update Imp-v1.5-3B-phi2
Browse files- LICENSE +201 -0
- README.md +97 -3
- added_tokens.json +42 -0
- config.json +63 -0
- configuration_imp.py +183 -0
- generation_config.json +6 -0
- images/bird.jpg +0 -0
- images/bus.jpg +0 -0
- images/car.jpg +0 -0
- merges.txt +0 -0
- model-00001-of-00007.safetensors +3 -0
- model-00002-of-00007.safetensors +3 -0
- model-00003-of-00007.safetensors +3 -0
- model-00004-of-00007.safetensors +3 -0
- model-00005-of-00007.safetensors +3 -0
- model-00006-of-00007.safetensors +3 -0
- model-00007-of-00007.safetensors +3 -0
- model.safetensors.index.json +755 -0
- modeling_imp.py +1262 -0
- special_tokens_map.json +5 -0
- tokenizer_config.json +363 -0
- vision_encoder.py +613 -0
- vocab.json +0 -0
LICENSE
ADDED
|
@@ -0,0 +1,201 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Apache License
|
| 2 |
+
Version 2.0, January 2004
|
| 3 |
+
http://www.apache.org/licenses/
|
| 4 |
+
|
| 5 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 6 |
+
|
| 7 |
+
1. Definitions.
|
| 8 |
+
|
| 9 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 10 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 11 |
+
|
| 12 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 13 |
+
the copyright owner that is granting the License.
|
| 14 |
+
|
| 15 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 16 |
+
other entities that control, are controlled by, or are under common
|
| 17 |
+
control with that entity. For the purposes of this definition,
|
| 18 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 19 |
+
direction or management of such entity, whether by contract or
|
| 20 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 21 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 22 |
+
|
| 23 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 24 |
+
exercising permissions granted by this License.
|
| 25 |
+
|
| 26 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 27 |
+
including but not limited to software source code, documentation
|
| 28 |
+
source, and configuration files.
|
| 29 |
+
|
| 30 |
+
"Object" form shall mean any form resulting from mechanical
|
| 31 |
+
transformation or translation of a Source form, including but
|
| 32 |
+
not limited to compiled object code, generated documentation,
|
| 33 |
+
and conversions to other media types.
|
| 34 |
+
|
| 35 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 36 |
+
Object form, made available under the License, as indicated by a
|
| 37 |
+
copyright notice that is included in or attached to the work
|
| 38 |
+
(an example is provided in the Appendix below).
|
| 39 |
+
|
| 40 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 41 |
+
form, that is based on (or derived from) the Work and for which the
|
| 42 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 43 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 44 |
+
of this License, Derivative Works shall not include works that remain
|
| 45 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 46 |
+
the Work and Derivative Works thereof.
|
| 47 |
+
|
| 48 |
+
"Contribution" shall mean any work of authorship, including
|
| 49 |
+
the original version of the Work and any modifications or additions
|
| 50 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 51 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 52 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 53 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 54 |
+
means any form of electronic, verbal, or written communication sent
|
| 55 |
+
to the Licensor or its representatives, including but not limited to
|
| 56 |
+
communication on electronic mailing lists, source code control systems,
|
| 57 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 58 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 59 |
+
excluding communication that is conspicuously marked or otherwise
|
| 60 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 61 |
+
|
| 62 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 63 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 64 |
+
subsequently incorporated within the Work.
|
| 65 |
+
|
| 66 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 67 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 68 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 69 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 70 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 71 |
+
Work and such Derivative Works in Source or Object form.
|
| 72 |
+
|
| 73 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 74 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 75 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 76 |
+
(except as stated in this section) patent license to make, have made,
|
| 77 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 78 |
+
where such license applies only to those patent claims licensable
|
| 79 |
+
by such Contributor that are necessarily infringed by their
|
| 80 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 81 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 82 |
+
institute patent litigation against any entity (including a
|
| 83 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 84 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 85 |
+
or contributory patent infringement, then any patent licenses
|
| 86 |
+
granted to You under this License for that Work shall terminate
|
| 87 |
+
as of the date such litigation is filed.
|
| 88 |
+
|
| 89 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 90 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 91 |
+
modifications, and in Source or Object form, provided that You
|
| 92 |
+
meet the following conditions:
|
| 93 |
+
|
| 94 |
+
(a) You must give any other recipients of the Work or
|
| 95 |
+
Derivative Works a copy of this License; and
|
| 96 |
+
|
| 97 |
+
(b) You must cause any modified files to carry prominent notices
|
| 98 |
+
stating that You changed the files; and
|
| 99 |
+
|
| 100 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 101 |
+
that You distribute, all copyright, patent, trademark, and
|
| 102 |
+
attribution notices from the Source form of the Work,
|
| 103 |
+
excluding those notices that do not pertain to any part of
|
| 104 |
+
the Derivative Works; and
|
| 105 |
+
|
| 106 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 107 |
+
distribution, then any Derivative Works that You distribute must
|
| 108 |
+
include a readable copy of the attribution notices contained
|
| 109 |
+
within such NOTICE file, excluding those notices that do not
|
| 110 |
+
pertain to any part of the Derivative Works, in at least one
|
| 111 |
+
of the following places: within a NOTICE text file distributed
|
| 112 |
+
as part of the Derivative Works; within the Source form or
|
| 113 |
+
documentation, if provided along with the Derivative Works; or,
|
| 114 |
+
within a display generated by the Derivative Works, if and
|
| 115 |
+
wherever such third-party notices normally appear. The contents
|
| 116 |
+
of the NOTICE file are for informational purposes only and
|
| 117 |
+
do not modify the License. You may add Your own attribution
|
| 118 |
+
notices within Derivative Works that You distribute, alongside
|
| 119 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 120 |
+
that such additional attribution notices cannot be construed
|
| 121 |
+
as modifying the License.
|
| 122 |
+
|
| 123 |
+
You may add Your own copyright statement to Your modifications and
|
| 124 |
+
may provide additional or different license terms and conditions
|
| 125 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 126 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 127 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 128 |
+
the conditions stated in this License.
|
| 129 |
+
|
| 130 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 131 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 132 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 133 |
+
this License, without any additional terms or conditions.
|
| 134 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 135 |
+
the terms of any separate license agreement you may have executed
|
| 136 |
+
with Licensor regarding such Contributions.
|
| 137 |
+
|
| 138 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 139 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 140 |
+
except as required for reasonable and customary use in describing the
|
| 141 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 142 |
+
|
| 143 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 144 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 145 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 146 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 147 |
+
implied, including, without limitation, any warranties or conditions
|
| 148 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 149 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 150 |
+
appropriateness of using or redistributing the Work and assume any
|
| 151 |
+
risks associated with Your exercise of permissions under this License.
|
| 152 |
+
|
| 153 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 154 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 155 |
+
unless required by applicable law (such as deliberate and grossly
|
| 156 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 157 |
+
liable to You for damages, including any direct, indirect, special,
|
| 158 |
+
incidental, or consequential damages of any character arising as a
|
| 159 |
+
result of this License or out of the use or inability to use the
|
| 160 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 161 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 162 |
+
other commercial damages or losses), even if such Contributor
|
| 163 |
+
has been advised of the possibility of such damages.
|
| 164 |
+
|
| 165 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 166 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 167 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 168 |
+
or other liability obligations and/or rights consistent with this
|
| 169 |
+
License. However, in accepting such obligations, You may act only
|
| 170 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 171 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 172 |
+
defend, and hold each Contributor harmless for any liability
|
| 173 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 174 |
+
of your accepting any such warranty or additional liability.
|
| 175 |
+
|
| 176 |
+
END OF TERMS AND CONDITIONS
|
| 177 |
+
|
| 178 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 179 |
+
|
| 180 |
+
To apply the Apache License to your work, attach the following
|
| 181 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 182 |
+
replaced with your own identifying information. (Don't include
|
| 183 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 184 |
+
comment syntax for the file format. We also recommend that a
|
| 185 |
+
file or class name and description of purpose be included on the
|
| 186 |
+
same "printed page" as the copyright notice for easier
|
| 187 |
+
identification within third-party archives.
|
| 188 |
+
|
| 189 |
+
Copyright [yyyy] [name of copyright owner]
|
| 190 |
+
|
| 191 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 192 |
+
you may not use this file except in compliance with the License.
|
| 193 |
+
You may obtain a copy of the License at
|
| 194 |
+
|
| 195 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 196 |
+
|
| 197 |
+
Unless required by applicable law or agreed to in writing, software
|
| 198 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 199 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 200 |
+
See the License for the specific language governing permissions and
|
| 201 |
+
limitations under the License.
|
README.md
CHANGED
|
@@ -1,3 +1,97 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
pipeline_tag: text-generation
|
| 4 |
+
datasets:
|
| 5 |
+
- liuhaotian/LLaVA-Pretrain
|
| 6 |
+
- liuhaotian/LLaVA-Instruct-150K
|
| 7 |
+
---
|
| 8 |
+
# 😈 Imp
|
| 9 |
+
|
| 10 |
+
> A very small man can cast a very large shadow.
|
| 11 |
+
>
|
| 12 |
+
> ——*George R.R. Martin, A Clash of Kings*
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
\[Technical report (coming soon)\] [[Demo](https://xmbot.net/imp/)\] [[Github](https://github.com/MILVLG/imp)\]
|
| 16 |
+
|
| 17 |
+
## Introduction
|
| 18 |
+
|
| 19 |
+
The Imp project aims to provide a family of a strong multimodal `small` language models (MSLMs). Our `Imp-v1.5-3B-Phi2` is a strong MSLM with only **3B** parameters, which is build upon a small yet powerful SLM [Phi-2 ](https://huggingface.co/microsoft/phi-2)(2.7B) and a powerful visual encoder [SigLIP ](https://huggingface.co/google/siglip-so400m-patch14-384)(0.4B), and trained on the [LLaVA-v1.5](https://github.com/haotian-liu/LLaVA) training set.
|
| 20 |
+
|
| 21 |
+
As shown in the Table below, `Imp-v1.5-3B-Phi2` significantly outperforms the counterparts of similar model sizes, and even achieves slightly better performance than the strong LLaVA-7B model on various multimodal benchmarks.
|
| 22 |
+
|
| 23 |
+
We release our model weights and provide an example below to run our model . Detailed technical report and corresponding training/evaluation code will be released soon on our [GitHub repo](https://github.com/MILVLG/imp). We will persistently improve our model and release the next versions to further improve model performance :)
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
## How to use
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
**Install dependencies**
|
| 30 |
+
```bash
|
| 31 |
+
pip install transformers # latest version is ok, but we recommend v4.31.0
|
| 32 |
+
pip install -q pillow accelerate einops
|
| 33 |
+
```
|
| 34 |
+
|
| 35 |
+
You can use the following code for model inference. The format of text instruction is similar to [LLaVA](https://github.com/haotian-liu/LLaVA). A Colab page to run this example is provided [here](https://colab.research.google.com/drive/1EBYky6xIPjnlPppo2gZaiNK6gEsjXgom?usp=drive_link#scrollTo=2-VpU6QzWCVZ). Note that the example can only be run on GPUs currently.
|
| 36 |
+
|
| 37 |
+
```Python
|
| 38 |
+
import torch
|
| 39 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 40 |
+
from PIL import Image
|
| 41 |
+
|
| 42 |
+
torch.set_default_device("cuda")
|
| 43 |
+
|
| 44 |
+
#Create model
|
| 45 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 46 |
+
"MILVLG/Imp-v1.5-3B-Phi2",
|
| 47 |
+
torch_dtype=torch.float16,
|
| 48 |
+
device_map="auto",
|
| 49 |
+
trust_remote_code=True)
|
| 50 |
+
tokenizer = AutoTokenizer.from_pretrained("MILVLG/Imp-v1.5-3B-Phi2", trust_remote_code=True)
|
| 51 |
+
|
| 52 |
+
#Set inputs
|
| 53 |
+
text = "A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\nWhat are the colors of the bus in the image? ASSISTANT:"
|
| 54 |
+
image = Image.open("images/bus.jpg")
|
| 55 |
+
|
| 56 |
+
input_ids = tokenizer(text, return_tensors='pt').input_ids
|
| 57 |
+
image_tensor = model.image_preprocess(image)
|
| 58 |
+
|
| 59 |
+
#Generate the answer
|
| 60 |
+
output_ids = model.generate(
|
| 61 |
+
input_ids,
|
| 62 |
+
max_new_tokens=100,
|
| 63 |
+
images=image_tensor,
|
| 64 |
+
use_cache=True)[0]
|
| 65 |
+
print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())
|
| 66 |
+
```
|
| 67 |
+
|
| 68 |
+
## Model evaluation
|
| 69 |
+
We conduct evaluation on 9 commonly-used benchmarks, including 5 academic VQA benchmarks and 4 popular MLLM benchmarks, to compare our Imp model with LLaVA (7B) and existing MSLMs of similar model sizes.
|
| 70 |
+
|
| 71 |
+
| Models | Size | VQAv2 | GQA | SQA(IMG) | TextVQA | POPE | MME(P) | MMB |MMBCN |MM-Vet|
|
| 72 |
+
|:--------:|:-----:|:----:|:-------------:|:--------:|:-----:|:----:|:-------:|:-------:|:-------:|:-------:|
|
| 73 |
+
| [LLaVA-v1.5-lora](https://huggingface.co/liuhaotian/llava-v1.5-7b) | 7B |79.10 | 63.00| 68.40 |58.20| 86.40 | 1476.9 | 66.10 |- |30.2|
|
| 74 |
+
| [TinyGPT-V](https://huggingface.co/Tyrannosaurus/TinyGPT-V) | 3B | - | 33.60 | - | - | -| - | - |- |-|
|
| 75 |
+
| [LLaVA-Phi](https://github.com/zhuyiche/llava-phi) | 3B | 71.40 | - | 68.40 | 48.60 | 85.00 | 1335.1 | 59.80 |-|28.9|
|
| 76 |
+
| [MobileVLM](https://huggingface.co/mtgv/MobileVLM-3B) | 3B | - | 59.00 | 61.00 | 47.50 | 84.90 | 1288.9 | 59.60 |- |-|
|
| 77 |
+
| [MC-LLaVA-3b](https://huggingface.co/visheratin/MC-LLaVA-3b) | 3B | 64.24 | 49.60 | - | 38.59 | 80.59 | - | - |- |-|
|
| 78 |
+
| **Imp-v1.5-3B-Phi2** | 3B | **81.18** | **63.54** | **72.78**| **59.84** | **88.87**| **1446.4** | **72.94**| 46.65 |**43.3**|
|
| 79 |
+
|
| 80 |
+
## License
|
| 81 |
+
This project is licensed under the Apache License 2.0 - see the [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) file for details.
|
| 82 |
+
|
| 83 |
+
## About us
|
| 84 |
+
This project is maintained by the [MILVLG](https://github.com/MILVLG)@Hangzhou Dianzi University (HDU) led by Prof. Zhou Yu and Jun Yu, and is mainly developed by Zhenwei Shao and Xuecheng Ouyang. We hope our model may serve as a strong baseline to inspire future research on MSLM, as well as its derivative applications on mobile devices and robots.
|
| 85 |
+
|
| 86 |
+
## Citation
|
| 87 |
+
|
| 88 |
+
If you use our model or refer our work in your studies, please cite:
|
| 89 |
+
|
| 90 |
+
```bibtex
|
| 91 |
+
@misc{imp2024,
|
| 92 |
+
author = {Shao, Zhenwei and Ouyang, Xuecheng and Yu, Zhou and Yu, Jun},
|
| 93 |
+
title = {Imp: An Emprical Study of Multimodal Small Language Models},
|
| 94 |
+
year = {2024},
|
| 95 |
+
url = {https://huggingface.co/MILVLG/imp-v1-3b}
|
| 96 |
+
}
|
| 97 |
+
```
|
added_tokens.json
ADDED
|
@@ -0,0 +1,42 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"\t\t": 50294,
|
| 3 |
+
"\t\t\t": 50293,
|
| 4 |
+
"\t\t\t\t": 50292,
|
| 5 |
+
"\t\t\t\t\t": 50291,
|
| 6 |
+
"\t\t\t\t\t\t": 50290,
|
| 7 |
+
"\t\t\t\t\t\t\t": 50289,
|
| 8 |
+
"\t\t\t\t\t\t\t\t": 50288,
|
| 9 |
+
"\t\t\t\t\t\t\t\t\t": 50287,
|
| 10 |
+
" ": 50286,
|
| 11 |
+
" ": 50285,
|
| 12 |
+
" ": 50284,
|
| 13 |
+
" ": 50283,
|
| 14 |
+
" ": 50282,
|
| 15 |
+
" ": 50281,
|
| 16 |
+
" ": 50280,
|
| 17 |
+
" ": 50279,
|
| 18 |
+
" ": 50278,
|
| 19 |
+
" ": 50277,
|
| 20 |
+
" ": 50276,
|
| 21 |
+
" ": 50275,
|
| 22 |
+
" ": 50274,
|
| 23 |
+
" ": 50273,
|
| 24 |
+
" ": 50272,
|
| 25 |
+
" ": 50271,
|
| 26 |
+
" ": 50270,
|
| 27 |
+
" ": 50269,
|
| 28 |
+
" ": 50268,
|
| 29 |
+
" ": 50267,
|
| 30 |
+
" ": 50266,
|
| 31 |
+
" ": 50265,
|
| 32 |
+
" ": 50264,
|
| 33 |
+
" ": 50263,
|
| 34 |
+
" ": 50262,
|
| 35 |
+
" ": 50261,
|
| 36 |
+
" ": 50260,
|
| 37 |
+
" ": 50259,
|
| 38 |
+
" ": 50258,
|
| 39 |
+
" ": 50257,
|
| 40 |
+
"</s>": 50295,
|
| 41 |
+
"<image>": 50296
|
| 42 |
+
}
|
config.json
ADDED
|
@@ -0,0 +1,63 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "MILVLG/Imp-v1.5-3B-Phi2",
|
| 3 |
+
"activation_function": "gelu_new",
|
| 4 |
+
"architectures": [
|
| 5 |
+
"ImpForCausalLM"
|
| 6 |
+
],
|
| 7 |
+
"attn_pdrop": 0.0,
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration_imp.ImpConfig",
|
| 10 |
+
"AutoModelForCausalLM": "modeling_imp.ImpForCausalLM"
|
| 11 |
+
},
|
| 12 |
+
"embd_pdrop": 0.0,
|
| 13 |
+
"eos_token_id": 50295,
|
| 14 |
+
"flash_attn": false,
|
| 15 |
+
"flash_rotary": false,
|
| 16 |
+
"freeze_mm_mlp_adapter": false,
|
| 17 |
+
"fused_dense": false,
|
| 18 |
+
"image_aspect_ratio": "square",
|
| 19 |
+
"image_token": "<image>",
|
| 20 |
+
"image_token_index": 50296,
|
| 21 |
+
"img_processor": null,
|
| 22 |
+
"initializer_range": 0.02,
|
| 23 |
+
"layer_norm_epsilon": 1e-05,
|
| 24 |
+
"mm_hidden_size": 1152,
|
| 25 |
+
"mm_projector_lr": 2e-05,
|
| 26 |
+
"mm_projector_type": "mlp2x_gelu",
|
| 27 |
+
"mm_use_im_patch_token": false,
|
| 28 |
+
"mm_use_im_start_end": false,
|
| 29 |
+
"mm_vision_select_feature": "patch",
|
| 30 |
+
"mm_vision_select_layer": -2,
|
| 31 |
+
"mm_vision_tower": "google/siglip-so400m-patch14-384",
|
| 32 |
+
"model_type": "imp",
|
| 33 |
+
"n_embd": 2560,
|
| 34 |
+
"n_head": 32,
|
| 35 |
+
"n_head_kv": null,
|
| 36 |
+
"n_inner": null,
|
| 37 |
+
"n_layer": 32,
|
| 38 |
+
"n_positions": 3072,
|
| 39 |
+
"pad_token_id": 50256,
|
| 40 |
+
"resid_pdrop": 0.1,
|
| 41 |
+
"rotary_dim": 32,
|
| 42 |
+
"tie_word_embeddings": false,
|
| 43 |
+
"tokenizer_model_max_length": 3072,
|
| 44 |
+
"tokenizer_padding_side": "right",
|
| 45 |
+
"torch_dtype": "float16",
|
| 46 |
+
"transformers_version": "4.31.0",
|
| 47 |
+
"use_cache": true,
|
| 48 |
+
"use_mm_proj": true,
|
| 49 |
+
"vision_tower_config": {
|
| 50 |
+
"attention_dropout": 0.0,
|
| 51 |
+
"hidden_act": "gelu_pytorch_tanh",
|
| 52 |
+
"hidden_size": 1152,
|
| 53 |
+
"image_size": 384,
|
| 54 |
+
"intermediate_size": 4304,
|
| 55 |
+
"layer_norm_eps": 1e-06,
|
| 56 |
+
"model_type": "siglip_vision_model",
|
| 57 |
+
"num_attention_heads": 16,
|
| 58 |
+
"num_channels": 3,
|
| 59 |
+
"num_hidden_layers": 27,
|
| 60 |
+
"patch_size": 14
|
| 61 |
+
},
|
| 62 |
+
"vocab_size": 51200
|
| 63 |
+
}
|
configuration_imp.py
ADDED
|
@@ -0,0 +1,183 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) MILVLG team.
