Imp-v1.5-4b update
Browse files- LICENSE +201 -0
- README copy.md +96 -0
- added_tokens.json +14 -0
- config.json +79 -0
- configuration_imp.py +313 -0
- generation_config.json +12 -0
- merges.txt +0 -0
- model-00001-of-00009.safetensors +3 -0
- model-00002-of-00009.safetensors +3 -0
- model-00003-of-00009.safetensors +3 -0
- model-00004-of-00009.safetensors +3 -0
- model-00005-of-00009.safetensors +3 -0
- model-00006-of-00009.safetensors +3 -0
- model-00007-of-00009.safetensors +3 -0
- model-00008-of-00009.safetensors +3 -0
- model-00009-of-00009.safetensors +3 -0
- model.safetensors.index.json +627 -0
- modeling_imp.py +1521 -0
- special_tokens_map.json +30 -0
- tokenizer.json +0 -0
- tokenizer_config.json +138 -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 copy.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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-v1.5-4b` is a strong MSLM with only **4B** parameters, which is build upon a small yet powerful SLM [Phi-3 ](https://huggingface.co/microsoft/Phi-3-mini-128k-instruct)(3.8B) and a powerful visual encoder [SigLIP ](https://huggingface.co/google/siglip-so400m-patch14-384)(0.4B), and trained on 1M mixed dataset.
|
| 20 |
+
|
| 21 |
+
As shown in the Table below, `imp-v1.5-4b` significantly outperforms the counterparts of similar model sizes on various multimodal benchmarks.
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
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 :)
|
| 25 |
+
|
| 26 |
+
|
| 27 |
+
## How to use
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
**Install dependencies**
|
| 31 |
+
```bash
|
| 32 |
+
pip install transformers # latest version is ok, but we recommend v4.36.0
|
| 33 |
+
pip install -q pillow accelerate einops
|
| 34 |
+
```
|
| 35 |
+
|
| 36 |
+
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.
|
| 37 |
+
|
| 38 |
+
```Python
|
| 39 |
+
import torch
|
| 40 |
+
from transformers import AutoModelForCausalLM, AutoTokenizer
|
| 41 |
+
from PIL import Image
|
| 42 |
+
|
| 43 |
+
torch.set_default_device("cuda")
|
| 44 |
+
|
| 45 |
+
#Create model
|
| 46 |
+
model = AutoModelForCausalLM.from_pretrained(
|
| 47 |
+
"MILVLG/imp-v1.5-4b",
|
| 48 |
+
torch_dtype=torch.float16,
|
| 49 |
+
device_map="auto",
|
| 50 |
+
trust_remote_code=True)
|
| 51 |
+
tokenizer = AutoTokenizer.from_pretrained("MILVLG/imp-v1.5-4b", trust_remote_code=True)
|
| 52 |
+
|
| 53 |
+
#Set inputs
|
| 54 |
+
text = "<|user|>\n<image>\nWhat are the colors of the bus in the image?\n<|end|>\n<|assistant|>\n"
|
| 55 |
+
image = Image.open("images/bus.jpg")
|
| 56 |
+
|
| 57 |
+
input_ids = tokenizer(text, return_tensors='pt').input_ids
|
| 58 |
+
image_tensor = model.image_preprocess(image)
|
| 59 |
+
|
| 60 |
+
#Generate the answer
|
| 61 |
+
output_ids = model.generate(
|
| 62 |
+
input_ids,
|
| 63 |
+
max_new_tokens=100,
|
| 64 |
+
images=image_tensor,
|
| 65 |
+
use_cache=True)[0]
|
| 66 |
+
print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())
|
| 67 |
+
```
|
| 68 |
+
|
| 69 |
+
## Model evaluation
|
| 70 |
+
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.
|
| 71 |
+
|
| 72 |
+
| Models | Size | VQAv2 | GQA | SQA(IMG) | TextVQA | POPE | MME(P) | MMB |MMB_CN|MM-Vet|
|
| 73 |
+
|:--------:|:-----:|:----:|:-------------:|:--------:|:-----:|:----:|:-------:|:-------:|:-------:|:-------:|
|
| 74 |
+
| imp-v1.5-4b| 4B | 81.46 | 63.51 | 77.99|60.16 | 86.86| 1507.7 |73.28 |61.08|44.6|
|
| 75 |
+
<!-- | [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| -->
|
| 76 |
+
|
| 77 |
+
|
| 78 |
+
|
| 79 |
+
## License
|
| 80 |
+
This project is licensed under the Apache License 2.0 - see the [LICENSE](https://www.apache.org/licenses/LICENSE-2.0) file for details.
|
| 81 |
+
|
| 82 |
+
## About us
|
| 83 |
+
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.
|
| 84 |
+
|
| 85 |
+
## Citation
|
| 86 |
+
|
| 87 |
+
If you use our model or refer our work in your studies, please cite:
|
| 88 |
+
|
| 89 |
+
```bibtex
|
| 90 |
+
@misc{imp2024,
|
| 91 |
+
author = {Shao, Zhenwei and Ouyang, Xuecheng and Yu, Zhou and Yu, Jun},
|
| 92 |
+
title = {Imp: An Emprical Study of Multimodal Small Language Models},
|
| 93 |
+
year = {2024},
|
| 94 |
+
url = {https://huggingface.co/MILVLG/imp-v1-3b}
|
| 95 |
+
}
|
| 96 |
+
```
|
added_tokens.json
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"<|assistant|>": 32001,
|
| 3 |
+
"<|endoftext|>": 32000,
|
| 4 |
+
"<|end|>": 32007,
|
| 5 |
+
"<|placeholder1|>": 32002,
|
| 6 |
+
"<|placeholder2|>": 32003,
|
| 7 |
+
"<|placeholder3|>": 32004,
|
| 8 |
+
"<|placeholder4|>": 32005,
|
| 9 |
+
"<|placeholder5|>": 32008,
|
| 10 |
+
"<|placeholder6|>": 32009,
|
| 11 |
+
"<|system|>": 32006,
|
| 12 |
+
"<|user|>": 32010,
|
| 13 |
+
"<image>": 32011
|
| 14 |
+
}
|
config.json
ADDED
|
@@ -0,0 +1,79 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_name_or_path": "MILVLG/imp-v1.5-4b",
|
| 3 |
+
"activation_function": "gelu_new",
|
| 4 |
+
"architectures": [
|
| 5 |
+
"ImpPhi3ForCausalLM"
|
| 6 |
+
],
|
| 7 |
+
"attn_pdrop": 0.0,
|
| 8 |
+
"auto_map": {
|
| 9 |
+
"AutoConfig": "configuration_imp.ImpPhi3Config",
|
| 10 |
+
"AutoModelForCausalLM": "modeling_imp.ImpPhi3ForCausalLM"
|
| 11 |
+
},
|
| 12 |
+
"embd_pdrop": 0.0,
|
| 13 |
+
"bos_token_id": 1,
|
| 14 |
+
"eos_token_id": 32007,
|
| 15 |
+
"flash_attn": false,
|
| 16 |
+
"flash_rotary": false,
|
| 17 |
+
"freeze_mm_mlp_adapter": false,
|
| 18 |
+
"hidden_act": "silu",
|
| 19 |
+
"hidden_size": 3072,
|
| 20 |
+
"image_aspect_ratio": "pad",
|
| 21 |
+
"initializer_range": 0.02,
|
| 22 |
+
"intermediate_size": 8192,
|
| 23 |
+
"max_position_embeddings": 4096,
|
| 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_phi3",
|
| 33 |
+
"num_attention_heads": 32,
|
| 34 |
+
"num_hidden_layers": 32,
|
| 35 |
+
"num_key_value_heads": 32,
|
| 36 |
+
"original_max_position_embeddings": 4096,
|
| 37 |
+
"pad_token_id": 32000,
|
| 38 |
+
"resid_pdrop": 0.0,
|
| 39 |
+
"rms_norm_eps": 1e-05,
|
| 40 |
+
"rope_scaling": null,
|
| 41 |
+
"rope_theta": 10000.0,
|
| 42 |
+
"sliding_window": 2047,
|
| 43 |
+
"tie_word_embeddings": false,
|
| 44 |
+
"tokenizer_model_max_length": 2560,
|
| 45 |
+
"tokenizer_padding_side": "right",
|
| 46 |
+
"torch_dtype": "bfloat16",
|
| 47 |
+
"transformers_version": "4.36.0",
|
| 48 |
+
"tune_mm_mlp_adapter": false,
|
| 49 |
+
"use_cache": true,
|
| 50 |
+
"use_mm_proj": true,
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
"fused_dense": false,
|
| 54 |
+
"image_token": "<image>",
|
| 55 |
+
"image_token_index": 32011,
|
| 56 |
+
"img_processor": null,
|
| 57 |
+
"layer_norm_epsilon": 1e-05,
|
| 58 |
+
"n_embd": 2560,
|
| 59 |
+
"n_head": 32,
|
| 60 |
+
"n_head_kv": null,
|
| 61 |
+
"n_inner": null,
|
| 62 |
+
"n_layer": 32,
|
| 63 |
+
"n_positions": 3072,
|
| 64 |
+
"rotary_dim": 32,
|
| 65 |
+
"vision_tower_config": {
|
| 66 |
+
"attention_dropout": 0.0,
|
| 67 |
+
"hidden_act": "gelu_pytorch_tanh",
|
| 68 |
+
"hidden_size": 1152,
|
| 69 |
+
"image_size": 384,
|
| 70 |
+
"intermediate_size": 4304,
|
| 71 |
+
"layer_norm_eps": 1e-06,
|
| 72 |
+
"model_type": "siglip_vision_model",
|
| 73 |
+
"num_attention_heads": 16,
|
| 74 |
+
"num_channels": 3,
|
| 75 |
+
"num_hidden_layers": 27,
|
| 76 |
+
"patch_size": 14
|
| 77 |
+
},
|
| 78 |
+
"vocab_size": 32064
|
| 79 |
+
}
|
configuration_imp.py
ADDED
|
@@ -0,0 +1,313 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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 Phi3Config(PretrainedConfig):
|
| 59 |
+
r"""
|
| 60 |
+
This is the configuration class to store the configuration of a [`Phi3Model`]. It is used to instantiate a Phi-3
|
| 61 |
+
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
|
| 62 |
+
defaults will yield a similar configuration to that of the
|
| 63 |
+
[microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct).
|
| 64 |
+
|
| 65 |
+
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
| 66 |
+
documentation from [`PretrainedConfig`] for more information.
|
| 67 |
+
|
| 68 |
+
Args:
|
| 69 |
+
vocab_size (`int`, *optional*, defaults to 32064):
|
| 70 |
+
Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the
|
| 71 |
+
`inputs_ids` passed when calling [`Phi3Model`].
|
| 72 |
+
hidden_size (`int`, *optional*, defaults to 3072):
|
| 73 |
+
Dimension of the hidden representations.
|
| 74 |
+
intermediate_size (`int`, *optional*, defaults to 8192):
|
| 75 |
+
Dimension of the MLP representations.
|
| 76 |
+
num_hidden_layers (`int`, *optional*, defaults to 32):
|
| 77 |
+
Number of hidden layers in the Transformer decoder.
|
| 78 |
+
num_attention_heads (`int`, *optional*, defaults to 32):
|
| 79 |
+
Number of attention heads for each attention layer in the Transformer decoder.
|
| 80 |
+
num_key_value_heads (`int`, *optional*):
|
| 81 |
+
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
| 82 |
+
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
| 83 |
+
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
| 84 |
+
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
| 85 |
+
by meanpooling all the original heads within that group. For more details checkout [this
|
| 86 |
+
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to
|
| 87 |
+
`num_attention_heads`.
|
| 88 |
+
resid_pdrop (`float`, *optional*, defaults to 0.0):
|
| 89 |
+
Dropout probability for mlp outputs.
|
| 90 |
+
embd_pdrop (`int`, *optional*, defaults to 0.0):
|
| 91 |
+
The dropout ratio for the embeddings.
|
| 92 |
+
attention_dropout (`float`, *optional*, defaults to 0.0):
|
| 93 |
+
The dropout ratio after computing the attention scores.
|
| 94 |
+
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
| 95 |
+
The non-linear activation function (function or string) in the decoder.
|
| 96 |
+
max_position_embeddings (`int`, *optional*, defaults to 4096):
|
| 97 |
+
The maximum sequence length that this model might ever be used with.
|
| 98 |
+
original_max_position_embeddings (`int`, *optional*, defaults to 4096):
|
| 99 |
+
The maximum sequence length that this model was trained with. This is used to determine the size of the
|
| 100 |
+
original RoPE embeddings when using long scaling.
|
| 101 |
+
initializer_range (`float`, *optional*, defaults to 0.02):
|
| 102 |
+
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
| 103 |
+
rms_norm_eps (`float`, *optional*, defaults to 1e-05):
|
| 104 |
+
The epsilon value used for the RMSNorm.
|
| 105 |
+
use_cache (`bool`, *optional*, defaults to `True`):
|
| 106 |
+
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
| 107 |
+
relevant if `config.is_decoder=True`. Whether to tie weight embeddings or not.
|
| 108 |
+
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
| 109 |
+
Whether to tie weight embeddings
|
| 110 |
+
rope_theta (`float`, *optional*, defaults to 10000.0):
|
| 111 |
+
The base period of the RoPE embeddings.
|
| 112 |
+
rope_scaling (`dict`, *optional*):
|
| 113 |
+
The scaling strategy for the RoPE embeddings. If `None`, no scaling is applied. If a dictionary, it must
|
| 114 |
+
contain the following keys: `type`, `short_factor` and `long_factor`. The `type` must be either `su` or `yarn` and
|
| 115 |
+
the `short_factor` and `long_factor` must be lists of numbers with the same length as the hidden size
|
| 116 |
+
divided by the number of attention heads divided by 2.
|
| 117 |
+
bos_token_id (`int`, *optional*, defaults to 1):
|
| 118 |
+
The id of the "beginning-of-sequence" token.
|
| 119 |
+
eos_token_id (`int`, *optional*, defaults to 32000):
|
| 120 |
+
The id of the "end-of-sequence" token.
|
| 121 |
+
pad_token_id (`int`, *optional*, defaults to 32000):
|
| 122 |
+
The id of the padding token.
|
| 123 |
+
sliding_window (`int`, *optional*):
|
| 124 |
+
Sliding window attention window size. If `None`, no sliding window is applied.
|
| 125 |
+
|
| 126 |
+
Example:
|
| 127 |
+
|
| 128 |
+
```python
|
| 129 |
+
>>> from transformers import Phi3Model, Phi3Config
|
| 130 |
+
|
| 131 |
+
>>> # Initializing a Phi-3 style configuration
|
| 132 |
+
>>> configuration = Phi3Config.from_pretrained("microsoft/Phi-3-mini-4k-instruct")
|
| 133 |
+
|
| 134 |
+
>>> # Initializing a model from the configuration
|
| 135 |
+
>>> model = Phi3Model(configuration)
|
| 136 |
+
|
| 137 |
+
>>> # Accessing the model configuration
|
| 138 |
+
>>> configuration = model.config
|
| 139 |
+
```"""
|
| 140 |
+
|
| 141 |
+
model_type = "phi3"
|
| 142 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
| 143 |
+
|
| 144 |
+
def __init__(
|
| 145 |
+
self,
|
| 146 |
+
vocab_size=32064,
|
| 147 |
+
hidden_size=3072,
|
| 148 |
+
intermediate_size=8192,
|
| 149 |
+
num_hidden_layers=32,
|
| 150 |
+
num_attention_heads=32,
|
| 151 |
+
num_key_value_heads=None,
|
| 152 |
+
resid_pdrop=0.0,
|
| 153 |
+
embd_pdrop=0.0,
|
| 154 |
+
attention_dropout=0.0,
|
| 155 |
+
hidden_act="silu",
|
| 156 |
+
max_position_embeddings=4096,
|
| 157 |
+
original_max_position_embeddings=4096,
|
| 158 |
+
initializer_range=0.02,
|
| 159 |
+
rms_norm_eps=1e-5,
|
| 160 |
+
use_cache=True,
|
| 161 |
+
tie_word_embeddings=False,
|
| 162 |
+
rope_theta=10000.0,
|
| 163 |
+
rope_scaling=None,
|
| 164 |
+
bos_token_id=1,
|
| 165 |
+
eos_token_id=32000,
|
| 166 |
+
pad_token_id=32000,
|
| 167 |
+
sliding_window=None,
|
| 168 |
+
**kwargs,
|
| 169 |
+
):
|
| 170 |
+
self.vocab_size = vocab_size
|
| 171 |
+
self.hidden_size = hidden_size
|
| 172 |
+
self.intermediate_size = intermediate_size
|
| 173 |
+
self.num_hidden_layers = num_hidden_layers
|
| 174 |
+
self.num_attention_heads = num_attention_heads
|
| 175 |
+
|
| 176 |
+
if num_key_value_heads is None:
|
| 177 |
+
num_key_value_heads = num_attention_heads
|
| 178 |
+
|
| 179 |
+
self.num_key_value_heads = num_key_value_heads
|
| 180 |
+
self.resid_pdrop = resid_pdrop
|
| 181 |
+
self.embd_pdrop = embd_pdrop
|
| 182 |
+
self.attention_dropout = attention_dropout
|
| 183 |
+
self.hidden_act = hidden_act
|
| 184 |
+
self.max_position_embeddings = max_position_embeddings
|
| 185 |
+
self.original_max_position_embeddings = original_max_position_embeddings
|
| 186 |
+
self.initializer_range = initializer_range
|
| 187 |
+
self.rms_norm_eps = rms_norm_eps
|
| 188 |
+
self.use_cache = use_cache
|
| 189 |
+
self.rope_theta = rope_theta
|
| 190 |
+
self.rope_scaling = rope_scaling
|
| 191 |
+
self._rope_scaling_validation()
|
| 192 |
+
self.sliding_window = sliding_window
|
| 193 |
+
|
| 194 |
+
super().__init__(
|
| 195 |
+
bos_token_id=bos_token_id,
|
| 196 |
+
eos_token_id=eos_token_id,
|
| 197 |
+
pad_token_id=pad_token_id,
|
| 198 |
+
tie_word_embeddings=tie_word_embeddings,
|
| 199 |
+
**kwargs,
|
| 200 |
+
)
|
| 201 |
+
|
| 202 |
+
def _rope_scaling_validation(self):
|
| 203 |
+
"""
|
| 204 |
+
Validate the `rope_scaling` configuration.
