Initial release: MOTIF vision encoder 7B (bf16 safetensors + modeling_motif.py)
Browse files- LICENSE +202 -0
- README.md +124 -0
- config.json +36 -0
- model-00001-of-00004.safetensors +3 -0
- model-00002-of-00004.safetensors +3 -0
- model-00003-of-00004.safetensors +3 -0
- model-00004-of-00004.safetensors +3 -0
- model.safetensors.index.json +780 -0
- modeling_motif.py +1671 -0
- preprocessor_config.json +27 -0
LICENSE
ADDED
|
@@ -0,0 +1,202 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
|
| 2 |
+
Apache License
|
| 3 |
+
Version 2.0, January 2004
|
| 4 |
+
http://www.apache.org/licenses/
|
| 5 |
+
|
| 6 |
+
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
| 7 |
+
|
| 8 |
+
1. Definitions.
|
| 9 |
+
|
| 10 |
+
"License" shall mean the terms and conditions for use, reproduction,
|
| 11 |
+
and distribution as defined by Sections 1 through 9 of this document.
|
| 12 |
+
|
| 13 |
+
"Licensor" shall mean the copyright owner or entity authorized by
|
| 14 |
+
the copyright owner that is granting the License.
|
| 15 |
+
|
| 16 |
+
"Legal Entity" shall mean the union of the acting entity and all
|
| 17 |
+
other entities that control, are controlled by, or are under common
|
| 18 |
+
control with that entity. For the purposes of this definition,
|
| 19 |
+
"control" means (i) the power, direct or indirect, to cause the
|
| 20 |
+
direction or management of such entity, whether by contract or
|
| 21 |
+
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
| 22 |
+
outstanding shares, or (iii) beneficial ownership of such entity.
|
| 23 |
+
|
| 24 |
+
"You" (or "Your") shall mean an individual or Legal Entity
|
| 25 |
+
exercising permissions granted by this License.
|
| 26 |
+
|
| 27 |
+
"Source" form shall mean the preferred form for making modifications,
|
| 28 |
+
including but not limited to software source code, documentation
|
| 29 |
+
source, and configuration files.
|
| 30 |
+
|
| 31 |
+
"Object" form shall mean any form resulting from mechanical
|
| 32 |
+
transformation or translation of a Source form, including but
|
| 33 |
+
not limited to compiled object code, generated documentation,
|
| 34 |
+
and conversions to other media types.
|
| 35 |
+
|
| 36 |
+
"Work" shall mean the work of authorship, whether in Source or
|
| 37 |
+
Object form, made available under the License, as indicated by a
|
| 38 |
+
copyright notice that is included in or attached to the work
|
| 39 |
+
(an example is provided in the Appendix below).
|
| 40 |
+
|
| 41 |
+
"Derivative Works" shall mean any work, whether in Source or Object
|
| 42 |
+
form, that is based on (or derived from) the Work and for which the
|
| 43 |
+
editorial revisions, annotations, elaborations, or other modifications
|
| 44 |
+
represent, as a whole, an original work of authorship. For the purposes
|
| 45 |
+
of this License, Derivative Works shall not include works that remain
|
| 46 |
+
separable from, or merely link (or bind by name) to the interfaces of,
|
| 47 |
+
the Work and Derivative Works thereof.
|
| 48 |
+
|
| 49 |
+
"Contribution" shall mean any work of authorship, including
|
| 50 |
+
the original version of the Work and any modifications or additions
|
| 51 |
+
to that Work or Derivative Works thereof, that is intentionally
|
| 52 |
+
submitted to Licensor for inclusion in the Work by the copyright owner
|
| 53 |
+
or by an individual or Legal Entity authorized to submit on behalf of
|
| 54 |
+
the copyright owner. For the purposes of this definition, "submitted"
|
| 55 |
+
means any form of electronic, verbal, or written communication sent
|
| 56 |
+
to the Licensor or its representatives, including but not limited to
|
| 57 |
+
communication on electronic mailing lists, source code control systems,
|
| 58 |
+
and issue tracking systems that are managed by, or on behalf of, the
|
| 59 |
+
Licensor for the purpose of discussing and improving the Work, but
|
| 60 |
+
excluding communication that is conspicuously marked or otherwise
|
| 61 |
+
designated in writing by the copyright owner as "Not a Contribution."
|
| 62 |
+
|
| 63 |
+
"Contributor" shall mean Licensor and any individual or Legal Entity
|
| 64 |
+
on behalf of whom a Contribution has been received by Licensor and
|
| 65 |
+
subsequently incorporated within the Work.
|
| 66 |
+
|
| 67 |
+
2. Grant of Copyright License. Subject to the terms and conditions of
|
| 68 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 69 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 70 |
+
copyright license to reproduce, prepare Derivative Works of,
|
| 71 |
+
publicly display, publicly perform, sublicense, and distribute the
|
| 72 |
+
Work and such Derivative Works in Source or Object form.
|
| 73 |
+
|
| 74 |
+
3. Grant of Patent License. Subject to the terms and conditions of
|
| 75 |
+
this License, each Contributor hereby grants to You a perpetual,
|
| 76 |
+
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
| 77 |
+
(except as stated in this section) patent license to make, have made,
|
| 78 |
+
use, offer to sell, sell, import, and otherwise transfer the Work,
|
| 79 |
+
where such license applies only to those patent claims licensable
|
| 80 |
+
by such Contributor that are necessarily infringed by their
|
| 81 |
+
Contribution(s) alone or by combination of their Contribution(s)
|
| 82 |
+
with the Work to which such Contribution(s) was submitted. If You
|
| 83 |
+
institute patent litigation against any entity (including a
|
| 84 |
+
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
| 85 |
+
or a Contribution incorporated within the Work constitutes direct
|
| 86 |
+
or contributory patent infringement, then any patent licenses
|
| 87 |
+
granted to You under this License for that Work shall terminate
|
| 88 |
+
as of the date such litigation is filed.
|
| 89 |
+
|
| 90 |
+
4. Redistribution. You may reproduce and distribute copies of the
|
| 91 |
+
Work or Derivative Works thereof in any medium, with or without
|
| 92 |
+
modifications, and in Source or Object form, provided that You
|
| 93 |
+
meet the following conditions:
|
| 94 |
+
|
| 95 |
+
(a) You must give any other recipients of the Work or
|
| 96 |
+
Derivative Works a copy of this License; and
|
| 97 |
+
|
| 98 |
+
(b) You must cause any modified files to carry prominent notices
|
| 99 |
+
stating that You changed the files; and
|
| 100 |
+
|
| 101 |
+
(c) You must retain, in the Source form of any Derivative Works
|
| 102 |
+
that You distribute, all copyright, patent, trademark, and
|
| 103 |
+
attribution notices from the Source form of the Work,
|
| 104 |
+
excluding those notices that do not pertain to any part of
|
| 105 |
+
the Derivative Works; and
|
| 106 |
+
|
| 107 |
+
(d) If the Work includes a "NOTICE" text file as part of its
|
| 108 |
+
distribution, then any Derivative Works that You distribute must
|
| 109 |
+
include a readable copy of the attribution notices contained
|
| 110 |
+
within such NOTICE file, excluding those notices that do not
|
| 111 |
+
pertain to any part of the Derivative Works, in at least one
|
| 112 |
+
of the following places: within a NOTICE text file distributed
|
| 113 |
+
as part of the Derivative Works; within the Source form or
|
| 114 |
+
documentation, if provided along with the Derivative Works; or,
|
| 115 |
+
within a display generated by the Derivative Works, if and
|
| 116 |
+
wherever such third-party notices normally appear. The contents
|
| 117 |
+
of the NOTICE file are for informational purposes only and
|
| 118 |
+
do not modify the License. You may add Your own attribution
|
| 119 |
+
notices within Derivative Works that You distribute, alongside
|
| 120 |
+
or as an addendum to the NOTICE text from the Work, provided
|
| 121 |
+
that such additional attribution notices cannot be construed
|
| 122 |
+
as modifying the License.
|
| 123 |
+
|
| 124 |
+
You may add Your own copyright statement to Your modifications and
|
| 125 |
+
may provide additional or different license terms and conditions
|
| 126 |
+
for use, reproduction, or distribution of Your modifications, or
|
| 127 |
+
for any such Derivative Works as a whole, provided Your use,
|
| 128 |
+
reproduction, and distribution of the Work otherwise complies with
|
| 129 |
+
the conditions stated in this License.
|
| 130 |
+
|
| 131 |
+
5. Submission of Contributions. Unless You explicitly state otherwise,
|
| 132 |
+
any Contribution intentionally submitted for inclusion in the Work
|
| 133 |
+
by You to the Licensor shall be under the terms and conditions of
|
| 134 |
+
this License, without any additional terms or conditions.