|
| 2 |
+
# Licensed under the Apache 2.0 license.
|
| 3 |
+
#
|
| 4 |
+
# Some code here is copied from the project Phi-2 (https://huggingface.co/microsoft/phi-2),
|
| 5 |
+
# SigLIP@transformers==4.37.0.dev0 (https://huggingface.co/google/siglip-so400m-patch14-384),
|
| 6 |
+
# and Llava (https://github.com/haotian-liu/LLaVA), and modified by
|
| 7 |
+
# Zhenwei Shao (shaozw@hdu.edu.cn) @ MILVLG. We thank them for their great works.
|
| 8 |
+
#
|
| 9 |
+
# We keep their original copyright statements as follows, which should be inherited:
|
| 10 |
+
# ------------------------------- Phi-2 ---------------------------------------------
|
| 11 |
+
# Copyright (c) Microsoft Corporation.
|
| 12 |
+
# Licensed under the MIT license.
|
| 13 |
+
# https://huggingface.co/google/siglip-so400m-patch14-384
|
| 14 |
+
#
|
| 15 |
+
# Copyright (c) 2022, Tri Dao, trid@cs.stanford.edu.
|
| 16 |
+
# Licensed under the BSD 3-Clause License.
|
| 17 |
+
# ------------------------------- SigLIP --------------------------------------------
|
| 18 |
+
# Copyright 2024 Google AI and The HuggingFace Team. All rights reserved.
|
| 19 |
+
#
|
| 20 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 21 |
+
# you may not use this file except in compliance with the License.
|
| 22 |
+
# You may obtain a copy of the License at
|
| 23 |
+
#
|
| 24 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 25 |
+
#
|
| 26 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 27 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 28 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 29 |
+
# See the License for the specific language governing permissions and
|
| 30 |
+
# limitations under the License.
|
| 31 |
+
# ------------------------------- Llava ---------------------------------------------
|
| 32 |
+
# Copyright 2023 Haotian Liu
|
| 33 |
+
#
|
| 34 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
| 35 |
+
# you may not use this file except in compliance with the License.
|
| 36 |
+
# You may obtain a copy of the License at
|
| 37 |
+
#
|
| 38 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
| 39 |
+
#
|
| 40 |
+
# Unless required by applicable law or agreed to in writing, software
|
| 41 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
| 42 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 43 |
+
# See the License for the specific language governing permissions and
|
| 44 |
+
# limitations under the License.
|
| 45 |
+
# -----------------------------------------------------------------------------------
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
import os
|
| 49 |
+
import math
|
| 50 |
+
from typing import Optional, Union
|
| 51 |
+
|
| 52 |
+
from transformers import PretrainedConfig
|
| 53 |
+
from transformers.utils import logging
|
| 54 |
+
|
| 55 |
+
logger = logging.get_logger(__name__)
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
class PhiConfig(PretrainedConfig):
|
| 59 |
+
"""Phi configuration."""
|
| 60 |
+
|
| 61 |
+
model_type = "phi-msft"
|
| 62 |
+
attribute_map = {
|
| 63 |
+
"max_position_embeddings": "n_positions",
|
| 64 |
+
"hidden_size": "n_embd",
|
| 65 |
+
"num_attention_heads": "n_head",
|
| 66 |
+
"num_hidden_layers": "n_layer",
|
| 67 |
+
}
|
| 68 |
+
|
| 69 |
+
def __init__(
|
| 70 |
+
self,
|
| 71 |
+
vocab_size: int = 50304,
|
| 72 |
+
n_positions: int = 2048,
|
| 73 |
+
n_embd: int = 1024,
|
| 74 |
+
n_layer: int = 20,
|
| 75 |
+
n_inner: Optional[int] = None,
|
| 76 |
+
n_head: int = 16,
|
| 77 |
+
n_head_kv: Optional[int] = None,
|
| 78 |
+
rotary_dim: Optional[int] = 32,
|
| 79 |
+
activation_function: Optional[str] = "gelu_new",
|
| 80 |
+
flash_attn: bool = False,
|
| 81 |
+
flash_rotary: bool = False,
|
| 82 |
+
fused_dense: bool = False,
|
| 83 |
+
attn_pdrop: float = 0.0,
|
| 84 |
+
embd_pdrop: float = 0.0,
|
| 85 |
+
resid_pdrop: float = 0.0,
|
| 86 |
+
layer_norm_epsilon: float = 1e-5,
|
| 87 |
+
initializer_range: float = 0.02,
|
| 88 |
+
tie_word_embeddings: bool = False,
|
| 89 |
+
pad_vocab_size_multiple: int = 64,
|
| 90 |
+
**kwargs
|
| 91 |
+
) -> None:
|
| 92 |
+
self.vocab_size = int(math.ceil(vocab_size / pad_vocab_size_multiple) * pad_vocab_size_multiple)
|
| 93 |
+
self.n_positions = n_positions
|
| 94 |
+
self.n_embd = n_embd
|
| 95 |
+
self.n_layer = n_layer
|
| 96 |
+
self.n_inner = n_inner
|
| 97 |
+
self.n_head = n_head
|
| 98 |
+
self.n_head_kv = n_head_kv
|
| 99 |
+
self.rotary_dim = min(rotary_dim, n_embd // n_head)
|
| 100 |
+
self.activation_function = activation_function
|
| 101 |
+
self.flash_attn = flash_attn
|
| 102 |
+
self.flash_rotary = flash_rotary
|
| 103 |
+
self.fused_dense = fused_dense
|
| 104 |
+
self.attn_pdrop = attn_pdrop
|
| 105 |
+
self.embd_pdrop = embd_pdrop
|
| 106 |
+
self.resid_pdrop = resid_pdrop
|
| 107 |
+
self.layer_norm_epsilon = layer_norm_epsilon
|
| 108 |
+
self.initializer_range = initializer_range
|
| 109 |
+
|
| 110 |
+
super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
|
| 111 |
+
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
class SiglipVisionConfig(PretrainedConfig):
|
| 115 |
+
|
| 116 |
+
model_type = "siglip_vision_model"
|
| 117 |
+
|
| 118 |
+
def __init__(
|
| 119 |
+
self,
|
| 120 |
+
hidden_size=768,
|
| 121 |
+
intermediate_size=3072,
|
| 122 |
+
num_hidden_layers=12,
|
| 123 |
+
num_attention_heads=12,
|
| 124 |
+
num_channels=3,
|
| 125 |
+
image_size=224,
|
| 126 |
+
patch_size=16,
|
| 127 |
+
hidden_act="gelu_pytorch_tanh",
|
| 128 |
+
layer_norm_eps=1e-6,
|
| 129 |
+
attention_dropout=0.0,
|
| 130 |
+
**kwargs,
|
| 131 |
+
):
|
| 132 |
+
super().__init__(**kwargs)
|
| 133 |
+
|
| 134 |
+
self.hidden_size = hidden_size
|
| 135 |
+
self.intermediate_size = intermediate_size
|
| 136 |
+
self.num_hidden_layers = num_hidden_layers
|
| 137 |
+
self.num_attention_heads = num_attention_heads
|
| 138 |
+
self.num_channels = num_channels
|
| 139 |
+
self.patch_size = patch_size
|
| 140 |
+
self.image_size = image_size
|
| 141 |
+
self.attention_dropout = attention_dropout
|
| 142 |
+
self.layer_norm_eps = layer_norm_eps
|
| 143 |
+
self.hidden_act = hidden_act
|
| 144 |
+
|
| 145 |
+
@classmethod
|
| 146 |
+
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
| 147 |
+
cls._set_token_in_kwargs(kwargs)
|
| 148 |
+
|
| 149 |
+
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
| 150 |
+
|
| 151 |
+
# get the vision config dict if we are loading from SiglipConfig
|
| 152 |
+
if config_dict.get("model_type") == "siglip":
|
| 153 |
+
config_dict = config_dict["vision_config"]
|
| 154 |
+
|
| 155 |
+
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
| 156 |
+
logger.warning(
|
| 157 |
+
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
| 158 |
+
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
| 159 |
+
)
|
| 160 |
+
|
| 161 |
+
return cls.from_dict(config_dict, **kwargs)
|
| 162 |
+
|
| 163 |
+
|
| 164 |
+
class ImpConfig(PhiConfig):
|
| 165 |
+
model_type = "imp"
|
| 166 |
+
|
| 167 |
+
def __init__(self, **kwargs):
|
| 168 |
+
super().__init__(**kwargs)
|
| 169 |
+
self.image_token_index = getattr(self, "image_token_index", 50296)
|
| 170 |
+
self.image_token = getattr(self, "image_token", "<image>")
|
| 171 |
+
|
| 172 |
+
if not hasattr(self, "vision_tower_config") and hasattr(self, "mm_vision_tower"):
|
| 173 |
+
vision_tower_config = SiglipVisionConfig.from_pretrained(self.mm_vision_tower)
|
| 174 |
+
self.vision_tower_config = vision_tower_config.to_diff_dict()
|
| 175 |
+
|
| 176 |
+
@property
|
| 177 |
+
def vision_tower_cfg(self):
|
| 178 |
+
cfg = SiglipVisionConfig.from_dict(self.vision_tower_config)
|
| 179 |
+
# imp-v1 only supports `patch` feature for now w/o cls token
|
| 180 |
+
# cfg.mm_vision_select_feature = self.mm_vision_select_feature
|
| 181 |
+
cfg.mm_vision_select_layer = self.mm_vision_select_layer
|
| 182 |
+
cfg.mm_vision_tower = self.mm_vision_tower
|
| 183 |
+
return cfg
|
generation_config.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"eos_token_id":50295,
|
| 3 |
+
"pad_token_id":50256,
|
| 4 |
+
"_from_model_config": true,
|
| 5 |
+
"transformers_version": "4.31.0"
|
| 6 |
+
}
|
images/bird.jpg
ADDED
|
images/bus.jpg
ADDED
|
images/car.jpg
ADDED
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:29f0e0859601ec012c410dc6471b8965c5d3d16875f84fa095782487cb83113f
|
| 3 |
+
size 996428776
|
model-00002-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:beda1bcee40de85570fd28e37eef166276cbe7fbf0835579f8f847e12494968a
|
| 3 |
+
size 996507088
|
model-00003-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:20f3b15038f8d7191a99835d3e21aa63927a9361ad35e6c778a63b3136830b82
|
| 3 |
+
size 996512312
|
model-00004-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:28422cb2f997225cf1a3a1771cb37e3b49e41c0312b989f98096fd9b4a31338a
|
| 3 |
+
size 996512088
|
model-00005-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c672e017ca169da27170ab02267b9a9deedb2e3b240fc6cefdc734aa6e262a13
|
| 3 |
+
size 996507152
|
model-00006-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:fa13bd6005ccb00d73a495f1c0d06009dbf0891ccaf659cca28af3a59a4e1722
|
| 3 |
+
size 1021447256
|
model-00007-of-00007.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7bc6ecad02a265ac62054118b43c5485b769f398963da4c9f73d2ece152fb027
|
| 3 |
+
size 370061024
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,755 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 6373874240
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"lm_head.linear.bias": "model-00007-of-00007.safetensors",
|
| 7 |
+
"lm_head.linear.weight": "model-00007-of-00007.safetensors",
|
| 8 |
+
"lm_head.ln.bias": "model-00007-of-00007.safetensors",
|
| 9 |
+
"lm_head.ln.weight": "model-00007-of-00007.safetensors",
|
| 10 |
+
"transformer.embd.wte.weight": "model-00001-of-00007.safetensors",
|
| 11 |
+
"transformer.h.0.ln.bias": "model-00001-of-00007.safetensors",
|
| 12 |
+
"transformer.h.0.ln.weight": "model-00001-of-00007.safetensors",
|
| 13 |
+
"transformer.h.0.mixer.Wqkv.bias": "model-00001-of-00007.safetensors",
|
| 14 |
+
"transformer.h.0.mixer.Wqkv.weight": "model-00001-of-00007.safetensors",
|
| 15 |
+
"transformer.h.0.mixer.out_proj.bias": "model-00001-of-00007.safetensors",
|
| 16 |
+
"transformer.h.0.mixer.out_proj.weight": "model-00001-of-00007.safetensors",
|
| 17 |
+
"transformer.h.0.mlp.fc1.bias": "model-00001-of-00007.safetensors",
|
| 18 |
+
"transformer.h.0.mlp.fc1.weight": "model-00001-of-00007.safetensors",
|
| 19 |
+
"transformer.h.0.mlp.fc2.bias": "model-00001-of-00007.safetensors",
|
| 20 |
+
"transformer.h.0.mlp.fc2.weight": "model-00001-of-00007.safetensors",
|
| 21 |
+
"transformer.h.1.ln.bias": "model-00001-of-00007.safetensors",
|
| 22 |
+
"transformer.h.1.ln.weight": "model-00001-of-00007.safetensors",
|
| 23 |
+
"transformer.h.1.mixer.Wqkv.bias": "model-00001-of-00007.safetensors",
|
| 24 |
+
"transformer.h.1.mixer.Wqkv.weight": "model-00001-of-00007.safetensors",
|
| 25 |
+
"transformer.h.1.mixer.out_proj.bias": "model-00001-of-00007.safetensors",
|
| 26 |
+
"transformer.h.1.mixer.out_proj.weight": "model-00001-of-00007.safetensors",
|
| 27 |
+
"transformer.h.1.mlp.fc1.bias": "model-00001-of-00007.safetensors",
|
| 28 |
+
"transformer.h.1.mlp.fc1.weight": "model-00001-of-00007.safetensors",
|
| 29 |
+
"transformer.h.1.mlp.fc2.bias": "model-00001-of-00007.safetensors",
|
| 30 |
+
"transformer.h.1.mlp.fc2.weight": "model-00001-of-00007.safetensors",
|
| 31 |
+
"transformer.h.10.ln.bias": "model-00002-of-00007.safetensors",
|
| 32 |
+
"transformer.h.10.ln.weight": "model-00002-of-00007.safetensors",
|
| 33 |
+
"transformer.h.10.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 34 |
+
"transformer.h.10.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 35 |
+
"transformer.h.10.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 36 |
+
"transformer.h.10.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 37 |
+
"transformer.h.10.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 38 |
+
"transformer.h.10.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 39 |
+
"transformer.h.10.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 40 |
+
"transformer.h.10.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 41 |
+
"transformer.h.11.ln.bias": "model-00002-of-00007.safetensors",
|
| 42 |
+
"transformer.h.11.ln.weight": "model-00002-of-00007.safetensors",
|
| 43 |
+
"transformer.h.11.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 44 |
+
"transformer.h.11.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 45 |
+
"transformer.h.11.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 46 |
+
"transformer.h.11.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 47 |
+
"transformer.h.11.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 48 |
+
"transformer.h.11.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 49 |
+
"transformer.h.11.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 50 |
+
"transformer.h.11.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 51 |
+
"transformer.h.12.ln.bias": "model-00003-of-00007.safetensors",
|
| 52 |
+
"transformer.h.12.ln.weight": "model-00003-of-00007.safetensors",
|
| 53 |
+
"transformer.h.12.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 54 |
+
"transformer.h.12.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 55 |
+
"transformer.h.12.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 56 |
+
"transformer.h.12.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 57 |
+
"transformer.h.12.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 58 |
+
"transformer.h.12.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 59 |
+
"transformer.h.12.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 60 |
+
"transformer.h.12.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 61 |
+
"transformer.h.13.ln.bias": "model-00003-of-00007.safetensors",
|
| 62 |
+
"transformer.h.13.ln.weight": "model-00003-of-00007.safetensors",
|
| 63 |
+
"transformer.h.13.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 64 |
+
"transformer.h.13.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 65 |
+
"transformer.h.13.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 66 |
+
"transformer.h.13.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 67 |
+
"transformer.h.13.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 68 |
+
"transformer.h.13.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 69 |
+
"transformer.h.13.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 70 |
+
"transformer.h.13.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 71 |
+
"transformer.h.14.ln.bias": "model-00003-of-00007.safetensors",
|
| 72 |
+
"transformer.h.14.ln.weight": "model-00003-of-00007.safetensors",
|
| 73 |
+
"transformer.h.14.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 74 |
+
"transformer.h.14.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 75 |
+
"transformer.h.14.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 76 |
+
"transformer.h.14.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 77 |
+
"transformer.h.14.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 78 |
+
"transformer.h.14.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 79 |
+
"transformer.h.14.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 80 |
+
"transformer.h.14.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 81 |
+
"transformer.h.15.ln.bias": "model-00003-of-00007.safetensors",
|
| 82 |
+
"transformer.h.15.ln.weight": "model-00003-of-00007.safetensors",
|
| 83 |
+
"transformer.h.15.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 84 |
+
"transformer.h.15.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 85 |
+
"transformer.h.15.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 86 |
+
"transformer.h.15.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 87 |
+
"transformer.h.15.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 88 |
+
"transformer.h.15.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 89 |
+
"transformer.h.15.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 90 |
+
"transformer.h.15.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 91 |
+
"transformer.h.16.ln.bias": "model-00003-of-00007.safetensors",
|
| 92 |
+
"transformer.h.16.ln.weight": "model-00003-of-00007.safetensors",
|
| 93 |
+
"transformer.h.16.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 94 |
+
"transformer.h.16.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 95 |
+
"transformer.h.16.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 96 |
+
"transformer.h.16.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 97 |
+
"transformer.h.16.mlp.fc1.bias": "model-00003-of-00007.safetensors",
|
| 98 |
+
"transformer.h.16.mlp.fc1.weight": "model-00003-of-00007.safetensors",
|
| 99 |
+
"transformer.h.16.mlp.fc2.bias": "model-00003-of-00007.safetensors",
|
| 100 |
+
"transformer.h.16.mlp.fc2.weight": "model-00003-of-00007.safetensors",
|
| 101 |
+
"transformer.h.17.ln.bias": "model-00003-of-00007.safetensors",
|
| 102 |
+
"transformer.h.17.ln.weight": "model-00003-of-00007.safetensors",
|
| 103 |
+
"transformer.h.17.mixer.Wqkv.bias": "model-00003-of-00007.safetensors",
|
| 104 |
+
"transformer.h.17.mixer.Wqkv.weight": "model-00003-of-00007.safetensors",
|
| 105 |
+
"transformer.h.17.mixer.out_proj.bias": "model-00003-of-00007.safetensors",
|
| 106 |
+
"transformer.h.17.mixer.out_proj.weight": "model-00003-of-00007.safetensors",
|
| 107 |
+
"transformer.h.17.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 108 |
+
"transformer.h.17.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 109 |
+
"transformer.h.17.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 110 |
+
"transformer.h.17.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 111 |
+
"transformer.h.18.ln.bias": "model-00004-of-00007.safetensors",
|
| 112 |
+
"transformer.h.18.ln.weight": "model-00004-of-00007.safetensors",
|
| 113 |
+
"transformer.h.18.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 114 |
+
"transformer.h.18.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 115 |
+
"transformer.h.18.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 116 |
+
"transformer.h.18.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 117 |
+
"transformer.h.18.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 118 |
+
"transformer.h.18.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 119 |
+
"transformer.h.18.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 120 |
+
"transformer.h.18.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 121 |
+
"transformer.h.19.ln.bias": "model-00004-of-00007.safetensors",
|
| 122 |
+
"transformer.h.19.ln.weight": "model-00004-of-00007.safetensors",
|
| 123 |
+
"transformer.h.19.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 124 |
+
"transformer.h.19.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 125 |
+
"transformer.h.19.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 126 |
+
"transformer.h.19.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 127 |
+
"transformer.h.19.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 128 |
+
"transformer.h.19.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 129 |
+
"transformer.h.19.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 130 |
+
"transformer.h.19.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 131 |
+
"transformer.h.2.ln.bias": "model-00001-of-00007.safetensors",
|
| 132 |
+
"transformer.h.2.ln.weight": "model-00001-of-00007.safetensors",
|
| 133 |
+
"transformer.h.2.mixer.Wqkv.bias": "model-00001-of-00007.safetensors",
|
| 134 |
+
"transformer.h.2.mixer.Wqkv.weight": "model-00001-of-00007.safetensors",
|
| 135 |
+
"transformer.h.2.mixer.out_proj.bias": "model-00001-of-00007.safetensors",
|
| 136 |
+
"transformer.h.2.mixer.out_proj.weight": "model-00001-of-00007.safetensors",
|
| 137 |
+
"transformer.h.2.mlp.fc1.bias": "model-00001-of-00007.safetensors",
|
| 138 |
+
"transformer.h.2.mlp.fc1.weight": "model-00001-of-00007.safetensors",
|
| 139 |
+
"transformer.h.2.mlp.fc2.bias": "model-00001-of-00007.safetensors",
|
| 140 |
+
"transformer.h.2.mlp.fc2.weight": "model-00001-of-00007.safetensors",
|
| 141 |
+
"transformer.h.20.ln.bias": "model-00004-of-00007.safetensors",
|
| 142 |
+
"transformer.h.20.ln.weight": "model-00004-of-00007.safetensors",
|
| 143 |
+
"transformer.h.20.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 144 |
+
"transformer.h.20.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 145 |
+
"transformer.h.20.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 146 |
+
"transformer.h.20.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 147 |
+
"transformer.h.20.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 148 |
+
"transformer.h.20.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 149 |
+
"transformer.h.20.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 150 |
+
"transformer.h.20.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 151 |
+
"transformer.h.21.ln.bias": "model-00004-of-00007.safetensors",
|
| 152 |
+
"transformer.h.21.ln.weight": "model-00004-of-00007.safetensors",
|
| 153 |
+
"transformer.h.21.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 154 |
+
"transformer.h.21.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 155 |
+
"transformer.h.21.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 156 |
+
"transformer.h.21.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 157 |
+
"transformer.h.21.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 158 |
+
"transformer.h.21.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 159 |
+
"transformer.h.21.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 160 |
+
"transformer.h.21.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 161 |
+
"transformer.h.22.ln.bias": "model-00004-of-00007.safetensors",
|
| 162 |
+
"transformer.h.22.ln.weight": "model-00004-of-00007.safetensors",
|
| 163 |
+
"transformer.h.22.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 164 |
+
"transformer.h.22.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 165 |
+
"transformer.h.22.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 166 |
+
"transformer.h.22.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 167 |
+
"transformer.h.22.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 168 |
+
"transformer.h.22.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 169 |
+
"transformer.h.22.mlp.fc2.bias": "model-00004-of-00007.safetensors",
|
| 170 |
+
"transformer.h.22.mlp.fc2.weight": "model-00004-of-00007.safetensors",
|
| 171 |
+
"transformer.h.23.ln.bias": "model-00004-of-00007.safetensors",
|
| 172 |
+
"transformer.h.23.ln.weight": "model-00004-of-00007.safetensors",
|
| 173 |
+
"transformer.h.23.mixer.Wqkv.bias": "model-00004-of-00007.safetensors",
|
| 174 |
+
"transformer.h.23.mixer.Wqkv.weight": "model-00004-of-00007.safetensors",
|
| 175 |
+
"transformer.h.23.mixer.out_proj.bias": "model-00004-of-00007.safetensors",
|
| 176 |
+
"transformer.h.23.mixer.out_proj.weight": "model-00004-of-00007.safetensors",
|
| 177 |
+
"transformer.h.23.mlp.fc1.bias": "model-00004-of-00007.safetensors",
|
| 178 |
+
"transformer.h.23.mlp.fc1.weight": "model-00004-of-00007.safetensors",
|
| 179 |
+
"transformer.h.23.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 180 |
+
"transformer.h.23.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 181 |
+
"transformer.h.24.ln.bias": "model-00005-of-00007.safetensors",
|
| 182 |
+
"transformer.h.24.ln.weight": "model-00005-of-00007.safetensors",
|
| 183 |
+
"transformer.h.24.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 184 |
+
"transformer.h.24.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 185 |
+
"transformer.h.24.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 186 |
+
"transformer.h.24.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 187 |
+
"transformer.h.24.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 188 |
+
"transformer.h.24.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 189 |
+
"transformer.h.24.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 190 |
+
"transformer.h.24.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 191 |
+
"transformer.h.25.ln.bias": "model-00005-of-00007.safetensors",
|
| 192 |
+
"transformer.h.25.ln.weight": "model-00005-of-00007.safetensors",
|
| 193 |
+
"transformer.h.25.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 194 |
+
"transformer.h.25.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 195 |
+
"transformer.h.25.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 196 |
+
"transformer.h.25.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 197 |
+
"transformer.h.25.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 198 |
+
"transformer.h.25.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 199 |
+
"transformer.h.25.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 200 |
+
"transformer.h.25.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 201 |
+
"transformer.h.26.ln.bias": "model-00005-of-00007.safetensors",
|
| 202 |
+
"transformer.h.26.ln.weight": "model-00005-of-00007.safetensors",
|
| 203 |
+
"transformer.h.26.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 204 |
+
"transformer.h.26.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 205 |
+
"transformer.h.26.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 206 |
+
"transformer.h.26.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 207 |
+
"transformer.h.26.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 208 |
+
"transformer.h.26.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 209 |
+
"transformer.h.26.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 210 |
+
"transformer.h.26.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 211 |
+
"transformer.h.27.ln.bias": "model-00005-of-00007.safetensors",
|
| 212 |
+
"transformer.h.27.ln.weight": "model-00005-of-00007.safetensors",
|
| 213 |
+
"transformer.h.27.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 214 |
+
"transformer.h.27.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 215 |
+
"transformer.h.27.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 216 |
+
"transformer.h.27.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 217 |
+
"transformer.h.27.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 218 |
+
"transformer.h.27.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 219 |
+
"transformer.h.27.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 220 |
+
"transformer.h.27.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 221 |
+
"transformer.h.28.ln.bias": "model-00005-of-00007.safetensors",
|
| 222 |
+
"transformer.h.28.ln.weight": "model-00005-of-00007.safetensors",
|
| 223 |
+
"transformer.h.28.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 224 |
+
"transformer.h.28.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 225 |
+
"transformer.h.28.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 226 |
+
"transformer.h.28.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 227 |
+
"transformer.h.28.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 228 |
+
"transformer.h.28.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 229 |
+
"transformer.h.28.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 230 |
+
"transformer.h.28.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 231 |
+
"transformer.h.29.ln.bias": "model-00005-of-00007.safetensors",
|
| 232 |
+
"transformer.h.29.ln.weight": "model-00005-of-00007.safetensors",
|
| 233 |
+
"transformer.h.29.mixer.Wqkv.bias": "model-00005-of-00007.safetensors",
|
| 234 |
+
"transformer.h.29.mixer.Wqkv.weight": "model-00005-of-00007.safetensors",
|
| 235 |
+
"transformer.h.29.mixer.out_proj.bias": "model-00005-of-00007.safetensors",
|
| 236 |
+
"transformer.h.29.mixer.out_proj.weight": "model-00005-of-00007.safetensors",
|
| 237 |
+
"transformer.h.29.mlp.fc1.bias": "model-00005-of-00007.safetensors",
|
| 238 |
+
"transformer.h.29.mlp.fc1.weight": "model-00005-of-00007.safetensors",
|
| 239 |
+
"transformer.h.29.mlp.fc2.bias": "model-00005-of-00007.safetensors",
|
| 240 |
+
"transformer.h.29.mlp.fc2.weight": "model-00005-of-00007.safetensors",
|
| 241 |
+
"transformer.h.3.ln.bias": "model-00001-of-00007.safetensors",
|
| 242 |
+
"transformer.h.3.ln.weight": "model-00001-of-00007.safetensors",
|
| 243 |
+
"transformer.h.3.mixer.Wqkv.bias": "model-00001-of-00007.safetensors",
|
| 244 |
+
"transformer.h.3.mixer.Wqkv.weight": "model-00001-of-00007.safetensors",
|
| 245 |
+
"transformer.h.3.mixer.out_proj.bias": "model-00001-of-00007.safetensors",
|
| 246 |
+
"transformer.h.3.mixer.out_proj.weight": "model-00001-of-00007.safetensors",
|
| 247 |
+
"transformer.h.3.mlp.fc1.bias": "model-00001-of-00007.safetensors",
|
| 248 |
+
"transformer.h.3.mlp.fc1.weight": "model-00001-of-00007.safetensors",
|
| 249 |
+
"transformer.h.3.mlp.fc2.bias": "model-00001-of-00007.safetensors",
|
| 250 |
+
"transformer.h.3.mlp.fc2.weight": "model-00001-of-00007.safetensors",
|
| 251 |
+
"transformer.h.30.ln.bias": "model-00005-of-00007.safetensors",
|
| 252 |
+
"transformer.h.30.ln.weight": "model-00005-of-00007.safetensors",
|
| 253 |
+
"transformer.h.30.mixer.Wqkv.bias": "model-00006-of-00007.safetensors",
|
| 254 |
+
"transformer.h.30.mixer.Wqkv.weight": "model-00006-of-00007.safetensors",
|
| 255 |
+
"transformer.h.30.mixer.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 256 |
+
"transformer.h.30.mixer.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 257 |
+
"transformer.h.30.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 258 |
+
"transformer.h.30.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 259 |
+
"transformer.h.30.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 260 |
+
"transformer.h.30.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 261 |
+
"transformer.h.31.ln.bias": "model-00006-of-00007.safetensors",
|
| 262 |
+
"transformer.h.31.ln.weight": "model-00006-of-00007.safetensors",
|
| 263 |
+
"transformer.h.31.mixer.Wqkv.bias": "model-00006-of-00007.safetensors",
|
| 264 |
+
"transformer.h.31.mixer.Wqkv.weight": "model-00006-of-00007.safetensors",
|
| 265 |
+
"transformer.h.31.mixer.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 266 |
+
"transformer.h.31.mixer.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 267 |
+
"transformer.h.31.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 268 |
+
"transformer.h.31.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 269 |
+
"transformer.h.31.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 270 |
+
"transformer.h.31.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 271 |
+
"transformer.h.4.ln.bias": "model-00001-of-00007.safetensors",
|
| 272 |
+
"transformer.h.4.ln.weight": "model-00001-of-00007.safetensors",
|
| 273 |
+
"transformer.h.4.mixer.Wqkv.bias": "model-00001-of-00007.safetensors",
|
| 274 |
+
"transformer.h.4.mixer.Wqkv.weight": "model-00001-of-00007.safetensors",
|
| 275 |
+
"transformer.h.4.mixer.out_proj.bias": "model-00001-of-00007.safetensors",
|
| 276 |
+
"transformer.h.4.mixer.out_proj.weight": "model-00001-of-00007.safetensors",
|
| 277 |
+
"transformer.h.4.mlp.fc1.bias": "model-00001-of-00007.safetensors",
|
| 278 |
+
"transformer.h.4.mlp.fc1.weight": "model-00001-of-00007.safetensors",
|
| 279 |
+
"transformer.h.4.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 280 |
+
"transformer.h.4.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 281 |
+
"transformer.h.5.ln.bias": "model-00002-of-00007.safetensors",