|
| 205 |
+
"""
|
| 206 |
+
if self.rope_scaling is None:
|
| 207 |
+
return
|
| 208 |
+
|
| 209 |
+
if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 3:
|
| 210 |
+
raise ValueError(
|
| 211 |
+
"`rope_scaling` must be a dictionary with three fields, `type`, `short_factor` and `long_factor`, "
|
| 212 |
+
f"got {self.rope_scaling}"
|
| 213 |
+
)
|
| 214 |
+
rope_scaling_type = self.rope_scaling.get("type", None)
|
| 215 |
+
rope_scaling_short_factor = self.rope_scaling.get("short_factor", None)
|
| 216 |
+
rope_scaling_long_factor = self.rope_scaling.get("long_factor", None)
|
| 217 |
+
if rope_scaling_type is None or rope_scaling_type not in ["su", "yarn"]:
|
| 218 |
+
raise ValueError(f"`rope_scaling`'s type field must be one of ['su', 'yarn'], got {rope_scaling_type}")
|
| 219 |
+
if not (
|
| 220 |
+
isinstance(rope_scaling_short_factor, list)
|
| 221 |
+
and all(isinstance(x, (int, float)) for x in rope_scaling_short_factor)
|
| 222 |
+
):
|
| 223 |
+
raise ValueError(
|
| 224 |
+
f"`rope_scaling`'s short_factor field must be a list of numbers, got {rope_scaling_short_factor}"
|
| 225 |
+
)
|
| 226 |
+
if not len(rope_scaling_short_factor) == self.hidden_size // self.num_attention_heads // 2:
|
| 227 |
+
raise ValueError(
|
| 228 |
+
f"`rope_scaling`'s short_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_short_factor)}"
|
| 229 |
+
)
|
| 230 |
+
if not (
|
| 231 |
+
isinstance(rope_scaling_long_factor, list)
|
| 232 |
+
and all(isinstance(x, (int, float)) for x in rope_scaling_long_factor)
|
| 233 |
+
):
|
| 234 |
+
raise ValueError(
|
| 235 |
+
f"`rope_scaling`'s long_factor field must be a list of numbers, got {rope_scaling_long_factor}"
|
| 236 |
+
)
|
| 237 |
+
if not len(rope_scaling_long_factor) == self.hidden_size // self.num_attention_heads // 2:
|
| 238 |
+
raise ValueError(
|
| 239 |
+
f"`rope_scaling`'s long_factor field must have length {self.hidden_size // self.num_attention_heads // 2}, got {len(rope_scaling_long_factor)}"
|
| 240 |
+
)
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
class SiglipVisionConfig(PretrainedConfig):
|
| 245 |
+
|
| 246 |
+
model_type = "siglip_vision_model"
|
| 247 |
+
|
| 248 |
+
def __init__(
|
| 249 |
+
self,
|
| 250 |
+
hidden_size=768,
|
| 251 |
+
intermediate_size=3072,
|
| 252 |
+
num_hidden_layers=12,
|
| 253 |
+
num_attention_heads=12,
|
| 254 |
+
num_channels=3,
|
| 255 |
+
image_size=224,
|
| 256 |
+
patch_size=16,
|
| 257 |
+
hidden_act="gelu_pytorch_tanh",
|
| 258 |
+
layer_norm_eps=1e-6,
|
| 259 |
+
attention_dropout=0.0,
|
| 260 |
+
**kwargs,
|
| 261 |
+
):
|
| 262 |
+
super().__init__(**kwargs)
|
| 263 |
+
|
| 264 |
+
self.hidden_size = hidden_size
|
| 265 |
+
self.intermediate_size = intermediate_size
|
| 266 |
+
self.num_hidden_layers = num_hidden_layers
|
| 267 |
+
self.num_attention_heads = num_attention_heads
|
| 268 |
+
self.num_channels = num_channels
|
| 269 |
+
self.patch_size = patch_size
|
| 270 |
+
self.image_size = image_size
|
| 271 |
+
self.attention_dropout = attention_dropout
|
| 272 |
+
self.layer_norm_eps = layer_norm_eps
|
| 273 |
+
self.hidden_act = hidden_act
|
| 274 |
+
|
| 275 |
+
@classmethod
|
| 276 |
+
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
| 277 |
+
cls._set_token_in_kwargs(kwargs)
|
| 278 |
+
|
| 279 |
+
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
| 280 |
+
|
| 281 |
+
# get the vision config dict if we are loading from SiglipConfig
|
| 282 |
+
if config_dict.get("model_type") == "siglip":
|
| 283 |
+
config_dict = config_dict["vision_config"]
|
| 284 |
+
|
| 285 |
+
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
| 286 |
+
logger.warning(
|
| 287 |
+
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
| 288 |
+
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
| 289 |
+
)
|
| 290 |
+
|
| 291 |
+
return cls.from_dict(config_dict, **kwargs)
|
| 292 |
+
|
| 293 |
+
|
| 294 |
+
class ImpPhi3Config(Phi3Config):
|
| 295 |
+
model_type = "imp_phi3"
|
| 296 |
+
|
| 297 |
+
def __init__(self, **kwargs):
|
| 298 |
+
super().__init__(**kwargs)
|
| 299 |
+
self.image_token_index = getattr(self, "image_token_index", 50296)
|
| 300 |
+
self.image_token = getattr(self, "image_token", "<image>")
|
| 301 |
+
|
| 302 |
+
if not hasattr(self, "vision_tower_config") and hasattr(self, "mm_vision_tower"):
|
| 303 |
+
vision_tower_config = SiglipVisionConfig.from_pretrained(self.mm_vision_tower)
|
| 304 |
+
self.vision_tower_config = vision_tower_config.to_diff_dict()
|
| 305 |
+
|
| 306 |
+
@property
|
| 307 |
+
def vision_tower_cfg(self):
|
| 308 |
+
cfg = SiglipVisionConfig.from_dict(self.vision_tower_config)
|
| 309 |
+
# imp-v1 only supports `patch` feature for now w/o cls token
|
| 310 |
+
# cfg.mm_vision_select_feature = self.mm_vision_select_feature
|
| 311 |
+
cfg.mm_vision_select_layer = self.mm_vision_select_layer
|
| 312 |
+
cfg.mm_vision_tower = self.mm_vision_tower
|
| 313 |
+
return cfg
|
generation_config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 1,
|
| 4 |
+
"eos_token_id": [
|
| 5 |
+
32000,
|
| 6 |
+
32001,
|
| 7 |
+
32007
|
| 8 |
+
],
|
| 9 |
+
"max_new_tokens": 2000,
|
| 10 |
+
"pad_token_id": 2,
|
| 11 |
+
"transformers_version": "4.36.0"
|
| 12 |
+
}
|
merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
model-00001-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7fc741406a274d9e77885cef89cc57d4e4ce75cb50c3b07d079e1835a27f2613
|
| 3 |
+
size 952015208
|
model-00002-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:582f5813afe4c1677272bc5812749ab2f3f108d3e2e75c4b440a180c050eb041
|
| 3 |
+
size 1006684976
|
model-00003-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:65c38648b356ffba04485d18588eb042fb96b6f0c17716924f0e83711490835b
|
| 3 |
+
size 975240328
|
model-00004-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6810ac18f6536369f2967f69bc8063a95d45fe961ef18ef8ea0e7c38a5ce0917
|
| 3 |
+
size 962644808
|
model-00005-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e7e18a5f170fddd42ea0ff4cc3765ec41f2661897ff3cbe20237b6ba2655152b
|
| 3 |
+
size 1006685000
|
model-00006-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:251073aa76eda06de96c588bd2baff4f707cbae92c057ba6bae3168f737acd60
|
| 3 |
+
size 975240344
|
model-00007-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:6be89bb18770df00837859bdf4a41ae2bc8abf28b01b1edddf5aeb05cdf91325
|
| 3 |
+
size 962644808
|
model-00008-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:c845e1653a1505dc8c39ea86e7ef14d5df40cb56e786a90e57cdfd6a56ee242c
|
| 3 |
+
size 1023878536
|
model-00009-of-00009.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d38fcded882b7e25c4fa224bf1fea6a675dd8abf3db234645a1a81d471654930
|
| 3 |
+
size 598672880
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,627 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 8463619136
|
| 4 |
+
},
|
| 5 |
+
"weight_map": {
|
| 6 |
+
"lm_head.weight": "model-00009-of-00009.safetensors",
|
| 7 |
+
"model.embed_tokens.weight": "model-00001-of-00009.safetensors",
|
| 8 |
+
"model.layers.0.input_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 9 |
+
"model.layers.0.mlp.down_proj.weight": "model-00001-of-00009.safetensors",
|
| 10 |
+
"model.layers.0.mlp.gate_up_proj.weight": "model-00001-of-00009.safetensors",
|
| 11 |
+
"model.layers.0.post_attention_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 12 |
+
"model.layers.0.self_attn.o_proj.weight": "model-00001-of-00009.safetensors",
|
| 13 |
+
"model.layers.0.self_attn.qkv_proj.weight": "model-00001-of-00009.safetensors",
|
| 14 |
+
"model.layers.1.input_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 15 |
+
"model.layers.1.mlp.down_proj.weight": "model-00001-of-00009.safetensors",
|
| 16 |
+
"model.layers.1.mlp.gate_up_proj.weight": "model-00001-of-00009.safetensors",
|
| 17 |
+
"model.layers.1.post_attention_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 18 |
+
"model.layers.1.self_attn.o_proj.weight": "model-00001-of-00009.safetensors",
|
| 19 |
+
"model.layers.1.self_attn.qkv_proj.weight": "model-00001-of-00009.safetensors",
|
| 20 |
+
"model.layers.10.input_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 21 |
+
"model.layers.10.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 22 |
+
"model.layers.10.mlp.gate_up_proj.weight": "model-00003-of-00009.safetensors",
|
| 23 |
+
"model.layers.10.post_attention_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 24 |
+
"model.layers.10.self_attn.o_proj.weight": "model-00003-of-00009.safetensors",
|
| 25 |
+
"model.layers.10.self_attn.qkv_proj.weight": "model-00003-of-00009.safetensors",
|
| 26 |
+
"model.layers.11.input_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 27 |
+
"model.layers.11.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 28 |
+
"model.layers.11.mlp.gate_up_proj.weight": "model-00003-of-00009.safetensors",
|
| 29 |
+
"model.layers.11.post_attention_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 30 |
+
"model.layers.11.self_attn.o_proj.weight": "model-00003-of-00009.safetensors",
|
| 31 |
+
"model.layers.11.self_attn.qkv_proj.weight": "model-00003-of-00009.safetensors",
|
| 32 |
+
"model.layers.12.input_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 33 |
+
"model.layers.12.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 34 |
+
"model.layers.12.mlp.gate_up_proj.weight": "model-00004-of-00009.safetensors",
|
| 35 |
+
"model.layers.12.post_attention_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 36 |
+
"model.layers.12.self_attn.o_proj.weight": "model-00003-of-00009.safetensors",
|
| 37 |
+
"model.layers.12.self_attn.qkv_proj.weight": "model-00004-of-00009.safetensors",
|
| 38 |
+
"model.layers.13.input_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 39 |
+
"model.layers.13.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 40 |
+
"model.layers.13.mlp.gate_up_proj.weight": "model-00004-of-00009.safetensors",
|
| 41 |
+
"model.layers.13.post_attention_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 42 |
+
"model.layers.13.self_attn.o_proj.weight": "model-00004-of-00009.safetensors",
|
| 43 |
+
"model.layers.13.self_attn.qkv_proj.weight": "model-00004-of-00009.safetensors",
|
| 44 |
+
"model.layers.14.input_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 45 |
+
"model.layers.14.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 46 |
+
"model.layers.14.mlp.gate_up_proj.weight": "model-00004-of-00009.safetensors",
|
| 47 |
+
"model.layers.14.post_attention_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 48 |
+
"model.layers.14.self_attn.o_proj.weight": "model-00004-of-00009.safetensors",
|
| 49 |
+
"model.layers.14.self_attn.qkv_proj.weight": "model-00004-of-00009.safetensors",
|
| 50 |
+
"model.layers.15.input_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 51 |
+
"model.layers.15.mlp.down_proj.weight": "model-00004-of-00009.safetensors",
|
| 52 |
+
"model.layers.15.mlp.gate_up_proj.weight": "model-00004-of-00009.safetensors",
|
| 53 |
+
"model.layers.15.post_attention_layernorm.weight": "model-00004-of-00009.safetensors",
|
| 54 |
+
"model.layers.15.self_attn.o_proj.weight": "model-00004-of-00009.safetensors",
|
| 55 |
+
"model.layers.15.self_attn.qkv_proj.weight": "model-00004-of-00009.safetensors",
|
| 56 |
+
"model.layers.16.input_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 57 |
+
"model.layers.16.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 58 |
+
"model.layers.16.mlp.gate_up_proj.weight": "model-00005-of-00009.safetensors",
|
| 59 |
+
"model.layers.16.post_attention_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 60 |
+
"model.layers.16.self_attn.o_proj.weight": "model-00004-of-00009.safetensors",
|
| 61 |
+
"model.layers.16.self_attn.qkv_proj.weight": "model-00004-of-00009.safetensors",
|
| 62 |
+
"model.layers.17.input_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 63 |
+
"model.layers.17.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 64 |
+
"model.layers.17.mlp.gate_up_proj.weight": "model-00005-of-00009.safetensors",
|
| 65 |
+
"model.layers.17.post_attention_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 66 |
+
"model.layers.17.self_attn.o_proj.weight": "model-00005-of-00009.safetensors",
|
| 67 |
+
"model.layers.17.self_attn.qkv_proj.weight": "model-00005-of-00009.safetensors",
|
| 68 |
+
"model.layers.18.input_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 69 |
+
"model.layers.18.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 70 |
+
"model.layers.18.mlp.gate_up_proj.weight": "model-00005-of-00009.safetensors",
|
| 71 |
+
"model.layers.18.post_attention_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 72 |
+
"model.layers.18.self_attn.o_proj.weight": "model-00005-of-00009.safetensors",
|
| 73 |
+
"model.layers.18.self_attn.qkv_proj.weight": "model-00005-of-00009.safetensors",
|
| 74 |
+
"model.layers.19.input_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 75 |
+
"model.layers.19.mlp.down_proj.weight": "model-00005-of-00009.safetensors",
|
| 76 |
+
"model.layers.19.mlp.gate_up_proj.weight": "model-00005-of-00009.safetensors",
|
| 77 |
+
"model.layers.19.post_attention_layernorm.weight": "model-00005-of-00009.safetensors",
|
| 78 |
+
"model.layers.19.self_attn.o_proj.weight": "model-00005-of-00009.safetensors",
|
| 79 |
+
"model.layers.19.self_attn.qkv_proj.weight": "model-00005-of-00009.safetensors",
|
| 80 |
+
"model.layers.2.input_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 81 |
+
"model.layers.2.mlp.down_proj.weight": "model-00001-of-00009.safetensors",
|
| 82 |
+
"model.layers.2.mlp.gate_up_proj.weight": "model-00001-of-00009.safetensors",
|
| 83 |
+
"model.layers.2.post_attention_layernorm.weight": "model-00001-of-00009.safetensors",
|
| 84 |
+
"model.layers.2.self_attn.o_proj.weight": "model-00001-of-00009.safetensors",
|
| 85 |
+
"model.layers.2.self_attn.qkv_proj.weight": "model-00001-of-00009.safetensors",
|
| 86 |
+
"model.layers.20.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 87 |
+
"model.layers.20.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 88 |
+
"model.layers.20.mlp.gate_up_proj.weight": "model-00005-of-00009.safetensors",
|
| 89 |
+
"model.layers.20.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 90 |
+
"model.layers.20.self_attn.o_proj.weight": "model-00005-of-00009.safetensors",
|
| 91 |
+
"model.layers.20.self_attn.qkv_proj.weight": "model-00005-of-00009.safetensors",
|
| 92 |
+
"model.layers.21.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 93 |
+
"model.layers.21.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 94 |
+
"model.layers.21.mlp.gate_up_proj.weight": "model-00006-of-00009.safetensors",
|
| 95 |
+
"model.layers.21.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 96 |
+
"model.layers.21.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 97 |
+
"model.layers.21.self_attn.qkv_proj.weight": "model-00006-of-00009.safetensors",
|
| 98 |
+
"model.layers.22.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 99 |
+
"model.layers.22.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 100 |
+
"model.layers.22.mlp.gate_up_proj.weight": "model-00006-of-00009.safetensors",
|
| 101 |
+
"model.layers.22.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 102 |
+
"model.layers.22.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 103 |
+
"model.layers.22.self_attn.qkv_proj.weight": "model-00006-of-00009.safetensors",
|
| 104 |
+
"model.layers.23.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 105 |
+
"model.layers.23.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 106 |
+
"model.layers.23.mlp.gate_up_proj.weight": "model-00006-of-00009.safetensors",
|
| 107 |
+
"model.layers.23.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 108 |
+
"model.layers.23.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 109 |
+
"model.layers.23.self_attn.qkv_proj.weight": "model-00006-of-00009.safetensors",
|
| 110 |
+
"model.layers.24.input_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 111 |
+
"model.layers.24.mlp.down_proj.weight": "model-00006-of-00009.safetensors",
|
| 112 |
+
"model.layers.24.mlp.gate_up_proj.weight": "model-00006-of-00009.safetensors",
|
| 113 |
+
"model.layers.24.post_attention_layernorm.weight": "model-00006-of-00009.safetensors",
|
| 114 |
+
"model.layers.24.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 115 |
+
"model.layers.24.self_attn.qkv_proj.weight": "model-00006-of-00009.safetensors",
|
| 116 |
+
"model.layers.25.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 117 |
+
"model.layers.25.mlp.down_proj.weight": "model-00007-of-00009.safetensors",
|
| 118 |
+
"model.layers.25.mlp.gate_up_proj.weight": "model-00007-of-00009.safetensors",
|
| 119 |
+
"model.layers.25.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 120 |
+
"model.layers.25.self_attn.o_proj.weight": "model-00006-of-00009.safetensors",
|
| 121 |
+
"model.layers.25.self_attn.qkv_proj.weight": "model-00007-of-00009.safetensors",
|
| 122 |
+
"model.layers.26.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 123 |
+
"model.layers.26.mlp.down_proj.weight": "model-00007-of-00009.safetensors",
|
| 124 |
+
"model.layers.26.mlp.gate_up_proj.weight": "model-00007-of-00009.safetensors",
|
| 125 |
+
"model.layers.26.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 126 |
+
"model.layers.26.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 127 |
+
"model.layers.26.self_attn.qkv_proj.weight": "model-00007-of-00009.safetensors",
|
| 128 |
+
"model.layers.27.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 129 |
+
"model.layers.27.mlp.down_proj.weight": "model-00007-of-00009.safetensors",
|
| 130 |
+
"model.layers.27.mlp.gate_up_proj.weight": "model-00007-of-00009.safetensors",
|
| 131 |
+
"model.layers.27.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 132 |
+
"model.layers.27.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 133 |
+
"model.layers.27.self_attn.qkv_proj.weight": "model-00007-of-00009.safetensors",
|
| 134 |
+
"model.layers.28.input_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 135 |
+
"model.layers.28.mlp.down_proj.weight": "model-00007-of-00009.safetensors",
|
| 136 |
+
"model.layers.28.mlp.gate_up_proj.weight": "model-00007-of-00009.safetensors",
|
| 137 |
+
"model.layers.28.post_attention_layernorm.weight": "model-00007-of-00009.safetensors",
|
| 138 |
+
"model.layers.28.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 139 |
+
"model.layers.28.self_attn.qkv_proj.weight": "model-00007-of-00009.safetensors",
|
| 140 |
+
"model.layers.29.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 141 |
+
"model.layers.29.mlp.down_proj.weight": "model-00008-of-00009.safetensors",
|
| 142 |
+
"model.layers.29.mlp.gate_up_proj.weight": "model-00008-of-00009.safetensors",
|
| 143 |
+
"model.layers.29.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 144 |
+
"model.layers.29.self_attn.o_proj.weight": "model-00007-of-00009.safetensors",
|
| 145 |
+
"model.layers.29.self_attn.qkv_proj.weight": "model-00007-of-00009.safetensors",
|
| 146 |
+
"model.layers.3.input_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 147 |
+
"model.layers.3.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 148 |
+
"model.layers.3.mlp.gate_up_proj.weight": "model-00002-of-00009.safetensors",
|
| 149 |
+
"model.layers.3.post_attention_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 150 |
+
"model.layers.3.self_attn.o_proj.weight": "model-00001-of-00009.safetensors",
|
| 151 |
+
"model.layers.3.self_attn.qkv_proj.weight": "model-00001-of-00009.safetensors",
|
| 152 |
+
"model.layers.30.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 153 |
+
"model.layers.30.mlp.down_proj.weight": "model-00008-of-00009.safetensors",
|
| 154 |
+
"model.layers.30.mlp.gate_up_proj.weight": "model-00008-of-00009.safetensors",
|
| 155 |
+
"model.layers.30.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 156 |
+
"model.layers.30.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 157 |
+
"model.layers.30.self_attn.qkv_proj.weight": "model-00008-of-00009.safetensors",
|
| 158 |
+
"model.layers.31.input_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 159 |
+
"model.layers.31.mlp.down_proj.weight": "model-00008-of-00009.safetensors",
|
| 160 |
+
"model.layers.31.mlp.gate_up_proj.weight": "model-00008-of-00009.safetensors",
|
| 161 |
+
"model.layers.31.post_attention_layernorm.weight": "model-00008-of-00009.safetensors",
|
| 162 |
+
"model.layers.31.self_attn.o_proj.weight": "model-00008-of-00009.safetensors",
|
| 163 |
+
"model.layers.31.self_attn.qkv_proj.weight": "model-00008-of-00009.safetensors",
|
| 164 |
+
"model.layers.4.input_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 165 |
+
"model.layers.4.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 166 |
+
"model.layers.4.mlp.gate_up_proj.weight": "model-00002-of-00009.safetensors",
|
| 167 |
+
"model.layers.4.post_attention_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 168 |
+
"model.layers.4.self_attn.o_proj.weight": "model-00002-of-00009.safetensors",
|
| 169 |
+
"model.layers.4.self_attn.qkv_proj.weight": "model-00002-of-00009.safetensors",
|
| 170 |
+
"model.layers.5.input_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 171 |
+
"model.layers.5.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 172 |
+
"model.layers.5.mlp.gate_up_proj.weight": "model-00002-of-00009.safetensors",
|
| 173 |
+
"model.layers.5.post_attention_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 174 |
+
"model.layers.5.self_attn.o_proj.weight": "model-00002-of-00009.safetensors",
|
| 175 |
+
"model.layers.5.self_attn.qkv_proj.weight": "model-00002-of-00009.safetensors",
|
| 176 |
+
"model.layers.6.input_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 177 |
+
"model.layers.6.mlp.down_proj.weight": "model-00002-of-00009.safetensors",
|
| 178 |
+
"model.layers.6.mlp.gate_up_proj.weight": "model-00002-of-00009.safetensors",
|
| 179 |
+
"model.layers.6.post_attention_layernorm.weight": "model-00002-of-00009.safetensors",
|
| 180 |
+
"model.layers.6.self_attn.o_proj.weight": "model-00002-of-00009.safetensors",
|
| 181 |
+
"model.layers.6.self_attn.qkv_proj.weight": "model-00002-of-00009.safetensors",
|
| 182 |
+
"model.layers.7.input_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 183 |
+
"model.layers.7.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 184 |
+
"model.layers.7.mlp.gate_up_proj.weight": "model-00002-of-00009.safetensors",
|
| 185 |
+
"model.layers.7.post_attention_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 186 |
+
"model.layers.7.self_attn.o_proj.weight": "model-00002-of-00009.safetensors",
|
| 187 |
+
"model.layers.7.self_attn.qkv_proj.weight": "model-00002-of-00009.safetensors",
|
| 188 |
+
"model.layers.8.input_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 189 |
+
"model.layers.8.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 190 |
+
"model.layers.8.mlp.gate_up_proj.weight": "model-00003-of-00009.safetensors",
|
| 191 |
+
"model.layers.8.post_attention_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 192 |
+
"model.layers.8.self_attn.o_proj.weight": "model-00003-of-00009.safetensors",
|
| 193 |
+
"model.layers.8.self_attn.qkv_proj.weight": "model-00003-of-00009.safetensors",
|
| 194 |
+
"model.layers.9.input_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 195 |
+
"model.layers.9.mlp.down_proj.weight": "model-00003-of-00009.safetensors",
|
| 196 |
+
"model.layers.9.mlp.gate_up_proj.weight": "model-00003-of-00009.safetensors",
|
| 197 |
+
"model.layers.9.post_attention_layernorm.weight": "model-00003-of-00009.safetensors",
|
| 198 |
+
"model.layers.9.self_attn.o_proj.weight": "model-00003-of-00009.safetensors",
|
| 199 |
+
"model.layers.9.self_attn.qkv_proj.weight": "model-00003-of-00009.safetensors",
|
| 200 |
+
"model.mm_projector.0.bias": "model-00009-of-00009.safetensors",
|
| 201 |
+
"model.mm_projector.0.weight": "model-00009-of-00009.safetensors",
|
| 202 |
+
"model.mm_projector.2.bias": "model-00009-of-00009.safetensors",
|
| 203 |
+
"model.mm_projector.2.weight": "model-00009-of-00009.safetensors",
|
| 204 |
+
"model.norm.weight": "model-00008-of-00009.safetensors",
|
| 205 |
+
"model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.bias": "model-00008-of-00009.safetensors",
|
| 206 |
+
"model.vision_tower.vision_tower.vision_model.embeddings.patch_embedding.weight": "model-00008-of-00009.safetensors",
|
| 207 |
+
"model.vision_tower.vision_tower.vision_model.embeddings.position_embedding.weight": "model-00008-of-00009.safetensors",
|