|
| 135 |
+
Notwithstanding the above, nothing herein shall supersede or modify
|
| 136 |
+
the terms of any separate license agreement you may have executed
|
| 137 |
+
with Licensor regarding such Contributions.
|
| 138 |
+
|
| 139 |
+
6. Trademarks. This License does not grant permission to use the trade
|
| 140 |
+
names, trademarks, service marks, or product names of the Licensor,
|
| 141 |
+
except as required for reasonable and customary use in describing the
|
| 142 |
+
origin of the Work and reproducing the content of the NOTICE file.
|
| 143 |
+
|
| 144 |
+
7. Disclaimer of Warranty. Unless required by applicable law or
|
| 145 |
+
agreed to in writing, Licensor provides the Work (and each
|
| 146 |
+
Contributor provides its Contributions) on an "AS IS" BASIS,
|
| 147 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
| 148 |
+
implied, including, without limitation, any warranties or conditions
|
| 149 |
+
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
| 150 |
+
PARTICULAR PURPOSE. You are solely responsible for determining the
|
| 151 |
+
appropriateness of using or redistributing the Work and assume any
|
| 152 |
+
risks associated with Your exercise of permissions under this License.
|
| 153 |
+
|
| 154 |
+
8. Limitation of Liability. In no event and under no legal theory,
|
| 155 |
+
whether in tort (including negligence), contract, or otherwise,
|
| 156 |
+
unless required by applicable law (such as deliberate and grossly
|
| 157 |
+
negligent acts) or agreed to in writing, shall any Contributor be
|
| 158 |
+
liable to You for damages, including any direct, indirect, special,
|
| 159 |
+
incidental, or consequential damages of any character arising as a
|
| 160 |
+
result of this License or out of the use or inability to use the
|
| 161 |
+
Work (including but not limited to damages for loss of goodwill,
|
| 162 |
+
work stoppage, computer failure or malfunction, or any and all
|
| 163 |
+
other commercial damages or losses), even if such Contributor
|
| 164 |
+
has been advised of the possibility of such damages.
|
| 165 |
+
|
| 166 |
+
9. Accepting Warranty or Additional Liability. While redistributing
|
| 167 |
+
the Work or Derivative Works thereof, You may choose to offer,
|
| 168 |
+
and charge a fee for, acceptance of support, warranty, indemnity,
|
| 169 |
+
or other liability obligations and/or rights consistent with this
|
| 170 |
+
License. However, in accepting such obligations, You may act only
|
| 171 |
+
on Your own behalf and on Your sole responsibility, not on behalf
|
| 172 |
+
of any other Contributor, and only if You agree to indemnify,
|
| 173 |
+
defend, and hold each Contributor harmless for any liability
|
| 174 |
+
incurred by, or claims asserted against, such Contributor by reason
|
| 175 |
+
of your accepting any such warranty or additional liability.
|
| 176 |
+
|
| 177 |
+
END OF TERMS AND CONDITIONS
|
| 178 |
+
|
| 179 |
+
APPENDIX: How to apply the Apache License to your work.
|
| 180 |
+
|
| 181 |
+
To apply the Apache License to your work, attach the following
|
| 182 |
+
boilerplate notice, with the fields enclosed by brackets "[]"
|
| 183 |
+
replaced with your own identifying information. (Don't include
|
| 184 |
+
the brackets!) The text should be enclosed in the appropriate
|
| 185 |
+
comment syntax for the file format. We also recommend that a
|
| 186 |
+
file or class name and description of purpose be included on the
|
| 187 |
+
same "printed page" as the copyright notice for easier
|
| 188 |
+
identification within third-party archives.
|
| 189 |
+
|
| 190 |
+
Copyright [yyyy] [name of copyright owner]
|
| 191 |
+
|
| 192 |
+
Licensed under the Apache License, Version 2.0 (the "License");
|
| 193 |
+
you may not use this file except in compliance with the License.
|
| 194 |
+
You may obtain a copy of the License at
|
| 195 |
+
|
| 196 |
+
http://www.apache.org/licenses/LICENSE-2.0
|
| 197 |
+
|
| 198 |
+
Unless required by applicable law or agreed to in writing, software
|
| 199 |
+
distributed under the License is distributed on an "AS IS" BASIS,
|
| 200 |
+
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
| 201 |
+
See the License for the specific language governing permissions and
|
| 202 |
+
limitations under the License.
|
README.md
ADDED
|
@@ -0,0 +1,124 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: transformers
|
| 4 |
+
pipeline_tag: image-feature-extraction
|
| 5 |
+
tags:
|
| 6 |
+
- motif
|
| 7 |
+
- vision-transformer
|
| 8 |
+
- self-supervised
|
| 9 |
+
- image-feature-extraction
|
| 10 |
+
- video
|
| 11 |
+
- custom_code
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
# MOTIF Vision Encoder
|
| 15 |
+
|
| 16 |
+
MOTIF Vision Encoder is a **unified image + video** self-supervised vision encoder on a ViT
|
| 17 |
+
backbone. A single **3D-convolutional tokenizer** ingests both modalities — an image is a
|
| 18 |
+
1-frame clip (`T=1`), a video is `T>1` — so the same weights produce **dense patch-level
|
| 19 |
+
features** and a **language-aligned global (CLS) representation**.
|
| 20 |
+
|
| 21 |
+
- **Architecture**: ViT-7B (embed 4096 / depth 40 / heads 32), patch 16, 3D axial RoPE
|
| 22 |
+
(`base=100`), SwiGLU FFN, LayerScale, per-head QK-norm, gated attention, 4 register tokens.
|
| 23 |
+
- **Tokenizer**: `Conv3d(kernel=stride=(tubelet, patch, patch))` — image `(B,3,H,W)` → `T=1`,
|
| 24 |
+
video `(B,T,3,H,W)`. Token layout `[CLS] + [register × 4] + [patch × N]`.
|
| 25 |
+
- **This repo**: inference-only. Weights + self-contained `modeling_motif.py` (loaded via
|
| 26 |
+
`trust_remote_code`). Training/eval code lives in the GitHub repo.
|
| 27 |
+
|
| 28 |
+
## Usage
|
| 29 |
+
|
| 30 |
+
The model ships a self-contained `modeling_motif.py`, so it loads with `trust_remote_code=True`.
|
| 31 |
+
|
| 32 |
+
### Image
|
| 33 |
+
|
| 34 |
+
```python
|
| 35 |
+
import torch
|
| 36 |
+
from transformers import AutoImageProcessor, AutoModel
|
| 37 |
+
from transformers.image_utils import load_image
|
| 38 |
+
|
| 39 |
+
url = "http://images.cocodataset.org/val2017/000000039769.jpg"
|
| 40 |
+
image = load_image(url)
|
| 41 |
+
|
| 42 |
+
repo = "Motif-Technologies/motif-vision-encoder-7B"
|
| 43 |
+
processor = AutoImageProcessor.from_pretrained(repo)
|
| 44 |
+
model = AutoModel.from_pretrained(repo, trust_remote_code=True, dtype=torch.bfloat16).to("cuda").eval()
|
| 45 |
+
|
| 46 |
+
inputs = processor(images=image, return_tensors="pt").to(model.device, torch.bfloat16)
|
| 47 |
+
with torch.inference_mode():
|
| 48 |
+
outputs = model(**inputs)
|
| 49 |
+
|
| 50 |
+
outputs.last_hidden_state # (1, 1 + 4 + N, 4096) CLS + registers + patch tokens
|
| 51 |
+
outputs.pooler_output # (1, 4096) global (CLS) representation
|
| 52 |
+
|
| 53 |
+
patch_tokens = outputs.last_hidden_state[:, 5:, :] # (1, N, 4096), N = (H/16)*(W/16)
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
The processor resizes the shorter side to 512, center-crops to 512×512, and normalizes with
|
| 57 |
+
ImageNet mean/std (BICUBIC). `H`/`W` must be multiples of 16.