|
| 282 |
+
"transformer.h.5.ln.weight": "model-00002-of-00007.safetensors",
|
| 283 |
+
"transformer.h.5.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 284 |
+
"transformer.h.5.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 285 |
+
"transformer.h.5.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 286 |
+
"transformer.h.5.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 287 |
+
"transformer.h.5.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 288 |
+
"transformer.h.5.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 289 |
+
"transformer.h.5.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 290 |
+
"transformer.h.5.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 291 |
+
"transformer.h.6.ln.bias": "model-00002-of-00007.safetensors",
|
| 292 |
+
"transformer.h.6.ln.weight": "model-00002-of-00007.safetensors",
|
| 293 |
+
"transformer.h.6.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 294 |
+
"transformer.h.6.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 295 |
+
"transformer.h.6.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 296 |
+
"transformer.h.6.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 297 |
+
"transformer.h.6.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 298 |
+
"transformer.h.6.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 299 |
+
"transformer.h.6.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 300 |
+
"transformer.h.6.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 301 |
+
"transformer.h.7.ln.bias": "model-00002-of-00007.safetensors",
|
| 302 |
+
"transformer.h.7.ln.weight": "model-00002-of-00007.safetensors",
|
| 303 |
+
"transformer.h.7.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 304 |
+
"transformer.h.7.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 305 |
+
"transformer.h.7.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 306 |
+
"transformer.h.7.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 307 |
+
"transformer.h.7.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 308 |
+
"transformer.h.7.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 309 |
+
"transformer.h.7.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 310 |
+
"transformer.h.7.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 311 |
+
"transformer.h.8.ln.bias": "model-00002-of-00007.safetensors",
|
| 312 |
+
"transformer.h.8.ln.weight": "model-00002-of-00007.safetensors",
|
| 313 |
+
"transformer.h.8.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 314 |
+
"transformer.h.8.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 315 |
+
"transformer.h.8.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 316 |
+
"transformer.h.8.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 317 |
+
"transformer.h.8.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 318 |
+
"transformer.h.8.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 319 |
+
"transformer.h.8.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 320 |
+
"transformer.h.8.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 321 |
+
"transformer.h.9.ln.bias": "model-00002-of-00007.safetensors",
|
| 322 |
+
"transformer.h.9.ln.weight": "model-00002-of-00007.safetensors",
|
| 323 |
+
"transformer.h.9.mixer.Wqkv.bias": "model-00002-of-00007.safetensors",
|
| 324 |
+
"transformer.h.9.mixer.Wqkv.weight": "model-00002-of-00007.safetensors",
|
| 325 |
+
"transformer.h.9.mixer.out_proj.bias": "model-00002-of-00007.safetensors",
|
| 326 |
+
"transformer.h.9.mixer.out_proj.weight": "model-00002-of-00007.safetensors",
|
| 327 |
+
"transformer.h.9.mlp.fc1.bias": "model-00002-of-00007.safetensors",
|
| 328 |
+
"transformer.h.9.mlp.fc1.weight": "model-00002-of-00007.safetensors",
|
| 329 |
+
"transformer.h.9.mlp.fc2.bias": "model-00002-of-00007.safetensors",
|
| 330 |
+
"transformer.h.9.mlp.fc2.weight": "model-00002-of-00007.safetensors",
|
| 331 |
+
"transformer.mm_projector.0.bias": "model-00007-of-00007.safetensors",
|
| 332 |
+
"transformer.mm_projector.0.weight": "model-00007-of-00007.safetensors",
|
| 333 |
+
"transformer.mm_projector.2.bias": "model-00007-of-00007.safetensors",
|
| 334 |
+
"transformer.mm_projector.2.weight": "model-00007-of-00007.safetensors",
|
| 335 |
+
"transformer.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.bias": "model-00006-of-00007.safetensors",
|
| 336 |
+
"transformer.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.weight": "model-00006-of-00007.safetensors",
|
| 337 |
+
"transformer.vision_tower.vision_tower.vision_model.embeddings.position_embedding.weight": "model-00006-of-00007.safetensors",
|
| 338 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 339 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 340 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 341 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 342 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 343 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 344 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 345 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 346 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 347 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 348 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 349 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 350 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 351 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 352 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 353 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 354 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 355 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 356 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 357 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 358 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 359 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 360 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 361 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 362 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 363 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 364 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 365 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 366 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 367 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 368 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 369 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 370 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 371 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 372 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 373 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 374 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 375 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 376 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 377 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 378 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 379 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 380 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 381 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 382 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 383 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 384 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 385 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 386 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 387 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 388 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 389 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 390 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 391 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 392 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 393 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 394 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 395 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 396 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 397 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 398 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 399 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 400 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 401 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 402 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 403 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 404 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 405 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 406 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 407 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 408 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 409 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 410 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 411 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 412 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 413 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 414 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 415 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 416 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 417 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 418 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 419 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 420 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 421 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 422 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 423 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 424 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 425 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 426 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 427 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 428 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 429 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 430 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 431 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 432 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 433 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 434 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 435 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 436 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 437 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 438 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 439 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 440 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 441 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 442 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 443 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 444 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 445 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 446 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 447 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 448 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 449 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 450 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 451 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 452 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 453 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 454 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 455 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 456 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 457 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 458 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 459 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 460 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 461 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 462 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 463 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 464 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 465 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 466 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 467 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 468 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 469 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 470 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 471 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 472 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 473 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 474 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 475 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 476 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 477 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 478 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 479 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 480 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 481 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 482 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 483 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 484 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 485 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 486 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 487 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 488 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 489 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 490 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 491 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 492 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 493 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 494 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 495 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 496 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 497 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 498 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 499 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 500 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 501 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 502 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 503 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 504 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 505 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 506 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 507 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 508 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 509 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 510 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 511 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 512 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 513 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 514 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 515 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 516 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 517 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 518 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 519 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 520 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 521 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 522 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 523 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 524 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 525 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 526 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 527 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 528 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 529 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 530 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 531 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 532 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 533 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 534 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 535 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 536 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 537 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 538 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 539 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 540 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 541 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 542 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 543 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 544 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 545 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 546 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 547 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 548 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 549 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 550 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 551 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 552 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 553 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 554 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 555 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 556 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 557 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 558 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 559 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 560 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 561 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 562 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 563 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 564 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 565 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 566 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 567 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 568 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 569 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 570 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 571 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 572 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 573 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 574 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 575 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 576 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 577 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 578 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 579 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 580 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 581 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 582 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 583 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 584 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 585 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 586 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 587 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 588 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 589 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 590 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 591 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 592 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 593 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 594 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm1.bias": "model-00007-of-00007.safetensors",
|
| 595 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm1.weight": "model-00007-of-00007.safetensors",
|
| 596 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm2.bias": "model-00007-of-00007.safetensors",
|
| 597 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm2.weight": "model-00007-of-00007.safetensors",
|
| 598 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc1.bias": "model-00007-of-00007.safetensors",
|
| 599 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc1.weight": "model-00007-of-00007.safetensors",
|
| 600 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc2.bias": "model-00007-of-00007.safetensors",
|
| 601 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc2.weight": "model-00007-of-00007.safetensors",
|
| 602 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 603 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 604 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
|
| 605 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
|
| 606 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
|
| 607 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
|
| 608 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
|
| 609 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
|
| 610 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm1.bias": "model-00007-of-00007.safetensors",
|
| 611 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm1.weight": "model-00007-of-00007.safetensors",
|
| 612 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm2.bias": "model-00007-of-00007.safetensors",
|
| 613 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm2.weight": "model-00007-of-00007.safetensors",
|
| 614 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc1.bias": "model-00007-of-00007.safetensors",
|
| 615 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc1.weight": "model-00007-of-00007.safetensors",
|
| 616 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc2.bias": "model-00007-of-00007.safetensors",
|
| 617 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc2.weight": "model-00007-of-00007.safetensors",
|
| 618 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
|
| 619 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
|
| 620 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
|
| 621 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
|
| 622 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
|
| 623 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
|
| 624 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
|
| 625 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
|
| 626 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm1.bias": "model-00007-of-00007.safetensors",
|
| 627 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm1.weight": "model-00007-of-00007.safetensors",
|
| 628 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm2.bias": "model-00007-of-00007.safetensors",
|
| 629 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm2.weight": "model-00007-of-00007.safetensors",
|
| 630 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc1.bias": "model-00007-of-00007.safetensors",
|
| 631 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc1.weight": "model-00007-of-00007.safetensors",
|
| 632 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc2.bias": "model-00007-of-00007.safetensors",
|
| 633 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc2.weight": "model-00007-of-00007.safetensors",
|
| 634 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.k_proj.bias": "model-00007-of-00007.safetensors",
|
| 635 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.k_proj.weight": "model-00007-of-00007.safetensors",
|
| 636 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.out_proj.bias": "model-00007-of-00007.safetensors",
|
| 637 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.out_proj.weight": "model-00007-of-00007.safetensors",
|
| 638 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.q_proj.bias": "model-00007-of-00007.safetensors",
|
| 639 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.q_proj.weight": "model-00007-of-00007.safetensors",
|
| 640 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.v_proj.bias": "model-00007-of-00007.safetensors",
|
| 641 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.v_proj.weight": "model-00007-of-00007.safetensors",
|
| 642 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 643 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 644 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 645 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 646 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 647 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 648 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 649 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 650 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 651 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 652 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 653 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 654 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 655 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 656 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 657 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 658 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 659 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 660 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 661 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 662 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 663 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 664 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 665 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 666 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 667 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 668 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 669 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 670 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 671 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 672 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 673 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 674 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 675 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 676 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 677 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 678 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 679 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 680 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 681 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 682 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 683 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 684 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 685 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 686 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 687 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 688 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 689 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 690 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 691 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 692 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 693 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 694 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 695 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 696 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 697 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 698 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 699 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 700 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 701 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 702 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 703 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 704 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 705 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 706 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 707 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 708 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 709 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 710 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 711 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 712 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 713 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 714 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 715 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 716 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 717 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 718 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 719 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 720 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 721 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 722 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 723 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 724 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 725 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 726 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 727 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 728 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 729 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 730 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 731 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 732 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 733 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 734 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 735 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 736 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 737 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.v_proj.weight": "model-00006-of-00007.safetensors",
|
| 738 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm1.bias": "model-00006-of-00007.safetensors",
|
| 739 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm1.weight": "model-00006-of-00007.safetensors",
|
| 740 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm2.bias": "model-00006-of-00007.safetensors",
|
| 741 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm2.weight": "model-00006-of-00007.safetensors",
|
| 742 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc1.bias": "model-00006-of-00007.safetensors",
|
| 743 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc1.weight": "model-00006-of-00007.safetensors",
|
| 744 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc2.bias": "model-00006-of-00007.safetensors",
|
| 745 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc2.weight": "model-00006-of-00007.safetensors",
|
| 746 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.k_proj.bias": "model-00006-of-00007.safetensors",
|
| 747 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.k_proj.weight": "model-00006-of-00007.safetensors",
|
| 748 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.out_proj.bias": "model-00006-of-00007.safetensors",
|
| 749 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.out_proj.weight": "model-00006-of-00007.safetensors",
|
| 750 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.q_proj.bias": "model-00006-of-00007.safetensors",
|
| 751 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.q_proj.weight": "model-00006-of-00007.safetensors",
|
| 752 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.v_proj.bias": "model-00006-of-00007.safetensors",
|
| 753 |
+
"transformer.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.v_proj.weight": "model-00006-of-00007.safetensors"
|
| 754 |
+
}
|
| 755 |
+
}
|
modeling_imp.py
ADDED
|
@@ -0,0 +1,1262 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) MILVLG team.