| 208 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 209 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 210 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 211 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 212 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 213 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 214 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 215 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 216 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 217 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 218 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 219 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 220 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 221 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 222 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 223 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.0.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 224 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 225 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 226 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 227 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 228 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 229 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 230 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 231 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 232 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 233 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 234 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 235 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 236 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 237 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 238 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 239 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.1.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 240 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 241 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 242 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 243 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 244 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 245 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 246 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 247 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 248 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 249 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 250 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 251 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 252 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 253 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 254 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 255 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.10.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 256 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 257 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 258 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 259 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 260 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 261 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 262 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 263 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 264 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 265 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 266 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 267 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 268 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 269 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 270 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 271 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.11.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 272 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 273 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 274 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 275 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 276 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 277 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 278 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 279 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 280 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 281 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 282 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 283 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 284 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 285 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 286 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 287 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.12.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 288 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 289 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 290 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 291 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 292 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 293 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 294 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 295 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 296 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 297 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 298 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 299 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 300 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 301 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 302 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 303 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.13.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 304 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 305 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 306 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 307 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 308 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 309 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 310 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 311 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 312 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 313 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 314 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 315 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 316 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 317 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 318 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 319 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.14.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 320 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 321 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 322 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 323 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 324 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 325 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 326 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 327 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 328 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 329 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 330 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 331 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 332 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 333 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 334 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 335 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.15.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 336 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 337 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 338 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 339 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 340 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 341 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 342 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 343 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 344 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 345 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 346 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 347 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 348 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 349 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 350 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 351 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.16.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 352 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 353 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 354 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 355 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 356 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 357 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 358 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 359 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 360 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 361 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 362 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 363 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 364 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 365 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 366 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 367 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.17.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 368 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 369 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 370 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 371 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 372 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 373 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 374 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 375 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 376 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 377 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 378 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 379 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 380 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 381 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 382 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 383 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.18.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 384 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 385 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 386 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 387 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 388 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 389 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 390 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 391 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 392 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 393 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 394 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 395 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 396 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 397 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 398 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 399 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.19.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 400 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 401 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 402 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 403 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 404 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 405 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 406 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 407 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 408 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 409 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 410 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 411 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 412 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 413 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 414 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 415 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.2.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 416 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 417 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 418 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 419 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 420 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 421 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 422 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 423 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 424 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 425 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 426 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 427 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 428 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 429 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 430 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 431 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.20.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 432 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 433 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 434 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 435 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 436 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 437 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 438 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 439 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 440 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 441 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 442 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 443 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 444 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 445 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 446 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 447 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.21.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 448 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 449 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 450 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 451 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 452 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 453 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 454 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 455 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 456 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 457 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 458 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 459 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 460 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 461 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 462 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 463 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.22.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 464 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 465 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 466 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 467 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 468 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 469 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 470 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 471 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 472 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 473 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 474 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 475 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 476 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 477 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 478 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 479 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.23.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 480 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 481 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 482 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 483 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 484 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 485 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 486 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 487 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 488 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 489 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 490 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 491 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 492 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 493 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 494 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 495 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.24.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 496 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm1.bias": "model-00009-of-00009.safetensors",
|
| 497 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm1.weight": "model-00009-of-00009.safetensors",
|
| 498 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm2.bias": "model-00009-of-00009.safetensors",
|
| 499 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.layer_norm2.weight": "model-00009-of-00009.safetensors",
|
| 500 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc1.bias": "model-00009-of-00009.safetensors",
|
| 501 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc1.weight": "model-00009-of-00009.safetensors",
|
| 502 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc2.bias": "model-00009-of-00009.safetensors",
|
| 503 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.mlp.fc2.weight": "model-00009-of-00009.safetensors",
|
| 504 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.k_proj.bias": "model-00009-of-00009.safetensors",
|
| 505 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.k_proj.weight": "model-00009-of-00009.safetensors",
|
| 506 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.out_proj.bias": "model-00009-of-00009.safetensors",
|
| 507 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.out_proj.weight": "model-00009-of-00009.safetensors",
|
| 508 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.q_proj.bias": "model-00009-of-00009.safetensors",
|
| 509 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.q_proj.weight": "model-00009-of-00009.safetensors",
|
| 510 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.v_proj.bias": "model-00009-of-00009.safetensors",
|
| 511 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.25.self_attn.v_proj.weight": "model-00009-of-00009.safetensors",
|
| 512 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 513 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 514 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 515 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 516 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 517 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 518 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 519 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 520 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 521 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 522 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 523 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 524 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 525 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 526 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 527 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.3.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 528 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 529 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 530 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 531 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 532 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 533 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 534 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 535 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 536 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 537 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 538 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 539 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 540 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 541 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 542 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 543 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.4.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 544 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 545 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 546 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 547 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 548 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 549 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 550 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 551 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 552 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 553 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 554 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 555 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 556 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 557 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 558 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 559 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.5.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 560 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 561 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 562 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 563 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 564 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 565 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 566 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 567 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 568 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 569 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 570 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 571 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 572 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 573 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 574 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 575 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.6.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 576 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 577 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 578 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 579 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 580 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 581 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 582 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 583 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 584 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 585 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 586 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 587 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 588 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 589 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 590 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 591 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.7.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 592 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 593 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 594 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 595 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 596 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 597 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 598 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 599 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 600 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 601 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 602 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 603 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 604 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 605 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 606 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 607 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.8.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 608 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm1.bias": "model-00008-of-00009.safetensors",
|
| 609 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm1.weight": "model-00008-of-00009.safetensors",
|
| 610 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm2.bias": "model-00008-of-00009.safetensors",
|
| 611 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.layer_norm2.weight": "model-00008-of-00009.safetensors",
|
| 612 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc1.bias": "model-00008-of-00009.safetensors",
|
| 613 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc1.weight": "model-00008-of-00009.safetensors",
|
| 614 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc2.bias": "model-00008-of-00009.safetensors",
|
| 615 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.mlp.fc2.weight": "model-00008-of-00009.safetensors",
|
| 616 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.k_proj.bias": "model-00008-of-00009.safetensors",
|
| 617 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.k_proj.weight": "model-00008-of-00009.safetensors",
|
| 618 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.out_proj.bias": "model-00008-of-00009.safetensors",
|
| 619 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.out_proj.weight": "model-00008-of-00009.safetensors",
|
| 620 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.q_proj.bias": "model-00008-of-00009.safetensors",
|
| 621 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.q_proj.weight": "model-00008-of-00009.safetensors",
|
| 622 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.v_proj.bias": "model-00008-of-00009.safetensors",
|
| 623 |
+
"model.vision_tower.vision_tower.vision_model.encoder.layers.9.self_attn.v_proj.weight": "model-00008-of-00009.safetensors",
|
| 624 |
+
"model.vision_tower.vision_tower.vision_model.post_layernorm.bias": "model-00009-of-00009.safetensors",
|
| 625 |
+
"model.vision_tower.vision_tower.vision_model.post_layernorm.weight": "model-00009-of-00009.safetensors"
|
| 626 |
+
}
|
| 627 |
+
}
|
modeling_imp.py
ADDED
|
@@ -0,0 +1,1521 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 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.cache_utils import Cache, DynamicCache
|
| 27 |
+
from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
|
| 28 |
+
from transformers import (
|
| 29 |
+
PretrainedConfig,
|
| 30 |
+
PreTrainedModel,
|
| 31 |
+
AutoConfig,
|
| 32 |
+
AutoModelForCausalLM
|
| 33 |
+
)
|
| 34 |
+
from transformers.modeling_utils import PreTrainedModel
|
| 35 |
+
from transformers.activations import ACT2FN
|
| 36 |
+
from transformers.modeling_outputs import (
|
| 37 |
+
BaseModelOutputWithPast,
|
| 38 |
+
CausalLMOutputWithPast,
|
| 39 |
+
SequenceClassifierOutputWithPast,
|
| 40 |
+
TokenClassifierOutput,
|
| 41 |
+
)
|
| 42 |
+
import sys
|
| 43 |
+
from .configuration_imp import Phi3Config, ImpPhi3Config
|
| 44 |
+
from .vision_encoder import VisionTower
|
| 45 |
+
# from .vision_encoder import CLIPVisionTower
|
| 46 |
+
|
| 47 |
+
try:
|
| 48 |
+
from flash_attn.bert_padding import pad_input, unpad_input
|
| 49 |
+
from flash_attn.layers.rotary import RotaryEmbedding as FlashRotaryEmbedding
|
| 50 |
+
from flash_attn.modules.mha import FlashCrossAttention, FlashSelfAttention
|
| 51 |
+
from flash_attn.ops.fused_dense import FusedDense
|
| 52 |
+
except:
|
| 53 |
+
pad_input, unpad_input = None, None
|
| 54 |
+
FlashRotaryEmbedding = None
|
| 55 |
+
FlashSelfAttention, FlashCrossAttention = None, None
|
| 56 |
+
FusedDense = None
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
@dataclass
|
| 60 |
+
class InferenceParams:
|
| 61 |
+
"""Inference parameters passed to model to efficiently calculate
|
| 62 |
+
and store context during inference.
