond_ids.to(img.device),
1336
+ timesteps=timesteps,
1337
+ img_cfg=img_cfg, txt_cfg=txt_cfg)
1338
+ else:
1339
+ x = self.denoise(img, img_ids, txt, txt_ids,
1340
+ timesteps=timesteps, cfg=txt_cfg,
1341
+ neg_txt=no_both_txt)
1342
+ x = unpack(x.float(), height, width)
1343
+
1344
+ with torch.autocast(device_type=torch_device.type, dtype=torch.float32):
1345
+ if hasattr(self.vae.config, 'scaling_factor') and self.vae.config.scaling_factor is not None:
1346
+ x = x / self.vae.config.scaling_factor
1347
+ if hasattr(self.vae.config, 'shift_factor') and self.vae.config.shift_factor is not None:
1348
+ x = x + self.vae.config.shift_factor
1349
+ x = self.vae.decode(x, return_dict=False)[0]
1350
+ # bring into PIL format and save
1351
+ x = x.clamp(-1, 1)
1352
+ x = rearrange(x, "b c h w -> b h w c")
1353
+ x = (127.5 * (x + 1.0)).cpu().byte().numpy()
1354
+ if output_type == "np":
1355
+ return x
1356
+ images = []
1357
+ for i in range(bs):
1358
+ img = Image.fromarray(x[i])
1359
+ images.append(img)
1360
+ return images
1361
+
1362
+
1363
+ def get_schedule(self,
1364
+ num_steps: int,
1365
+ image_seq_len: int,
1366
+ base_shift: float = 0.5,
1367
+ max_shift: float = 1.15,
1368
+ shift: bool = True,
1369
+ ) -> list[float]:
1370
+ # extra step for zero
1371
+ timesteps = torch.linspace(1, 0, num_steps + 1)
1372
+ # shifting the schedule to favor high timesteps for higher signal images
1373
+ if shift:
1374
+ # eastimate mu based on linear estimation between two points
1375
+ mu = get_lin_function(y1=base_shift, y2=max_shift)(image_seq_len)
1376
+ timesteps = time_shift(mu, 1.0, timesteps)
1377
+
1378
+ return timesteps.tolist()
1379
+
1380
+ def denoise(self,
1381
+ input_img: Tensor,
1382
+ img_ids: Tensor,
1383
+ txt: Tensor,
1384
+ txt_ids: Tensor,
1385
+ # sampling parameters
1386
+ timesteps: list[float],
1387
+ cfg: float = 1.0,
1388
+ neg_txt = None):
1389
+ bs = input_img.shape[0]
1390
+ for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:]):
1391
+ t_vec = torch.full((bs,), t_curr, dtype=input_img.dtype, device=input_img.device)
1392
+ txt_ids = torch.zeros(bs, txt.shape[1], 3).to(txt.device)
1393
+ cond_eps = self.backbone(
1394
+ img=input_img,
1395
+ img_ids=img_ids,
1396
+ txt=txt,
1397
+ txt_ids=txt_ids,
1398
+ timesteps=t_vec,
1399
+ )
1400
+ txt_ids = torch.zeros(bs, neg_txt.shape[1], 3).to(neg_txt.device)
1401
+ uncond_eps = self.backbone(
1402
+ img=input_img,
1403
+ img_ids=img_ids,
1404
+ txt=neg_txt,
1405
+ txt_ids=txt_ids,
1406
+ timesteps=t_vec,
1407
+ )
1408
+ pred = uncond_eps + cfg * (cond_eps - uncond_eps)
1409
+ input_img = input_img + (t_prev - t_curr) * pred
1410
+ return input_img
1411
+
1412
+ def edit_denoise(self,
1413
+ input_img: Tensor,
1414
+ img_ids: Tensor,
1415
+ txt: Tensor,
1416
+ txt_ids: Tensor,
1417
+ no_txt_txt: Tensor,
1418
+ no_both_txt: Tensor,
1419
+ img_cond,
1420