|
| 58 |
+
|
| 59 |
+
### Video
|
| 60 |
+
|
| 61 |
+
An image is a 1-frame clip; a video is the same call with a `(B, T, 3, H, W)` tensor. Apply the
|
| 62 |
+
same per-frame transform (resize → center-crop → ImageNet norm) and stack over time:
|
| 63 |
+
|
| 64 |
+
```python
|
| 65 |
+
import torch
|
| 66 |
+
|
| 67 |
+
video = torch.randn(1, 8, 3, 256, 256, device="cuda", dtype=torch.bfloat16) # (B, T, 3, H, W)
|
| 68 |
+
with torch.inference_mode():
|
| 69 |
+
outputs = model(pixel_values=video)
|
| 70 |
+
```
|
| 71 |
+
|
| 72 |
+
## Model details
|
| 73 |
+
|
| 74 |
+
| | |
|
| 75 |
+
|---|---|
|
| 76 |
+
| Backbone | ViT-7B, patch 16, embed 4096, depth 40, heads 32, SwiGLU |
|
| 77 |
+
| Register tokens | 4 |
|
| 78 |
+
| Position encoding | 3D axial RoPE (T,H,W), `base=100.0` |
|
| 79 |
+
| Video tokenizer | 3D Conv, tubelet size 2 |
|
| 80 |
+
| Precision | bf16 weights |
|
| 81 |
+
| Training | DINO + iBOT + KoLeo self-distillation, Gram anchoring, contrastive caption alignment |
|
| 82 |
+
|
| 83 |
+
Outputs (`BaseModelOutputWithPooling`): `last_hidden_state` `(B, 1+4+N, 4096)`,
|
| 84 |
+
`pooler_output` `(B, 4096)`.
|
| 85 |
+
|
| 86 |
+
## Evaluation
|
| 87 |
+
|
| 88 |
+
Compared against the strongest publicly reported self-supervised / vision backbones. Higher is
|
| 89 |
+
better for every column **except KITTI depth MSE** (lower is better). Best value per column in
|
| 90 |
+
**bold**.
|
| 91 |
+
|
| 92 |
+
| Model | Params | ImageNet-1K<br>lin. probe ↑ | ADE20K<br>mIoU ↑ | DAVIS<br>J&F ↑ | K400 ↑ | KITTI<br>depth MSE ↓ |
|
| 93 |
+
|---|---|---|---|---|---|---|
|
| 94 |
+
| **Motif Vision Encoder** | **7B** | 87.2 | 52.0 | **71.7** | *in progress* | *in progress* |
|
| 95 |
+
| DINOv3 | 7B | **88.2** | **55.9** | 71.1 | **87.8** | **2.3** |
|
| 96 |
+
| DINOv2 | 1.1B | 86.5 | 49.0 | 63.9 | 84.4 | – |
|
| 97 |
+
| V-JEPA 2.1 | 2B | 85.5 | 47.9 | 69.0 | 87.7 | 3.x |
|
| 98 |
+
| SigLIP2 | 2B | 84.5 | 45.4 | 56.1 | 86.9 | – |
|
| 99 |
+
|
| 100 |
+
- **DAVIS video segmentation (J&F 71.7)** — best in the table, surpassing DINOv3 7B (71.1),
|
| 101 |
+
reflecting the encoder's dense, temporally-coherent patch features on video.
|
| 102 |
+
- **ImageNet-1K linear probe (87.2)** and **ADE20K semantic segmentation (52.0 mIoU)** are
|
| 103 |
+
second only to DINOv3 7B while ahead of DINOv2, V-JEPA 2.1, and SigLIP2.
|
| 104 |
+
- K400 action recognition and KITTI depth estimation are still being evaluated and will be added
|
| 105 |
+
as the runs complete.
|
| 106 |
+
|
| 107 |
+
Protocol: DINOv3-style linear/attentive probes for image tasks; V-JEPA 2-style protocol for
|
| 108 |
+
video. Comparison numbers are the best figures reported by each model's authors.
|
| 109 |
+
|
| 110 |
+
## License
|
| 111 |
+
|
| 112 |
+
Released under **Apache-2.0** (see `LICENSE`). The model was trained on data governed by the
|
| 113 |
+
respective dataset licenses; downstream users are responsible for compliance with those terms.
|
| 114 |
+
|
| 115 |
+
## Citation
|
| 116 |
+
|
| 117 |
+
```bibtex
|
| 118 |
+
@misc{motif_vision_encoder,
|
| 119 |
+
title = {MOTIF Vision Encoder: a unified image/video self-supervised vision encoder},
|
| 120 |
+
author = {Motif Technologies},
|
| 121 |
+
year = {2026},
|
| 122 |
+
howpublished = {\url{https://huggingface.co/Motif-Technologies/motif-vision-encoder-7B}}
|
| 123 |
+
}
|
| 124 |
+
```
|
config.json
ADDED
|
@@ -0,0 +1,36 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"MotifVisionModel"
|
| 4 |
+
],
|
| 5 |
+
"auto_map": {
|
| 6 |
+
"AutoConfig": "modeling_motif.MotifVisionConfig",
|
| 7 |
+
"AutoModel": "modeling_motif.MotifVisionModel"
|
| 8 |
+
},
|
| 9 |
+
"depth": 40,
|
| 10 |
+
"drop_path_rate": 0.0,
|
| 11 |
+
"dtype": "bfloat16",
|
| 12 |
+
"embed_dim": 4096,
|
| 13 |
+
"ffn_bias": true,
|
| 14 |
+
"ffn_layer": "swiglu64",
|
| 15 |
+
"ffn_ratio": 3.0,
|
| 16 |
+
"gated_attention": "elementwise",
|
| 17 |
+
"img_size": 512,
|
| 18 |
+
"in_chans": 3,
|
| 19 |
+
"layerscale_init": 1e-05,
|
| 20 |
+
"mask_k_bias": true,
|
| 21 |
+
"model_type": "motif_vision",
|
| 22 |
+
"n_storage_tokens": 4,
|
| 23 |
+
"norm_layer": "layernormbf16",
|
| 24 |
+
"num_frames": 1,
|
| 25 |
+
"num_heads": 32,
|
| 26 |
+
"patch_size": 16,
|
| 27 |
+
"pos_embed_rope_base": 100.0,
|
| 28 |
+
"pos_embed_rope_rescale_coords": 2.0,
|
| 29 |
+
"proj_bias": true,
|
| 30 |
+
"qk_norm": true,
|
| 31 |
+
"qkv_bias": false,
|
| 32 |
+
"transformers_version": "5.8.1",
|
| 33 |
+
"tubelet_size": 2,
|
| 34 |
+
"untie_cls_and_patch_norms": false,
|
| 35 |
+
"untie_global_and_local_cls_norm": true
|
| 36 |
+
}
|
model-00001-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7199e2a02cd8547f6e2d4271f51aab282eb3686d938e8eb54bb2af242e81cc8a
|
| 3 |
+
size 3939675808
|
model-00002-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:ae96352d3961197861bc7b8147d46d589b2d08557f6a2090157f576895a8e202
|
| 3 |
+
size 3994160376
|
model-00003-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:02a2a4bab1dc9dacc740b8c25aa90a21028a2e0feadfa077a1cd1268e4a7f09c
|
| 3 |
+
size 3994134520
|
model-00004-of-00004.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:04f1a7bf84c7c84cff9687f30c07e3982f9411618083d9e4d58c51dac78d202f
|
| 3 |
+
size 2853015560
|
model.safetensors.index.json
ADDED
|
@@ -0,0 +1,780 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_parameters": 7390451712,
|
| 4 |
+
"total_size": 14780903552
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"backbone.blocks.0.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 8 |
+
"backbone.blocks.0.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 9 |
+
"backbone.blocks.0.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 10 |
+
"backbone.blocks.0.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 11 |
+
"backbone.blocks.0.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 12 |
+
"backbone.blocks.0.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 13 |
+
"backbone.blocks.0.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 14 |
+
"backbone.blocks.0.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 15 |
+
"backbone.blocks.0.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 16 |
+
"backbone.blocks.0.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 17 |
+
"backbone.blocks.0.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 18 |
+
"backbone.blocks.0.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 19 |
+
"backbone.blocks.0.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 20 |
+
"backbone.blocks.0.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 21 |
+
"backbone.blocks.0.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 22 |
+
"backbone.blocks.0.norm1.bias": "model-00001-of-00004.safetensors",
|
| 23 |
+
"backbone.blocks.0.norm1.weight": "model-00001-of-00004.safetensors",
|
| 24 |
+
"backbone.blocks.0.norm2.bias": "model-00001-of-00004.safetensors",
|
| 25 |
+
"backbone.blocks.0.norm2.weight": "model-00001-of-00004.safetensors",
|
| 26 |
+
"backbone.blocks.1.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 27 |
+
"backbone.blocks.1.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 28 |
+
"backbone.blocks.1.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 29 |
+
"backbone.blocks.1.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 30 |
+
"backbone.blocks.1.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 31 |
+
"backbone.blocks.1.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 32 |
+
"backbone.blocks.1.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 33 |
+
"backbone.blocks.1.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 34 |
+
"backbone.blocks.1.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 35 |
+
"backbone.blocks.1.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 36 |
+
"backbone.blocks.1.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 37 |
+
"backbone.blocks.1.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 38 |
+
"backbone.blocks.1.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 39 |
+
"backbone.blocks.1.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 40 |
+
"backbone.blocks.1.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 41 |
+
"backbone.blocks.1.norm1.bias": "model-00001-of-00004.safetensors",
|
| 42 |
+
"backbone.blocks.1.norm1.weight": "model-00001-of-00004.safetensors",
|
| 43 |
+
"backbone.blocks.1.norm2.bias": "model-00001-of-00004.safetensors",
|
| 44 |
+
"backbone.blocks.1.norm2.weight": "model-00001-of-00004.safetensors",
|
| 45 |
+
"backbone.blocks.10.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 46 |
+
"backbone.blocks.10.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 47 |
+
"backbone.blocks.10.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 48 |
+
"backbone.blocks.10.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 49 |
+
"backbone.blocks.10.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 50 |
+
"backbone.blocks.10.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 51 |
+
"backbone.blocks.10.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 52 |
+
"backbone.blocks.10.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 53 |
+
"backbone.blocks.10.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 54 |
+
"backbone.blocks.10.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 55 |
+
"backbone.blocks.10.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 56 |
+
"backbone.blocks.10.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 57 |
+
"backbone.blocks.10.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 58 |
+
"backbone.blocks.10.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 59 |
+
"backbone.blocks.10.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 60 |
+
"backbone.blocks.10.norm1.bias": "model-00001-of-00004.safetensors",
|
| 61 |
+
"backbone.blocks.10.norm1.weight": "model-00001-of-00004.safetensors",
|
| 62 |
+
"backbone.blocks.10.norm2.bias": "model-00001-of-00004.safetensors",
|
| 63 |
+
"backbone.blocks.10.norm2.weight": "model-00001-of-00004.safetensors",
|
| 64 |
+
"backbone.blocks.11.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 65 |
+
"backbone.blocks.11.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 66 |
+
"backbone.blocks.11.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 67 |
+
"backbone.blocks.11.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 68 |