|
| 63 |
+
|
| 64 |
+
Reference:
|
| 65 |
+
https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/utils/generation.py.
|
| 66 |
+
|
| 67 |
+
Args:
|
| 68 |
+
max_seqlen: Maximum sequence length.
|
| 69 |
+
max_batch_size: Maximum batch size.
|
| 70 |
+
seqlen_offset: Sequence length offset.
|
| 71 |
+
batch_size_offset: Batch size offset.
|
| 72 |
+
key_value_memory_dict: Key value memory dictionary.
|
| 73 |
+
lengths_per_sample: Lengths per sample.
|
| 74 |
+
|
| 75 |
+
"""
|
| 76 |
+
|
| 77 |
+
max_seqlen: int = field(metadata={"help": "Maximum sequence length."})
|
| 78 |
+
|
| 79 |
+
max_batch_size: int = field(metadata={"help": "Maximum batch size."})
|
| 80 |
+
|
| 81 |
+
seqlen_offset: int = field(default=0, metadata={"help": "Sequence length offset."})
|
| 82 |
+
|
| 83 |
+
batch_size_offset: int = field(default=0, metadata={"help": "Batch size offset."})
|
| 84 |
+
|
| 85 |
+
key_value_memory_dict: Dict[str, Any] = field(
|
| 86 |
+
default_factory=dict, metadata={"help": "Key value memory dictionary."}
|
| 87 |
+
)
|
| 88 |
+
|
| 89 |
+
lengths_per_sample: torch.Tensor = field(default=None, metadata={"help": "Lengths per sample."})
|
| 90 |
+
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
|
| 94 |
+
# Copied from transformers.models.gemma.modeling_gemma.GemmaRotaryEmbedding with gemma->phi3, Gemma->Phi3
|
| 95 |
+
class Phi3RotaryEmbedding(nn.Module):
|
| 96 |
+
def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None):
|
| 97 |
+
super().__init__()
|
| 98 |
+
|
| 99 |
+
self.dim = dim
|
| 100 |
+
self.max_position_embeddings = max_position_embeddings
|
| 101 |
+
self.base = base
|
| 102 |
+
self.register_buffer("inv_freq", None, persistent=False)
|
| 103 |
+
|
| 104 |
+
@torch.no_grad()
|
| 105 |
+
def forward(self, x, position_ids, seq_len=None):
|
| 106 |
+
# x: [bs, num_attention_heads, seq_len, head_size]
|
| 107 |
+
if self.inv_freq is None:
|
| 108 |
+
self.inv_freq = 1.0 / (
|
| 109 |
+
self.base ** (torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim)
|
| 110 |
+
)
|
| 111 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
| 112 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 113 |
+
# Force float32 since bfloat16 loses precision on long contexts
|
| 114 |
+
# See https://github.com/huggingface/transformers/pull/29285
|
| 115 |
+
device_type = x.device.type
|
| 116 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
| 117 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
| 118 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 119 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 120 |
+
cos = emb.cos()
|
| 121 |
+
sin = emb.sin()
|
| 122 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 123 |
+
|
| 124 |
+
|
| 125 |
+
class Phi3SuScaledRotaryEmbedding(Phi3RotaryEmbedding):
|
| 126 |
+
def __init__(self, dim, config, device=None):
|
| 127 |
+
super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
|
| 128 |
+
|
| 129 |
+
self.short_factor = config.rope_scaling["short_factor"]
|
| 130 |
+
self.long_factor = config.rope_scaling["long_factor"]
|
| 131 |
+
self.original_max_position_embeddings = config.original_max_position_embeddings
|
| 132 |
+
|
| 133 |
+
@torch.no_grad()
|
| 134 |
+
def forward(self, x, position_ids, seq_len=None):
|
| 135 |
+
seq_len = torch.max(position_ids) + 1
|
| 136 |
+
if seq_len > self.original_max_position_embeddings:
|
| 137 |
+
ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
|
| 138 |
+
else:
|
| 139 |
+
ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
|
| 140 |
+
|
| 141 |
+
inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
|
| 142 |
+
self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
|
| 143 |
+
|
| 144 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
| 145 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 146 |
+
|
| 147 |
+
# Force float32 since bfloat16 loses precision on long contexts
|
| 148 |
+
# See https://github.com/huggingface/transformers/pull/29285
|
| 149 |
+
device_type = x.device.type
|
| 150 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
| 151 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
| 152 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 153 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 154 |
+
|
| 155 |
+
scale = self.max_position_embeddings / self.original_max_position_embeddings
|
| 156 |
+
if scale <= 1.0:
|
| 157 |
+
scaling_factor = 1.0
|
| 158 |
+
else:
|
| 159 |
+
scaling_factor = math.sqrt(1 + math.log(scale) / math.log(self.original_max_position_embeddings))
|
| 160 |
+
|
| 161 |
+
cos = emb.cos() * scaling_factor
|
| 162 |
+
sin = emb.sin() * scaling_factor
|
| 163 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
class Phi3YarnScaledRotaryEmbedding(Phi3RotaryEmbedding):
|
| 167 |
+
def __init__(self, dim, config, device=None):
|
| 168 |
+
super().__init__(dim, config.max_position_embeddings, config.rope_theta, device)
|
| 169 |
+
|
| 170 |
+
self.short_factor = config.rope_scaling["short_factor"]
|
| 171 |
+
self.long_factor = config.rope_scaling["long_factor"]
|
| 172 |
+
self.original_max_position_embeddings = config.original_max_position_embeddings
|
| 173 |
+
|
| 174 |
+
@torch.no_grad()
|
| 175 |
+
def forward(self, x, position_ids, seq_len=None):
|
| 176 |
+
seq_len = torch.max(position_ids) + 1
|
| 177 |
+
if seq_len > self.original_max_position_embeddings:
|
| 178 |
+
ext_factors = torch.tensor(self.long_factor, dtype=torch.float32, device=x.device)
|
| 179 |
+
else:
|
| 180 |
+
ext_factors = torch.tensor(self.short_factor, dtype=torch.float32, device=x.device)
|
| 181 |
+
|
| 182 |
+
inv_freq_shape = torch.arange(0, self.dim, 2, dtype=torch.int64, device=x.device).float() / self.dim
|
| 183 |
+
self.inv_freq = 1.0 / (ext_factors * self.base**inv_freq_shape)
|
| 184 |
+
|
| 185 |
+
inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
|
| 186 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 187 |
+
|
| 188 |
+
# Force float32 since bfloat16 loses precision on long contexts
|
| 189 |
+
# See https://github.com/huggingface/transformers/pull/29285
|
| 190 |
+
device_type = x.device.type
|
| 191 |
+
device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
|
| 192 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
| 193 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 194 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 195 |
+
|
| 196 |
+
scale = self.max_position_embeddings / self.original_max_position_embeddings
|
| 197 |
+
if scale <= 1.0:
|
| 198 |
+
scaling_factor = 1.0
|
| 199 |
+
else:
|
| 200 |
+
scaling_factor = 0.1 * math.log(scale) + 1.0
|
| 201 |
+
|
| 202 |
+
cos = emb.cos() * scaling_factor
|
| 203 |
+
sin = emb.sin() * scaling_factor
|
| 204 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 205 |
+
|
| 206 |
+
|
| 207 |
+
# Copied from transformers.models.llama.modeling_llama.rotate_half
|
| 208 |
+
def rotate_half(x):
|
| 209 |
+
"""Rotates half the hidden dims of the input."""
|
| 210 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 211 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 212 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb
|
| 216 |
+
def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
|
| 217 |
+
"""Applies Rotary Position Embedding to the query and key tensors.
|
| 218 |
+
|
| 219 |
+
Args:
|
| 220 |
+
q (`torch.Tensor`): The query tensor.
|
| 221 |
+
k (`torch.Tensor`): The key tensor.
|
| 222 |
+
cos (`torch.Tensor`): The cosine part of the rotary embedding.
|
| 223 |
+
sin (`torch.Tensor`): The sine part of the rotary embedding.
|
| 224 |
+
position_ids (`torch.Tensor`, *optional*):
|
| 225 |
+
Deprecated and unused.
|
| 226 |
+
unsqueeze_dim (`int`, *optional*, defaults to 1):
|
| 227 |
+
The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
|
| 228 |
+
sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
|
| 229 |
+
that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
|
| 230 |
+
k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
|
| 231 |
+
cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
|
| 232 |
+
the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
|
| 233 |
+
Returns:
|
| 234 |
+
`tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
|
| 235 |
+
"""
|
| 236 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
| 237 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
| 238 |
+
q_embed = (q * cos) + (rotate_half(q) * sin)
|
| 239 |
+
k_embed = (k * cos) + (rotate_half(k) * sin)
|
| 240 |
+
return q_embed, k_embed
|
| 241 |
+
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
class Phi3MLP(nn.Module):
|
| 245 |
+
def __init__(self, config):
|
| 246 |
+
super().__init__()
|
| 247 |
+
|
| 248 |
+
self.config = config
|
| 249 |
+
self.gate_up_proj = nn.Linear(config.hidden_size, 2 * config.intermediate_size, bias=False)
|
| 250 |
+
self.down_proj = nn.Linear(config.intermediate_size, config.hidden_size, bias=False)
|
| 251 |
+
|
| 252 |
+
self.activation_fn = ACT2FN[config.hidden_act]
|
| 253 |
+
|
| 254 |
+
def forward(self, hidden_states: torch.FloatTensor) -> torch.FloatTensor:
|
| 255 |
+
up_states = self.gate_up_proj(hidden_states)
|
| 256 |
+
|
| 257 |
+
gate, up_states = up_states.chunk(2, dim=-1)
|
| 258 |
+
up_states = up_states * self.activation_fn(gate)
|
| 259 |
+
|
| 260 |
+
return self.down_proj(up_states)
|
| 261 |
+
|
| 262 |
+
class Phi3RMSNorm(nn.Module):
|
| 263 |
+
def __init__(self, hidden_size, eps=1e-6):
|
| 264 |
+
"""
|
| 265 |
+
Phi3RMSNorm is equivalent to T5LayerNorm
|
| 266 |
+
"""
|
| 267 |
+
super().__init__()
|
| 268 |
+
self.weight = nn.Parameter(torch.ones(hidden_size))
|
| 269 |
+
self.variance_epsilon = eps
|
| 270 |
+
|
| 271 |
+
def forward(self, hidden_states):
|
| 272 |
+
input_dtype = hidden_states.dtype
|
| 273 |
+
hidden_states = hidden_states.to(torch.float32)
|
| 274 |
+
variance = hidden_states.pow(2).mean(-1, keepdim=True)
|
| 275 |
+
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
|
| 276 |
+
return self.weight * hidden_states.to(input_dtype)
|
| 277 |
+
|
| 278 |
+
|
| 279 |
+
|
| 280 |
+
# Copied from transformers.models.llama.modeling_llama.repeat_kv with llama->phi
|
| 281 |
+
def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
|
| 282 |
+
"""
|
| 283 |
+
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
|
| 284 |
+
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
|
| 285 |
+
"""
|
| 286 |
+
batch, num_key_value_heads, slen, head_dim = hidden_states.shape
|
| 287 |
+
if n_rep == 1:
|
| 288 |
+
return hidden_states
|
| 289 |
+
hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
|
| 290 |
+
return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
|
| 291 |
+
|
| 292 |
+
|
| 293 |
+
|
| 294 |
+
class Phi3Attention(nn.Module):
|
| 295 |
+
"""Multi-headed attention from 'Attention Is All You Need' paper"""
|
| 296 |
+
|
| 297 |
+
def __init__(self, config: Phi3Config, layer_idx: Optional[int] = None):
|
| 298 |
+
super().__init__()
|
| 299 |
+
self.config = config
|
| 300 |
+
self.layer_idx = layer_idx
|
| 301 |
+
if layer_idx is None:
|
| 302 |
+
# logger.warning_once(
|
| 303 |
+
# f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will "
|
| 304 |
+
# "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` "
|
| 305 |
+
# "when creating this class."
|
| 306 |
+
# )
|
| 307 |
+
pass
|
| 308 |
+
|
| 309 |
+
self.attention_dropout = config.attention_dropout
|
| 310 |
+
self.hidden_size = config.hidden_size
|
| 311 |
+
self.num_heads = config.num_attention_heads
|
| 312 |
+
self.head_dim = self.hidden_size // self.num_heads
|
| 313 |
+
self.num_key_value_heads = config.num_key_value_heads
|
| 314 |
+
self.num_key_value_groups = self.num_heads // self.num_key_value_heads
|
| 315 |
+
self.max_position_embeddings = config.max_position_embeddings
|
| 316 |
+
self.original_max_position_embeddings = config.original_max_position_embeddings
|
| 317 |
+
self.rope_theta = config.rope_theta
|
| 318 |
+
self.rope_scaling = config.rope_scaling
|
| 319 |
+
self.is_causal = True
|
| 320 |
+
|
| 321 |
+
if (self.head_dim * self.num_heads) != self.hidden_size:
|
| 322 |
+
raise ValueError(
|
| 323 |
+
f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
|
| 324 |
+
f" and `num_heads`: {self.num_heads})."