+ cond_img_ids,
1421
+ # sampling parameters
1422
+ timesteps: list[float],
1423
+ img_cfg: float = 1.0,
1424
+ txt_cfg: float = 1.0,):
1425
+ bs = input_img.shape[0]
1426
+ for t_curr, t_prev in zip(timesteps[:-1], timesteps[1:]):
1427
+ t_vec = torch.full((bs * 1,), t_curr, dtype=input_img.dtype, device=input_img.device)
1428
+ txt_ids = torch.zeros(bs, txt.shape[1], 3).to(txt.device)
1429
+ cond_eps = self.backbone(
1430
+ img=input_img,
1431
+ img_ids=img_ids,
1432
+ txt=txt,
1433
+ txt_ids=txt_ids,
1434
+ timesteps=t_vec,
1435
+ cond_img=img_cond,
1436
+ cond_img_ids=cond_img_ids,
1437
+ )
1438
+ txt_ids = torch.zeros(bs, no_both_txt.shape[1], 3).to(no_both_txt.device)
1439
+ no_both_eps = self.backbone(
1440
+ img=input_img,
1441
+ img_ids=img_ids,
1442
+ txt=no_both_txt,
1443
+ txt_ids=txt_ids,
1444
+ timesteps=t_vec,
1445
+ )
1446
+ txt_ids = torch.zeros(bs, no_txt_txt.shape[1], 3).to(no_txt_txt.device)
1447
+ no_txt_eps = self.backbone(
1448
+ img=input_img,
1449
+ img_ids=img_ids,
1450
+ txt=no_txt_txt,
1451
+ txt_ids=txt_ids,
1452
+ timesteps=t_vec,
1453
+ cond_img=img_cond,
1454
+ cond_img_ids=cond_img_ids,
1455
+ )
1456
+ pred = no_both_eps
1457
+ pred += img_cfg * (no_txt_eps - no_both_eps)
1458
+ pred += txt_cfg * (cond_eps - no_txt_eps)
1459
+ input_img = input_img + (t_prev - t_curr) * pred
1460
+ return input_img
1461
+
preprocessor_config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "crop_size": {
3
+ "height": -1,
4
+ "width": -1
5
+ },
6
+ "do_center_crop": false,
7
+ "do_convert_rgb": true,
8
+ "do_normalize": true,
9
+ "do_rescale": true,
10
+ "do_resize": true,
11
+ "hidden_stride": 2,
12
+ "image_mean": [
13
+ 0.48145466,
14
+ 0.4578275,
15
+ 0.40821073
16
+ ],
17
+ "image_processor_type": "CLIPImageProcessor",
18
+ "image_std": [
19
+ 0.26862954,
20
+ 0.26130258,
21
+ 0.27577711
22
+ ],
23
+ "max_pixels": 2408448,
24
+ "min_pixels": 200704,
25
+ "patch_size": 14,
26
+ "resample": 3,
27
+ "rescale_factor": 0.00392156862745098,
28
+ "size": {
29
+ "shortest_edge": -1
30
+ },
31
+ "temporal_patch_size": 1
32
+ }
special_tokens_map.json ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "additional_special_tokens": [
3
+ "<|im_start|>",
4
+ "<|im_end|>",
5
+ "<|object_ref_start|>",
6
+ "<|object_ref_end|>",
7
+ "<|box_start|>",
8
+ "<|box_end|>",
9
+ "<|quad_start|>",
10
+ "<|quad_end|>",
11
+ "<|vision_start|>",
12
+ "<|vision_end|>",
13
+ "<|vision_pad|>",
14
+ "<|image_pad|>",
15
+ "<|video_pad|>"
16
+ ],
17
+ "eos_token": {
18
+ "content": "<|im_end|>",
19
+ "lstrip": false,
20
+ "normalized": false,
21
+ "rstrip": false,
22
+ "single_word": false
23
+ },
24
+ "pad_token": {
25
+ "content": "<|endoftext|>",
26
+ "lstrip": false,
27
+ "normalized": false,
28
+ "rstrip": false,
29
+ "single_word": false
30
+ }
31
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:aeb13307a71acd8fe81861d94ad54ab689df773318809eed3cbe794b4492dae4