|
| 2 |
+
# Licensed under the Apache 2.0 license.
|
| 3 |
+
#
|
| 4 |
+
# Some code here is copied from the project Phi-2 (https://huggingface.co/microsoft/phi-2),
|
| 5 |
+
# SigLIP@transformers==4.37.0.dev0 (https://huggingface.co/google/siglip-so400m-patch14-384),
|
| 6 |
+
# and Llava (https://github.com/haotian-liu/LLaVA), and modified by
|
| 7 |
+
# Zhenwei Shao (shaozw@hdu.edu.cn) @ MILVLG. We thank them for their great works.
|
| 8 |
+
# And their original licenses and copyright should be inherited (see the statements
|
| 9 |
+
# in `configuration_imp.py` for more details).
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
# Be careful: The way how `past_key_values.seqlen_offset` is updated is modified from
|
| 13 |
+
# the implementation of original Phi-2. See the comments below for details.
|
| 14 |
+
|
| 15 |
+
from __future__ import annotations
|
| 16 |
+
import os
|
| 17 |
+
import math
|
| 18 |
+
import re
|
| 19 |
+
from dataclasses import dataclass, field
|
| 20 |
+
from typing import Any, Dict, Optional, Tuple, Union, List
|
| 21 |
+
from abc import ABC, abstractmethod
|
| 22 |
+
|
| 23 |
+
import torch
|
| 24 |
+
import torch.nn as nn
|
| 25 |
+
from einops import rearrange, repeat
|
| 26 |
+
from transformers import (
|
| 27 |
+
PretrainedConfig,
|
| 28 |
+
PreTrainedModel,
|
| 29 |
+
AutoConfig,
|
| 30 |
+
AutoModelForCausalLM
|
| 31 |
+
)
|
| 32 |
+
from transformers.activations import ACT2FN
|
| 33 |
+
from transformers.modeling_outputs import CausalLMOutputWithPast
|
| 34 |
+
import sys
|
| 35 |
+
from .configuration_imp import PhiConfig, ImpConfig
|
| 36 |
+
from .vision_encoder import VisionTower
|
| 37 |
+
|
| 38 |
+
try:
|
| 39 |
+
from flash_attn.bert_padding import pad_input, unpad_input
|
| 40 |
+
from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
|
| 41 |
+
from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
|
| 42 |
+
from flash_attn.ops.fused_dense import FusedDense
|
| 43 |
+
except:
|
| 44 |
+
pad_input, unpad_input = None, None
|
| 45 |
+
FlashRotaryEmbedding = None
|
| 46 |
+
FlashSelfAttention, FlashCrossAttention = None, None
|
| 47 |
+
FusedDense = None
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
@dataclass
|
| 51 |
+
class InferenceParams:
|
| 52 |
+
"""Inference parameters passed to model to efficiently calculate
|
| 53 |
+
and store context during inference.
|
| 54 |
+
|
| 55 |
+
Reference:
|
| 56 |
+
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
|
| 57 |
+
|
| 58 |
+
Args:
|
| 59 |
+
max_seqlen: Maximum sequence length.
|
| 60 |
+
max_batch_size: Maximum batch size.
|
| 61 |
+
seqlen_offset: Sequence length offset.
|
| 62 |
+
batch_size_offset: Batch size offset.
|
| 63 |
+
key_value_memory_dict: Key value memory dictionary.
|
| 64 |
+
lengths_per_sample: Lengths per sample.
|
| 65 |
+
|
| 66 |
+
"""
|
| 67 |
+
|
| 68 |
+
max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
|
| 69 |
+
|
| 70 |
+
max_batch_size: int = field(metadata={"help": "Maximum batch size."})
|
| 71 |
+
|
| 72 |
+
seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
|
| 73 |
+
|
| 74 |
+
batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
|
| 75 |
+
|
| 76 |
+
key_value_memory_dict: Dict[str, Any] = field(
|
| 77 |
+
default_factory=dict, metadata={"help": "Key value memory dictionary."}
|
| 78 |
+
)
|
| 79 |
+
|
| 80 |
+
lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
class Embedding(nn.Module):
|
| 84 |
+
"""Token embedding with dropout."""
|
| 85 |
+
|
| 86 |
+
def __init__(self, config: PretrainedConfig) -> None:
|
| 87 |
+
super().__init__()
|
| 88 |
+
|
| 89 |
+
self.wte = nn.Embedding(config.vocab_size, config.n_embd)
|
| 90 |
+
self.drop = nn.Dropout(config.embd_pdrop)
|
| 91 |
+
|
| 92 |
+
def forward(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
|
| 93 |
+
input_shape = input_ids.size()
|
| 94 |
+
input_ids = input_ids.view(-1, input_shape[-1])
|
| 95 |
+
|
| 96 |
+
hidden_states = self.wte(input_ids)
|
| 97 |
+
hidden_states = self.drop(hidden_states)
|
| 98 |
+
|
| 99 |
+
return hidden_states
|
| 100 |
+
|
| 101 |
+
|
| 102 |
+
|
| 103 |
+
def _apply_rotary_emb(
|
| 104 |
+
x: torch.FloatTensor,
|
| 105 |
+
cos: torch.FloatTensor,
|
| 106 |
+
sin: torch.FloatTensor,
|
| 107 |
+
) -> torch.FloatTensor:
|
| 108 |
+
_, seqlen, _, _ = x.shape
|
| 109 |
+
_, rotary_dim = cos.shape
|
| 110 |
+
rotary_dim *= 2
|
| 111 |
+
|
| 112 |
+
x_rot = x[:, :, :, :rotary_dim]
|
| 113 |
+
x_pass = x[:, :, :, rotary_dim:]
|
| 114 |
+
|
| 115 |
+
x1, x2 = x_rot.chunk(2, dim=-1)
|
| 116 |
+
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
| 117 |
+
x1, x2, c, s = [t.to(dtype=torch.float32) for t in [x1, x2, c, s]]
|
| 118 |
+
|
| 119 |
+
x_rot = torch.cat([x1 * c - x2 * s, x1 * s + x2 * c], axis=-1).to(x.dtype)
|
| 120 |
+
|
| 121 |
+
return torch.cat([x_rot, x_pass], axis=-1)
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def _apply_rotary_emb_kv(
|
| 125 |
+
kv: torch.FloatTensor,
|
| 126 |
+
cos: torch.FloatTensor,
|
| 127 |
+
sin: torch.FloatTensor,
|
| 128 |
+
cos_k: Optional[torch.FloatTensor] = None,
|
| 129 |
+
sin_k: Optional[torch.FloatTensor] = None,
|
| 130 |
+
) -> torch.FloatTensor:
|
| 131 |
+
_, seqlen, _, _, _ = kv.shape
|
| 132 |
+
_, rotary_dim = cos.shape
|
| 133 |
+
rotary_dim *= 2
|
| 134 |
+
|
| 135 |
+
k_rot = kv[:, :, 0, :, :rotary_dim]
|
| 136 |
+
k_pass = kv[:, :, 0, :, rotary_dim:]
|
| 137 |
+
|
| 138 |
+
k1, k2 = k_rot.chunk(2, dim=-1)
|
| 139 |
+
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
| 140 |
+
k1, k2, c, s = [t.to(dtype=torch.float32) for t in [k1, k2, c, s]]
|
| 141 |
+
|
| 142 |
+
k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(kv.dtype)
|
| 143 |
+
|
| 144 |
+
return torch.cat(
|
| 145 |
+
[
|
| 146 |
+
torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
|
| 147 |
+
kv[:, :, 1:2, :, :],
|
| 148 |
+
],
|
| 149 |
+
axis=2,
|
| 150 |
+
)
|
| 151 |
+
|
| 152 |
+
|
| 153 |
+
def _apply_rotary_emb_qkv(
|
| 154 |
+
qkv: torch.FloatTensor,
|
| 155 |
+
cos: torch.FloatTensor,
|
| 156 |
+
sin: torch.FloatTensor,
|
| 157 |
+
cos_k: Optional[torch.FloatTensor] = None,
|
| 158 |
+
sin_k: Optional[torch.FloatTensor] = None,
|
| 159 |
+
) -> torch.FloatTensor:
|
| 160 |
+
_, seqlen, _, _, _ = qkv.shape
|
| 161 |
+
_, rotary_dim = cos.shape
|
| 162 |
+
rotary_dim *= 2
|
| 163 |
+
|
| 164 |
+
q_rot = qkv[:, :, 0, :, :rotary_dim]
|
| 165 |
+
q_pass = qkv[:, :, 0, :, rotary_dim:]
|
| 166 |
+
|
| 167 |
+
k_rot = qkv[:, :, 1, :, :rotary_dim]
|
| 168 |
+
k_pass = qkv[:, :, 1, :, rotary_dim:]
|
| 169 |
+
|
| 170 |
+
q1, q2 = q_rot.chunk(2, dim=-1)
|
| 171 |
+
k1, k2 = k_rot.chunk(2, dim=-1)
|
| 172 |
+
c, s = rearrange(cos[:seqlen], "s d -> s 1 d"), rearrange(sin[:seqlen], "s d -> s 1 d")
|
| 173 |
+
q1, q2, k1, k2, c, s = [t.to(dtype=torch.float32) for t in [q1, q2, k1, k2, c, s]]
|
| 174 |
+
|
| 175 |
+
q_rot = torch.cat([q1 * c - q2 * s, q1 * s + q2 * c], axis=-1).to(qkv.dtype)
|
| 176 |
+
k_rot = torch.cat([k1 * c - k2 * s, k1 * s + k2 * c], axis=-1).to(qkv.dtype)
|
| 177 |
+
|
| 178 |
+
return torch.cat(
|
| 179 |
+
[
|
| 180 |
+
torch.cat([q_rot, q_pass], axis=-1).unsqueeze(2),
|
| 181 |
+
torch.cat([k_rot, k_pass], axis=-1).unsqueeze(2),
|
| 182 |
+
qkv[:, :, 2:3, :, :],
|
| 183 |
+
],
|
| 184 |
+
axis=2,
|
| 185 |
+
)
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
class RotaryEmbedding(nn.Module):
|
| 189 |
+
"""Rotary positional embedding (RoPE).
|
| 190 |
+
|
| 191 |
+
Reference:
|
| 192 |
+
RoFormer: Enhanced Transformer with Rotary Position Embedding.
|
| 193 |
+
https://arxiv.org/pdf/2104.09864.pdf.
|
| 194 |
+
|
| 195 |
+
"""
|
| 196 |
+
|
| 197 |
+
def __init__(
|
| 198 |
+
self,
|
| 199 |
+
dim: int,
|
| 200 |
+
base: int = 10000,
|
| 201 |
+
scale_base: Optional[float] = None,
|
| 202 |
+
pos_idx_in_fp32: bool = True,
|
| 203 |
+
max_position_embeddings: int = 2048,
|
| 204 |
+
device: Optional[str] = None,
|
| 205 |
+
**kwargs,
|
| 206 |
+
) -> None:
|
| 207 |
+
super().__init__()
|
| 208 |
+
|
| 209 |
+
if scale_base is not None:
|
| 210 |
+
raise NotImplementedError
|
| 211 |
+
|
| 212 |
+
self.dim = dim
|
| 213 |
+
self.base = float(base)
|
| 214 |
+
self.scale_base = scale_base
|
| 215 |
+
self.pos_idx_in_fp32 = pos_idx_in_fp32
|
| 216 |
+
self.max_position_embeddings = max_position_embeddings
|
| 217 |
+
self.device = device
|
| 218 |
+
|
| 219 |
+
# Generate and save the inverse frequency buffer (non-trainable)
|
| 220 |
+
inv_freq = self._compute_inv_freq(device)
|
| 221 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 222 |
+
|
| 223 |
+
# Generate and save the scale buffer (non-trainable)
|
| 224 |
+
scale = (
|
| 225 |
+
(torch.arange(0, dim, 2, device=device, dtype=torch.float32) + 0.4 * dim) / (1.4 * dim)
|
| 226 |
+
if scale_base is not None
|
| 227 |
+
else None
|
| 228 |
+
)
|
| 229 |
+
self.register_buffer("scale", scale, persistent=False)
|
| 230 |
+
|
| 231 |
+
# Initialize cached attributes since ONNX can't rely on dynamic initialization
|
| 232 |
+
self._update_cos_sin_cache(max_position_embeddings, device=device, dtype=torch.float32)
|
| 233 |
+
|
| 234 |
+
def _compute_inv_freq(self, device: Optional[str] = None) -> torch.FloatTensor:
|
| 235 |
+
return 1.0 / (self.base ** (torch.arange(0, self.dim, 2, device=device, dtype=torch.float32) / self.dim))
|
| 236 |
+
|
| 237 |
+
def _update_cos_sin_cache(
|
| 238 |
+
self,
|
| 239 |
+
seqlen: int,
|
| 240 |
+
device: Optional[str] = None,
|
| 241 |
+
dtype: Optional[torch.dtype] = None,
|
| 242 |
+
) -> None:
|
| 243 |
+
self._seq_len_cached = seqlen
|
| 244 |
+
|
| 245 |
+
# fp32 is preferred since the output of `torch.arange` can be quite large
|
| 246 |
+
# and bf16 would lose a lot of precision
|
| 247 |
+
if self.pos_idx_in_fp32:
|
| 248 |
+
t = torch.arange(seqlen, device=device, dtype=torch.float32)
|
| 249 |
+
if self.inv_freq.dtype != torch.float32:
|
| 250 |
+
inv_freq = self._compute_inv_freq(device=device)
|
| 251 |
+
else:
|
| 252 |
+
inv_freq = self.inv_freq
|
| 253 |
+
else:
|
| 254 |
+
t = torch.arange(seqlen, device=device, dtype=self.inv_freq.dtype)
|
| 255 |
+
inv_freq = self.inv_freq
|
| 256 |
+
|
| 257 |
+
# `torch.outer` is preferred since `torch.einsum` converts from fp32 to fp16 if used with AMP
|
| 258 |
+
freqs = torch.outer(t, inv_freq)
|
| 259 |
+
if self.scale is None:
|
| 260 |
+
self._cos_cached = torch.cos(freqs).to(dtype)
|
| 261 |
+
self._sin_cached = torch.sin(freqs).to(dtype)
|
| 262 |
+
else:
|
| 263 |
+
power = (
|
| 264 |
+
torch.arange(seqlen, dtype=self.scale.dtype, device=self.scale.device) - seqlen // 2
|
| 265 |
+
) / self.scale_base
|
| 266 |
+
scale = self.scale.to(device=power.device) ** rearrange(power, "s -> s 1")
|
| 267 |
+
|
| 268 |
+
# Force the scale multiplication to happen in fp32
|
| 269 |
+
self._cos_cached = (torch.cos(freqs) * scale).to(dtype)
|
| 270 |
+
self._sin_cached = (torch.sin(freqs) * scale).to(dtype)
|
| 271 |
+
self._cos_k_cached = (torch.cos(freqs) / scale).to(dtype)
|
| 272 |
+
self._sin_k_cached = (torch.sin(freqs) / scale).to(dtype)
|
| 273 |
+
|
| 274 |
+
def forward(
|
| 275 |
+
self,
|
| 276 |
+
qkv: torch.Tensor,
|
| 277 |
+
kv: Optional[torch.Tensor] = None,
|
| 278 |
+
seqlen_offset: int = 0,
|
| 279 |
+
**kwargs,
|
| 280 |
+
) -> Tuple[torch.Tensor, torch.Tensor]:
|
| 281 |
+
if (
|
| 282 |
+
self._seq_len_cached < qkv.shape[1] + seqlen_offset
|
| 283 |
+
or self._cos_cached.device != qkv.device
|
| 284 |
+
or self._cos_cached.dtype != qkv.dtype
|
| 285 |
+
or (self.training and self._cos_cached.is_inference())
|
| 286 |
+
):
|
| 287 |
+
self._update_cos_sin_cache(qkv.shape[1] + seqlen_offset, device=qkv.device, dtype=qkv.dtype)
|
| 288 |
+
|
| 289 |
+
if kv is None:
|
| 290 |
+
return _apply_rotary_emb_qkv(
|
| 291 |
+
qkv,
|
| 292 |
+
self._cos_cached[seqlen_offset:],
|
| 293 |
+
self._sin_cached[seqlen_offset:],
|
| 294 |
+
)
|
| 295 |
+
else:
|
| 296 |
+
q = _apply_rotary_emb(
|
| 297 |
+
qkv,
|
| 298 |
+
self._cos_cached[seqlen_offset:],
|
| 299 |
+
self._sin_cached[seqlen_offset:],
|
| 300 |
+
)
|
| 301 |
+
kv = _apply_rotary_emb_kv(
|
| 302 |
+
kv,
|
| 303 |
+
self._cos_cached[seqlen_offset:],
|
| 304 |
+
self._sin_cached[seqlen_offset:],
|
| 305 |
+
)
|
| 306 |
+
|
| 307 |
+
return q, kv
|
| 308 |
+
|
| 309 |
+
|
| 310 |
+
class MLP(nn.Module):
|
| 311 |
+
"""Multi-Layer Perceptron.
|
| 312 |
+
|
| 313 |
+
Reference:
|
| 314 |
+
Attention Is All You Need.
|
| 315 |
+
https://arxiv.org/pdf/1706.03762.pdf.
|
| 316 |
+
|
| 317 |
+
"""
|
| 318 |
+
|
| 319 |
+
def __init__(
|
| 320 |
+
self,
|
| 321 |
+
config: PretrainedConfig,
|
| 322 |
+
n_inner: Optional[int] = None,
|
| 323 |
+
act_fn: Optional[str] = None,
|
| 324 |
+
) -> None:
|
| 325 |
+
super().__init__()
|
| 326 |
+
|
| 327 |
+
act_fn = config.activation_function if act_fn is None else act_fn
|
| 328 |
+
|
| 329 |
+
n_inner = getattr(config, "n_inner", None) if n_inner is None else n_inner
|
| 330 |
+
n_inner = n_inner if n_inner is not None else 4 * config.n_embd
|
| 331 |
+
|
| 332 |
+
self.fc1 = nn.Linear(config.n_embd, n_inner)
|
| 333 |
+
self.fc2 = nn.Linear(n_inner, config.n_embd)
|
| 334 |
+
self.act = ACT2FN[act_fn]
|
| 335 |
+
|
| 336 |
+
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
| 337 |
+
hidden_states = self.fc1(hidden_states)
|
| 338 |
+
hidden_states = self.act(hidden_states)
|
| 339 |
+
hidden_states = self.fc2(hidden_states)
|
| 340 |
+
|
| 341 |
+
return hidden_states
|
| 342 |
+
|
| 343 |
+
|
| 344 |
+
class SelfAttention(nn.Module):
|
| 345 |
+
"""Self-attention layer (compatible with PyTorch).
|
| 346 |
+
|
| 347 |
+
Reference:
|
| 348 |
+
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
|
| 349 |
+
|
| 350 |
+
"""
|
| 351 |
+
|
| 352 |
+
def __init__(
|
| 353 |
+
self,
|
| 354 |
+
causal: bool = True,
|
| 355 |
+
softmax_scale: Optional[float] = None,
|
| 356 |
+
attention_dropout: float = 0.0,
|
| 357 |
+
) -> None:
|
| 358 |
+
super().__init__()
|
| 359 |
+
|
| 360 |
+
self.causal = causal
|
| 361 |
+
self.softmax_scale = softmax_scale
|
| 362 |
+
self.drop = nn.Dropout(attention_dropout)
|
| 363 |
+
|
| 364 |
+
@torch.autocast("cpu", enabled=False)
|
| 365 |
+
@torch.autocast("cuda", enabled=False)
|
| 366 |
+
def forward(
|
| 367 |
+
self,
|
| 368 |
+
qkv: torch.FloatTensor,
|
| 369 |
+
causal: bool = None,
|
| 370 |
+
key_padding_mask: Optional[torch.BoolTensor] = None,
|
| 371 |
+
**kwargs,
|
| 372 |
+
) -> torch.FloatTensor:
|
| 373 |
+
batch_size, seqlen = qkv.shape[0], qkv.shape[1]
|
| 374 |
+
q, k, v = qkv.unbind(dim=2)
|
| 375 |
+
|
| 376 |
+
q = q.to(torch.float32)
|
| 377 |
+
k = k.to(torch.float32)
|
| 378 |
+
|
| 379 |
+
causal = self.causal if causal is None else causal
|
| 380 |
+
softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
|
| 381 |
+
|
| 382 |
+
# Autocast is manually disabled to avoid `torch.einsum` performing the operation
|
| 383 |
+
# using float16, which might lead to overflow
|
| 384 |
+
scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
|
| 385 |
+
|
| 386 |
+
if key_padding_mask is not None:
|
| 387 |
+
padding_mask = torch.full((batch_size, seqlen), -10000.0, dtype=scores.dtype, device=scores.device)
|
| 388 |
+
padding_mask.masked_fill_(key_padding_mask, 0.0)
|
| 389 |
+
|
| 390 |
+
scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
|
| 391 |
+
|
| 392 |
+
if causal:
|
| 393 |
+
causal_mask = torch.triu(torch.full((seqlen, seqlen), -10000.0, device=scores.device), 1)
|
| 394 |
+
scores = scores + causal_mask.to(dtype=scores.dtype)
|
| 395 |
+
|
| 396 |
+
attention = torch.softmax(scores, dim=-1).to(v.dtype)
|
| 397 |
+
attention = self.drop(attention)
|
| 398 |
+
|
| 399 |
+
output = torch.einsum("bhts,bshd->bthd", attention, v)
|
| 400 |
+
|
| 401 |
+
return output
|
| 402 |
+
|
| 403 |
+
|
| 404 |
+
class CrossAttention(nn.Module):
|
| 405 |
+
"""Cross-attention layer (compatible with PyTorch).