|
| 325 |
+
)
|
| 326 |
+
|
| 327 |
+
op_size = self.num_heads * self.head_dim + 2 * (self.num_key_value_heads * self.head_dim)
|
| 328 |
+
self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
|
| 329 |
+
self.qkv_proj = nn.Linear(self.hidden_size, op_size, bias=False)
|
| 330 |
+
self._init_rope()
|
| 331 |
+
|
| 332 |
+
def _init_rope(self):
|
| 333 |
+
if self.rope_scaling is None:
|
| 334 |
+
self.rotary_emb = Phi3RotaryEmbedding(
|
| 335 |
+
self.head_dim,
|
| 336 |
+
max_position_embeddings=self.max_position_embeddings,
|
| 337 |
+
base=self.rope_theta,
|
| 338 |
+
)
|
| 339 |
+
else:
|
| 340 |
+
scaling_type = self.config.rope_scaling["type"]
|
| 341 |
+
if scaling_type == "su":
|
| 342 |
+
self.rotary_emb = Phi3SuScaledRotaryEmbedding(self.head_dim, self.config)
|
| 343 |
+
elif scaling_type == "yarn":
|
| 344 |
+
self.rotary_emb = Phi3YarnScaledRotaryEmbedding(self.head_dim, self.config)
|
| 345 |
+
else:
|
| 346 |
+
raise ValueError(f"Unknown RoPE scaling type {scaling_type}")
|
| 347 |
+
|
| 348 |
+
def forward(
|
| 349 |
+
self,
|
| 350 |
+
hidden_states: torch.Tensor,
|
| 351 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 352 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 353 |
+
past_key_value: Optional[Cache] = None,
|
| 354 |
+
output_attentions: bool = False,
|
| 355 |
+
use_cache: bool = False,
|
| 356 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 357 |
+
# logger.warning_once("You are not running the flash-attention implementation, expect numerical differences.")
|
| 358 |
+
|
| 359 |
+
bsz, q_len, _ = hidden_states.size()
|
| 360 |
+
|
| 361 |
+
qkv = self.qkv_proj(hidden_states)
|
| 362 |
+
query_pos = self.num_heads * self.head_dim
|
| 363 |
+
query_states = qkv[..., :query_pos]
|
| 364 |
+
key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
|
| 365 |
+
value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
|
| 366 |
+
|
| 367 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 368 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 369 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 370 |
+
|
| 371 |
+
kv_seq_len = key_states.shape[-2]
|
| 372 |
+
if past_key_value is not None:
|
| 373 |
+
if self.layer_idx is None:
|
| 374 |
+
raise ValueError(
|
| 375 |
+
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 376 |
+
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 377 |
+
"with a layer index."
|
| 378 |
+
)
|
| 379 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 380 |
+
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
|
| 381 |
+
|
| 382 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 383 |
+
|
| 384 |
+
if past_key_value is not None:
|
| 385 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 386 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 387 |
+
|
| 388 |
+
# repeat k/v heads if n_kv_heads < n_heads
|
| 389 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 390 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 391 |
+
|
| 392 |
+
attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
|
| 393 |
+
|
| 394 |
+
if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
|
| 395 |
+
raise ValueError(
|
| 396 |
+
f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is"
|
| 397 |
+
f" {attn_weights.size()}"
|
| 398 |
+
)
|
| 399 |
+
|
| 400 |
+
if attention_mask is not None:
|
| 401 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
| 402 |
+
raise ValueError(
|
| 403 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 404 |
+
)
|
| 405 |
+
attn_weights = attn_weights + attention_mask
|
| 406 |
+
|
| 407 |
+
# upcast attention to fp32
|
| 408 |
+
attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(value_states.dtype)
|
| 409 |
+
attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
|
| 410 |
+
|
| 411 |
+
attn_output = torch.matmul(attn_weights, value_states)
|
| 412 |
+
|
| 413 |
+
if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
|
| 414 |
+
raise ValueError(
|
| 415 |
+
f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
|
| 416 |
+
f" {attn_output.size()}"
|
| 417 |
+
)
|
| 418 |
+
|
| 419 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 420 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
|
| 421 |
+
|
| 422 |
+
attn_output = self.o_proj(attn_output)
|
| 423 |
+
|
| 424 |
+
if not output_attentions:
|
| 425 |
+
attn_weights = None
|
| 426 |
+
|
| 427 |
+
return attn_output, attn_weights, past_key_value
|
| 428 |
+
|
| 429 |
+
|
| 430 |
+
class Phi3FlashAttention2(Phi3Attention):
|
| 431 |
+
"""
|
| 432 |
+
Phi-3 flash attention module. This module inherits from `Phi3Attention` as the weights of the module stays
|
| 433 |
+
untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
|
| 434 |
+
flash attention and deal with padding tokens in case the input contains any of them.
|
| 435 |
+
"""
|
| 436 |
+
|
| 437 |
+
# Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__
|
| 438 |
+
def __init__(self, *args, **kwargs):
|
| 439 |
+
super().__init__(*args, **kwargs)
|
| 440 |
+
|
| 441 |
+
# TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
|
| 442 |
+
# flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
|
| 443 |
+
# Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
|
| 444 |
+
self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
|
| 445 |
+
|
| 446 |
+
def forward(
|
| 447 |
+
self,
|
| 448 |
+
hidden_states: torch.Tensor,
|
| 449 |
+
attention_mask: Optional[torch.LongTensor] = None,
|
| 450 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 451 |
+
past_key_value: Optional[Cache] = None,
|
| 452 |
+
output_attentions: bool = False,
|
| 453 |
+
use_cache: bool = False,
|
| 454 |
+
**kwargs,
|
| 455 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 456 |
+
# Phi3FlashAttention2 attention does not support output_attentions
|
| 457 |
+
|
| 458 |
+
if not _flash_supports_window_size:
|
| 459 |
+
# logger.warning_once(
|
| 460 |
+
# "The current flash attention version does not support sliding window attention. Please use `attn_implementation='eager'` or upgrade flash-attn library."
|
| 461 |
+
# )
|
| 462 |
+
raise ValueError("The current flash attention version does not support sliding window attention.")
|
| 463 |
+
|
| 464 |
+
output_attentions = False
|
| 465 |
+
|
| 466 |
+
if "padding_mask" in kwargs:
|
| 467 |
+
warnings.warn(
|
| 468 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 469 |
+
)
|
| 470 |
+
|
| 471 |
+
# overwrite attention_mask with padding_mask
|
| 472 |
+
attention_mask = kwargs.pop("padding_mask")
|
| 473 |
+
|
| 474 |
+
bsz, q_len, _ = hidden_states.size()
|
| 475 |
+
|
| 476 |
+
qkv = self.qkv_proj(hidden_states)
|
| 477 |
+
query_pos = self.num_heads * self.head_dim
|
| 478 |
+
query_states = qkv[..., :query_pos]
|
| 479 |
+
key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
|
| 480 |
+
value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
|
| 481 |
+
|
| 482 |
+
# Flash attention requires the input to have the shape
|
| 483 |
+
# batch_size x seq_length x head_dim x hidden_dim
|
| 484 |
+
# therefore we just need to keep the original shape
|
| 485 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 486 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 487 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 488 |
+
|
| 489 |
+
kv_seq_len = key_states.shape[-2]
|
| 490 |
+
if past_key_value is not None:
|
| 491 |
+
if self.layer_idx is None:
|
| 492 |
+
raise ValueError(
|
| 493 |
+
f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
|
| 494 |
+
"for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
|
| 495 |
+
"with a layer index."
|
| 496 |
+
)
|
| 497 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 498 |
+
|
| 499 |
+
# Because the input can be padded, the absolute sequence length depends on the max position id.
|
| 500 |
+
rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1
|
| 501 |
+
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=rotary_seq_len)
|
| 502 |
+
|
| 503 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 504 |
+
|
| 505 |
+
use_sliding_windows = (
|
| 506 |
+
_flash_supports_window_size
|
| 507 |
+
and getattr(self.config, "sliding_window", None) is not None
|
| 508 |
+
and kv_seq_len > self.config.sliding_window
|
| 509 |
+
)
|
| 510 |
+
|
| 511 |
+
if past_key_value is not None:
|
| 512 |
+
# Activate slicing cache only if the config has a value `sliding_windows` attribute
|
| 513 |
+
cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
|
| 514 |
+
if (
|
| 515 |
+
getattr(self.config, "sliding_window", None) is not None
|
| 516 |
+
and kv_seq_len > self.config.sliding_window
|
| 517 |
+
and cache_has_contents
|
| 518 |
+
):
|
| 519 |
+
slicing_tokens = 1 - self.config.sliding_window
|
| 520 |
+
|
| 521 |
+
past_key = past_key_value[self.layer_idx][0]
|
| 522 |
+
past_value = past_key_value[self.layer_idx][1]
|
| 523 |
+
|
| 524 |
+
past_key = past_key[:, :, slicing_tokens:, :].contiguous()
|
| 525 |
+
past_value = past_value[:, :, slicing_tokens:, :].contiguous()
|
| 526 |
+
|
| 527 |
+
if past_key.shape[-2] != self.config.sliding_window - 1:
|
| 528 |
+
raise ValueError(
|
| 529 |
+
f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
|
| 530 |
+
f" {past_key.shape}"
|
| 531 |
+
)
|
| 532 |
+
|
| 533 |
+
if attention_mask is not None:
|
| 534 |
+
attention_mask = attention_mask[:, slicing_tokens:]
|
| 535 |
+
attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1)
|
| 536 |
+
|
| 537 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 538 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 539 |
+
|
| 540 |
+
# repeat k/v heads if n_kv_heads < n_heads
|
| 541 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 542 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 543 |
+
|
| 544 |
+
attn_dropout = self.attention_dropout if self.training else 0.0
|
| 545 |
+
|
| 546 |
+
# In PEFT, usually we cast the layer norms in float32 for training stability reasons
|
| 547 |
+
# therefore the input hidden states gets silently casted in float32. Hence, we need
|
| 548 |
+
# cast them back in the correct dtype just to be sure everything works as expected.
|
| 549 |
+
# This might slowdown training & inference so it is recommended to not cast the LayerNorms
|
| 550 |
+
# in fp32.
|
| 551 |
+
|
| 552 |
+
if query_states.dtype == torch.float32:
|
| 553 |
+
if torch.is_autocast_enabled():
|
| 554 |
+
target_dtype = torch.get_autocast_gpu_dtype()
|
| 555 |
+
# Handle the case where the model is quantized
|
| 556 |
+
elif hasattr(self.config, "_pre_quantization_dtype"):
|
| 557 |
+
target_dtype = self.config._pre_quantization_dtype
|
| 558 |
+
else:
|
| 559 |
+
target_dtype = self.qkv_proj.weight.dtype
|
| 560 |
+
|
| 561 |
+
# logger.warning_once(
|
| 562 |
+
# f"The input hidden states seems to be silently casted in float32, this might be related to"
|
| 563 |
+
# f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
|
| 564 |
+
# f" {target_dtype}."
|
| 565 |
+
# )
|
| 566 |
+
|
| 567 |
+
query_states = query_states.to(target_dtype)
|
| 568 |
+
key_states = key_states.to(target_dtype)
|
| 569 |
+
value_states = value_states.to(target_dtype)
|
| 570 |
+
|
| 571 |
+
# Reashape to the expected shape for Flash Attention
|
| 572 |
+
query_states = query_states.transpose(1, 2)
|
| 573 |
+
key_states = key_states.transpose(1, 2)
|
| 574 |
+
value_states = value_states.transpose(1, 2)
|
| 575 |
+
|
| 576 |
+
attn_output = self._flash_attention_forward(
|
| 577 |
+
query_states,
|
| 578 |
+
key_states,
|
| 579 |
+
value_states,
|
| 580 |
+
attention_mask,
|
| 581 |
+
q_len,
|
| 582 |
+
dropout=attn_dropout,
|
| 583 |
+
use_sliding_windows=use_sliding_windows,
|
| 584 |
+
)
|
| 585 |
+
|
| 586 |
+
attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
|
| 587 |
+
attn_output = self.o_proj(attn_output)
|
| 588 |
+
|
| 589 |
+
if not output_attentions:
|
| 590 |
+
attn_weights = None
|
| 591 |
+
|
| 592 |
+
return attn_output, attn_weights, past_key_value
|
| 593 |
+
|
| 594 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._flash_attention_forward
|
| 595 |
+
def _flash_attention_forward(
|
| 596 |
+
self,
|
| 597 |
+
query_states,
|
| 598 |
+
key_states,
|
| 599 |
+
value_states,
|
| 600 |
+
attention_mask,
|
| 601 |
+
query_length,
|
| 602 |
+
dropout=0.0,
|
| 603 |
+
softmax_scale=None,
|
| 604 |
+
use_sliding_windows=False,
|
| 605 |
+
):
|
| 606 |
+
"""
|
| 607 |
+
Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
|
| 608 |
+
first unpad the input, then computes the attention scores and pad the final attention scores.
|
| 609 |
+
|
| 610 |
+
Args:
|
| 611 |
+
query_states (`torch.Tensor`):
|
| 612 |
+
Input query states to be passed to Flash Attention API
|
| 613 |
+
key_states (`torch.Tensor`):
|
| 614 |
+
Input key states to be passed to Flash Attention API
|
| 615 |
+
value_states (`torch.Tensor`):
|
| 616 |
+
Input value states to be passed to Flash Attention API
|
| 617 |
+
attention_mask (`torch.Tensor`):
|
| 618 |
+
The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
|
| 619 |
+
position of padding tokens and 1 for the position of non-padding tokens.
|
| 620 |
+
dropout (`float`):
|
| 621 |
+
Attention dropout
|
| 622 |
+
softmax_scale (`float`, *optional*):
|
| 623 |
+
The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
|
| 624 |
+
use_sliding_windows (`bool`, *optional*):
|
| 625 |
+
Whether to activate sliding window attention.
|
| 626 |
+
"""
|
| 627 |
+
if not self._flash_attn_uses_top_left_mask:
|
| 628 |
+
causal = self.is_causal
|
| 629 |
+
else:
|
| 630 |
+
# TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__.
|
| 631 |
+
causal = self.is_causal and query_length != 1
|
| 632 |
+
|
| 633 |
+
# Contains at least one padding token in the sequence
|
| 634 |
+
if attention_mask is not None:
|
| 635 |
+
batch_size = query_states.shape[0]
|
| 636 |
+
query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
|
| 637 |
+
query_states, key_states, value_states, attention_mask, query_length
|
| 638 |
+
)
|
| 639 |
+
|
| 640 |
+
cu_seqlens_q, cu_seqlens_k = cu_seq_lens
|
| 641 |
+
max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
|
| 642 |
+
|
| 643 |
+
if not use_sliding_windows:
|
| 644 |
+
attn_output_unpad = flash_attn_varlen_func(
|
| 645 |
+
query_states,
|
| 646 |
+
key_states,
|
| 647 |
+
value_states,
|
| 648 |
+
cu_seqlens_q=cu_seqlens_q,
|
| 649 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 650 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
| 651 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
| 652 |
+
dropout_p=dropout,
|
| 653 |
+
softmax_scale=softmax_scale,
|
| 654 |
+
causal=causal,
|
| 655 |
+
)
|
| 656 |
+
else:
|
| 657 |
+
attn_output_unpad = flash_attn_varlen_func(
|
| 658 |
+
query_states,
|
| 659 |
+
key_states,
|
| 660 |
+
value_states,
|
| 661 |
+
cu_seqlens_q=cu_seqlens_q,
|
| 662 |
+
cu_seqlens_k=cu_seqlens_k,
|
| 663 |
+
max_seqlen_q=max_seqlen_in_batch_q,
|
| 664 |
+
max_seqlen_k=max_seqlen_in_batch_k,
|
| 665 |
+
dropout_p=dropout,
|
| 666 |
+
softmax_scale=softmax_scale,
|
| 667 |
+
causal=causal,
|
| 668 |
+
window_size=(self.config.sliding_window, self.config.sliding_window),
|
| 669 |
+
)
|
| 670 |
+
|
| 671 |
+
attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
|
| 672 |
+
else:
|
| 673 |
+
if not use_sliding_windows:
|
| 674 |
+
attn_output = flash_attn_func(
|
| 675 |
+
query_states,
|
| 676 |
+
key_states,
|
| 677 |
+
value_states,
|
| 678 |
+
dropout,
|
| 679 |
+
softmax_scale=softmax_scale,
|
| 680 |
+
causal=causal,
|
| 681 |
+
)
|
| 682 |
+
else:
|
| 683 |
+
attn_output = flash_attn_func(
|
| 684 |
+
query_states,
|
| 685 |
+
key_states,
|
| 686 |
+
value_states,
|
| 687 |
+
dropout,
|
| 688 |
+
softmax_scale=softmax_scale,
|
| 689 |
+
causal=causal,
|
| 690 |
+
window_size=(self.config.sliding_window, self.config.sliding_window),
|
| 691 |
+
)
|
| 692 |
+
|
| 693 |
+
return attn_output
|
| 694 |
+
|
| 695 |
+
# Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input
|
| 696 |
+
def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
|
| 697 |
+
batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape
|
| 698 |
+
|
| 699 |
+
# On the first iteration we need to properly re-create the padding mask
|
| 700 |
+
# by slicing it on the proper place
|
| 701 |
+
if kv_seq_len != attention_mask.shape[-1]:
|
| 702 |
+
attention_mask_num_tokens = attention_mask.shape[-1]
|
| 703 |
+
attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :]
|
| 704 |
+
|
| 705 |
+
indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
|
| 706 |
+
|
| 707 |
+
key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
|
| 708 |
+
value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k)
|
| 709 |
+
|
| 710 |
+
if query_length == kv_seq_len:
|
| 711 |
+
query_layer = index_first_axis(
|
| 712 |
+
query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k
|
| 713 |
+
)
|
| 714 |
+
cu_seqlens_q = cu_seqlens_k
|
| 715 |
+
max_seqlen_in_batch_q = max_seqlen_in_batch_k
|
| 716 |
+
indices_q = indices_k
|
| 717 |
+
elif query_length == 1:
|
| 718 |
+
max_seqlen_in_batch_q = 1
|
| 719 |
+
cu_seqlens_q = torch.arange(
|
| 720 |
+
batch_size + 1, dtype=torch.int32, device=query_layer.device
|
| 721 |
+
) # There is a memcpy here, that is very bad.