+
"backbone.blocks.11.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 69 |
+
"backbone.blocks.11.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 70 |
+
"backbone.blocks.11.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 71 |
+
"backbone.blocks.11.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 72 |
+
"backbone.blocks.11.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 73 |
+
"backbone.blocks.11.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 74 |
+
"backbone.blocks.11.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 75 |
+
"backbone.blocks.11.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 76 |
+
"backbone.blocks.11.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 77 |
+
"backbone.blocks.11.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 78 |
+
"backbone.blocks.11.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 79 |
+
"backbone.blocks.11.norm1.bias": "model-00002-of-00004.safetensors",
|
| 80 |
+
"backbone.blocks.11.norm1.weight": "model-00002-of-00004.safetensors",
|
| 81 |
+
"backbone.blocks.11.norm2.bias": "model-00002-of-00004.safetensors",
|
| 82 |
+
"backbone.blocks.11.norm2.weight": "model-00002-of-00004.safetensors",
|
| 83 |
+
"backbone.blocks.12.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 84 |
+
"backbone.blocks.12.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 85 |
+
"backbone.blocks.12.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 86 |
+
"backbone.blocks.12.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 87 |
+
"backbone.blocks.12.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 88 |
+
"backbone.blocks.12.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 89 |
+
"backbone.blocks.12.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 90 |
+
"backbone.blocks.12.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 91 |
+
"backbone.blocks.12.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 92 |
+
"backbone.blocks.12.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 93 |
+
"backbone.blocks.12.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 94 |
+
"backbone.blocks.12.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 95 |
+
"backbone.blocks.12.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 96 |
+
"backbone.blocks.12.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 97 |
+
"backbone.blocks.12.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 98 |
+
"backbone.blocks.12.norm1.bias": "model-00002-of-00004.safetensors",
|
| 99 |
+
"backbone.blocks.12.norm1.weight": "model-00002-of-00004.safetensors",
|
| 100 |
+
"backbone.blocks.12.norm2.bias": "model-00002-of-00004.safetensors",
|
| 101 |
+
"backbone.blocks.12.norm2.weight": "model-00002-of-00004.safetensors",
|
| 102 |
+
"backbone.blocks.13.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 103 |
+
"backbone.blocks.13.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 104 |
+
"backbone.blocks.13.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 105 |
+
"backbone.blocks.13.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 106 |
+
"backbone.blocks.13.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 107 |
+
"backbone.blocks.13.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 108 |
+
"backbone.blocks.13.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 109 |
+
"backbone.blocks.13.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 110 |
+
"backbone.blocks.13.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 111 |
+
"backbone.blocks.13.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 112 |
+
"backbone.blocks.13.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 113 |
+
"backbone.blocks.13.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 114 |
+
"backbone.blocks.13.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 115 |
+
"backbone.blocks.13.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 116 |
+
"backbone.blocks.13.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 117 |
+
"backbone.blocks.13.norm1.bias": "model-00002-of-00004.safetensors",
|
| 118 |
+
"backbone.blocks.13.norm1.weight": "model-00002-of-00004.safetensors",
|
| 119 |
+
"backbone.blocks.13.norm2.bias": "model-00002-of-00004.safetensors",
|
| 120 |
+
"backbone.blocks.13.norm2.weight": "model-00002-of-00004.safetensors",
|
| 121 |
+
"backbone.blocks.14.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 122 |
+
"backbone.blocks.14.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 123 |
+
"backbone.blocks.14.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 124 |
+
"backbone.blocks.14.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 125 |
+
"backbone.blocks.14.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 126 |
+
"backbone.blocks.14.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 127 |
+
"backbone.blocks.14.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 128 |
+
"backbone.blocks.14.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 129 |
+
"backbone.blocks.14.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 130 |
+
"backbone.blocks.14.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 131 |
+
"backbone.blocks.14.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 132 |
+
"backbone.blocks.14.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 133 |
+
"backbone.blocks.14.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 134 |
+
"backbone.blocks.14.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 135 |
+
"backbone.blocks.14.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 136 |
+
"backbone.blocks.14.norm1.bias": "model-00002-of-00004.safetensors",
|
| 137 |
+
"backbone.blocks.14.norm1.weight": "model-00002-of-00004.safetensors",
|
| 138 |
+
"backbone.blocks.14.norm2.bias": "model-00002-of-00004.safetensors",
|
| 139 |
+
"backbone.blocks.14.norm2.weight": "model-00002-of-00004.safetensors",
|
| 140 |
+
"backbone.blocks.15.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 141 |
+
"backbone.blocks.15.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 142 |
+
"backbone.blocks.15.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 143 |
+
"backbone.blocks.15.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 144 |
+
"backbone.blocks.15.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 145 |
+
"backbone.blocks.15.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 146 |
+
"backbone.blocks.15.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 147 |
+
"backbone.blocks.15.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 148 |
+
"backbone.blocks.15.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 149 |
+
"backbone.blocks.15.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 150 |
+
"backbone.blocks.15.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 151 |
+
"backbone.blocks.15.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 152 |
+
"backbone.blocks.15.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 153 |
+
"backbone.blocks.15.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 154 |
+
"backbone.blocks.15.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 155 |
+
"backbone.blocks.15.norm1.bias": "model-00002-of-00004.safetensors",
|
| 156 |
+
"backbone.blocks.15.norm1.weight": "model-00002-of-00004.safetensors",
|
| 157 |
+
"backbone.blocks.15.norm2.bias": "model-00002-of-00004.safetensors",
|
| 158 |
+
"backbone.blocks.15.norm2.weight": "model-00002-of-00004.safetensors",
|
| 159 |
+
"backbone.blocks.16.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 160 |
+
"backbone.blocks.16.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 161 |
+
"backbone.blocks.16.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 162 |
+
"backbone.blocks.16.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 163 |
+
"backbone.blocks.16.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 164 |
+
"backbone.blocks.16.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 165 |
+
"backbone.blocks.16.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 166 |
+
"backbone.blocks.16.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 167 |
+
"backbone.blocks.16.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 168 |
+
"backbone.blocks.16.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 169 |
+
"backbone.blocks.16.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 170 |
+
"backbone.blocks.16.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 171 |
+
"backbone.blocks.16.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 172 |
+
"backbone.blocks.16.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 173 |
+
"backbone.blocks.16.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 174 |
+
"backbone.blocks.16.norm1.bias": "model-00002-of-00004.safetensors",
|
| 175 |
+
"backbone.blocks.16.norm1.weight": "model-00002-of-00004.safetensors",
|
| 176 |
+
"backbone.blocks.16.norm2.bias": "model-00002-of-00004.safetensors",
|
| 177 |
+
"backbone.blocks.16.norm2.weight": "model-00002-of-00004.safetensors",
|
| 178 |
+
"backbone.blocks.17.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 179 |
+
"backbone.blocks.17.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 180 |
+
"backbone.blocks.17.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 181 |
+
"backbone.blocks.17.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 182 |
+
"backbone.blocks.17.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 183 |
+
"backbone.blocks.17.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 184 |
+
"backbone.blocks.17.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 185 |
+
"backbone.blocks.17.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 186 |
+
"backbone.blocks.17.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 187 |
+
"backbone.blocks.17.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 188 |
+
"backbone.blocks.17.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 189 |
+
"backbone.blocks.17.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 190 |
+
"backbone.blocks.17.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 191 |
+
"backbone.blocks.17.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 192 |
+
"backbone.blocks.17.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 193 |
+
"backbone.blocks.17.norm1.bias": "model-00002-of-00004.safetensors",
|
| 194 |
+
"backbone.blocks.17.norm1.weight": "model-00002-of-00004.safetensors",