3
+ size 11422654
tokenizer_config.json ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_bos_token": false,
3
+ "add_prefix_space": false,
4
+ "added_tokens_decoder": {
5
+ "151643": {
6
+ "content": "<|endoftext|>",
7
+ "lstrip": false,
8
+ "normalized": false,
9
+ "rstrip": false,
10
+ "single_word": false,
11
+ "special": true
12
+ },
13
+ "151644": {
14
+ "content": "<|im_start|>",
15
+ "lstrip": false,
16
+ "normalized": false,
17
+ "rstrip": false,
18
+ "single_word": false,
19
+ "special": true
20
+ },
21
+ "151645": {
22
+ "content": "<|im_end|>",
23
+ "lstrip": false,
24
+ "normalized": false,
25
+ "rstrip": false,
26
+ "single_word": false,
27
+ "special": true
28
+ },
29
+ "151646": {
30
+ "content": "<|object_ref_start|>",
31
+ "lstrip": false,
32
+ "normalized": false,
33
+ "rstrip": false,
34
+ "single_word": false,
35
+ "special": true
36
+ },
37
+ "151647": {
38
+ "content": "<|object_ref_end|>",
39
+ "lstrip": false,
40
+ "normalized": false,
41
+ "rstrip": false,
42
+ "single_word": false,
43
+ "special": true
44
+ },
45
+ "151648": {
46
+ "content": "<|box_start|>",
47
+ "lstrip": false,
48
+ "normalized": false,
49
+ "rstrip": false,
50
+ "single_word": false,
51
+ "special": true
52
+ },
53
+ "151649": {
54
+ "content": "<|box_end|>",
55
+ "lstrip": false,
56
+ "normalized": false,
57
+ "rstrip": false,
58
+ "single_word": false,
59
+ "special": true
60
+ },
61
+ "151650": {
62
+ "content": "<|quad_start|>",
63
+ "lstrip": false,
64
+ "normalized": false,
65
+ "rstrip": false,
66
+ "single_word": false,
67
+ "special": true
68
+ },
69
+ "151651": {
70
+ "content": "<|quad_end|>",
71
+ "lstrip": false,
72
+ "normalized": false,
73
+ "rstrip": false,
74
+ "single_word": false,
75
+ "special": true
76
+ },
77
+ "151652": {
78
+ "content": "<|vision_start|>",
79
+ "lstrip": false,
80
+ "normalized": false,
81
+ "rstrip": false,
82
+ "single_word": false,
83
+ "special": true
84
+ },
85
+ "151653": {
86
+ "content": "<|vision_end|>",
87
+ "lstrip": false,
88
+ "normalized": false,
89
+ "rstrip": false,
90
+ "single_word": false,
91
+ "special": true
92
+ },
93
+ "151654": {
94
+ "content": "<|vision_pad|>",
95
+ "lstrip": false,
96
+ "normalized": false,
97
+ "rstrip": false,
98
+ "single_word": false,
99
+ "special": true
100
+ },
101
+ "151655": {
102
+ "content": "<|image_pad|>",
103
+ "lstrip": false,
104
+ "normalized": false,
105
+ "rstrip": false,
106
+ "single_word": false,
107
+ "special": true
108
+ },
109
+ "151656": {
110
+ "content": "<|video_pad|>",
111
+ "lstrip": false,
112
+ "normalized": false,
113
+ "rstrip": false,
114
+ "single_word": false,
115
+ "special": true
116
+ },
117
+ "151657": {
118
+ "content": "<tool_call>",
119
+ "lstrip": false,
120
+ "normalized": false,
121
+ "rstrip": false,
122
+ "single_word": false,
123
+ "special": false
124
+ },