|
| 722 |
+
indices_q = cu_seqlens_q[:-1]
|
| 723 |
+
query_layer = query_layer.squeeze(1)
|
| 724 |
+
else:
|
| 725 |
+
# The -q_len: slice assumes left padding.
|
| 726 |
+
attention_mask = attention_mask[:, -query_length:]
|
| 727 |
+
query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
|
| 728 |
+
|
| 729 |
+
return (
|
| 730 |
+
query_layer,
|
| 731 |
+
key_layer,
|
| 732 |
+
value_layer,
|
| 733 |
+
indices_q,
|
| 734 |
+
(cu_seqlens_q, cu_seqlens_k),
|
| 735 |
+
(max_seqlen_in_batch_q, max_seqlen_in_batch_k),
|
| 736 |
+
)
|
| 737 |
+
|
| 738 |
+
|
| 739 |
+
# copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Phi3
|
| 740 |
+
# TODO @Arthur no longer copied from LLama after static cache
|
| 741 |
+
class Phi3SdpaAttention(Phi3Attention):
|
| 742 |
+
"""
|
| 743 |
+
Phi3 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
|
| 744 |
+
`Phi3Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
|
| 745 |
+
SDPA API.
|
| 746 |
+
"""
|
| 747 |
+
|
| 748 |
+
# Adapted from Phi3Attention.forward
|
| 749 |
+
def forward(
|
| 750 |
+
self,
|
| 751 |
+
hidden_states: torch.Tensor,
|
| 752 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 753 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 754 |
+
past_key_value: Optional[Cache] = None,
|
| 755 |
+
output_attentions: bool = False,
|
| 756 |
+
use_cache: bool = False,
|
| 757 |
+
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
|
| 758 |
+
if output_attentions:
|
| 759 |
+
# TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented.
|
| 760 |
+
# logger.warning_once(
|
| 761 |
+
# "Phi3Model is using Phi3SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
|
| 762 |
+
# 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
|
| 763 |
+
# )
|
| 764 |
+
return super().forward(
|
| 765 |
+
hidden_states=hidden_states,
|
| 766 |
+
attention_mask=attention_mask,
|
| 767 |
+
position_ids=position_ids,
|
| 768 |
+
past_key_value=past_key_value,
|
| 769 |
+
output_attentions=output_attentions,
|
| 770 |
+
use_cache=use_cache,
|
| 771 |
+
)
|
| 772 |
+
|
| 773 |
+
bsz, q_len, _ = hidden_states.size()
|
| 774 |
+
|
| 775 |
+
qkv = self.qkv_proj(hidden_states)
|
| 776 |
+
query_pos = self.num_heads * self.head_dim
|
| 777 |
+
query_states = qkv[..., :query_pos]
|
| 778 |
+
key_states = qkv[..., query_pos : query_pos + self.num_key_value_heads * self.head_dim]
|
| 779 |
+
value_states = qkv[..., query_pos + self.num_key_value_heads * self.head_dim :]
|
| 780 |
+
|
| 781 |
+
query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
|
| 782 |
+
key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 783 |
+
value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
|
| 784 |
+
|
| 785 |
+
kv_seq_len = key_states.shape[-2]
|
| 786 |
+
if past_key_value is not None:
|
| 787 |
+
kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
|
| 788 |
+
cos, sin = self.rotary_emb(value_states, position_ids, seq_len=kv_seq_len)
|
| 789 |
+
|
| 790 |
+
query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
|
| 791 |
+
|
| 792 |
+
if past_key_value is not None:
|
| 793 |
+
cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models
|
| 794 |
+
key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
|
| 795 |
+
|
| 796 |
+
key_states = repeat_kv(key_states, self.num_key_value_groups)
|
| 797 |
+
value_states = repeat_kv(value_states, self.num_key_value_groups)
|
| 798 |
+
|
| 799 |
+
if attention_mask is not None:
|
| 800 |
+
if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
|
| 801 |
+
raise ValueError(
|
| 802 |
+
f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
|
| 803 |
+
)
|
| 804 |
+
|
| 805 |
+
# SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
|
| 806 |
+
# Reference: https://github.com/pytorch/pytorch/issues/112577.
|
| 807 |
+
if query_states.device.type == "cuda" and attention_mask is not None:
|
| 808 |
+
query_states = query_states.contiguous()
|
| 809 |
+
key_states = key_states.contiguous()
|
| 810 |
+
value_states = value_states.contiguous()
|
| 811 |
+
|
| 812 |
+
attn_output = torch.nn.functional.scaled_dot_product_attention(
|
| 813 |
+
query_states,
|
| 814 |
+
key_states,
|
| 815 |
+
value_states,
|
| 816 |
+
attn_mask=attention_mask,
|
| 817 |
+
dropout_p=self.attention_dropout if self.training else 0.0,
|
| 818 |
+
# The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
|
| 819 |
+
is_causal=self.is_causal and attention_mask is None and q_len > 1,
|
| 820 |
+
)
|
| 821 |
+
|
| 822 |
+
attn_output = attn_output.transpose(1, 2).contiguous()
|
| 823 |
+
attn_output = attn_output.view(bsz, q_len, self.hidden_size)
|
| 824 |
+
|
| 825 |
+
attn_output = self.o_proj(attn_output)
|
| 826 |
+
|
| 827 |
+
return attn_output, None, past_key_value
|
| 828 |
+
|
| 829 |
+
|
| 830 |
+
|
| 831 |
+
|
| 832 |
+
PHI3_ATTENTION_CLASSES = {
|
| 833 |
+
"eager": Phi3Attention,
|
| 834 |
+
"flash_attention_2": Phi3FlashAttention2,
|
| 835 |
+
"sdpa": Phi3SdpaAttention,
|
| 836 |
+
}
|
| 837 |
+
|
| 838 |
+
class Phi3DecoderLayer(nn.Module):
|
| 839 |
+
def __init__(self, config: Phi3Config, layer_idx: int):
|
| 840 |
+
super().__init__()
|
| 841 |
+
|
| 842 |
+
self.config = config
|
| 843 |
+
self.self_attn = PHI3_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx=layer_idx)
|
| 844 |
+
|
| 845 |
+
self.mlp = Phi3MLP(config)
|
| 846 |
+
self.input_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 847 |
+
|
| 848 |
+
self.resid_attn_dropout = nn.Dropout(config.resid_pdrop)
|
| 849 |
+
self.resid_mlp_dropout = nn.Dropout(config.resid_pdrop)
|
| 850 |
+
self.post_attention_layernorm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 851 |
+
|
| 852 |
+
def forward(
|
| 853 |
+
self,
|
| 854 |
+
hidden_states: torch.Tensor,
|
| 855 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 856 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 857 |
+
past_key_value: Optional[Tuple[torch.Tensor]] = None,
|
| 858 |
+
output_attentions: Optional[bool] = False,
|
| 859 |
+
use_cache: Optional[bool] = False,
|
| 860 |
+
**kwargs,
|
| 861 |
+
) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
|
| 862 |
+
if "padding_mask" in kwargs:
|
| 863 |
+
warnings.warn(
|
| 864 |
+
"Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`"
|
| 865 |
+
)
|
| 866 |
+
"""
|
| 867 |
+
Args:
|
| 868 |
+
hidden_states (`torch.FloatTensor`):
|
| 869 |
+
input to the layer of shape `(batch, seq_len, embed_dim)`
|
| 870 |
+
attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
|
| 871 |
+
`(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
|
| 872 |
+
position_ids (`torch.LongTensor` of shape `({0})`, *optional*):
|
| 873 |
+
Indices of positions of each input sequence tokens in the position embeddings. Selected in the range
|
| 874 |
+
`[0, config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
|
| 875 |
+
output_attentions (`bool`, *optional*):
|
| 876 |
+
Whether or not to return the attentions tensors of all attention layers. See `attentions` under
|
| 877 |
+
returned tensors for more detail.
|
| 878 |
+
use_cache (`bool`, *optional*):
|
| 879 |
+
If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
|
| 880 |
+
(see `past_key_values`).
|
| 881 |
+
past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
|
| 882 |
+
"""
|
| 883 |
+
|
| 884 |
+
residual = hidden_states
|
| 885 |
+
|
| 886 |
+
hidden_states = self.input_layernorm(hidden_states)
|
| 887 |
+
|
| 888 |
+
# Self Attention
|
| 889 |
+
attn_outputs, self_attn_weights, present_key_value = self.self_attn(
|
| 890 |
+
hidden_states=hidden_states,
|
| 891 |
+
attention_mask=attention_mask,
|
| 892 |
+
position_ids=position_ids,
|
| 893 |
+
past_key_value=past_key_value,
|
| 894 |
+
output_attentions=output_attentions,
|
| 895 |
+
use_cache=use_cache,
|
| 896 |
+
)
|
| 897 |
+
|
| 898 |
+
hidden_states = residual + self.resid_attn_dropout(attn_outputs)
|
| 899 |
+
|
| 900 |
+
residual = hidden_states
|
| 901 |
+
hidden_states = self.post_attention_layernorm(hidden_states)
|
| 902 |
+
hidden_states = self.mlp(hidden_states)
|
| 903 |
+
hidden_states = residual + self.resid_mlp_dropout(hidden_states)
|
| 904 |
+
|
| 905 |
+
outputs = (hidden_states,)
|
| 906 |
+
|
| 907 |
+
if output_attentions:
|
| 908 |
+
outputs += (self_attn_weights,)
|
| 909 |
+
|
| 910 |
+
if use_cache:
|
| 911 |
+
outputs += (present_key_value,)
|
| 912 |
+
|
| 913 |
+
return outputs
|
| 914 |
+
|
| 915 |
+
|
| 916 |
+
|
| 917 |
+
class Phi3PreTrainedModel(PreTrainedModel):
|
| 918 |
+
config_class = Phi3Config
|
| 919 |
+
base_model_prefix = "model"
|
| 920 |
+
supports_gradient_checkpointing = True
|
| 921 |
+
_no_split_modules = ["Phi3DecoderLayer"]
|
| 922 |
+
_skip_keys_device_placement = "past_key_values"
|
| 923 |
+
_supports_flash_attn_2 = True
|
| 924 |
+
_supports_sdpa = False
|
| 925 |
+
_supports_cache_class = True
|
| 926 |
+
|
| 927 |
+
_version = "0.0.5"
|
| 928 |
+
|
| 929 |
+
def _init_weights(self, module):
|
| 930 |
+
std = self.config.initializer_range
|
| 931 |
+
if isinstance(module, nn.Linear):
|
| 932 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 933 |
+
if module.bias is not None:
|
| 934 |
+
module.bias.data.zero_()
|
| 935 |
+
elif isinstance(module, nn.Embedding):
|
| 936 |
+
module.weight.data.normal_(mean=0.0, std=std)
|
| 937 |
+
if module.padding_idx is not None:
|
| 938 |
+
module.weight.data[module.padding_idx].zero_()
|
| 939 |
+
|
| 940 |
+
def prepare_inputs_for_generation(
|
| 941 |
+
self,
|
| 942 |
+
input_ids: torch.LongTensor,
|
| 943 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 944 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 945 |
+
attention_mask: Optional[Union[torch.LongTensor, torch.BoolTensor]] = None,
|
| 946 |
+
**kwargs,
|
| 947 |
+
) -> Dict[str, Any]:
|
| 948 |
+
if past_key_values is not None:
|
| 949 |
+
if isinstance(past_key_values, Cache):
|
| 950 |
+
cache_length = past_key_values.get_seq_length()
|
| 951 |
+
past_length = past_key_values.seen_tokens
|
| 952 |
+
max_cache_length = past_key_values.get_max_length()
|
| 953 |
+
else:
|
| 954 |
+
cache_length = past_length = past_key_values[0][0].shape[2]
|
| 955 |
+
max_cache_length = None
|
| 956 |
+
|
| 957 |
+
# Keep only the unprocessed tokens:
|
| 958 |
+
# 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where
|
| 959 |
+
# some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as
|
| 960 |
+
# input)
|
| 961 |
+
if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]:
|
| 962 |
+
input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :]
|
| 963 |
+
# 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard
|
| 964 |
+
# input_ids based on the past_length.
|
| 965 |
+
elif past_length < input_ids.shape[1]:
|
| 966 |
+
input_ids = input_ids[:, past_length:]
|
| 967 |
+
# 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens.
|
| 968 |
+
|
| 969 |
+
# If we are about to go beyond the maximum cache length, we need to crop the input attention mask.
|
| 970 |
+
if (
|
| 971 |
+
max_cache_length is not None
|
| 972 |
+
and attention_mask is not None
|
| 973 |
+
and cache_length + input_ids.shape[1] > max_cache_length
|
| 974 |
+
):
|
| 975 |
+
attention_mask = attention_mask[:, -max_cache_length:]
|
| 976 |
+
|
| 977 |
+
position_ids = kwargs.get("position_ids", None)
|
| 978 |
+
if attention_mask is not None and position_ids is None:
|
| 979 |
+
# create position_ids on the fly for batch generation
|
| 980 |
+
position_ids = attention_mask.long().cumsum(-1) - 1
|
| 981 |
+
position_ids.masked_fill_(attention_mask == 0, 1)
|
| 982 |
+
if past_key_values:
|
| 983 |
+
position_ids = position_ids[:, -input_ids.shape[1] :]
|
| 984 |
+
|
| 985 |
+
# if `inputs_embeds` are passed, we only want to use them in the 1st generation step
|
| 986 |
+
if inputs_embeds is not None and past_key_values is None:
|
| 987 |
+
model_inputs = {"inputs_embeds": inputs_embeds}
|
| 988 |
+
else:
|
| 989 |
+
model_inputs = {"input_ids": input_ids}
|
| 990 |
+
|
| 991 |
+
model_inputs.update(
|
| 992 |
+
{
|
| 993 |
+
"position_ids": position_ids,
|
| 994 |
+
"past_key_values": past_key_values,
|
| 995 |
+
"use_cache": kwargs.get("use_cache"),
|
| 996 |
+
"attention_mask": attention_mask,
|
| 997 |
+
}
|
| 998 |
+
)
|
| 999 |
+
return model_inputs
|
| 1000 |
+
|
| 1001 |
+
|
| 1002 |
+
|
| 1003 |
+
|
| 1004 |
+
class LlavaMetaModel(ABC):
|
| 1005 |
+
"""
|
| 1006 |
+
Define the APIs for building components that are related to image perceiving.
|
| 1007 |
+
This implementation is based on the implementation from the Llave project.