|
| 195 |
+
"backbone.blocks.17.norm2.bias": "model-00002-of-00004.safetensors",
|
| 196 |
+
"backbone.blocks.17.norm2.weight": "model-00002-of-00004.safetensors",
|
| 197 |
+
"backbone.blocks.18.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 198 |
+
"backbone.blocks.18.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 199 |
+
"backbone.blocks.18.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 200 |
+
"backbone.blocks.18.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 201 |
+
"backbone.blocks.18.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 202 |
+
"backbone.blocks.18.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 203 |
+
"backbone.blocks.18.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 204 |
+
"backbone.blocks.18.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 205 |
+
"backbone.blocks.18.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 206 |
+
"backbone.blocks.18.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 207 |
+
"backbone.blocks.18.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 208 |
+
"backbone.blocks.18.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 209 |
+
"backbone.blocks.18.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 210 |
+
"backbone.blocks.18.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 211 |
+
"backbone.blocks.18.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 212 |
+
"backbone.blocks.18.norm1.bias": "model-00002-of-00004.safetensors",
|
| 213 |
+
"backbone.blocks.18.norm1.weight": "model-00002-of-00004.safetensors",
|
| 214 |
+
"backbone.blocks.18.norm2.bias": "model-00002-of-00004.safetensors",
|
| 215 |
+
"backbone.blocks.18.norm2.weight": "model-00002-of-00004.safetensors",
|
| 216 |
+
"backbone.blocks.19.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 217 |
+
"backbone.blocks.19.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 218 |
+
"backbone.blocks.19.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 219 |
+
"backbone.blocks.19.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 220 |
+
"backbone.blocks.19.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 221 |
+
"backbone.blocks.19.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 222 |
+
"backbone.blocks.19.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 223 |
+
"backbone.blocks.19.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 224 |
+
"backbone.blocks.19.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 225 |
+
"backbone.blocks.19.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 226 |
+
"backbone.blocks.19.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 227 |
+
"backbone.blocks.19.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 228 |
+
"backbone.blocks.19.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 229 |
+
"backbone.blocks.19.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 230 |
+
"backbone.blocks.19.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 231 |
+
"backbone.blocks.19.norm1.bias": "model-00002-of-00004.safetensors",
|
| 232 |
+
"backbone.blocks.19.norm1.weight": "model-00002-of-00004.safetensors",
|
| 233 |
+
"backbone.blocks.19.norm2.bias": "model-00002-of-00004.safetensors",
|
| 234 |
+
"backbone.blocks.19.norm2.weight": "model-00002-of-00004.safetensors",
|
| 235 |
+
"backbone.blocks.2.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 236 |
+
"backbone.blocks.2.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 237 |
+
"backbone.blocks.2.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 238 |
+
"backbone.blocks.2.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 239 |
+
"backbone.blocks.2.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 240 |
+
"backbone.blocks.2.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 241 |
+
"backbone.blocks.2.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 242 |
+
"backbone.blocks.2.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 243 |
+
"backbone.blocks.2.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 244 |
+
"backbone.blocks.2.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 245 |
+
"backbone.blocks.2.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 246 |
+
"backbone.blocks.2.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 247 |
+
"backbone.blocks.2.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 248 |
+
"backbone.blocks.2.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 249 |
+
"backbone.blocks.2.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 250 |
+
"backbone.blocks.2.norm1.bias": "model-00001-of-00004.safetensors",
|
| 251 |
+
"backbone.blocks.2.norm1.weight": "model-00001-of-00004.safetensors",
|
| 252 |
+
"backbone.blocks.2.norm2.bias": "model-00001-of-00004.safetensors",
|
| 253 |
+
"backbone.blocks.2.norm2.weight": "model-00001-of-00004.safetensors",
|
| 254 |
+
"backbone.blocks.20.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 255 |
+
"backbone.blocks.20.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 256 |
+
"backbone.blocks.20.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 257 |
+
"backbone.blocks.20.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 258 |
+
"backbone.blocks.20.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 259 |
+
"backbone.blocks.20.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 260 |
+
"backbone.blocks.20.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 261 |
+
"backbone.blocks.20.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 262 |
+
"backbone.blocks.20.ls2.gamma": "model-00002-of-00004.safetensors",
|
| 263 |
+
"backbone.blocks.20.mlp.w1.bias": "model-00002-of-00004.safetensors",
|
| 264 |
+
"backbone.blocks.20.mlp.w1.weight": "model-00002-of-00004.safetensors",
|
| 265 |
+
"backbone.blocks.20.mlp.w2.bias": "model-00002-of-00004.safetensors",
|
| 266 |
+
"backbone.blocks.20.mlp.w2.weight": "model-00002-of-00004.safetensors",
|
| 267 |
+
"backbone.blocks.20.mlp.w3.bias": "model-00002-of-00004.safetensors",
|
| 268 |
+
"backbone.blocks.20.mlp.w3.weight": "model-00002-of-00004.safetensors",
|
| 269 |
+
"backbone.blocks.20.norm1.bias": "model-00002-of-00004.safetensors",
|
| 270 |
+
"backbone.blocks.20.norm1.weight": "model-00002-of-00004.safetensors",
|
| 271 |
+
"backbone.blocks.20.norm2.bias": "model-00002-of-00004.safetensors",
|
| 272 |
+
"backbone.blocks.20.norm2.weight": "model-00002-of-00004.safetensors",
|
| 273 |
+
"backbone.blocks.21.attn.gate_proj.bias": "model-00002-of-00004.safetensors",
|
| 274 |
+
"backbone.blocks.21.attn.gate_proj.weight": "model-00002-of-00004.safetensors",
|
| 275 |
+
"backbone.blocks.21.attn.k_norm.weight": "model-00002-of-00004.safetensors",
|
| 276 |
+
"backbone.blocks.21.attn.proj.bias": "model-00002-of-00004.safetensors",
|
| 277 |
+
"backbone.blocks.21.attn.proj.weight": "model-00002-of-00004.safetensors",
|
| 278 |
+
"backbone.blocks.21.attn.q_norm.weight": "model-00002-of-00004.safetensors",
|
| 279 |
+
"backbone.blocks.21.attn.qkv.weight": "model-00002-of-00004.safetensors",
|
| 280 |
+
"backbone.blocks.21.ls1.gamma": "model-00002-of-00004.safetensors",
|
| 281 |
+
"backbone.blocks.21.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 282 |
+
"backbone.blocks.21.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 283 |
+
"backbone.blocks.21.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 284 |
+
"backbone.blocks.21.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 285 |
+
"backbone.blocks.21.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 286 |
+
"backbone.blocks.21.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 287 |
+
"backbone.blocks.21.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 288 |
+
"backbone.blocks.21.norm1.bias": "model-00002-of-00004.safetensors",
|
| 289 |
+
"backbone.blocks.21.norm1.weight": "model-00002-of-00004.safetensors",
|
| 290 |
+
"backbone.blocks.21.norm2.bias": "model-00002-of-00004.safetensors",
|
| 291 |
+
"backbone.blocks.21.norm2.weight": "model-00002-of-00004.safetensors",
|
| 292 |
+
"backbone.blocks.22.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 293 |
+
"backbone.blocks.22.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 294 |
+
"backbone.blocks.22.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 295 |
+
"backbone.blocks.22.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 296 |
+
"backbone.blocks.22.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 297 |
+
"backbone.blocks.22.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 298 |
+
"backbone.blocks.22.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 299 |
+
"backbone.blocks.22.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 300 |
+
"backbone.blocks.22.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 301 |
+
"backbone.blocks.22.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 302 |
+
"backbone.blocks.22.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 303 |
+
"backbone.blocks.22.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 304 |
+
"backbone.blocks.22.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 305 |
+
"backbone.blocks.22.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 306 |
+
"backbone.blocks.22.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 307 |
+
"backbone.blocks.22.norm1.bias": "model-00003-of-00004.safetensors",
|
| 308 |
+
"backbone.blocks.22.norm1.weight": "model-00003-of-00004.safetensors",
|
| 309 |
+
"backbone.blocks.22.norm2.bias": "model-00003-of-00004.safetensors",
|
| 310 |
+
"backbone.blocks.22.norm2.weight": "model-00003-of-00004.safetensors",
|
| 311 |
+
"backbone.blocks.23.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 312 |
+
"backbone.blocks.23.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 313 |
+
"backbone.blocks.23.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 314 |
+
"backbone.blocks.23.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 315 |
+
"backbone.blocks.23.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 316 |
+
"backbone.blocks.23.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 317 |
+
"backbone.blocks.23.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 318 |
+
"backbone.blocks.23.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 319 |
+
"backbone.blocks.23.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 320 |
+