125
+ "151658": {
126
+ "content": "</tool_call>",
127
+ "lstrip": false,
128
+ "normalized": false,
129
+ "rstrip": false,
130
+ "single_word": false,
131
+ "special": false
132
+ },
133
+ "151659": {
134
+ "content": "<|fim_prefix|>",
135
+ "lstrip": false,
136
+ "normalized": false,
137
+ "rstrip": false,
138
+ "single_word": false,
139
+ "special": false
140
+ },
141
+ "151660": {
142
+ "content": "<|fim_middle|>",
143
+ "lstrip": false,
144
+ "normalized": false,
145
+ "rstrip": false,
146
+ "single_word": false,
147
+ "special": false
148
+ },
149
+ "151661": {
150
+ "content": "<|fim_suffix|>",
151
+ "lstrip": false,
152
+ "normalized": false,
153
+ "rstrip": false,
154
+ "single_word": false,
155
+ "special": false
156
+ },
157
+ "151662": {
158
+ "content": "<|fim_pad|>",
159
+ "lstrip": false,
160
+ "normalized": false,
161
+ "rstrip": false,
162
+ "single_word": false,
163
+ "special": false
164
+ },
165
+ "151663": {
166
+ "content": "<|repo_name|>",
167
+ "lstrip": false,
168
+ "normalized": false,
169
+ "rstrip": false,
170
+ "single_word": false,
171
+ "special": false
172
+ },
173
+ "151664": {
174
+ "content": "<|file_sep|>",
175
+ "lstrip": false,
176
+ "normalized": false,
177
+ "rstrip": false,
178
+ "single_word": false,
179
+ "special": false
180
+ },
181
+ "151665": {
182
+ "content": "<tool_response>",
183
+ "lstrip": false,
184
+ "normalized": false,
185
+ "rstrip": false,
186
+ "single_word": false,
187
+ "special": false
188
+ },
189
+ "151666": {
190
+ "content": "</tool_response>",
191
+ "lstrip": false,
192
+ "normalized": false,
193
+ "rstrip": false,
194
+ "single_word": false,
195
+ "special": false
196
+ },
197
+ "151667": {
198
+ "content": "<think>",
199
+ "lstrip": false,
200
+ "normalized": false,
201
+ "rstrip": false,
202
+ "single_word": false,
203
+ "special": false
204
+ },
205
+ "151668": {
206
+ "content": "</think>",
207
+ "lstrip": false,
208
+ "normalized": false,
209
+ "rstrip": false,
210
+ "single_word": false,
211
+ "special": false
212
+ }
213
+ },
214
+ "additional_special_tokens": [
215
+ "<|im_start|>",
216
+ "<|im_end|>",
217
+ "<|object_ref_start|>",
218
+ "<|object_ref_end|>",
219
+ "<|box_start|>",
220
+ "<|box_end|>",
221
+ "<|quad_start|>",
222
+ "<|quad_end|>",
223
+ "<|vision_start|>",
224
+ "<|vision_end|>",
225
+ "<|vision_pad|>",
226
+ "<|image_pad|>",
227
+ "<|video_pad|>"
228
+ ],
229
+ "bos_token": null,
230
+ "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 %}",
231
+ "clean_up_tokenization_spaces": false,
232
+ "eos_token": "<|im_end|>",
233
+ "errors": "replace",
234
+ "extra_special_tokens": {},
235
+ "model_max_length": 131072,
236
+ "pad_token": "<|endoftext|>",
237
+ "split_special_tokens": false,
238
+ "tokenizer_class": "Qwen2Tokenizer",
239
+ "unk_token": null
240
+ }
vocab.json ADDED
The diff for this file is too large to render. See raw diff