|
| 406 |
+
|
| 407 |
+
Reference:
|
| 408 |
+
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/modules/mha.py.
|
| 409 |
+
|
| 410 |
+
"""
|
| 411 |
+
|
| 412 |
+
def __init__(
|
| 413 |
+
self,
|
| 414 |
+
causal: bool = True,
|
| 415 |
+
softmax_scale: Optional[float] = None,
|
| 416 |
+
attention_dropout: float = 0.0,
|
| 417 |
+
) -> None:
|
| 418 |
+
super().__init__()
|
| 419 |
+
|
| 420 |
+
self.causal = causal
|
| 421 |
+
self.softmax_scale = softmax_scale
|
| 422 |
+
self.drop = nn.Dropout(attention_dropout)
|
| 423 |
+
|
| 424 |
+
@torch.autocast("cpu", enabled=False)
|
| 425 |
+
@torch.autocast("cuda", enabled=False)
|
| 426 |
+
def forward(
|
| 427 |
+
self,
|
| 428 |
+
q: torch.FloatTensor,
|
| 429 |
+
kv: torch.FloatTensor,
|
| 430 |
+
causal: bool = None,
|
| 431 |
+
key_padding_mask: Optional[torch.BoolTensor] = None,
|
| 432 |
+
**kwargs,
|
| 433 |
+
) -> torch.FloatTensor:
|
| 434 |
+
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
| 435 |
+
seqlen_k = kv.shape[1]
|
| 436 |
+
|
| 437 |
+
if kv.shape[3] != q.shape[2]:
|
| 438 |
+
kv = repeat(kv, "... hkv d -> ... (hkv g) d", g=q.shape[2] // kv.shape[3])
|
| 439 |
+
k, v = kv.unbind(dim=2)
|
| 440 |
+
|
| 441 |
+
q = q.to(torch.float32)
|
| 442 |
+
k = k.to(torch.float32)
|
| 443 |
+
|
| 444 |
+
causal = self.causal if causal is None else causal
|
| 445 |
+
softmax_scale = self.softmax_scale or 1.0 / math.sqrt(q.shape[-1])
|
| 446 |
+
|
| 447 |
+
# Autocast is manually disabled to avoid `torch.einsum` performing the operation
|
| 448 |
+
# using float16, which might lead to overflow
|
| 449 |
+
scores = torch.einsum("bthd,bshd->bhts", q, k * softmax_scale)
|
| 450 |
+
|
| 451 |
+
if key_padding_mask is not None:
|
| 452 |
+
padding_mask = torch.full(
|
| 453 |
+
(batch_size, seqlen_k),
|
| 454 |
+
-10000.0,
|
| 455 |
+
dtype=scores.dtype,
|
| 456 |
+
device=scores.device,
|
| 457 |
+
)
|
| 458 |
+
padding_mask.masked_fill_(key_padding_mask, 0.0)
|
| 459 |
+
|
| 460 |
+
scores = scores + rearrange(padding_mask, "b s -> b 1 1 s")
|
| 461 |
+
|
| 462 |
+
if causal:
|
| 463 |
+
rows = rearrange(torch.arange(seqlen_q, device=q.device, dtype=torch.long), "s -> s 1")
|
| 464 |
+
cols = torch.arange(seqlen_k, device=k.device, dtype=torch.long)
|
| 465 |
+
causal_mask = cols > rows + seqlen_k - seqlen_q
|
| 466 |
+
|
| 467 |
+
scores = scores.masked_fill(causal_mask, -10000.0)
|
| 468 |
+
|
| 469 |
+
attention = torch.softmax(scores, dim=-1).to(v.dtype)
|
| 470 |
+
attention = self.drop(attention)
|
| 471 |
+
|
| 472 |
+
output = torch.einsum("bhts,bshd->bthd", attention, v)
|
| 473 |
+
|
| 474 |
+
return output
|
| 475 |
+
|
| 476 |
+
|
| 477 |
+
def _find_mha_dims(
|
| 478 |
+
config: PretrainedConfig,
|
| 479 |
+
n_head: Optional[int] = None,
|
| 480 |
+
n_head_kv: Optional[int] = None,
|
| 481 |
+
head_dim: Optional[int] = None,
|
| 482 |
+
) -> Tuple[int, int]:
|
| 483 |
+
if n_head is None and head_dim is None:
|
| 484 |
+
head_dim = config.n_embd // config.n_head
|
| 485 |
+
n_head = config.n_head
|
| 486 |
+
elif n_head is None or head_dim is None:
|
| 487 |
+
raise ValueError("`n_head` and `head_dim` must be both specified or `None`.")
|
| 488 |
+
|
| 489 |
+
if n_head_kv is None:
|
| 490 |
+
n_head_kv = getattr(config, "n_head_kv", None) or n_head
|
| 491 |
+
|
| 492 |
+
return n_head, n_head_kv, head_dim
|
| 493 |
+
|
| 494 |
+
|
| 495 |
+
def _update_kv_cache(kv: torch.FloatTensor, inference_params: InferenceParams, layer_idx: int) -> torch.FloatTensor:
|
| 496 |
+
num_heads, head_dim = kv.shape[-2:]
|
| 497 |
+
|
| 498 |
+
if layer_idx not in inference_params.key_value_memory_dict:
|
| 499 |
+
inference_params.key_value_memory_dict[layer_idx] = torch.empty(
|
| 500 |
+
inference_params.max_batch_size,
|
| 501 |
+
inference_params.max_seqlen,
|
| 502 |
+
2,
|
| 503 |
+
num_heads,
|
| 504 |
+
head_dim,
|
| 505 |
+
dtype=kv.dtype,
|
| 506 |
+
device=kv.device,
|
| 507 |
+
)
|
| 508 |
+
|
| 509 |
+
batch_start = inference_params.batch_size_offset
|
| 510 |
+
batch_end = batch_start + kv.shape[0]
|
| 511 |
+
|
| 512 |
+
sequence_start = inference_params.seqlen_offset
|
| 513 |
+
sequence_end = sequence_start + kv.shape[1]
|
| 514 |
+
|
| 515 |
+
# When the current sequence length is equal to or larger than the maximum sequence length,
|
| 516 |
+
# we need to concatenate the current `kv` with the cached `kv` to expand its length
|
| 517 |
+
if sequence_end >= inference_params.max_seqlen:
|
| 518 |
+
inference_params.key_value_memory_dict[layer_idx] = torch.concatenate((inference_params.key_value_memory_dict[layer_idx], kv), dim=1)
|
| 519 |
+
|
| 520 |
+
inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, sequence_start:sequence_end, ...] = kv
|
| 521 |
+
kv = inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, :sequence_end, ...]
|
| 522 |
+
|
| 523 |
+
return kv
|
| 524 |
+
|
| 525 |
+
|
| 526 |
+
class MHA(nn.Module):
|
| 527 |
+
"""Multi-head attention layer."""
|
| 528 |
+
|
| 529 |
+
def __init__(
|
| 530 |
+
self,
|
| 531 |
+
config: PretrainedConfig,
|
| 532 |
+
dtype: Optional[torch.dtype] = None,
|
| 533 |
+
device: Optional[str] = None,
|
| 534 |
+
rotary_dim: Optional[int] = None,
|
| 535 |
+
rotary_base: float = 10000.0,
|
| 536 |
+
rotary_scale_base: Optional[float] = None,
|
| 537 |
+
n_head: Optional[int] = None,
|
| 538 |
+
n_head_kv: Optional[int] = None,
|
| 539 |
+
head_dim: Optional[int] = None,
|
| 540 |
+
bias: bool = True,
|
| 541 |
+
causal: bool = True,
|
| 542 |
+
softmax_scale: Optional[float] = None,
|
| 543 |
+
layer_idx: Optional[int] = None,
|
| 544 |
+
return_residual: bool = False,
|
| 545 |
+
checkpointing: bool = False,
|
| 546 |
+
) -> None:
|
| 547 |
+
super().__init__()
|
| 548 |
+
|
| 549 |
+
# Rotary embedding
|
| 550 |
+
self.rotary_dim = rotary_dim if rotary_dim is not None else getattr(config, "rotary_dim", 0)
|
| 551 |
+
if self.rotary_dim > 0:
|
| 552 |
+
rotary_cls = FlashRotaryEmbedding if config.flash_rotary else RotaryEmbedding
|
| 553 |
+
if rotary_cls is None:
|
| 554 |
+
rotary_cls = RotaryEmbedding
|
| 555 |
+
|
| 556 |
+
rotary_kwargs = {}
|
| 557 |
+
if rotary_cls is RotaryEmbedding:
|
| 558 |
+
rotary_kwargs["max_position_embeddings"] = config.n_positions
|
| 559 |
+
|
| 560 |
+
self.rotary_emb = rotary_cls(
|
| 561 |
+
self.rotary_dim,
|
| 562 |
+
base=rotary_base,
|
| 563 |
+
scale_base=rotary_scale_base,
|
| 564 |
+
device=device,
|
| 565 |
+
**rotary_kwargs,
|
| 566 |
+
)
|
| 567 |
+
|
| 568 |
+
# MLP
|
| 569 |
+
self.n_head, self.n_head_kv, self.head_dim = _find_mha_dims(
|
| 570 |
+
config, n_head=n_head, n_head_kv=n_head_kv, head_dim=head_dim
|
| 571 |
+
)
|
| 572 |
+
op_size = self.head_dim * (self.n_head + 2 * self.n_head_kv)
|
| 573 |
+
hidden_size = config.n_embd
|
| 574 |
+
|
| 575 |
+
linear_cls = FusedDense if config.fused_dense else nn.Linear
|
| 576 |
+
if linear_cls is None:
|
| 577 |
+
linear_cls = nn.Linear
|
| 578 |
+
|
| 579 |
+
self.Wqkv = linear_cls(hidden_size, op_size, bias=bias, device=device, dtype=dtype)
|
| 580 |
+
self.out_proj = linear_cls(hidden_size, hidden_size, bias=bias, device=device, dtype=dtype)
|
| 581 |
+
|
| 582 |
+
# Attention
|
| 583 |
+
attn_cls = FlashSelfAttention if config.flash_attn else SelfAttention
|
| 584 |
+
if attn_cls is None:
|
| 585 |
+
attn_cls = SelfAttention
|
| 586 |
+
|
| 587 |
+
cross_attn_cls = FlashCrossAttention if config.flash_attn else CrossAttention
|
| 588 |
+
if cross_attn_cls is None:
|
| 589 |
+
cross_attn_cls = CrossAttention
|
| 590 |
+
|
| 591 |
+
self.inner_attn = attn_cls(
|
| 592 |
+
causal=causal,
|
| 593 |
+
softmax_scale=softmax_scale,
|
| 594 |
+
attention_dropout=config.attn_pdrop,
|
| 595 |
+
)
|
| 596 |
+
self.inner_cross_attn = cross_attn_cls(
|
| 597 |
+
causal=causal,
|
| 598 |
+
softmax_scale=softmax_scale,
|
| 599 |
+
attention_dropout=config.attn_pdrop,
|
| 600 |
+
)
|
| 601 |
+
|
| 602 |
+
self.flash_attn = config.flash_attn and attn_cls is FlashSelfAttention
|
| 603 |
+
self.layer_idx = layer_idx
|
| 604 |
+
self.return_residual = return_residual
|
| 605 |
+
self.checkpointing = checkpointing
|
| 606 |
+
|
| 607 |
+
def _forward_self_attn(
|
| 608 |
+
self, x: torch.FloatTensor, key_padding_mask: Optional[torch.BoolTensor]
|
| 609 |
+
) -> torch.FloatTensor:
|
| 610 |
+
qkv = self.Wqkv(x)
|
| 611 |
+
qkv = rearrange(qkv, "... (three h d) -> ... three h d", three=3, d=self.head_dim)
|
| 612 |
+
|
| 613 |
+
if self.rotary_dim > 0:
|
| 614 |
+
qkv = self.rotary_emb(qkv)
|
| 615 |
+
|
| 616 |
+
if self.flash_attn:
|
| 617 |
+
batch_size, seqlen = qkv.shape[0], qkv.shape[1]
|
| 618 |
+
|
| 619 |
+
cu_seqlens, max_seqlen = None, None
|
| 620 |
+
if key_padding_mask is not None:
|
| 621 |
+
# If `key_padding_mask` is supplied, we need to unpad the input and retrieve
|
| 622 |
+
# the `cu_seqlens` and `max_seqlen` to be used by `flash-attn`
|
| 623 |
+
qkv, indices, cu_seqlens, max_seqlen = unpad_input(qkv, key_padding_mask)
|
| 624 |
+
|
| 625 |
+
if self.checkpointing:
|
| 626 |
+
attn_output = torch.utils.checkpoint.checkpoint(
|
| 627 |
+
self.inner_attn, qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen
|
| 628 |
+
)
|
| 629 |
+
else:
|
| 630 |
+
attn_output = self.inner_attn(qkv, cu_seqlens=cu_seqlens, max_seqlen=max_seqlen).to(qkv.device)
|
| 631 |
+
|
| 632 |
+
# If `key_padding_mask` is supplied, we need to pad the output back to the original shape
|
| 633 |
+
return pad_input(attn_output, indices, batch_size, seqlen) if key_padding_mask is not None else attn_output
|
| 634 |
+
|
| 635 |
+
if self.checkpointing:
|
| 636 |
+
return torch.utils.checkpoint.checkpoint(self.inner_attn, qkv, key_padding_mask=key_padding_mask)
|
| 637 |
+
|
| 638 |
+
return self.inner_attn(qkv, key_padding_mask=key_padding_mask)
|
| 639 |
+
|
| 640 |
+
def _forward_cross_attn(
|
| 641 |
+
self,
|
| 642 |
+
x: torch.FloatTensor,
|
| 643 |
+
past_key_values: Optional[InferenceParams],
|
| 644 |
+
key_padding_mask: Optional[torch.BoolTensor],
|
| 645 |
+
) -> torch.FloatTensor:
|
| 646 |
+
batch_size = x.shape[0]
|
| 647 |
+
|
| 648 |
+
qkv = self.Wqkv(x)
|
| 649 |
+
|
| 650 |
+
q = qkv[..., : self.n_head * self.head_dim]
|
| 651 |
+
q = rearrange(q, "... (h d) -> ... h d", d=self.head_dim)
|
| 652 |
+
|
| 653 |
+
kv = qkv[..., self.n_head * self.head_dim :]
|
| 654 |
+
kv = rearrange(kv, "... (two hkv d) -> ... two hkv d", two=2, d=self.head_dim)
|
| 655 |
+
|
| 656 |
+
seqlen_offset = past_key_values.seqlen_offset if past_key_values is not None else 0
|
| 657 |
+
causal = None if seqlen_offset == 0 else False
|
| 658 |
+
if self.rotary_dim > 0:
|
| 659 |
+
q, kv = self.rotary_emb(q, kv=kv, seqlen_offset=seqlen_offset)
|
| 660 |
+
|
| 661 |
+
if past_key_values is not None:
|
| 662 |
+
kv = _update_kv_cache(kv, past_key_values, self.layer_idx)
|
| 663 |
+
|
| 664 |
+
if self.flash_attn:
|
| 665 |
+
batch_size, seqlen_q = q.shape[0], q.shape[1]
|
| 666 |
+
seqlen_k = kv.shape[1]
|
| 667 |
+
|
| 668 |
+
cu_seqlens_q, cu_seqlens_k, max_seqlen_q, max_seqlen_k = (
|
| 669 |
+
None,
|
| 670 |
+
None,
|
| 671 |
+
None,
|
| 672 |
+
None,
|
| 673 |
+
)
|
| 674 |
+
if key_padding_mask is not None:
|
| 675 |
+
kv, _, cu_seqlens_k, max_seqlen_k = unpad_input(kv, key_padding_mask)
|
| 676 |
+
|
| 677 |
+
if seqlen_q == 1:
|
| 678 |
+
key_padding_mask = torch.ones(batch_size, 1, device=q.device)
|
| 679 |
+
elif seqlen_q != seqlen_k:
|
| 680 |
+
key_padding_mask = key_padding_mask[:, -seqlen_q:]
|
| 681 |
+
|
| 682 |
+
q, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, key_padding_mask)
|
| 683 |
+
|
| 684 |
+
if self.checkpointing:
|
| 685 |
+
attn_output = torch.utils.checkpoint.checkpoint(
|
| 686 |
+
self.inner_cross_attn,
|
| 687 |
+
q,
|
| 688 |
+
kv,
|
| 689 |
+
causal=causal,
|
| 690 |
+
cu_seqlens=cu_seqlens_q,
|
| 691 |
+
max_seqlen=max_seqlen_q,
|
| 692 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 693 |
+
max_seqlen_k=max_seqlen_k,
|
| 694 |
+
)
|
| 695 |
+
else:
|
| 696 |
+
attn_output = self.inner_cross_attn(
|
| 697 |
+
q,
|
| 698 |
+
kv,
|
| 699 |
+
causal=causal,
|
| 700 |
+
cu_seqlens=cu_seqlens_q,
|
| 701 |
+
max_seqlen=max_seqlen_q,
|
| 702 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 703 |
+
max_seqlen_k=max_seqlen_k,
|
| 704 |
+
)
|
| 705 |
+
|
| 706 |
+
return (
|
| 707 |
+
pad_input(attn_output, indices_q, batch_size, max_seqlen_q)
|
| 708 |
+
if key_padding_mask is not None
|
| 709 |
+
else attn_output
|
| 710 |
+
)
|
| 711 |
+
|
| 712 |
+
if self.checkpointing:
|
| 713 |
+
return torch.utils.checkpoint.checkpoint(
|
| 714 |
+
self.inner_cross_attn,
|
| 715 |
+
q,
|
| 716 |
+
kv,
|
| 717 |
+
key_padding_mask=key_padding_mask,
|
| 718 |
+
causal=causal,
|
| 719 |
+
)
|
| 720 |
+
|
| 721 |
+
return self.inner_cross_attn(q, kv, key_padding_mask=key_padding_mask, causal=causal)
|
| 722 |
+
|
| 723 |
+
def forward(
|
| 724 |
+
self,
|
| 725 |
+
x: torch.FloatTensor,
|
| 726 |
+
past_key_values: Optional[InferenceParams] = None,
|
| 727 |
+
attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
|
| 728 |
+
**kwargs,
|
| 729 |
+
) -> Tuple[torch.FloatTensor, torch.FloatTensor]:
|
| 730 |
+
if attention_mask is not None:
|
| 731 |
+
attention_mask = attention_mask.bool()
|
| 732 |
+
else:
|
| 733 |
+
attention_mask = None
|
| 734 |
+
|
| 735 |
+
# MHA
|
| 736 |
+
if self.n_head == self.n_head_kv:
|
| 737 |
+
if past_key_values is None:
|
| 738 |
+
# If `past_key_values` are not supplied, we run self-attention
|
| 739 |
+
attn_output = self._forward_self_attn(x, attention_mask)
|
| 740 |
+
else:
|
| 741 |
+
# If `past_key_values` are supplied, it means that we might have cached values and
|
| 742 |
+
# could take advantage of cross-attention
|
| 743 |
+
attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
|
| 744 |
+
# MQA / GQA
|
| 745 |
+
else:
|
| 746 |
+
# Regardless of `past_key_values` being supplied or not, it always use cross-attention
|
| 747 |
+
# because `q` and `kv` lengths might be different
|
| 748 |
+
attn_output = self._forward_cross_attn(x, past_key_values, attention_mask)
|
| 749 |
+
|
| 750 |
+
output = rearrange(attn_output, "... h d -> ... (h d)")
|
| 751 |
+
output = self.out_proj(output)
|
| 752 |
+
|
| 753 |
+
return output if not self.return_residual else (output, x)
|
| 754 |
+
|
| 755 |
+
|
| 756 |
+
class ParallelBlock(nn.Module):
|
| 757 |
+
"""Parallel block.
|
| 758 |
+
|
| 759 |
+
This block applies parallel mixer and MLP layers to the input (used in GPT-J and CodeGen).
|
| 760 |
+
|
| 761 |
+
"""
|
| 762 |
+
|
| 763 |
+
def __init__(
|
| 764 |
+
self,
|
| 765 |
+
config: PretrainedConfig,
|
| 766 |
+
block_idx: Optional[int] = None,
|
| 767 |
+
) -> None:
|
| 768 |
+
super().__init__()
|
| 769 |
+
|
| 770 |
+
self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
| 771 |
+
self.resid_dropout = nn.Dropout(config.resid_pdrop)
|
| 772 |
+
self.block_idx = block_idx
|
| 773 |
+
|
| 774 |
+
self.mixer = MHA(config, layer_idx=block_idx)
|
| 775 |
+
self.mlp = MLP(config)
|
| 776 |
+
|
| 777 |
+
def forward(
|
| 778 |
+
self,
|
| 779 |
+
hidden_states: torch.FloatTensor,
|
| 780 |
+
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
| 781 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
| 782 |
+
**kwargs,
|
| 783 |
+
) -> torch.FloatTensor:
|
| 784 |
+
residual = hidden_states
|
| 785 |
+
hidden_states = self.ln(hidden_states)
|
| 786 |
+
|
| 787 |
+
attn_outputs = self.mixer(
|
| 788 |
+
hidden_states,
|
| 789 |
+
past_key_values=past_key_values,
|
| 790 |
+
attention_mask=attention_mask,
|
| 791 |
+
)
|
| 792 |
+
if isinstance(attn_outputs, tuple):
|
| 793 |
+
attn_outputs = attn_outputs[0]
|
| 794 |
+
|
| 795 |
+
attn_outputs = self.resid_dropout(attn_outputs)
|
| 796 |
+
feed_forward_hidden_states = self.resid_dropout(self.mlp(hidden_states))
|
| 797 |
+
|
| 798 |
+
hidden_states = attn_outputs + feed_forward_hidden_states + residual
|
| 799 |
+
|
| 800 |
+
return hidden_states
|
| 801 |
+
|
| 802 |
+
|
| 803 |
+
class CausalLMHead(nn.Module):
|
| 804 |
+
"""Causal Language Modeling head.