"backbone.blocks.23.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 321 |
+
"backbone.blocks.23.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 322 |
+
"backbone.blocks.23.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 323 |
+
"backbone.blocks.23.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 324 |
+
"backbone.blocks.23.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 325 |
+
"backbone.blocks.23.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 326 |
+
"backbone.blocks.23.norm1.bias": "model-00003-of-00004.safetensors",
|
| 327 |
+
"backbone.blocks.23.norm1.weight": "model-00003-of-00004.safetensors",
|
| 328 |
+
"backbone.blocks.23.norm2.bias": "model-00003-of-00004.safetensors",
|
| 329 |
+
"backbone.blocks.23.norm2.weight": "model-00003-of-00004.safetensors",
|
| 330 |
+
"backbone.blocks.24.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 331 |
+
"backbone.blocks.24.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 332 |
+
"backbone.blocks.24.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 333 |
+
"backbone.blocks.24.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 334 |
+
"backbone.blocks.24.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 335 |
+
"backbone.blocks.24.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 336 |
+
"backbone.blocks.24.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 337 |
+
"backbone.blocks.24.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 338 |
+
"backbone.blocks.24.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 339 |
+
"backbone.blocks.24.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 340 |
+
"backbone.blocks.24.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 341 |
+
"backbone.blocks.24.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 342 |
+
"backbone.blocks.24.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 343 |
+
"backbone.blocks.24.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 344 |
+
"backbone.blocks.24.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 345 |
+
"backbone.blocks.24.norm1.bias": "model-00003-of-00004.safetensors",
|
| 346 |
+
"backbone.blocks.24.norm1.weight": "model-00003-of-00004.safetensors",
|
| 347 |
+
"backbone.blocks.24.norm2.bias": "model-00003-of-00004.safetensors",
|
| 348 |
+
"backbone.blocks.24.norm2.weight": "model-00003-of-00004.safetensors",
|
| 349 |
+
"backbone.blocks.25.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 350 |
+
"backbone.blocks.25.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 351 |
+
"backbone.blocks.25.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 352 |
+
"backbone.blocks.25.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 353 |
+
"backbone.blocks.25.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 354 |
+
"backbone.blocks.25.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 355 |
+
"backbone.blocks.25.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 356 |
+
"backbone.blocks.25.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 357 |
+
"backbone.blocks.25.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 358 |
+
"backbone.blocks.25.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 359 |
+
"backbone.blocks.25.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 360 |
+
"backbone.blocks.25.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 361 |
+
"backbone.blocks.25.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 362 |
+
"backbone.blocks.25.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 363 |
+
"backbone.blocks.25.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 364 |
+
"backbone.blocks.25.norm1.bias": "model-00003-of-00004.safetensors",
|
| 365 |
+
"backbone.blocks.25.norm1.weight": "model-00003-of-00004.safetensors",
|
| 366 |
+
"backbone.blocks.25.norm2.bias": "model-00003-of-00004.safetensors",
|
| 367 |
+
"backbone.blocks.25.norm2.weight": "model-00003-of-00004.safetensors",
|
| 368 |
+
"backbone.blocks.26.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 369 |
+
"backbone.blocks.26.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 370 |
+
"backbone.blocks.26.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 371 |
+
"backbone.blocks.26.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 372 |
+
"backbone.blocks.26.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 373 |
+
"backbone.blocks.26.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 374 |
+
"backbone.blocks.26.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 375 |
+
"backbone.blocks.26.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 376 |
+
"backbone.blocks.26.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 377 |
+
"backbone.blocks.26.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 378 |
+
"backbone.blocks.26.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 379 |
+
"backbone.blocks.26.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 380 |
+
"backbone.blocks.26.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 381 |
+
"backbone.blocks.26.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 382 |
+
"backbone.blocks.26.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 383 |
+
"backbone.blocks.26.norm1.bias": "model-00003-of-00004.safetensors",
|
| 384 |
+
"backbone.blocks.26.norm1.weight": "model-00003-of-00004.safetensors",
|
| 385 |
+
"backbone.blocks.26.norm2.bias": "model-00003-of-00004.safetensors",
|
| 386 |
+
"backbone.blocks.26.norm2.weight": "model-00003-of-00004.safetensors",
|
| 387 |
+
"backbone.blocks.27.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 388 |
+
"backbone.blocks.27.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 389 |
+
"backbone.blocks.27.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 390 |
+
"backbone.blocks.27.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 391 |
+
"backbone.blocks.27.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 392 |
+
"backbone.blocks.27.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 393 |
+
"backbone.blocks.27.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 394 |
+
"backbone.blocks.27.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 395 |
+
"backbone.blocks.27.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 396 |
+
"backbone.blocks.27.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 397 |
+
"backbone.blocks.27.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 398 |
+
"backbone.blocks.27.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 399 |
+
"backbone.blocks.27.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 400 |
+
"backbone.blocks.27.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 401 |
+
"backbone.blocks.27.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 402 |
+
"backbone.blocks.27.norm1.bias": "model-00003-of-00004.safetensors",
|
| 403 |
+
"backbone.blocks.27.norm1.weight": "model-00003-of-00004.safetensors",
|
| 404 |
+
"backbone.blocks.27.norm2.bias": "model-00003-of-00004.safetensors",
|
| 405 |
+
"backbone.blocks.27.norm2.weight": "model-00003-of-00004.safetensors",
|
| 406 |
+
"backbone.blocks.28.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 407 |
+
"backbone.blocks.28.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 408 |
+
"backbone.blocks.28.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 409 |
+
"backbone.blocks.28.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 410 |
+
"backbone.blocks.28.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 411 |
+
"backbone.blocks.28.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 412 |
+
"backbone.blocks.28.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 413 |
+
"backbone.blocks.28.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 414 |
+
"backbone.blocks.28.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 415 |
+
"backbone.blocks.28.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 416 |
+
"backbone.blocks.28.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 417 |
+
"backbone.blocks.28.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 418 |
+
"backbone.blocks.28.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 419 |
+
"backbone.blocks.28.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 420 |
+
"backbone.blocks.28.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 421 |
+
"backbone.blocks.28.norm1.bias": "model-00003-of-00004.safetensors",
|
| 422 |
+
"backbone.blocks.28.norm1.weight": "model-00003-of-00004.safetensors",
|
| 423 |
+
"backbone.blocks.28.norm2.bias": "model-00003-of-00004.safetensors",
|
| 424 |
+
"backbone.blocks.28.norm2.weight": "model-00003-of-00004.safetensors",
|
| 425 |
+
"backbone.blocks.29.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 426 |
+
"backbone.blocks.29.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 427 |
+
"backbone.blocks.29.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 428 |
+
"backbone.blocks.29.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 429 |
+
"backbone.blocks.29.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 430 |
+
"backbone.blocks.29.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 431 |
+
"backbone.blocks.29.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 432 |
+
"backbone.blocks.29.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 433 |
+
"backbone.blocks.29.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 434 |
+
"backbone.blocks.29.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 435 |
+
"backbone.blocks.29.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 436 |
+
"backbone.blocks.29.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 437 |
+
"backbone.blocks.29.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 438 |
+
"backbone.blocks.29.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 439 |
+
"backbone.blocks.29.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 440 |
+
"backbone.blocks.29.norm1.bias": "model-00003-of-00004.safetensors",
|
| 441 |
+
"backbone.blocks.29.norm1.weight": "model-00003-of-00004.safetensors",
|
| 442 |
+
"backbone.blocks.29.norm2.bias": "model-00003-of-00004.safetensors",
|
| 443 |
+
"backbone.blocks.29.norm2.weight": "model-00003-of-00004.safetensors",
|
| 444 |
+
"backbone.blocks.3.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 445 |
+
"backbone.blocks.3.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 446 |
+
"backbone.blocks.3.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 447 |