|
| 1008 |
+
"""
|
| 1009 |
+
|
| 1010 |
+
def get_vision_tower(self):
|
| 1011 |
+
vision_tower = getattr(self, 'vision_tower', None)
|
| 1012 |
+
if type(vision_tower) is list:
|
| 1013 |
+
vision_tower = vision_tower[0]
|
| 1014 |
+
return vision_tower
|
| 1015 |
+
|
| 1016 |
+
def build_vision_tower(self, config):
|
| 1017 |
+
self.vision_tower = VisionTower(config.vision_tower_cfg)
|
| 1018 |
+
# self.vision_tower = CLIPVisionTower(config.vision_tower_cfg)
|
| 1019 |
+
|
| 1020 |
+
def build_vision_projector(self, config):
|
| 1021 |
+
projector_type = getattr(config, 'mm_projector_type', 'linear')
|
| 1022 |
+
|
| 1023 |
+
if projector_type == 'linear':
|
| 1024 |
+
self.mm_projector = nn.Linear(config.mm_hidden_size, config.hidden_size)
|
| 1025 |
+
return
|
| 1026 |
+
|
| 1027 |
+
mlp_gelu_match = re.match(r'^mlp(\d+)x_gelu$', projector_type)
|
| 1028 |
+
if mlp_gelu_match:
|
| 1029 |
+
mlp_depth = int(mlp_gelu_match.group(1))
|
| 1030 |
+
modules = [nn.Linear(config.mm_hidden_size, config.hidden_size)]
|
| 1031 |
+
for _ in range(1, mlp_depth):
|
| 1032 |
+
modules.append(nn.GELU())
|
| 1033 |
+
modules.append(nn.Linear(config.hidden_size, config.hidden_size))
|
| 1034 |
+
self.mm_projector = nn.Sequential(*modules)
|
| 1035 |
+
return
|
| 1036 |
+
|
| 1037 |
+
if projector_type == 'identity':
|
| 1038 |
+
self.mm_projector = nn.Identity()
|
| 1039 |
+
return
|
| 1040 |
+
|
| 1041 |
+
raise ValueError(f'Unknown projector type: {projector_type}')
|
| 1042 |
+
|
| 1043 |
+
|
| 1044 |
+
class ImpPhi3Model(Phi3PreTrainedModel, LlavaMetaModel):
|
| 1045 |
+
"""Imp model. This implementation is modified from the implementation of Phi-2"""
|
| 1046 |
+
|
| 1047 |
+
config_class = ImpPhi3Config
|
| 1048 |
+
|
| 1049 |
+
def __init__(self, config: ImpPhi3Config) -> None:
|
| 1050 |
+
super().__init__(config)
|
| 1051 |
+
self.padding_idx = config.pad_token_id
|
| 1052 |
+
self.vocab_size = config.vocab_size
|
| 1053 |
+
|
| 1054 |
+
self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
|
| 1055 |
+
self.embed_dropout = nn.Dropout(config.embd_pdrop)
|
| 1056 |
+
self.layers = nn.ModuleList(
|
| 1057 |
+
[Phi3DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
|
| 1058 |
+
)
|
| 1059 |
+
self._attn_implementation = config._attn_implementation
|
| 1060 |
+
self.norm = Phi3RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
|
| 1061 |
+
|
| 1062 |
+
self.gradient_checkpointing = False
|
| 1063 |
+
|
| 1064 |
+
if hasattr(config, "mm_vision_tower"):
|
| 1065 |
+
self.build_vision_tower(config)
|
| 1066 |
+
self.build_vision_projector(config)
|
| 1067 |
+
# Initialize weights and apply final processing
|
| 1068 |
+
self.post_init()
|
| 1069 |
+
|
| 1070 |
+
|
| 1071 |
+
def get_input_embeddings(self) -> nn.Embedding:
|
| 1072 |
+
return self.embed_tokens
|
| 1073 |
+
|
| 1074 |
+
def set_input_embeddings(self, new_embeddings: nn.Embedding) -> None:
|
| 1075 |
+
self.embed_tokens = value
|
| 1076 |
+
|
| 1077 |
+
def forward(
|
| 1078 |
+
self,
|
| 1079 |
+
input_ids: torch.LongTensor = None,
|
| 1080 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1081 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1082 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1083 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1084 |
+
use_cache: Optional[bool] = None,
|
| 1085 |
+
output_attentions: Optional[bool] = None,
|
| 1086 |
+
output_hidden_states: Optional[bool] = None,
|
| 1087 |
+
return_dict: Optional[bool] = None,
|
| 1088 |
+
) -> Union[Tuple, BaseModelOutputWithPast]:
|
| 1089 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1090 |
+
output_hidden_states = (
|
| 1091 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1092 |
+
)
|
| 1093 |
+
use_cache = use_cache if use_cache is not None else self.config.use_cache
|
| 1094 |
+
|
| 1095 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1096 |
+
|
| 1097 |
+
# retrieve input_ids and inputs_embeds
|
| 1098 |
+
if input_ids is not None and inputs_embeds is not None:
|
| 1099 |
+
raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
|
| 1100 |
+
elif input_ids is not None:
|
| 1101 |
+
batch_size, seq_length = input_ids.shape[:2]
|
| 1102 |
+
elif inputs_embeds is not None:
|
| 1103 |
+
batch_size, seq_length = inputs_embeds.shape[:2]
|
| 1104 |
+
else:
|
| 1105 |
+
raise ValueError("You have to specify either input_ids or inputs_embeds")
|
| 1106 |
+
|
| 1107 |
+
past_key_values_length = 0
|
| 1108 |
+
|
| 1109 |
+
if self.gradient_checkpointing and self.training:
|
| 1110 |
+
if use_cache:
|
| 1111 |
+
# logger.warning_once(
|
| 1112 |
+
# "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
|
| 1113 |
+
# )
|
| 1114 |
+
use_cache = False
|
| 1115 |
+
|
| 1116 |
+
if use_cache:
|
| 1117 |
+
use_legacy_cache = not isinstance(past_key_values, Cache)
|
| 1118 |
+
if use_legacy_cache:
|
| 1119 |
+
past_key_values = DynamicCache.from_legacy_cache(past_key_values)
|
| 1120 |
+
past_key_values_length = past_key_values.get_usable_length(seq_length)
|
| 1121 |
+
|
| 1122 |
+
if position_ids is None:
|
| 1123 |
+
device = input_ids.device if input_ids is not None else inputs_embeds.device
|
| 1124 |
+
position_ids = torch.arange(
|
| 1125 |
+
past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device
|
| 1126 |
+
)
|
| 1127 |
+
position_ids = position_ids.unsqueeze(0).view(-1, seq_length)
|
| 1128 |
+
else:
|
| 1129 |
+
position_ids = position_ids.view(-1, seq_length).long()
|
| 1130 |
+
|
| 1131 |
+
if inputs_embeds is None:
|
| 1132 |
+
inputs_embeds = self.embed_tokens(input_ids)
|
| 1133 |
+
|
| 1134 |
+
if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache:
|
| 1135 |
+
is_padding_right = attention_mask[:, -1].sum().item() != batch_size
|
| 1136 |
+
if is_padding_right:
|
| 1137 |
+
raise ValueError(
|
| 1138 |
+
"You are attempting to perform batched generation with padding_side='right'"
|
| 1139 |
+
" this may lead to unexpected behaviour for Flash Attention version of Phi3. Make sure to "
|
| 1140 |
+
" call `tokenizer.padding_side = 'left'` before tokenizing the input. "
|
| 1141 |
+
)
|
| 1142 |
+
|
| 1143 |
+
if self._attn_implementation == "flash_attention_2":
|
| 1144 |
+
# 2d mask is passed through the layers
|
| 1145 |
+
attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None
|
| 1146 |
+
else:
|
| 1147 |
+
# 4d mask is passed through the layers
|
| 1148 |
+
attention_mask = _prepare_4d_causal_attention_mask(
|
| 1149 |
+
attention_mask,
|
| 1150 |
+
(batch_size, seq_length),
|
| 1151 |
+
inputs_embeds,
|
| 1152 |
+
past_key_values_length,
|
| 1153 |
+
sliding_window=self.config.sliding_window,
|
| 1154 |
+
)
|
| 1155 |
+
|
| 1156 |
+
hidden_states = inputs_embeds
|
| 1157 |
+
|
| 1158 |
+
# decoder layers
|
| 1159 |
+
all_hidden_states = () if output_hidden_states else None
|
| 1160 |
+
all_self_attns = () if output_attentions else None
|
| 1161 |
+
next_decoder_cache = None
|
| 1162 |
+
|
| 1163 |
+
for decoder_layer in self.layers:
|
| 1164 |
+
if output_hidden_states:
|
| 1165 |
+
all_hidden_states += (hidden_states,)
|
| 1166 |
+
|
| 1167 |
+
if self.gradient_checkpointing and self.training:
|
| 1168 |
+
layer_outputs = self._gradient_checkpointing_func(
|
| 1169 |
+
decoder_layer.__call__,
|
| 1170 |
+
hidden_states,
|
| 1171 |
+
attention_mask,
|
| 1172 |
+
position_ids,
|
| 1173 |
+
past_key_values,
|
| 1174 |
+
output_attentions,
|
| 1175 |
+
use_cache,
|
| 1176 |
+
)
|
| 1177 |
+
else:
|
| 1178 |
+
layer_outputs = decoder_layer(
|
| 1179 |
+
hidden_states,
|
| 1180 |
+
attention_mask=attention_mask,
|
| 1181 |
+
position_ids=position_ids,
|
| 1182 |
+
past_key_value=past_key_values,
|
| 1183 |
+
output_attentions=output_attentions,
|
| 1184 |
+
use_cache=use_cache,
|
| 1185 |
+
)
|
| 1186 |
+
|
| 1187 |
+
hidden_states = layer_outputs[0]
|
| 1188 |
+
|
| 1189 |
+
if use_cache:
|
| 1190 |
+
next_decoder_cache = layer_outputs[2 if output_attentions else 1]
|
| 1191 |
+
|
| 1192 |
+
if output_attentions:
|
| 1193 |
+
all_self_attns += (layer_outputs[1],)
|
| 1194 |
+
|
| 1195 |
+
hidden_states = self.norm(hidden_states)
|
| 1196 |
+
|
| 1197 |
+
# add hidden states from the last decoder layer
|
| 1198 |
+
if output_hidden_states:
|
| 1199 |
+
all_hidden_states += (hidden_states,)
|
| 1200 |
+
|
| 1201 |
+
next_cache = None
|
| 1202 |
+
if use_cache:
|
| 1203 |
+
next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache
|
| 1204 |
+
if not return_dict:
|
| 1205 |
+
return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None)
|
| 1206 |
+
return BaseModelOutputWithPast(
|
| 1207 |
+
last_hidden_state=hidden_states,
|
| 1208 |
+
past_key_values=next_cache,
|
| 1209 |
+
hidden_states=all_hidden_states,
|
| 1210 |
+
attentions=all_self_attns,
|
| 1211 |
+
)
|
| 1212 |
+
|
| 1213 |
+
|
| 1214 |
+
|
| 1215 |
+
class LlavaMetaForCausalLM(ABC):
|
| 1216 |
+
"""This implementation is based on the implementation from the Llave project."""
|
| 1217 |
+
|
| 1218 |
+
def init_constants(self, config):
|
| 1219 |
+
self.IGNORE_INDEX = getattr(config, 'ignore_index', -100)
|
| 1220 |
+
self.IMAGE_TOKEN_INDEX = getattr(config, 'image_token_index', 50296)
|
| 1221 |
+
self.DEFAULT_IMAGE_TOKEN = getattr(config, 'image_token', "<image>")
|
| 1222 |
+
|
| 1223 |
+
@abstractmethod
|
| 1224 |
+
def get_model(self):
|
| 1225 |
+
pass
|
| 1226 |
+
|
| 1227 |
+
def get_vision_tower(self):
|
| 1228 |
+
return self.get_model().get_vision_tower()
|
| 1229 |
+
|
| 1230 |
+
def encode_images(self, images):
|
| 1231 |
+
image_features = self.get_model().get_vision_tower()(images)
|
| 1232 |
+
image_features = self.get_model().mm_projector(image_features)
|
| 1233 |
+
return image_features
|
| 1234 |
+
|
| 1235 |
+
def prepare_inputs_labels_for_multimodal(
|
| 1236 |
+
self, input_ids, position_ids, attention_mask, past_key_values, labels, images
|
| 1237 |
+
):
|
| 1238 |
+
vision_tower = self.get_vision_tower()
|
| 1239 |
+
# if vision_tower is None or images is None or past_key_values.seqlen_offset != 0:
|
| 1240 |
+
if past_key_values is not None:
|
| 1241 |
+
target_shape = past_key_values[0][0].shape[2] + 1
|
| 1242 |
+
attention_mask = torch.ones(
|
| 1243 |
+
(attention_mask.shape[0], target_shape),
|
| 1244 |
+
dtype=attention_mask.dtype,
|
| 1245 |
+
device=attention_mask.device
|
| 1246 |
+
)
|
| 1247 |
+
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
|
| 1248 |
+
# print('position_ids', position_ids.shape)
|
| 1249 |
+
# print(input_ids[:, -1:].item())
|
| 1250 |
+
return input_ids[:, -1:], position_ids, attention_mask, past_key_values, None, labels
|
| 1251 |
+
|
| 1252 |
+
if type(images) is list or images.ndim == 5:
|
| 1253 |
+
concat_images = torch.cat([image for image in images], dim=0)
|
| 1254 |
+
# concat_images.requires_grad_(True)
|
| 1255 |
+
image_features = self.encode_images(concat_images)
|
| 1256 |
+
split_sizes = [image.shape[0] for image in images]
|
| 1257 |
+
image_features = torch.split(image_features, split_sizes, dim=0)
|
| 1258 |
+
image_features = [x.flatten(0, 1).to(self.device) for x in image_features]
|
| 1259 |
+
else:
|
| 1260 |
+
# images.requires_grad_(True)
|
| 1261 |
+
image_features = self.encode_images(images).to(self.device)
|
| 1262 |
+
|
| 1263 |
+
# TODO: image start / end is not implemented here to support pretraining.
|
| 1264 |
+
if getattr(self.config, 'tune_mm_mlp_adapter', False) and getattr(self.config, 'mm_use_im_start_end', False):