|
| 805 |
+
|
| 806 |
+
Reference:
|
| 807 |
+
Improving Language Understanding by Generative Pre-Training.
|
| 808 |
+
https://cdn.openai.com/research-covers/language-unsupervised/language_understanding_paper.pdf.
|
| 809 |
+
|
| 810 |
+
"""
|
| 811 |
+
|
| 812 |
+
def __init__(self, config: PretrainedConfig) -> None:
|
| 813 |
+
super().__init__()
|
| 814 |
+
|
| 815 |
+
self.ln = nn.LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
|
| 816 |
+
self.linear = nn.Linear(config.n_embd, config.vocab_size)
|
| 817 |
+
|
| 818 |
+
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
| 819 |
+
hidden_states = self.ln(hidden_states)
|
| 820 |
+
logits = self.linear(hidden_states).to(torch.float32)
|
| 821 |
+
|
| 822 |
+
return logits
|
| 823 |
+
|
| 824 |
+
|
| 825 |
+
class PhiPreTrainedModel(PreTrainedModel):
|
| 826 |
+
"""Phi pre-trained model."""
|
| 827 |
+
|
| 828 |
+
config_class = PhiConfig
|
| 829 |
+
base_model_prefix = "transformer"
|
| 830 |
+
supports_gradient_checkpointing = True
|
| 831 |
+
_no_split_modules = ["ParallelBlock", "CLIPEncoderLayer", "Block"]
|
| 832 |
+
|
| 833 |
+
def __init__(self, *inputs, **kwargs) -> None:
|
| 834 |
+
super().__init__(*inputs, **kwargs)
|
| 835 |
+
|
| 836 |
+
def _init_weights(self, module: nn.Module) -> None:
|
| 837 |
+
if isinstance(module, (nn.Linear,)):
|
| 838 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 839 |
+
if module.bias is not None:
|
| 840 |
+
module.bias.data.zero_()
|
| 841 |
+
elif isinstance(module, nn.Embedding):
|
| 842 |
+
module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
|
| 843 |
+
if module.padding_idx is not None:
|
| 844 |
+
module.weight.data[module.padding_idx].zero_()
|
| 845 |
+
elif isinstance(module, nn.LayerNorm):
|
| 846 |
+
if module.bias is not None:
|
| 847 |
+
module.bias.data.zero_()
|
| 848 |
+
module.weight.data.fill_(1.0)
|
| 849 |
+
|
| 850 |
+
def prepare_inputs_for_generation(
|
| 851 |
+
self,
|
| 852 |
+
input_ids: torch.LongTensor,
|
| 853 |
+
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
| 854 |
+
attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
|
| 855 |
+
**kwargs,
|
| 856 |
+
) -> Dict[str, Any]:
|
| 857 |
+
if past_key_values is None or not (isinstance(past_key_values, InferenceParams)):
|
| 858 |
+
past_key_values = InferenceParams(
|
| 859 |
+
max_seqlen=self.config.n_positions,
|
| 860 |
+
max_batch_size=input_ids.shape[0],
|
| 861 |
+
seqlen_offset=0,
|
| 862 |
+
batch_size_offset=0,
|
| 863 |
+
key_value_memory_dict={},
|
| 864 |
+
lengths_per_sample=None,
|
| 865 |
+
)
|
| 866 |
+
else:
|
| 867 |
+
# ======================================================================
|
| 868 |
+
# Assume that `past_key_values` has cached all tokens up to the last token in `input_ids`
|
| 869 |
+
# inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, sequence_start:sequence_end, ...]
|
| 870 |
+
# past_key_values.seqlen_offset = input_ids.shape[1] - 1
|
| 871 |
+
# ======================================================================
|
| 872 |
+
# I change the way of updating `past_key_values.seqlen_offset` to make the inference of imp work.
|
| 873 |
+
# [Edited by zhenwei - 2024-01-20 21:15]
|
| 874 |
+
input_ids = input_ids[:, -1].unsqueeze(-1)
|
| 875 |
+
|
| 876 |
+
return {
|
| 877 |
+
"input_ids": input_ids,
|
| 878 |
+
"past_key_values": past_key_values,
|
| 879 |
+
"attention_mask": attention_mask,
|
| 880 |
+
}
|
| 881 |
+
|
| 882 |
+
|
| 883 |
+
class LlavaMetaModel(ABC):
|
| 884 |
+
"""
|
| 885 |
+
Define the APIs for building components that are related to image perceiving.
|
| 886 |
+
This implementation is based on the implementation from the Llave project.
|
| 887 |
+
"""
|
| 888 |
+
|
| 889 |
+
def get_vision_tower(self):
|
| 890 |
+
vision_tower = getattr(self, 'vision_tower', None)
|
| 891 |
+
if type(vision_tower) is list:
|
| 892 |
+
vision_tower = vision_tower[0]
|
| 893 |
+
return vision_tower
|
| 894 |
+
|
| 895 |
+
def build_vision_tower(self, config):
|
| 896 |
+
self.vision_tower = VisionTower(config.vision_tower_cfg)
|
| 897 |
+
|
| 898 |
+
def build_vision_projector(self, config):
|
| 899 |
+
projector_type = getattr(config, 'mm_projector_type', 'linear')
|
| 900 |
+
|
| 901 |
+
if projector_type == 'linear':
|
| 902 |
+
self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)
|
| 903 |
+
return
|
| 904 |
+
|
| 905 |
+
mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
|
| 906 |
+
if mlp_gelu_match:
|
| 907 |
+
mlp_depth = int(mlp_gelu_match.group(1))
|
| 908 |
+
modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
|
| 909 |
+
for _ in range(1, mlp_depth):
|
| 910 |
+
modules.append(nn.GELU())
|
| 911 |
+
modules.append(nn.Linear(config.hidden_size, config.hidden_size))
|
| 912 |
+
self.mm_projector = nn.Sequential(*modules)
|
| 913 |
+
return
|
| 914 |
+
|
| 915 |
+
if projector_type == 'identity':
|
| 916 |
+
self.mm_projector = nn.Identity()
|
| 917 |
+
return
|
| 918 |
+
|
| 919 |
+
raise ValueError(f'Unknown projector type: {projector_type}')
|
| 920 |
+
|
| 921 |
+
|
| 922 |
+
class ImpModel(PhiPreTrainedModel, LlavaMetaModel):
|
| 923 |
+
"""Imp model. This implementation is modified from the implementation of Phi-2"""
|
| 924 |
+
|
| 925 |
+
config_class = ImpConfig
|
| 926 |
+
# _keys_to_ignore_on_load_missing = [""]
|
| 927 |
+
# _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
|
| 928 |
+
|
| 929 |
+
def __init__(self, config: ImpConfig) -> None:
|
| 930 |
+
super().__init__(config)
|
| 931 |
+
|
| 932 |
+
self.embd = Embedding(config)
|
| 933 |
+
self.h = nn.ModuleList([ParallelBlock(config, block_idx=i) for i in range(config.n_layer)])
|
| 934 |
+
self.gradient_checkpointing = False
|
| 935 |
+
|
| 936 |
+
if hasattr(config, "mm_vision_tower"):
|
| 937 |
+
self.build_vision_tower(config)
|
| 938 |
+
self.build_vision_projector(config)
|
| 939 |
+
|
| 940 |
+
self.post_init()
|
| 941 |
+
|
| 942 |
+
def embed_tokens(self, input_ids: torch.LongTensor) -> torch.FloatTensor:
|
| 943 |
+
return self.embd(input_ids)[0]
|
| 944 |
+
|
| 945 |
+
def get_input_embeddings(self) -> nn.Embedding:
|
| 946 |
+
return self.embd.wte
|
| 947 |
+
|
| 948 |
+
def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
|
| 949 |
+
self.embd.wte = new_embeddings
|
| 950 |
+
|
| 951 |
+
def forward(
|
| 952 |
+
self,
|
| 953 |
+
input_ids: torch.LongTensor,
|
| 954 |
+
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
| 955 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
| 956 |
+
inputs_embeds: Optional[torch.FloatTensor] = None
|
| 957 |
+
) -> torch.FloatTensor:
|
| 958 |
+
|
| 959 |
+
if inputs_embeds is None:
|
| 960 |
+
hidden_states = self.embd(input_ids)
|
| 961 |
+
else:
|
| 962 |
+
hidden_states = inputs_embeds
|
| 963 |
+
|
| 964 |
+
for layer in self.h:
|
| 965 |
+
if self.gradient_checkpointing and self.training:
|
| 966 |
+
|
| 967 |
+
def create_custom_forward(module):
|
| 968 |
+
def custom_forward(*inputs):
|
| 969 |
+
# None for past_key_value
|
| 970 |
+
return module(*inputs)
|
| 971 |
+
|
| 972 |
+
return custom_forward
|
| 973 |
+
|
| 974 |
+
hidden_states = torch.utils.checkpoint.checkpoint(
|
| 975 |
+
create_custom_forward(layer),
|
| 976 |
+
hidden_states,
|
| 977 |
+
None,
|
| 978 |
+
attention_mask,
|
| 979 |
+
)
|
| 980 |
+
else:
|
| 981 |
+
hidden_states = layer(
|
| 982 |
+
hidden_states,
|
| 983 |
+
past_key_values=past_key_values,
|
| 984 |
+
attention_mask=attention_mask,
|
| 985 |
+
)
|
| 986 |
+
|
| 987 |
+
# I change the way of updating `past_key_values.seqlen_offset` to make the inference of imp work.
|
| 988 |
+
# [Edited by zhenwei - 2024-01-20 21:15]
|
| 989 |
+
if past_key_values is not None: # FIXME: when multi-batch inference, it is a bug
|
| 990 |
+
past_key_values.seqlen_offset += hidden_states.shape[1]
|
| 991 |
+
|
| 992 |
+
return hidden_states
|
| 993 |
+
|
| 994 |
+
|
| 995 |
+
class LlavaMetaForCausalLM(ABC):
|
| 996 |
+
"""This implementation is based on the implementation from the Llave project."""
|
| 997 |
+
|
| 998 |
+
def init_constants(self, config):
|
| 999 |
+
self.IGNORE_INDEX = getattr(config, 'ignore_index', -100)
|
| 1000 |
+
self.IMAGE_TOKEN_INDEX = getattr(config, 'image_token_index', 50296)
|
| 1001 |
+
self.DEFAULT_IMAGE_TOKEN = getattr(config, 'image_token', "<image>")
|
| 1002 |
+
|
| 1003 |
+
@abstractmethod
|
| 1004 |
+
def get_model(self):
|
| 1005 |
+
pass
|
| 1006 |
+
|
| 1007 |
+
def get_vision_tower(self):
|
| 1008 |
+
return self.get_model().get_vision_tower()
|
| 1009 |
+
|
| 1010 |
+
def encode_images(self, images):
|
| 1011 |
+
image_features = self.get_model().get_vision_tower()(images)
|
| 1012 |
+
image_features = self.get_model().mm_projector(image_features)
|
| 1013 |
+
return image_features
|
| 1014 |
+
|
| 1015 |
+
def prepare_inputs_labels_for_multimodal(
|
| 1016 |
+
self, input_ids, position_ids, attention_mask, past_key_values, labels, images
|
| 1017 |
+
):
|
| 1018 |
+
vision_tower = self.get_vision_tower()
|
| 1019 |
+
# if vision_tower is None or images is None or past_key_values.seqlen_offset != 0:
|
| 1020 |
+
if vision_tower is None or images is None or input_ids.shape[1] == 1:
|
| 1021 |
+
if past_key_values is not None and vision_tower is not None and images is not None and input_ids.shape[1] == 1:
|
| 1022 |
+
target_shape = past_key_values.seqlen_offset + 1
|
| 1023 |
+
# inference_params.key_value_memory_dict[layer_idx][batch_start:batch_end, sequence_start:sequence_end, ...]
|
| 1024 |
+
attention_mask = torch.cat((attention_mask, torch.ones(
|
| 1025 |
+
(attention_mask.shape[0], target_shape - attention_mask.shape[1]),
|
| 1026 |
+
dtype=attention_mask.dtype,
|
| 1027 |
+
device=attention_mask.device
|
| 1028 |
+
)), dim=1)
|
| 1029 |
+
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
|
| 1030 |
+
return input_ids, position_ids, attention_mask, past_key_values, None, labels
|
| 1031 |
+
|
| 1032 |
+
if type(images) is list or images.ndim == 5:
|
| 1033 |
+
concat_images = torch.cat([image for image in images], dim=0)
|
| 1034 |
+
concat_images = concat_images.to(device=self.device, dtype=vision_tower.dtype)
|
| 1035 |
+
image_features = self.encode_images(concat_images)
|
| 1036 |
+
split_sizes = [image.shape[0] for image in images]
|
| 1037 |
+
image_features = torch.split(image_features, split_sizes, dim=0)
|
| 1038 |
+
image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
|
| 1039 |
+
else:
|
| 1040 |
+
images = images.to(device=self.device, dtype=vision_tower.dtype)
|
| 1041 |
+
image_features = self.encode_images(images).to(self.device)
|
| 1042 |
+
|
| 1043 |
+
# TODO: image start / end is not implemented here to support pretraining.
|
| 1044 |
+
if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
|
| 1045 |
+
raise NotImplementedError
|
| 1046 |
+
|
| 1047 |
+
# Let's just add dummy tensors if they do not exist,
|
| 1048 |
+
# it is a headache to deal with None all the time.
|
| 1049 |
+
# But it is not ideal, and if you have a better idea,
|
| 1050 |
+
# please open an issue / submit a PR, thanks.
|
| 1051 |
+
_labels = labels
|
| 1052 |
+
_position_ids = position_ids
|
| 1053 |
+
_attention_mask = attention_mask
|
| 1054 |
+
if attention_mask is None:
|
| 1055 |
+
attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
|
| 1056 |
+
else:
|
| 1057 |
+
attention_mask = attention_mask.bool()
|
| 1058 |
+
if position_ids is None:
|
| 1059 |
+
position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
|
| 1060 |
+
if labels is None:
|
| 1061 |
+
labels = torch.full_like(input_ids, self.IGNORE_INDEX)
|
| 1062 |
+
|
| 1063 |
+
# remove the padding using attention_mask -- TODO: double check
|
| 1064 |
+
input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
|
| 1065 |
+
labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
|
| 1066 |
+
|
| 1067 |
+
new_input_embeds = []
|
| 1068 |
+
new_labels = []
|
| 1069 |
+
cur_image_idx = 0
|
| 1070 |
+
for batch_idx, cur_input_ids in enumerate(input_ids):
|
| 1071 |
+
num_images = (cur_input_ids == self.IMAGE_TOKEN_INDEX).sum()
|
| 1072 |
+
if num_images == 0:
|
| 1073 |
+
cur_image_features = image_features[cur_image_idx]
|
| 1074 |
+
cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
|
| 1075 |
+
cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
|
| 1076 |
+
new_input_embeds.append(cur_input_embeds)
|
| 1077 |
+
new_labels.append(labels[batch_idx])
|
| 1078 |
+
cur_image_idx += 1
|
| 1079 |
+
continue
|
| 1080 |
+
|
| 1081 |
+
image_token_indices = [-1] + torch.where(cur_input_ids == self.IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
|
| 1082 |
+
cur_input_ids_noim = []
|
| 1083 |
+
cur_labels = labels[batch_idx]
|
| 1084 |
+
cur_labels_noim = []
|
| 1085 |
+
for i in range(len(image_token_indices) - 1):
|
| 1086 |
+
cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
|
| 1087 |
+
cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
|
| 1088 |
+
split_sizes = [x.shape[0] for x in cur_labels_noim]
|
| 1089 |
+
cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
|
| 1090 |
+
# print(cur_input_embeds.shape)
|
| 1091 |
+
cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
|
| 1092 |
+
cur_new_input_embeds = []
|
| 1093 |
+
cur_new_labels = []
|
| 1094 |
+
|
| 1095 |
+
for i in range(num_images + 1):
|
| 1096 |
+
cur_new_input_embeds.append(cur_input_embeds_no_im[i])
|
| 1097 |
+
cur_new_labels.append(cur_labels_noim[i])
|
| 1098 |
+
if i < num_images:
|
| 1099 |
+
cur_image_features = image_features[cur_image_idx]
|
| 1100 |
+
cur_image_idx += 1
|
| 1101 |
+
cur_new_input_embeds.append(cur_image_features)
|
| 1102 |
+
cur_new_labels.append(torch.full((cur_image_features.shape[0],), self.IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
|
| 1103 |
+
|
| 1104 |
+
cur_new_input_embeds = torch.cat(cur_new_input_embeds)
|
| 1105 |
+
cur_new_labels = torch.cat(cur_new_labels)
|
| 1106 |
+
|
| 1107 |
+
new_input_embeds.append(cur_new_input_embeds)
|
| 1108 |
+
new_labels.append(cur_new_labels)
|
| 1109 |
+
|
| 1110 |
+
# Truncate sequences to max length as image embeddings can make the sequence longer
|
| 1111 |
+
tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
|
| 1112 |
+
if tokenizer_model_max_length is not None:
|
| 1113 |
+
new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
|
| 1114 |
+
new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
|
| 1115 |
+
|
| 1116 |
+
# Combine them
|
| 1117 |
+
max_len = max(x.shape[0] for x in new_input_embeds)
|
| 1118 |
+
batch_size = len(new_input_embeds)
|
| 1119 |
+
|
| 1120 |
+
new_input_embeds_padded = []
|
| 1121 |
+
new_labels_padded = torch.full((batch_size, max_len), self.IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
|
| 1122 |
+
attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
|
| 1123 |
+
position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
|
| 1124 |
+
|
| 1125 |
+
for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
|
| 1126 |
+
cur_len = cur_new_embed.shape[0]
|
| 1127 |
+
if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
|
| 1128 |
+
new_input_embeds_padded.append(torch.cat((
|
| 1129 |
+
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
|
| 1130 |
+
cur_new_embed
|
| 1131 |
+
), dim=0))
|
| 1132 |
+
if cur_len > 0:
|
| 1133 |
+
new_labels_padded[i, -cur_len:] = cur_new_labels
|
| 1134 |
+
attention_mask[i, -cur_len:] = True
|
| 1135 |
+
position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
| 1136 |
+
else:
|
| 1137 |
+
new_input_embeds_padded.append(torch.cat((
|
| 1138 |
+
cur_new_embed,
|
| 1139 |
+
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
|
| 1140 |
+
), dim=0))
|
| 1141 |
+
if cur_len > 0:
|
| 1142 |
+
new_labels_padded[i, :cur_len] = cur_new_labels
|
| 1143 |
+
attention_mask[i, :cur_len] = True
|
| 1144 |
+
position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
| 1145 |
+
|
| 1146 |
+
new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
|
| 1147 |
+
|
| 1148 |
+
if _labels is None:
|
| 1149 |
+
new_labels = None
|
| 1150 |
+
else:
|
| 1151 |
+
new_labels = new_labels_padded
|
| 1152 |
+
|
| 1153 |
+
if _attention_mask is None:
|
| 1154 |
+
attention_mask = None
|
| 1155 |
+
else:
|
| 1156 |
+
attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
|
| 1157 |
+
|
| 1158 |
+
if _position_ids is None:
|
| 1159 |
+
position_ids = None
|
| 1160 |
+
|
| 1161 |
+
return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
|
| 1162 |
+
|
| 1163 |
+
|
| 1164 |
+
class ImpForCausalLM(PhiPreTrainedModel, LlavaMetaForCausalLM):
|
| 1165 |
+
"""Imp for Causal Language Modeling."""