+
"backbone.blocks.3.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 448 |
+
"backbone.blocks.3.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 449 |
+
"backbone.blocks.3.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 450 |
+
"backbone.blocks.3.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 451 |
+
"backbone.blocks.3.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 452 |
+
"backbone.blocks.3.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 453 |
+
"backbone.blocks.3.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 454 |
+
"backbone.blocks.3.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 455 |
+
"backbone.blocks.3.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 456 |
+
"backbone.blocks.3.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 457 |
+
"backbone.blocks.3.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 458 |
+
"backbone.blocks.3.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 459 |
+
"backbone.blocks.3.norm1.bias": "model-00001-of-00004.safetensors",
|
| 460 |
+
"backbone.blocks.3.norm1.weight": "model-00001-of-00004.safetensors",
|
| 461 |
+
"backbone.blocks.3.norm2.bias": "model-00001-of-00004.safetensors",
|
| 462 |
+
"backbone.blocks.3.norm2.weight": "model-00001-of-00004.safetensors",
|
| 463 |
+
"backbone.blocks.30.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 464 |
+
"backbone.blocks.30.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 465 |
+
"backbone.blocks.30.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 466 |
+
"backbone.blocks.30.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 467 |
+
"backbone.blocks.30.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 468 |
+
"backbone.blocks.30.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 469 |
+
"backbone.blocks.30.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 470 |
+
"backbone.blocks.30.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 471 |
+
"backbone.blocks.30.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 472 |
+
"backbone.blocks.30.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 473 |
+
"backbone.blocks.30.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 474 |
+
"backbone.blocks.30.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 475 |
+
"backbone.blocks.30.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 476 |
+
"backbone.blocks.30.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 477 |
+
"backbone.blocks.30.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 478 |
+
"backbone.blocks.30.norm1.bias": "model-00003-of-00004.safetensors",
|
| 479 |
+
"backbone.blocks.30.norm1.weight": "model-00003-of-00004.safetensors",
|
| 480 |
+
"backbone.blocks.30.norm2.bias": "model-00003-of-00004.safetensors",
|
| 481 |
+
"backbone.blocks.30.norm2.weight": "model-00003-of-00004.safetensors",
|
| 482 |
+
"backbone.blocks.31.attn.gate_proj.bias": "model-00003-of-00004.safetensors",
|
| 483 |
+
"backbone.blocks.31.attn.gate_proj.weight": "model-00003-of-00004.safetensors",
|
| 484 |
+
"backbone.blocks.31.attn.k_norm.weight": "model-00003-of-00004.safetensors",
|
| 485 |
+
"backbone.blocks.31.attn.proj.bias": "model-00003-of-00004.safetensors",
|
| 486 |
+
"backbone.blocks.31.attn.proj.weight": "model-00003-of-00004.safetensors",
|
| 487 |
+
"backbone.blocks.31.attn.q_norm.weight": "model-00003-of-00004.safetensors",
|
| 488 |
+
"backbone.blocks.31.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 489 |
+
"backbone.blocks.31.ls1.gamma": "model-00003-of-00004.safetensors",
|
| 490 |
+
"backbone.blocks.31.ls2.gamma": "model-00003-of-00004.safetensors",
|
| 491 |
+
"backbone.blocks.31.mlp.w1.bias": "model-00003-of-00004.safetensors",
|
| 492 |
+
"backbone.blocks.31.mlp.w1.weight": "model-00003-of-00004.safetensors",
|
| 493 |
+
"backbone.blocks.31.mlp.w2.bias": "model-00003-of-00004.safetensors",
|
| 494 |
+
"backbone.blocks.31.mlp.w2.weight": "model-00003-of-00004.safetensors",
|
| 495 |
+
"backbone.blocks.31.mlp.w3.bias": "model-00003-of-00004.safetensors",
|
| 496 |
+
"backbone.blocks.31.mlp.w3.weight": "model-00003-of-00004.safetensors",
|
| 497 |
+
"backbone.blocks.31.norm1.bias": "model-00003-of-00004.safetensors",
|
| 498 |
+
"backbone.blocks.31.norm1.weight": "model-00003-of-00004.safetensors",
|
| 499 |
+
"backbone.blocks.31.norm2.bias": "model-00003-of-00004.safetensors",
|
| 500 |
+
"backbone.blocks.31.norm2.weight": "model-00003-of-00004.safetensors",
|
| 501 |
+
"backbone.blocks.32.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 502 |
+
"backbone.blocks.32.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 503 |
+
"backbone.blocks.32.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 504 |
+
"backbone.blocks.32.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 505 |
+
"backbone.blocks.32.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 506 |
+
"backbone.blocks.32.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 507 |
+
"backbone.blocks.32.attn.qkv.weight": "model-00003-of-00004.safetensors",
|
| 508 |
+
"backbone.blocks.32.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 509 |
+
"backbone.blocks.32.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 510 |
+
"backbone.blocks.32.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 511 |
+
"backbone.blocks.32.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 512 |
+
"backbone.blocks.32.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 513 |
+
"backbone.blocks.32.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 514 |
+
"backbone.blocks.32.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 515 |
+
"backbone.blocks.32.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 516 |
+
"backbone.blocks.32.norm1.bias": "model-00003-of-00004.safetensors",
|
| 517 |
+
"backbone.blocks.32.norm1.weight": "model-00003-of-00004.safetensors",
|
| 518 |
+
"backbone.blocks.32.norm2.bias": "model-00004-of-00004.safetensors",
|
| 519 |
+
"backbone.blocks.32.norm2.weight": "model-00004-of-00004.safetensors",
|
| 520 |
+
"backbone.blocks.33.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 521 |
+
"backbone.blocks.33.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 522 |
+
"backbone.blocks.33.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 523 |
+
"backbone.blocks.33.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 524 |
+
"backbone.blocks.33.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 525 |
+
"backbone.blocks.33.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 526 |
+
"backbone.blocks.33.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 527 |
+
"backbone.blocks.33.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 528 |
+
"backbone.blocks.33.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 529 |
+
"backbone.blocks.33.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 530 |
+
"backbone.blocks.33.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 531 |
+
"backbone.blocks.33.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 532 |
+
"backbone.blocks.33.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 533 |
+
"backbone.blocks.33.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 534 |
+
"backbone.blocks.33.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 535 |
+
"backbone.blocks.33.norm1.bias": "model-00004-of-00004.safetensors",
|
| 536 |
+
"backbone.blocks.33.norm1.weight": "model-00004-of-00004.safetensors",
|
| 537 |
+
"backbone.blocks.33.norm2.bias": "model-00004-of-00004.safetensors",
|
| 538 |
+
"backbone.blocks.33.norm2.weight": "model-00004-of-00004.safetensors",
|
| 539 |
+
"backbone.blocks.34.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 540 |
+
"backbone.blocks.34.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 541 |
+
"backbone.blocks.34.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 542 |
+
"backbone.blocks.34.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 543 |
+
"backbone.blocks.34.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 544 |
+
"backbone.blocks.34.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 545 |
+
"backbone.blocks.34.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 546 |
+
"backbone.blocks.34.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 547 |
+
"backbone.blocks.34.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 548 |
+
"backbone.blocks.34.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 549 |
+
"backbone.blocks.34.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 550 |
+
"backbone.blocks.34.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 551 |
+
"backbone.blocks.34.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 552 |
+
"backbone.blocks.34.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 553 |
+
"backbone.blocks.34.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 554 |
+
"backbone.blocks.34.norm1.bias": "model-00004-of-00004.safetensors",
|
| 555 |
+
"backbone.blocks.34.norm1.weight": "model-00004-of-00004.safetensors",
|
| 556 |
+
"backbone.blocks.34.norm2.bias": "model-00004-of-00004.safetensors",
|
| 557 |
+
"backbone.blocks.34.norm2.weight": "model-00004-of-00004.safetensors",
|
| 558 |
+
"backbone.blocks.35.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 559 |
+
"backbone.blocks.35.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 560 |
+
"backbone.blocks.35.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 561 |
+
"backbone.blocks.35.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 562 |
+
"backbone.blocks.35.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 563 |
+
"backbone.blocks.35.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 564 |
+
"backbone.blocks.35.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 565 |
+
"backbone.blocks.35.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 566 |
+
"backbone.blocks.35.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 567 |
+
"backbone.blocks.35.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 568 |
+
"backbone.blocks.35.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 569 |
+
"backbone.blocks.35.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 570 |
+
"backbone.blocks.35.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 571 |
+
"backbone.blocks.35.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 572 |
+
"backbone.blocks.35.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 573 |