|
| 1265 |
+
raise NotImplementedError
|
| 1266 |
+
|
| 1267 |
+
# Let's just add dummy tensors if they do not exist,
|
| 1268 |
+
# it is a headache to deal with None all the time.
|
| 1269 |
+
# But it is not ideal, and if you have a better idea,
|
| 1270 |
+
# please open an issue / submit a PR, thanks.
|
| 1271 |
+
_labels = labels
|
| 1272 |
+
_position_ids = position_ids
|
| 1273 |
+
_attention_mask = attention_mask
|
| 1274 |
+
if attention_mask is None:
|
| 1275 |
+
attention_mask = torch.ones_like(input_ids, dtype=torch.bool)
|
| 1276 |
+
else:
|
| 1277 |
+
attention_mask = attention_mask.bool()
|
| 1278 |
+
if position_ids is None:
|
| 1279 |
+
position_ids = torch.arange(0, input_ids.shape[1], dtype=torch.long, device=input_ids.device)
|
| 1280 |
+
if labels is None:
|
| 1281 |
+
labels = torch.full_like(input_ids, self.IGNORE_INDEX)
|
| 1282 |
+
|
| 1283 |
+
# remove the padding using attention_mask -- TODO: double check
|
| 1284 |
+
input_ids = [cur_input_ids[cur_attention_mask] for cur_input_ids, cur_attention_mask in zip(input_ids, attention_mask)]
|
| 1285 |
+
labels = [cur_labels[cur_attention_mask] for cur_labels, cur_attention_mask in zip(labels, attention_mask)]
|
| 1286 |
+
|
| 1287 |
+
new_input_embeds = []
|
| 1288 |
+
new_labels = []
|
| 1289 |
+
cur_image_idx = 0
|
| 1290 |
+
for batch_idx, cur_input_ids in enumerate(input_ids):
|
| 1291 |
+
num_images = (cur_input_ids == self.IMAGE_TOKEN_INDEX).sum()
|
| 1292 |
+
if num_images == 0:
|
| 1293 |
+
cur_image_features = image_features[cur_image_idx]
|
| 1294 |
+
cur_input_embeds_1 = self.get_model().embed_tokens(cur_input_ids)
|
| 1295 |
+
cur_input_embeds = torch.cat([cur_input_embeds_1, cur_image_features[0:0]], dim=0)
|
| 1296 |
+
new_input_embeds.append(cur_input_embeds)
|
| 1297 |
+
new_labels.append(labels[batch_idx])
|
| 1298 |
+
cur_image_idx += 1
|
| 1299 |
+
continue
|
| 1300 |
+
|
| 1301 |
+
image_token_indices = [-1] + torch.where(cur_input_ids == self.IMAGE_TOKEN_INDEX)[0].tolist() + [cur_input_ids.shape[0]]
|
| 1302 |
+
cur_input_ids_noim = []
|
| 1303 |
+
cur_labels = labels[batch_idx]
|
| 1304 |
+
cur_labels_noim = []
|
| 1305 |
+
for i in range(len(image_token_indices) - 1):
|
| 1306 |
+
cur_input_ids_noim.append(cur_input_ids[image_token_indices[i]+1:image_token_indices[i+1]])
|
| 1307 |
+
cur_labels_noim.append(cur_labels[image_token_indices[i]+1:image_token_indices[i+1]])
|
| 1308 |
+
split_sizes = [x.shape[0] for x in cur_labels_noim]
|
| 1309 |
+
cur_input_embeds = self.get_model().embed_tokens(torch.cat(cur_input_ids_noim))
|
| 1310 |
+
# print(cur_input_embeds.shape)
|
| 1311 |
+
cur_input_embeds_no_im = torch.split(cur_input_embeds, split_sizes, dim=0)
|
| 1312 |
+
cur_new_input_embeds = []
|
| 1313 |
+
cur_new_labels = []
|
| 1314 |
+
|
| 1315 |
+
for i in range(num_images + 1):
|
| 1316 |
+
cur_new_input_embeds.append(cur_input_embeds_no_im[i])
|
| 1317 |
+
cur_new_labels.append(cur_labels_noim[i])
|
| 1318 |
+
if i < num_images:
|
| 1319 |
+
cur_image_features = image_features[cur_image_idx]
|
| 1320 |
+
cur_image_idx += 1
|
| 1321 |
+
cur_new_input_embeds.append(cur_image_features)
|
| 1322 |
+
cur_new_labels.append(torch.full((cur_image_features.shape[0],), self.IGNORE_INDEX, device=cur_labels.device, dtype=cur_labels.dtype))
|
| 1323 |
+
|
| 1324 |
+
cur_new_input_embeds = torch.cat(cur_new_input_embeds)
|
| 1325 |
+
cur_new_labels = torch.cat(cur_new_labels)
|
| 1326 |
+
|
| 1327 |
+
new_input_embeds.append(cur_new_input_embeds)
|
| 1328 |
+
new_labels.append(cur_new_labels)
|
| 1329 |
+
|
| 1330 |
+
# Truncate sequences to max length as image embeddings can make the sequence longer
|
| 1331 |
+
tokenizer_model_max_length = getattr(self.config, 'tokenizer_model_max_length', None)
|
| 1332 |
+
if tokenizer_model_max_length is not None:
|
| 1333 |
+
new_input_embeds = [x[:tokenizer_model_max_length] for x in new_input_embeds]
|
| 1334 |
+
new_labels = [x[:tokenizer_model_max_length] for x in new_labels]
|
| 1335 |
+
|
| 1336 |
+
# Combine them
|
| 1337 |
+
max_len = max(x.shape[0] for x in new_input_embeds)
|
| 1338 |
+
batch_size = len(new_input_embeds)
|
| 1339 |
+
|
| 1340 |
+
new_input_embeds_padded = []
|
| 1341 |
+
new_labels_padded = torch.full((batch_size, max_len), self.IGNORE_INDEX, dtype=new_labels[0].dtype, device=new_labels[0].device)
|
| 1342 |
+
attention_mask = torch.zeros((batch_size, max_len), dtype=attention_mask.dtype, device=attention_mask.device)
|
| 1343 |
+
position_ids = torch.zeros((batch_size, max_len), dtype=position_ids.dtype, device=position_ids.device)
|
| 1344 |
+
|
| 1345 |
+
for i, (cur_new_embed, cur_new_labels) in enumerate(zip(new_input_embeds, new_labels)):
|
| 1346 |
+
cur_len = cur_new_embed.shape[0]
|
| 1347 |
+
if getattr(self.config, 'tokenizer_padding_side', 'right') == "left":
|
| 1348 |
+
new_input_embeds_padded.append(torch.cat((
|
| 1349 |
+
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
|
| 1350 |
+
cur_new_embed
|
| 1351 |
+
), dim=0))
|
| 1352 |
+
if cur_len > 0:
|
| 1353 |
+
new_labels_padded[i, -cur_len:] = cur_new_labels
|
| 1354 |
+
attention_mask[i, -cur_len:] = True
|
| 1355 |
+
position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
| 1356 |
+
else:
|
| 1357 |
+
new_input_embeds_padded.append(torch.cat((
|
| 1358 |
+
cur_new_embed,
|
| 1359 |
+
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device)
|
| 1360 |
+
), dim=0))
|
| 1361 |
+
if cur_len > 0:
|
| 1362 |
+
new_labels_padded[i, :cur_len] = cur_new_labels
|
| 1363 |
+
attention_mask[i, :cur_len] = True
|
| 1364 |
+
position_ids[i, :cur_len] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
| 1365 |
+
|
| 1366 |
+
new_input_embeds = torch.stack(new_input_embeds_padded, dim=0)
|
| 1367 |
+
|
| 1368 |
+
if new_input_embeds.shape[-2] > 2000:
|
| 1369 |
+
self.need_clear_cache = True
|
| 1370 |
+
|
| 1371 |
+
if _labels is None:
|
| 1372 |
+
new_labels = None
|
| 1373 |
+
else:
|
| 1374 |
+
new_labels = new_labels_padded
|
| 1375 |
+
|
| 1376 |
+
if _attention_mask is None:
|
| 1377 |
+
attention_mask = None
|
| 1378 |
+
else:
|
| 1379 |
+
attention_mask = attention_mask.to(dtype=_attention_mask.dtype)
|
| 1380 |
+
|
| 1381 |
+
if _position_ids is None:
|
| 1382 |
+
position_ids = None
|
| 1383 |
+
|
| 1384 |
+
return None, position_ids, attention_mask, past_key_values, new_input_embeds, new_labels
|
| 1385 |
+
|
| 1386 |
+
|
| 1387 |
+
class ImpPhi3ForCausalLM(Phi3PreTrainedModel, LlavaMetaForCausalLM):
|
| 1388 |
+
"""Impphi3 for Causal Language Modeling."""
|
| 1389 |
+
|
| 1390 |
+
config_class = ImpPhi3Config
|
| 1391 |
+
|
| 1392 |
+
def __init__(self, config: ImpPhi3Config) -> None:
|
| 1393 |
+
super().__init__(config)
|
| 1394 |
+
|
| 1395 |
+
self.model = ImpPhi3Model(config)
|
| 1396 |
+
self.vocab_size = config.vocab_size
|
| 1397 |
+
self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
|
| 1398 |
+
self.need_clear_cache = False
|
| 1399 |
+
|
| 1400 |
+
self.post_init()
|
| 1401 |
+
self.init_constants(config)
|
| 1402 |
+
|
| 1403 |
+
# def get_output_embeddings(self) -> nn.Linear:
|
| 1404 |
+
# return self.lm_head
|
| 1405 |
+
|
| 1406 |
+
# def set_output_embeddings(self, new_embeddings: nn.Linear) -> None:
|
| 1407 |
+
# self.lm_head = new_embeddings
|
| 1408 |
+
|
| 1409 |
+
def get_model(self):
|
| 1410 |
+
return self.model
|
| 1411 |
+
|
| 1412 |
+
|
| 1413 |
+
def image_preprocess(self, images):
|
| 1414 |
+
return self.get_vision_tower().image_processor(images)['pixel_values']
|
| 1415 |
+
|
| 1416 |
+
|
| 1417 |
+
def forward(
|
| 1418 |
+
self,
|
| 1419 |
+
input_ids: torch.LongTensor = None,
|
| 1420 |
+
attention_mask: Optional[torch.Tensor] = None,
|
| 1421 |
+
position_ids: Optional[torch.LongTensor] = None,
|
| 1422 |
+
past_key_values: Optional[List[torch.FloatTensor]] = None,
|
| 1423 |
+
inputs_embeds: Optional[torch.FloatTensor] = None,
|
| 1424 |
+
labels: Optional[torch.LongTensor] = None,
|
| 1425 |
+
use_cache: Optional[bool] = None,
|
| 1426 |
+
output_attentions: Optional[bool] = None,
|
| 1427 |
+
output_hidden_states: Optional[bool] = None,
|
| 1428 |
+
images: Optional[torch.FloatTensor] = None,
|
| 1429 |
+
return_dict: Optional[bool] = None,
|
| 1430 |
+
) -> Union[Tuple, CausalLMOutputWithPast]:
|
| 1431 |
+
|
| 1432 |
+
if inputs_embeds is None:
|
| 1433 |
+
(
|
| 1434 |
+
input_ids,
|
| 1435 |
+
position_ids,
|
| 1436 |
+
attention_mask,
|
| 1437 |
+
past_key_values,
|
| 1438 |
+
inputs_embeds,
|
| 1439 |
+
labels
|
| 1440 |
+
) = self.prepare_inputs_labels_for_multimodal(
|
| 1441 |
+
input_ids,
|
| 1442 |
+
position_ids,
|
| 1443 |
+
attention_mask,
|
| 1444 |
+
past_key_values,
|
| 1445 |
+
labels,
|
| 1446 |
+
images
|
| 1447 |
+
)
|
| 1448 |
+
|
| 1449 |
+
output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
|
| 1450 |
+
output_hidden_states = (
|
| 1451 |
+
output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
|
| 1452 |
+
)
|
| 1453 |
+
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
|
| 1454 |
+
|
| 1455 |
+
outputs = self.model(
|
| 1456 |
+
input_ids=input_ids,
|
| 1457 |
+
attention_mask=attention_mask,
|
| 1458 |
+
position_ids=position_ids,
|
| 1459 |
+
past_key_values=past_key_values,
|
| 1460 |
+
inputs_embeds=inputs_embeds,
|
| 1461 |
+
use_cache=use_cache,
|
| 1462 |
+
output_attentions=output_attentions,
|
| 1463 |
+
output_hidden_states=output_hidden_states,
|
| 1464 |
+
return_dict=return_dict,
|
| 1465 |
+
)
|
| 1466 |
+
|
| 1467 |
+
hidden_states = outputs[0]
|
| 1468 |
+
logits = self.lm_head(hidden_states)
|
| 1469 |
+
# logits = logits.float()
|
| 1470 |
+
|
| 1471 |
+
loss = None
|
| 1472 |
+
if labels is not None:
|
| 1473 |
+
# Shift so that tokens < n predict n
|
| 1474 |
+
shift_logits = logits[..., :-1, :].contiguous()
|
| 1475 |
+
shift_labels = labels[..., 1:].contiguous()
|
| 1476 |
+
# Flatten the tokens
|
| 1477 |
+
loss_fct = CrossEntropyLoss()
|
| 1478 |
+
shift_logits = shift_logits.view(-1, self.config.vocab_size)
|
| 1479 |
+
shift_labels = shift_labels.view(-1)
|
| 1480 |
+
# Enable model parallelism
|
| 1481 |
+
shift_labels = shift_labels.to(shift_logits.device)
|
| 1482 |
+
loss = loss_fct(shift_logits, shift_labels)
|
| 1483 |
+
|
| 1484 |
+
if not return_dict:
|
| 1485 |
+
output = (logits,) + outputs[1:]
|
| 1486 |
+
return (loss,) + output if loss is not None else output
|
| 1487 |
+
|
| 1488 |
+
return CausalLMOutputWithPast(
|
| 1489 |
+
loss=loss,
|
| 1490 |
+
logits=logits,
|
| 1491 |
+
past_key_values=outputs.past_key_values,
|
| 1492 |
+
hidden_states=outputs.hidden_states,
|
| 1493 |
+
attentions=outputs.attentions,
|
| 1494 |
+
)
|
| 1495 |
+
|
| 1496 |
+
|
| 1497 |
+
# # inputs_embeds.requires_grad_(True)
|
| 1498 |
+
# return super().forward(
|
| 1499 |
+
# input_ids=input_ids,
|
| 1500 |
+
# attention_mask=attention_mask,
|
| 1501 |
+
# position_ids=position_ids,
|
| 1502 |
+
# past_key_values=past_key_values,
|
| 1503 |
+
# inputs_embeds=inputs_embeds,
|
| 1504 |
+
# labels=labels,
|
| 1505 |
+
# use_cache=use_cache,
|
| 1506 |
+
# output_attentions=output_attentions,
|
| 1507 |
+
# output_hidden_states=output_hidden_states,
|
| 1508 |
+
# return_dict=return_dict
|
| 1509 |
+
# )
|
| 1510 |
+
|
| 1511 |
+
def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
|
| 1512 |
+
images = kwargs.pop("images", None)
|
| 1513 |
+
_inputs = super().prepare_inputs_for_generation(
|
| 1514 |
+
input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
|
| 1515 |
+
)
|
| 1516 |
+
if images is not None:
|
| 1517 |
+
_inputs['images'] = images
|
| 1518 |
+
return _inputs
|
| 1519 |
+
|
| 1520 |
+
AutoConfig.register("imp_phi3", ImpPhi3Config)
|
| 1521 |
+
AutoModelForCausalLM.register(ImpPhi3Config, ImpPhi3ForCausalLM)
|
special_tokens_map.json
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"bos_token": {
|
| 3 |
+
"content": "<s>",
|
| 4 |
+
"lstrip": false,
|
| 5 |
+
"normalized": false,
|
| 6 |
+
"rstrip": false,
|
| 7 |
+
"single_word": false
|
| 8 |
+
},
|
| 9 |
+
"eos_token": {
|
| 10 |
+
"content": "</s>",
|
| 11 |
+
"lstrip": false,
|
| 12 |
+
"normalized": false,
|
| 13 |
+
"rstrip": false,
|
| 14 |
+
"single_word": false
|
| 15 |
+
},
|
| 16 |
+
"pad_token": {
|
| 17 |
+
"content": "<|endoftext|>",
|
| 18 |
+
"lstrip": false,
|
| 19 |
+
"normalized": false,
|
| 20 |
+
"rstrip": false,
|
| 21 |
+
"single_word": false
|
| 22 |
+
},
|
| 23 |
+
"unk_token": {
|
| 24 |
+
"content": "<unk>",
|
| 25 |
+
"lstrip": false,
|
| 26 |
+
"normalized": false,
|
| 27 |
+
"rstrip": false,
|
| 28 |
+
"single_word": false
|
| 29 |
+
}
|
| 30 |
+
}
|
tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tokenizer_config.json
ADDED
|
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": true,
|
| 3 |
+
"add_eos_token": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"32011": {
|
| 6 |
+
"content": "<image>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"0": {
|
| 14 |
+
"content": "<unk>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"1": {
|
| 22 |
+
"content": "<s>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"2": {
|
| 30 |
+
"content": "</s>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"32000": {
|
| 38 |
+
"content": "<|endoftext|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"32001": {
|
| 46 |
+
"content": "<|assistant|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"32002": {
|
| 54 |
+
"content": "<|placeholder1|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"32003": {
|
| 62 |
+
"content": "<|placeholder2|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"32004": {
|
| 70 |
+
"content": "<|placeholder3|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"32005": {
|
| 78 |
+
"content": "<|placeholder4|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"32006": {
|
| 86 |
+
"content": "<|system|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"32007": {
|
| 94 |
+
"content": "<|end|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"32008": {
|
| 102 |
+
"content": "<|placeholder5|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"32009": {
|
| 110 |
+
"content": "<|placeholder6|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"32010": {
|
| 118 |
+
"content": "<|user|>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": true
|
| 124 |
+
}
|
| 125 |
+
},
|
| 126 |
+
"bos_token": "<s>",
|
| 127 |
+
"chat_template": "{{ bos_token }}{% for message in messages %}{% if (message['role'] == 'user') %}{{'<|user|>' + '\n' + message['content'] + '<|end|>' + '\n' + '<|assistant|>' + '\n'}}{% elif (message['role'] == 'assistant') %}{{message['content'] + '<|end|>' + '\n'}}{% endif %}{% endfor %}",
|
| 128 |
+
"clean_up_tokenization_spaces": false,
|
| 129 |
+
"eos_token": "</s>",
|
| 130 |
+
"legacy": false,
|
| 131 |
+
"model_max_length": 4096,
|
| 132 |
+
"pad_token": "<|endoftext|>",
|
| 133 |
+
"padding_side": "left",
|
| 134 |
+
"sp_model_kwargs": {},
|
| 135 |
+
"tokenizer_class": "LlamaTokenizer",
|
| 136 |
+
"unk_token": "<unk>",
|
| 137 |
+
"use_default_system_prompt": false
|
| 138 |
+
}
|
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
|
|
|