|
| 1166 |
+
|
| 1167 |
+
# _keys_to_ignore_on_load_missing = [""]
|
| 1168 |
+
# _keys_to_ignore_on_load_unexpected = [r"transformer\.h\.\d+\.mlp.(fc_in|fc_out)\.(weight|bias)"]
|
| 1169 |
+
config_class = ImpConfig
|
| 1170 |
+
|
| 1171 |
+
def __init__(self, config: ImpConfig) -> None:
|
| 1172 |
+
super().__init__(config)
|
| 1173 |
+
|
| 1174 |
+
self.transformer = ImpModel(config)
|
| 1175 |
+
self.lm_head = CausalLMHead(config)
|
| 1176 |
+
|
| 1177 |
+
self.post_init()
|
| 1178 |
+
self.init_constants(config)
|
| 1179 |
+
|
| 1180 |
+
def get_output_embeddings(self) -> nn.Linear:
|
| 1181 |
+
return self.lm_head.linear
|
| 1182 |
+
|
| 1183 |
+
def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
|
| 1184 |
+
self.lm_head.linear = new_embeddings
|
| 1185 |
+
|
| 1186 |
+
def get_model(self):
|
| 1187 |
+
return self.transformer
|
| 1188 |
+
|
| 1189 |
+
def image_preprocess(self, images):
|
| 1190 |
+
return self.get_vision_tower().image_processor(images)['pixel_values']
|
| 1191 |
+
|
| 1192 |
+
def backbone_forward(
|
| 1193 |
+
self,
|
| 1194 |
+
input_ids: torch.LongTensor,
|
| 1195 |
+
past_key_values: Optional[Union[torch.FloatTensor, InferenceParams]] = None,
|
| 1196 |
+
attention_mask: Optional[torch.BoolTensor] = None,
|
| 1197 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1198 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1199 |
+
**kwargs,
|
| 1200 |
+
) -> CausalLMOutputWithPast:
|
| 1201 |
+
hidden_states = self.transformer(input_ids, past_key_values=past_key_values, attention_mask=attention_mask, inputs_embeds=inputs_embeds)
|
| 1202 |
+
lm_logits = self.lm_head(hidden_states)
|
| 1203 |
+
|
| 1204 |
+
return CausalLMOutputWithPast(loss=None, logits=lm_logits, past_key_values=past_key_values)
|
| 1205 |
+
|
| 1206 |
+
def forward(
|
| 1207 |
+
self,
|
| 1208 |
+
input_ids: torch.LongTensor = None,
|
| 1209 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1210 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1211 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1212 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1213 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1214 |
+
use_cache: Optional[bool] = None,
|
| 1215 |
+
output_attentions: Optional[bool] = None,
|
| 1216 |
+
output_hidden_states: Optional[bool] = None,
|
| 1217 |
+
images: Optional[torch.FloatTensor] = None,
|
| 1218 |
+
return_dict: Optional[bool] = None,
|
| 1219 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
| 1220 |
+
|
| 1221 |
+
if inputs_embeds is None:
|
| 1222 |
+
(
|
| 1223 |
+
input_ids,
|
| 1224 |
+
position_ids,
|
| 1225 |
+
attention_mask,
|
| 1226 |
+
past_key_values,
|
| 1227 |
+
inputs_embeds,
|
| 1228 |
+
labels
|
| 1229 |
+
) = self.prepare_inputs_labels_for_multimodal(
|
| 1230 |
+
input_ids,
|
| 1231 |
+
position_ids,
|
| 1232 |
+
attention_mask,
|
| 1233 |
+
past_key_values,
|
| 1234 |
+
labels,
|
| 1235 |
+
images
|
| 1236 |
+
)
|
| 1237 |
+
|
| 1238 |
+
return self.backbone_forward(
|
| 1239 |
+
input_ids=input_ids,
|
| 1240 |
+
attention_mask=attention_mask,
|
| 1241 |
+
position_ids=position_ids,
|
| 1242 |
+
past_key_values=past_key_values,
|
| 1243 |
+
inputs_embeds=inputs_embeds,
|
| 1244 |
+
labels=labels,
|
| 1245 |
+
use_cache=use_cache,
|
| 1246 |
+
output_attentions=output_attentions,
|
| 1247 |
+
output_hidden_states=output_hidden_states,
|
| 1248 |
+
return_dict=return_dict
|
| 1249 |
+
)
|
| 1250 |
+
|
| 1251 |
+
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
|
| 1252 |
+
images = kwargs.pop("images", None)
|
| 1253 |
+
_inputs = super().prepare_inputs_for_generation(
|
| 1254 |
+
input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
|
| 1255 |
+
)
|
| 1256 |
+
if images is not None:
|
| 1257 |
+
_inputs['images'] = images
|
| 1258 |
+
return _inputs
|
| 1259 |
+
|
| 1260 |
+
|
| 1261 |
+
AutoConfig.register("imp", ImpConfig)
|
| 1262 |
+
AutoModelForCausalLM.register(ImpConfig, ImpForCausalLM)
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": "<|endoftext|>",
|
| 3 |
+
"eos_token": "</s>",
|
| 4 |
+
"unk_token": "<|endoftext|>"
|
| 5 |
+
}
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,363 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"50296": {
|
| 6 |
+
"content": "<image>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"50295": {
|
| 14 |
+
"content": "</s>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"50256": {
|
| 22 |
+
"content": "<|endoftext|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"50257": {
|
| 30 |
+
"content": " ",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": true,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": false
|
| 36 |
+
},
|
| 37 |
+
"50258": {
|
| 38 |
+
"content": " ",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": true,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": false
|
| 44 |
+
},
|
| 45 |
+
"50259": {
|
| 46 |
+
"content": " ",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": true,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": false
|
| 52 |
+
},
|
| 53 |
+
"50260": {
|
| 54 |
+
"content": " ",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": true,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": false
|
| 60 |
+
},
|
| 61 |
+
"50261": {
|
| 62 |
+
"content": " ",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": true,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": false
|
| 68 |
+
},
|
| 69 |
+
"50262": {
|
| 70 |
+
"content": " ",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": true,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": false
|
| 76 |
+
},
|
| 77 |
+
"50263": {
|
| 78 |
+
"content": " ",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": true,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": false
|
| 84 |
+
},
|
| 85 |
+
"50264": {
|
| 86 |
+
"content": " ",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": true,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": false
|
| 92 |
+
},
|
| 93 |
+
"50265": {
|
| 94 |
+
"content": " ",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": true,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": false
|
| 100 |
+
},
|
| 101 |
+
"50266": {
|
| 102 |
+
"content": " ",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": true,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": false
|
| 108 |
+
},
|
| 109 |
+
"50267": {
|
| 110 |
+
"content": " ",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": true,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": false
|
| 116 |
+
},
|
| 117 |
+
"50268": {
|
| 118 |
+
"content": " ",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": true,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"50269": {
|
| 126 |
+
"content": " ",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": true,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"50270": {
|
| 134 |
+
"content": " ",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": true,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"50271": {
|
| 142 |
+
"content": " ",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": true,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"50272": {
|
| 150 |
+
"content": " ",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": true,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"50273": {
|
| 158 |
+
"content": " ",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": true,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"50274": {
|
| 166 |
+
"content": " ",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": true,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"50275": {
|
| 174 |
+
"content": " ",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": true,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"50276": {
|
| 182 |
+
"content": " ",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": true,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": false
|
| 188 |
+
},
|
| 189 |
+
"50277": {
|
| 190 |
+
"content": " ",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": true,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": false
|
| 196 |
+
},
|
| 197 |
+
"50278": {
|
| 198 |
+
"content": " ",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": true,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": false
|
| 204 |
+
},
|
| 205 |
+
"50279": {
|
| 206 |
+
"content": " ",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": true,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": false
|
| 212 |
+
},
|
| 213 |
+
"50280": {
|
| 214 |
+
"content": " ",
|
| 215 |
+
"lstrip": false,
|
| 216 |
+
"normalized": true,
|
| 217 |
+
"rstrip": false,
|
| 218 |
+
"single_word": false,
|
| 219 |
+
"special": false
|
| 220 |
+
},
|
| 221 |
+
"50281": {
|
| 222 |
+
"content": " ",
|
| 223 |
+
"lstrip": false,
|
| 224 |
+
"normalized": true,
|
| 225 |
+
"rstrip": false,
|
| 226 |
+
"single_word": false,
|
| 227 |
+
"special": false
|
| 228 |
+
},
|
| 229 |
+
"50282": {
|
| 230 |
+
"content": " ",
|
| 231 |
+
"lstrip": false,
|
| 232 |
+
"normalized": true,
|
| 233 |
+
"rstrip": false,
|
| 234 |
+
"single_word": false,
|
| 235 |
+
"special": false
|
| 236 |
+
},
|
| 237 |
+
"50283": {
|
| 238 |
+
"content": " ",
|
| 239 |
+
"lstrip": false,
|
| 240 |
+
"normalized": true,
|
| 241 |
+
"rstrip": false,
|
| 242 |
+
"single_word": false,
|
| 243 |
+
"special": false
|
| 244 |
+
},
|
| 245 |
+
"50284": {
|
| 246 |
+
"content": " ",
|
| 247 |
+
"lstrip": false,
|
| 248 |
+
"normalized": true,
|
| 249 |
+
"rstrip": false,
|
| 250 |
+
"single_word": false,
|
| 251 |
+
"special": false
|
| 252 |
+
},
|
| 253 |
+
"50285": {
|
| 254 |
+
"content": " ",
|
| 255 |
+
"lstrip": false,
|
| 256 |
+
"normalized": true,
|
| 257 |
+
"rstrip": false,
|
| 258 |
+
"single_word": false,
|
| 259 |
+
"special": false
|
| 260 |
+
},
|
| 261 |
+
"50286": {
|
| 262 |
+
"content": " ",
|
| 263 |
+
"lstrip": false,
|
| 264 |
+
"normalized": true,
|
| 265 |
+
"rstrip": false,
|
| 266 |
+
"single_word": false,
|
| 267 |
+
"special": false
|
| 268 |
+
},
|
| 269 |
+
"50287": {
|
| 270 |
+
"content": "\t\t\t\t\t\t\t\t\t",
|
| 271 |
+
"lstrip": false,
|
| 272 |
+
"normalized": true,
|
| 273 |
+
"rstrip": false,
|
| 274 |
+
"single_word": false,
|
| 275 |
+
"special": false
|
| 276 |
+
},
|
| 277 |
+
"50288": {
|
| 278 |
+
"content": "\t\t\t\t\t\t\t\t",
|
| 279 |
+
"lstrip": false,
|
| 280 |
+
"normalized": true,
|
| 281 |
+
"rstrip": false,
|
| 282 |
+
"single_word": false,
|
| 283 |
+
"special": false
|
| 284 |
+
},
|
| 285 |
+
"50289": {
|
| 286 |
+
"content": "\t\t\t\t\t\t\t",
|
| 287 |
+
"lstrip": false,
|
| 288 |
+
"normalized": true,
|
| 289 |
+
"rstrip": false,
|
| 290 |
+
"single_word": false,
|
| 291 |
+
"special": false
|
| 292 |
+
},
|
| 293 |
+
"50290": {
|
| 294 |
+
"content": "\t\t\t\t\t\t",
|
| 295 |
+
"lstrip": false,
|
| 296 |
+
"normalized": true,
|
| 297 |
+
"rstrip": false,
|
| 298 |
+
"single_word": false,
|
| 299 |
+
"special": false
|
| 300 |
+
},
|
| 301 |
+
"50291": {
|
| 302 |
+
"content": "\t\t\t\t\t",
|
| 303 |
+
"lstrip": false,
|
| 304 |
+
"normalized": true,
|
| 305 |
+
"rstrip": false,
|
| 306 |
+
"single_word": false,
|
| 307 |
+
"special": false
|
| 308 |
+
},
|
| 309 |
+
"50292": {
|
| 310 |
+
"content": "\t\t\t\t",
|
| 311 |
+
"lstrip": false,
|
| 312 |
+
"normalized": true,
|
| 313 |
+
"rstrip": false,
|
| 314 |
+
"single_word": false,
|
| 315 |
+
"special": false
|
| 316 |
+
},
|
| 317 |
+
"50293": {
|
| 318 |
+
"content": "\t\t\t",
|
| 319 |
+
"lstrip": false,
|
| 320 |
+
"normalized": true,
|
| 321 |
+
"rstrip": false,
|
| 322 |
+
"single_word": false,
|
| 323 |
+
"special": false
|
| 324 |
+
},
|
| 325 |
+
"50294": {
|
| 326 |
+
"content": "\t\t",
|
| 327 |
+
"lstrip": false,
|
| 328 |
+
"normalized": true,
|
| 329 |
+
"rstrip": false,
|
| 330 |
+
"single_word": false,
|
| 331 |
+
"special": false
|
| 332 |
+
}
|
| 333 |
+
},
|
| 334 |
+
"bos_token": {
|
| 335 |
+
"__type": "AddedToken",
|
| 336 |
+
"content": "<|endoftext|>",
|
| 337 |
+
"lstrip": false,
|
| 338 |
+
"normalized": true,
|
| 339 |
+
"rstrip": false,
|
| 340 |
+
"single_word": false
|
| 341 |
+
},
|
| 342 |
+
"clean_up_tokenization_spaces": true,
|
| 343 |
+
"eos_token": {
|
| 344 |
+
"__type": "AddedToken",
|
| 345 |
+
"content": "<|endoftext|>",
|
| 346 |
+
"lstrip": false,
|
| 347 |
+
"normalized": true,
|
| 348 |
+
"rstrip": false,
|
| 349 |
+
"single_word": false
|
| 350 |
+
},
|
| 351 |
+
"errors": "replace",
|
| 352 |
+
"model_max_length": 3072,
|
| 353 |
+
"pad_token": null,
|
| 354 |
+
"tokenizer_class": "CodeGenTokenizer",
|
| 355 |
+
"unk_token": {
|
| 356 |
+
"__type": "AddedToken",
|
| 357 |
+
"content": "<|endoftext|>",
|
| 358 |
+
"lstrip": false,
|
| 359 |
+
"normalized": true,
|
| 360 |
+
"rstrip": false,
|
| 361 |
+
"single_word": false
|
| 362 |
+
}
|
| 363 |
+
}
|
vision_encoder.py
ADDED
|
@@ -0,0 +1,613 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) MILVLG team.
|
| 2 |
+
# Licensed under the Apache 2.0 license.
|
| 3 |
+
#
|
| 4 |
+
# Some code here is copied from the project Phi-2 (https://huggingface.co/microsoft/phi-2),
|
| 5 |
+
# SigLIP@transformers==4.37.0.dev0 (https://huggingface.co/google/siglip-so400m-patch14-384),
|
| 6 |
+
# and Llava (https://github.com/haotian-liu/LLaVA), and modified by
|
| 7 |
+
# Zhenwei Shao (shaozw@hdu.edu.cn) @ MILVLG. We thank them for their great works.
|
| 8 |
+
# And their original licenses and copyright should be inherited (see the statements
|
| 9 |
+
# in `configuration_imp.py` for more details).
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
from typing import Any, Optional, Tuple, Union, List, Dict
|
| 13 |
+
from dataclasses import dataclass
|
| 14 |
+
import math
|
| 15 |
+
import warnings
|
| 16 |
+
from functools import partial, reduce
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
import numpy as np
|
| 20 |
+
from PIL import Image
|
| 21 |
+
import torch
|
| 22 |
+
import torch.utils.checkpoint
|
| 23 |
+
from torch import nn
|
| 24 |
+
|
| 25 |
+
from transformers.image_processing_utils import BatchFeature
|
| 26 |
+
from transformers.image_transforms import (
|
| 27 |
+
convert_to_rgb,
|
| 28 |
+
normalize,
|
| 29 |
+
rescale,
|
| 30 |
+
resize,
|
| 31 |
+
to_channel_dimension_format,
|
| 32 |
+
)
|
| 33 |
+
from transformers.image_utils import (
|
| 34 |
+
ChannelDimension,
|
| 35 |
+
PILImageResampling,
|
| 36 |
+
to_numpy_array,
|
| 37 |
+
)
|
| 38 |
+
from transformers.activations import ACT2FN
|
| 39 |
+
from transformers.modeling_outputs import BaseModelOutput, BaseModelOutputWithPooling
|
| 40 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 41 |
+
from transformers.utils import ModelOutput
|
| 42 |
+
|
| 43 |
+
from .configuration_imp import SiglipVisionConfig
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
# ============================================================================
|
| 47 |
+
# A simple image preprocessor for SigLIP models.
|
| 48 |
+
# ============================================================================
|
| 49 |
+
|
| 50 |
+
def expand2square(pil_img, background_color):
|
| 51 |
+
width, height = pil_img.size
|
| 52 |
+
if width == height:
|
| 53 |
+
return pil_img
|
| 54 |
+
elif width > height:
|
| 55 |
+
result = Image.new(pil_img.mode, (width, width), background_color)
|
| 56 |
+
result.paste(pil_img, (0, (width - height) // 2))
|
| 57 |
+
return result
|
| 58 |
+
else:
|
| 59 |
+
result = Image.new(pil_img.mode, (height, height), background_color)
|
| 60 |
+
result.paste(pil_img, ((height - width) // 2, 0))
|
| 61 |
+
return result
|
| 62 |
+
|
| 63 |
+
def simple_image_processor(
|
| 64 |
+
images,
|
| 65 |
+
image_mean=(0.5, 0.5, 0.5),
|
| 66 |
+
image_std=(0.5, 0.5, 0.5),
|
| 67 |
+
size=(384, 384),
|
| 68 |
+
resample=PILImageResampling.BICUBIC,
|
| 69 |
+
rescale_factor=1 / 255,
|
| 70 |
+
data_format=ChannelDimension.FIRST,
|
| 71 |
+
return_tensors="pt"
|
| 72 |
+
):
|
| 73 |
+
|
| 74 |
+
if isinstance(images, Image.Image):
|
| 75 |
+
images = [images]
|
| 76 |
+
else:
|
| 77 |
+
assert isinstance(images, list)
|
| 78 |
+
|
| 79 |
+
new_images = []
|
| 80 |
+
for image in images:
|
| 81 |
+
image = expand2square(image, tuple(int(x*255) for x in image_mean))
|
| 82 |
+
new_images.append(image)
|
| 83 |
+
images=new_images
|
| 84 |
+
|
| 85 |
+
transforms = [
|
| 86 |
+
convert_to_rgb,
|
| 87 |
+
to_numpy_array,
|
| 88 |
+
partial(resize, size=size, resample=resample, data_format=data_format),
|
| 89 |
+
partial(rescale, scale=rescale_factor, data_format=data_format),
|
| 90 |
+
partial(normalize, mean=image_mean, std=image_std, data_format=data_format),
|
| 91 |
+
partial(to_channel_dimension_format, channel_dim=data_format, input_channel_dim=data_format),
|
| 92 |
+
]
|
| 93 |
+
|
| 94 |
+
images = reduce(lambda x, f: [*map(f, x)], transforms, images)
|
| 95 |
+
data = {"pixel_values": images}
|
| 96 |
+
|
| 97 |
+
return BatchFeature(data=data, tensor_type=return_tensors)
|
| 98 |
+
|
| 99 |
+
# ============================================================================
|
| 100 |
+
# Definitions for SigLIP models.
|
| 101 |
+
# ============================================================================
|
| 102 |
+
|
| 103 |
+
@dataclass
|
| 104 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPVisionModelOutput with CLIP->Siglip
|
| 105 |
+
class SiglipVisionModelOutput(ModelOutput):
|
| 106 |
+
"""
|
| 107 |
+
Base class for vision model's outputs that also contains image embeddings of the pooling of the last hidden states.
|
| 108 |
+
|
| 109 |
+
Args:
|
| 110 |
+
image_embeds (`torch.FloatTensor` of shape `(batch_size, output_dim)` *optional* returned when model is initialized with `with_projection=True`):
|
| 111 |
+
The image embeddings obtained by applying the projection layer to the pooler_output.
|
| 112 |
+
last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 113 |
+
Sequence of hidden-states at the output of the last layer of the model.
|
| 114 |
+
hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
|
| 115 |
+
Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
|
| 116 |
+
one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
|
| 117 |
+
|
| 118 |
+
Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
|
| 119 |
+
attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
|
| 120 |
+
Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
|
| 121 |
+
sequence_length)`.
|
| 122 |
+
|
| 123 |
+
Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
|
| 124 |
+
heads.
|
| 125 |
+
"""
|
| 126 |
+
|
| 127 |
+
image_embeds: Optional[torch.FloatTensor] = None
|
| 128 |
+
last_hidden_state: torch.FloatTensor = None
|
| 129 |
+
hidden_states: Optional[Tuple[torch.FloatTensor]] = None
|
| 130 |
+
attentions: Optional[Tuple[torch.FloatTensor]] = None
|
| 131 |
+
|
| 132 |
+
|
| 133 |
+
class SiglipVisionEmbeddings(nn.Module):
|
| 134 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 135 |
+
super().__init__()
|
| 136 |
+
self.config = config
|
| 137 |
+
self.embed_dim = config.hidden_size
|
| 138 |
+
self.image_size = config.image_size
|
| 139 |
+
self.patch_size = config.patch_size
|
| 140 |
+
|
| 141 |
+
self.patch_embedding = nn.Conv2d(
|
| 142 |
+
in_channels=config.num_channels,
|
| 143 |
+
out_channels=self.embed_dim,
|
| 144 |
+
kernel_size=self.patch_size,
|
| 145 |
+
stride=self.patch_size,
|
| 146 |
+
padding="valid",
|
| 147 |
+
)
|
| 148 |
+
|
| 149 |
+
self.num_patches = (self.image_size // self.patch_size) ** 2
|
| 150 |
+
self.num_positions = self.num_patches
|
| 151 |
+
self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
|
| 152 |
+
self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False)
|
| 153 |
+
|
| 154 |
+
def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor:
|
| 155 |
+
patch_embeds = self.patch_embedding(pixel_values) # shape = [*, width, grid, grid]
|
| 156 |
+
embeddings = patch_embeds.flatten(2).transpose(1, 2)
|
| 157 |
+
|
| 158 |
+
embeddings = embeddings + self.position_embedding(self.position_ids)
|
| 159 |
+
return embeddings
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
|
| 163 |
+
class SiglipAttention(nn.Module):
|
| 164 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 165 |
+
|
| 166 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
|
| 167 |
+
def __init__(self, config):
|
| 168 |
+
super().__init__()
|
| 169 |
+
self.config = config
|
| 170 |
+
self.embed_dim = config.hidden_size
|
| 171 |
+
self.num_heads = config.num_attention_heads
|
| 172 |
+
self.head_dim = self.embed_dim // self.num_heads
|
| 173 |
+
if self.head_dim * self.num_heads != self.embed_dim:
|
| 174 |
+
raise ValueError(
|
| 175 |
+
f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
|
| 176 |
+
f" {self.num_heads})."
|
| 177 |
+
)
|
| 178 |
+
self.scale = self.head_dim**-0.5
|
| 179 |
+
self.dropout = config.attention_dropout
|
| 180 |
+
|
| 181 |
+
self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 182 |
+
self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 183 |
+
self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 184 |
+
self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
|
| 185 |
+
|
| 186 |
+
def forward(
|
| 187 |
+
self,
|
| 188 |
+
hidden_states: torch.Tensor,
|
| 189 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 190 |
+
output_attentions: Optional[bool] = False,
|
| 191 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 192 |
+
"""Input shape: Batch x Time x Channel"""
|
| 193 |
+
|
| 194 |
+
batch_size, q_len, _ = hidden_states.size()
|
| 195 |
+
|
| 196 |
+
query_states = self.q_proj(hidden_states)
|
| 197 |
+
key_states = self.k_proj(hidden_states)
|
| 198 |
+
value_states = self.v_proj(hidden_states)
|
| 199 |
+
|
| 200 |
+
query_states = query_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 201 |
+
key_states = key_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 202 |
+
value_states = value_states.view(batch_size, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 203 |
+
|
| 204 |
+
k_v_seq_len = key_states.shape[-2]
|
| 205 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) * self.scale
|
| 206 |
+
|
| 207 |
+
if attn_weights.size() != (batch_size, self.num_heads, q_len, k_v_seq_len):
|
| 208 |
+
raise ValueError(
|
| 209 |
+
f"Attention weights should be of size {(batch_size, self.num_heads, q_len, k_v_seq_len)}, but is"
|
| 210 |
+
f" {attn_weights.size()}"
|
| 211 |
+
)
|
| 212 |
+
|
| 213 |
+
if attention_mask is not None:
|
| 214 |
+
if attention_mask.size() != (batch_size, 1, q_len, k_v_seq_len):
|
| 215 |
+
raise ValueError(
|
| 216 |
+
f"Attention mask should be of size {(batch_size, 1, q_len, k_v_seq_len)}, but is {attention_mask.size()}"
|
| 217 |
+
)
|
| 218 |
+
attn_weights = attn_weights + attention_mask
|
| 219 |
+
|
| 220 |
+
# upcast attention to fp32
|
| 221 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
|
| 222 |
+
attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
|
| 223 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 224 |
+
|
| 225 |
+
if attn_output.size() != (batch_size, self.num_heads, q_len, self.head_dim):
|
| 226 |
+
raise ValueError(
|
| 227 |
+
f"`attn_output` should be of size {(batch_size, self.num_heads, q_len, self.head_dim)}, but is"
|
| 228 |
+
f" {attn_output.size()}"
|
| 229 |
+
)
|
| 230 |
+
|
| 231 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 232 |
+
attn_output = attn_output.reshape(batch_size, q_len, self.embed_dim)
|
| 233 |
+
|
| 234 |
+
attn_output = self.out_proj(attn_output)
|
| 235 |
+
|
| 236 |
+
return attn_output, attn_weights
|
| 237 |
+
|
| 238 |
+
|
| 239 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Siglip
|
| 240 |
+
class SiglipMLP(nn.Module):
|
| 241 |
+
def __init__(self, config):
|
| 242 |
+
super().__init__()
|
| 243 |
+
self.config = config
|
| 244 |
+
self.activation_fn = ACT2FN[config.hidden_act]
|
| 245 |
+
self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
|
| 246 |
+
self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
|
| 247 |
+
|
| 248 |
+
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
| 249 |
+
hidden_states = self.fc1(hidden_states)
|
| 250 |
+
hidden_states = self.activation_fn(hidden_states)
|
| 251 |
+
hidden_states = self.fc2(hidden_states)
|
| 252 |
+
return hidden_states
|
| 253 |
+
|
| 254 |
+
|
| 255 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPEncoderLayer with CLIP->Siglip
|
| 256 |
+
class SiglipEncoderLayer(nn.Module):
|
| 257 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 258 |
+
super().__init__()
|
| 259 |
+
self.embed_dim = config.hidden_size
|
| 260 |
+
self.self_attn = SiglipAttention(config)
|
| 261 |
+
self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
| 262 |
+
self.mlp = SiglipMLP(config)
|
| 263 |
+
self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
|
| 264 |
+
|
| 265 |
+
# Ignore copy
|
| 266 |
+
def forward(
|
| 267 |
+
self,
|
| 268 |
+
hidden_states: torch.Tensor,
|
| 269 |
+
attention_mask: torch.Tensor,
|
| 270 |
+
output_attentions: Optional[bool] = False,
|
| 271 |
+
) -> Tuple[torch.FloatTensor]:
|
| 272 |
+
"""
|
| 273 |
+
Args:
|
| 274 |
+
hidden_states (`torch.FloatTensor`):
|
| 275 |
+
Input to the layer of shape `(batch, seq_len, embed_dim)`.
|
| 276 |
+
attention_mask (`torch.FloatTensor`):
|
| 277 |
+
Attention mask of shape `(batch, 1, q_len, k_v_seq_len)` where padding elements are indicated by very large negative values.
|
| 278 |
+
output_attentions (`bool`, *optional*, defaults to `False`):
|
| 279 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 280 |
+
returned tensors for more detail.
|
| 281 |
+
"""
|
| 282 |
+
residual = hidden_states
|
| 283 |
+
|
| 284 |
+
hidden_states = self.layer_norm1(hidden_states)
|
| 285 |
+
hidden_states, attn_weights = self.self_attn(
|
| 286 |
+
hidden_states=hidden_states,
|
| 287 |
+
attention_mask=attention_mask,
|
| 288 |
+
output_attentions=output_attentions,
|
| 289 |
+
)
|
| 290 |
+
hidden_states = residual + hidden_states
|
| 291 |
+
|
| 292 |
+
residual = hidden_states
|
| 293 |
+
hidden_states = self.layer_norm2(hidden_states)
|
| 294 |
+
hidden_states = self.mlp(hidden_states)
|
| 295 |
+
hidden_states = residual + hidden_states
|
| 296 |
+
|
| 297 |
+
outputs = (hidden_states,)
|
| 298 |
+
|
| 299 |
+
if output_attentions:
|
| 300 |
+
outputs += (attn_weights,)
|
| 301 |
+
|
| 302 |
+
return outputs
|
| 303 |
+
|
| 304 |
+
|
| 305 |
+
class SiglipPreTrainedModel(PreTrainedModel):
|
| 306 |
+
"""
|
| 307 |
+
An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
|
| 308 |
+
models.
|
| 309 |
+
"""
|
| 310 |
+
|
| 311 |
+
config_class = SiglipVisionConfig
|
| 312 |
+
base_model_prefix = "siglip"
|
| 313 |
+
supports_gradient_checkpointing = True
|
| 314 |
+
|
| 315 |
+
def _init_weights(self, module):
|
| 316 |
+
"""Initialize the weights"""
|
| 317 |
+
pass
|
| 318 |
+
|
| 319 |
+
# Copied from transformers.models.clip.modeling_clip.CLIPEncoder with CLIP->Siglip
|
| 320 |
+
class SiglipEncoder(nn.Module):
|
| 321 |
+
"""
|
| 322 |
+
Transformer encoder consisting of `config.num_hidden_layers` self attention layers. Each layer is a
|
| 323 |
+
[`SiglipEncoderLayer`].
|
| 324 |
+
|
| 325 |
+
Args:
|
| 326 |
+
config: SiglipVisionConfig
|
| 327 |
+
"""
|
| 328 |
+
|
| 329 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 330 |
+
super().__init__()
|
| 331 |
+
self.config = config
|
| 332 |
+
self.layers = nn.ModuleList([SiglipEncoderLayer(config) for _ in range(config.num_hidden_layers)])
|
| 333 |
+
self.gradient_checkpointing = False
|
| 334 |
+
|
| 335 |
+
# Ignore copy
|
| 336 |
+
def forward(
|
| 337 |
+
self,
|
| 338 |
+
inputs_embeds,
|
| 339 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 340 |
+
output_attentions: Optional[bool] = None,
|
| 341 |
+
output_hidden_states: Optional[bool] = None,
|
| 342 |
+
return_dict: Optional[bool] = None,
|
| 343 |
+
) -> Union[Tuple, BaseModelOutput]:
|
| 344 |
+
r"""
|
| 345 |
+
Args:
|
| 346 |
+
inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`):
|
| 347 |
+
Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation.
|
| 348 |
+
This is useful if you want more control over how to convert `input_ids` indices into associated vectors
|
| 349 |
+
than the model's internal embedding lookup matrix.
|
| 350 |
+
attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
|
| 351 |
+
Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
|
| 352 |
+
|
| 353 |
+
- 1 for tokens that are **not masked**,
|
| 354 |
+
- 0 for tokens that are **masked**.
|
| 355 |
+
|
| 356 |
+
[What are attention masks?](../glossary#attention-mask)
|
| 357 |
+
output_attentions (`bool`, *optional*):
|
| 358 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 359 |
+
returned tensors for more detail.
|
| 360 |
+
output_hidden_states (`bool`, *optional*):
|
| 361 |
+
Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
|
| 362 |
+
for more detail.
|
| 363 |
+
return_dict (`bool`, *optional*):
|
| 364 |
+
Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
|
| 365 |
+
"""
|
| 366 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 367 |
+
output_hidden_states = (
|
| 368 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 369 |
+
)
|
| 370 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 371 |
+
|
| 372 |
+
encoder_states = () if output_hidden_states else None
|
| 373 |
+
all_attentions = () if output_attentions else None
|
| 374 |
+
|
| 375 |
+
hidden_states = inputs_embeds
|
| 376 |
+
for encoder_layer in self.layers:
|
| 377 |
+
if output_hidden_states:
|
| 378 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 379 |
+
if self.gradient_checkpointing and self.training:
|
| 380 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 381 |
+
encoder_layer.__call__,
|
| 382 |
+
hidden_states,
|
| 383 |
+
attention_mask,
|
| 384 |
+
output_attentions,
|
| 385 |
+
)
|
| 386 |
+
else:
|
| 387 |
+
layer_outputs = encoder_layer(
|
| 388 |
+
hidden_states,
|
| 389 |
+
attention_mask,
|
| 390 |
+
output_attentions=output_attentions,
|
| 391 |
+
)
|
| 392 |
+
|
| 393 |
+
hidden_states = layer_outputs[0]
|
| 394 |
+
|
| 395 |
+
if output_attentions:
|
| 396 |
+
all_attentions = all_attentions + (layer_outputs[1],)
|
| 397 |
+
|
| 398 |
+
if output_hidden_states:
|
| 399 |
+
encoder_states = encoder_states + (hidden_states,)
|
| 400 |
+
|
| 401 |
+
if not return_dict:
|
| 402 |
+
return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
|
| 403 |
+
return BaseModelOutput(
|
| 404 |
+
last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
|
| 405 |
+
)
|
| 406 |
+
|
| 407 |
+
|
| 408 |
+
class SiglipVisionTransformer(nn.Module):
|
| 409 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 410 |
+
super().__init__()
|
| 411 |
+
self.config = config
|
| 412 |
+
embed_dim = config.hidden_size
|
| 413 |
+
|
| 414 |
+
self.embeddings = SiglipVisionEmbeddings(config)
|
| 415 |
+
self.encoder = SiglipEncoder(config)
|
| 416 |
+
self.post_layernorm = nn.LayerNorm(embed_dim, eps=config.layer_norm_eps)
|
| 417 |
+
self.head = SiglipMultiheadAttentionPoolingHead(config)
|
| 418 |
+
|
| 419 |
+
def forward(
|
| 420 |
+
self,
|
| 421 |
+
pixel_values,
|
| 422 |
+
output_attentions: Optional[bool] = None,
|
| 423 |
+
output_hidden_states: Optional[bool] = None,
|
| 424 |
+
return_dict: Optional[bool] = None,
|
| 425 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 426 |
+
r"""
|
| 427 |
+
Returns:
|
| 428 |
+
|
| 429 |
+
"""
|
| 430 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 431 |
+
output_hidden_states = (
|
| 432 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 433 |
+
)
|
| 434 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 435 |
+
|
| 436 |
+
hidden_states = self.embeddings(pixel_values)
|
| 437 |
+
|
| 438 |
+
encoder_outputs = self.encoder(
|
| 439 |
+
inputs_embeds=hidden_states,
|
| 440 |
+
output_attentions=output_attentions,
|
| 441 |
+
output_hidden_states=output_hidden_states,
|
| 442 |
+
return_dict=return_dict,
|
| 443 |
+
)
|
| 444 |
+
|
| 445 |
+
last_hidden_state = encoder_outputs[0]
|
| 446 |
+
last_hidden_state = self.post_layernorm(last_hidden_state)
|
| 447 |
+
|
| 448 |
+
pooled_output = self.head(last_hidden_state)
|
| 449 |
+
|
| 450 |
+
if not return_dict:
|
| 451 |
+
return (last_hidden_state, pooled_output) + encoder_outputs[1:]
|
| 452 |
+
|
| 453 |
+
return BaseModelOutputWithPooling(
|
| 454 |
+
last_hidden_state=last_hidden_state,
|
| 455 |
+
pooler_output=pooled_output,
|
| 456 |
+
hidden_states=encoder_outputs.hidden_states,
|
| 457 |
+
attentions=encoder_outputs.attentions,
|
| 458 |
+
)
|
| 459 |
+
|
| 460 |
+
|
| 461 |
+
class SiglipMultiheadAttentionPoolingHead(nn.Module):
|
| 462 |
+
"""Multihead Attention Pooling."""
|
| 463 |
+
|
| 464 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 465 |
+
super().__init__()
|
| 466 |
+
|
| 467 |
+
self.probe = nn.Parameter(torch.randn(1, 1, config.hidden_size))
|
| 468 |
+
self.attention = torch.nn.MultiheadAttention(config.hidden_size, config.num_attention_heads, batch_first=True)
|
| 469 |
+
self.layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
|
| 470 |
+
self.mlp = SiglipMLP(config)
|
| 471 |
+
|
| 472 |
+
def forward(self, hidden_state):
|
| 473 |
+
batch_size = hidden_state.shape[0]
|
| 474 |
+
probe = self.probe.repeat(batch_size, 1, 1)
|
| 475 |
+
|
| 476 |
+
hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
|
| 477 |
+
|
| 478 |
+
residual = hidden_state
|
| 479 |
+
hidden_state = self.layernorm(hidden_state)
|
| 480 |
+
hidden_state = residual + self.mlp(hidden_state)
|
| 481 |
+
|
| 482 |
+
return hidden_state[:, 0]
|
| 483 |
+
|
| 484 |
+
|
| 485 |
+
class SiglipVisionModel(SiglipPreTrainedModel):
|
| 486 |
+
config_class = SiglipVisionConfig
|
| 487 |
+
main_input_name = "pixel_values"
|
| 488 |
+
_no_split_modules = ["SiglipEncoderLayer"]
|
| 489 |
+
|
| 490 |
+
def __init__(self, config: SiglipVisionConfig):
|
| 491 |
+
super().__init__(config)
|
| 492 |
+
|
| 493 |
+
self.vision_model = SiglipVisionTransformer(config)
|
| 494 |
+
|
| 495 |
+
# Initialize weights and apply final processing
|
| 496 |
+
self.post_init()
|
| 497 |
+
|
| 498 |
+
def get_input_embeddings(self) -> nn.Module:
|
| 499 |
+
return self.vision_model.embeddings.patch_embedding
|
| 500 |
+
|
| 501 |
+
def forward(
|
| 502 |
+
self,
|
| 503 |
+
pixel_values,
|
| 504 |
+
output_attentions: Optional[bool] = None,
|
| 505 |
+
output_hidden_states: Optional[bool] = None,
|
| 506 |
+
return_dict: Optional[bool] = None,
|
| 507 |
+
) -> Union[Tuple, BaseModelOutputWithPooling]:
|
| 508 |
+
r"""
|
| 509 |
+
Returns:
|
| 510 |
+
|
| 511 |
+
Examples:
|
| 512 |
+
|
| 513 |
+
```python
|
| 514 |
+
>>> from PIL import Image
|
| 515 |
+
>>> import requests
|
| 516 |
+
>>> from transformers import AutoProcessor, SiglipVisionModel
|
| 517 |
+
|
| 518 |
+
>>> model = SiglipVisionModel.from_pretrained("google/siglip-base-patch16-224")
|
| 519 |
+
>>> processor = AutoProcessor.from_pretrained("google/siglip-base-patch16-224")
|
| 520 |
+
|
| 521 |
+
>>> url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 522 |
+
>>> image = Image.open(requests.get(url, stream=True).raw)
|
| 523 |
+
|
| 524 |
+
>>> inputs = processor(images=image, return_tensors="pt")
|
| 525 |
+
|
| 526 |
+
>>> outputs = model(**inputs)
|
| 527 |
+
>>> last_hidden_state = outputs.last_hidden_state
|
| 528 |
+
>>> pooled_output = outputs.pooler_output # pooled features
|
| 529 |
+
```"""
|
| 530 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 531 |
+
|
| 532 |
+
return self.vision_model(
|
| 533 |
+
pixel_values=pixel_values,
|
| 534 |
+
output_attentions=output_attentions,
|
| 535 |
+
output_hidden_states=output_hidden_states,
|
| 536 |
+
return_dict=return_dict,
|
| 537 |
+
)
|
| 538 |
+
|
| 539 |
+
|
| 540 |
+
# ============================================================================
|
| 541 |
+
# VisionTower module for Imp
|
| 542 |
+
# ============================================================================
|
| 543 |
+
|
| 544 |
+
class VisionTower(nn.Module):
|
| 545 |
+
def __init__(self, vision_tower_cfg, delay_load=False):
|
| 546 |
+
super().__init__()
|
| 547 |
+
|
| 548 |
+
self.is_loaded = False
|
| 549 |
+
|
| 550 |
+
self.config = vision_tower_cfg
|
| 551 |
+
self.vision_tower_name = vision_tower_cfg.mm_vision_tower
|
| 552 |
+
self.select_layer = vision_tower_cfg.mm_vision_select_layer
|
| 553 |
+
# self.select_feature = getattr(vision_tower_cfg, 'mm_vision_select_feature', 'patch')
|
| 554 |
+
|
| 555 |
+
self.image_processor = simple_image_processor
|
| 556 |
+
|
| 557 |
+
if not delay_load:
|
| 558 |
+
self.load_model()
|
| 559 |
+
else:
|
| 560 |
+
raise NotImplementedError("delay load is not implemented yet.")
|
| 561 |
+
|
| 562 |
+
def load_model(self):
|
| 563 |
+
if self.is_loaded:
|
| 564 |
+
return
|
| 565 |
+
|
| 566 |
+
# "google/siglip-so400m-patch14-384"
|
| 567 |
+
# self.vision_tower = SiglipVisionModel.from_pretrained(self.vision_tower_name)
|
| 568 |
+
self.vision_tower = SiglipVisionModel(self.config)
|
| 569 |
+
del self.vision_tower.vision_model.encoder.layers[(self.select_layer + 1):]
|
| 570 |
+
self.vision_tower.vision_model.head = nn.Identity()
|
| 571 |
+
self.vision_tower.vision_model.post_layernorm=nn.Identity()
|
| 572 |
+
self.vision_tower.requires_grad_(False)
|
| 573 |
+
self.vision_tower.eval()
|
| 574 |
+
|
| 575 |
+
self.is_loaded = True
|
| 576 |
+
|
| 577 |
+
@torch.no_grad()
|
| 578 |
+
def forward(self, images):
|
| 579 |
+
if type(images) is list:
|
| 580 |
+
image_features = []
|
| 581 |
+
for image in images:
|
| 582 |
+
image_forward_out = self.vision_tower(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
|
| 583 |
+
image_feature = image_forward_out.hidden_states[-1].to(image.dtype)
|
| 584 |
+
assert image_features.shape[-2] == 729
|
| 585 |
+
image_features.append(image_feature)
|
| 586 |
+
else:
|
| 587 |
+
image_forward_outs = self.vision_tower(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
|
| 588 |
+
image_features = image_forward_outs.hidden_states[-1].to(images.dtype)
|
| 589 |
+
assert image_features.shape[-2] == 729
|
| 590 |
+
|
| 591 |
+
return image_features
|
| 592 |
+
|
| 593 |
+
@property
|
| 594 |
+
def dummy_feature(self):
|
| 595 |
+
return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
|
| 596 |
+
|
| 597 |
+
@property
|
| 598 |
+
def dtype(self):
|
| 599 |
+
for p in self.vision_tower.parameters():
|
| 600 |
+
return p.dtype
|
| 601 |
+
|
| 602 |
+
@property
|
| 603 |
+
def device(self):
|
| 604 |
+
for p in self.vision_tower.parameters():
|
| 605 |
+
return p.device
|
| 606 |
+
|
| 607 |
+
@property
|
| 608 |
+
def hidden_size(self):
|
| 609 |
+
return self.config.hidden_size
|
| 610 |
+
|
| 611 |
+
@property
|
| 612 |
+
def num_patches(self):
|
| 613 |
+
return (self.config.image_size // self.config.patch_size) ** 2
|
vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|