+
"backbone.blocks.35.norm1.bias": "model-00004-of-00004.safetensors",
|
| 574 |
+
"backbone.blocks.35.norm1.weight": "model-00004-of-00004.safetensors",
|
| 575 |
+
"backbone.blocks.35.norm2.bias": "model-00004-of-00004.safetensors",
|
| 576 |
+
"backbone.blocks.35.norm2.weight": "model-00004-of-00004.safetensors",
|
| 577 |
+
"backbone.blocks.36.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 578 |
+
"backbone.blocks.36.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 579 |
+
"backbone.blocks.36.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 580 |
+
"backbone.blocks.36.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 581 |
+
"backbone.blocks.36.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 582 |
+
"backbone.blocks.36.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 583 |
+
"backbone.blocks.36.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 584 |
+
"backbone.blocks.36.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 585 |
+
"backbone.blocks.36.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 586 |
+
"backbone.blocks.36.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 587 |
+
"backbone.blocks.36.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 588 |
+
"backbone.blocks.36.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 589 |
+
"backbone.blocks.36.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 590 |
+
"backbone.blocks.36.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 591 |
+
"backbone.blocks.36.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 592 |
+
"backbone.blocks.36.norm1.bias": "model-00004-of-00004.safetensors",
|
| 593 |
+
"backbone.blocks.36.norm1.weight": "model-00004-of-00004.safetensors",
|
| 594 |
+
"backbone.blocks.36.norm2.bias": "model-00004-of-00004.safetensors",
|
| 595 |
+
"backbone.blocks.36.norm2.weight": "model-00004-of-00004.safetensors",
|
| 596 |
+
"backbone.blocks.37.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 597 |
+
"backbone.blocks.37.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 598 |
+
"backbone.blocks.37.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 599 |
+
"backbone.blocks.37.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 600 |
+
"backbone.blocks.37.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 601 |
+
"backbone.blocks.37.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 602 |
+
"backbone.blocks.37.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 603 |
+
"backbone.blocks.37.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 604 |
+
"backbone.blocks.37.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 605 |
+
"backbone.blocks.37.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 606 |
+
"backbone.blocks.37.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 607 |
+
"backbone.blocks.37.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 608 |
+
"backbone.blocks.37.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 609 |
+
"backbone.blocks.37.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 610 |
+
"backbone.blocks.37.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 611 |
+
"backbone.blocks.37.norm1.bias": "model-00004-of-00004.safetensors",
|
| 612 |
+
"backbone.blocks.37.norm1.weight": "model-00004-of-00004.safetensors",
|
| 613 |
+
"backbone.blocks.37.norm2.bias": "model-00004-of-00004.safetensors",
|
| 614 |
+
"backbone.blocks.37.norm2.weight": "model-00004-of-00004.safetensors",
|
| 615 |
+
"backbone.blocks.38.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 616 |
+
"backbone.blocks.38.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 617 |
+
"backbone.blocks.38.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 618 |
+
"backbone.blocks.38.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 619 |
+
"backbone.blocks.38.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 620 |
+
"backbone.blocks.38.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 621 |
+
"backbone.blocks.38.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 622 |
+
"backbone.blocks.38.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 623 |
+
"backbone.blocks.38.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 624 |
+
"backbone.blocks.38.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 625 |
+
"backbone.blocks.38.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 626 |
+
"backbone.blocks.38.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 627 |
+
"backbone.blocks.38.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 628 |
+
"backbone.blocks.38.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 629 |
+
"backbone.blocks.38.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 630 |
+
"backbone.blocks.38.norm1.bias": "model-00004-of-00004.safetensors",
|
| 631 |
+
"backbone.blocks.38.norm1.weight": "model-00004-of-00004.safetensors",
|
| 632 |
+
"backbone.blocks.38.norm2.bias": "model-00004-of-00004.safetensors",
|
| 633 |
+
"backbone.blocks.38.norm2.weight": "model-00004-of-00004.safetensors",
|
| 634 |
+
"backbone.blocks.39.attn.gate_proj.bias": "model-00004-of-00004.safetensors",
|
| 635 |
+
"backbone.blocks.39.attn.gate_proj.weight": "model-00004-of-00004.safetensors",
|
| 636 |
+
"backbone.blocks.39.attn.k_norm.weight": "model-00004-of-00004.safetensors",
|
| 637 |
+
"backbone.blocks.39.attn.proj.bias": "model-00004-of-00004.safetensors",
|
| 638 |
+
"backbone.blocks.39.attn.proj.weight": "model-00004-of-00004.safetensors",
|
| 639 |
+
"backbone.blocks.39.attn.q_norm.weight": "model-00004-of-00004.safetensors",
|
| 640 |
+
"backbone.blocks.39.attn.qkv.weight": "model-00004-of-00004.safetensors",
|
| 641 |
+
"backbone.blocks.39.ls1.gamma": "model-00004-of-00004.safetensors",
|
| 642 |
+
"backbone.blocks.39.ls2.gamma": "model-00004-of-00004.safetensors",
|
| 643 |
+
"backbone.blocks.39.mlp.w1.bias": "model-00004-of-00004.safetensors",
|
| 644 |
+
"backbone.blocks.39.mlp.w1.weight": "model-00004-of-00004.safetensors",
|
| 645 |
+
"backbone.blocks.39.mlp.w2.bias": "model-00004-of-00004.safetensors",
|
| 646 |
+
"backbone.blocks.39.mlp.w2.weight": "model-00004-of-00004.safetensors",
|
| 647 |
+
"backbone.blocks.39.mlp.w3.bias": "model-00004-of-00004.safetensors",
|
| 648 |
+
"backbone.blocks.39.mlp.w3.weight": "model-00004-of-00004.safetensors",
|
| 649 |
+
"backbone.blocks.39.norm1.bias": "model-00004-of-00004.safetensors",
|
| 650 |
+
"backbone.blocks.39.norm1.weight": "model-00004-of-00004.safetensors",
|
| 651 |
+
"backbone.blocks.39.norm2.bias": "model-00004-of-00004.safetensors",
|
| 652 |
+
"backbone.blocks.39.norm2.weight": "model-00004-of-00004.safetensors",
|
| 653 |
+
"backbone.blocks.4.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 654 |
+
"backbone.blocks.4.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 655 |
+
"backbone.blocks.4.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 656 |
+
"backbone.blocks.4.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 657 |
+
"backbone.blocks.4.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 658 |
+
"backbone.blocks.4.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 659 |
+
"backbone.blocks.4.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 660 |
+
"backbone.blocks.4.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 661 |
+
"backbone.blocks.4.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 662 |
+
"backbone.blocks.4.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 663 |
+
"backbone.blocks.4.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 664 |
+
"backbone.blocks.4.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 665 |
+
"backbone.blocks.4.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 666 |
+
"backbone.blocks.4.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 667 |
+
"backbone.blocks.4.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 668 |
+
"backbone.blocks.4.norm1.bias": "model-00001-of-00004.safetensors",
|
| 669 |
+
"backbone.blocks.4.norm1.weight": "model-00001-of-00004.safetensors",
|
| 670 |
+
"backbone.blocks.4.norm2.bias": "model-00001-of-00004.safetensors",
|
| 671 |
+
"backbone.blocks.4.norm2.weight": "model-00001-of-00004.safetensors",
|
| 672 |
+
"backbone.blocks.5.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 673 |
+
"backbone.blocks.5.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 674 |
+
"backbone.blocks.5.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 675 |
+
"backbone.blocks.5.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 676 |
+
"backbone.blocks.5.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 677 |
+
"backbone.blocks.5.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 678 |
+
"backbone.blocks.5.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 679 |
+
"backbone.blocks.5.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 680 |
+
"backbone.blocks.5.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 681 |
+
"backbone.blocks.5.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 682 |
+
"backbone.blocks.5.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 683 |
+
"backbone.blocks.5.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 684 |
+
"backbone.blocks.5.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 685 |
+
"backbone.blocks.5.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 686 |
+
"backbone.blocks.5.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 687 |
+
"backbone.blocks.5.norm1.bias": "model-00001-of-00004.safetensors",
|
| 688 |
+
"backbone.blocks.5.norm1.weight": "model-00001-of-00004.safetensors",
|
| 689 |
+
"backbone.blocks.5.norm2.bias": "model-00001-of-00004.safetensors",
|
| 690 |
+
"backbone.blocks.5.norm2.weight": "model-00001-of-00004.safetensors",
|
| 691 |
+
"backbone.blocks.6.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 692 |
+
"backbone.blocks.6.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 693 |
+
"backbone.blocks.6.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 694 |
+
"backbone.blocks.6.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 695 |
+
"backbone.blocks.6.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 696 |
+
"backbone.blocks.6.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 697 |
+
"backbone.blocks.6.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 698 |
+
"backbone.blocks.6.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 699 |
+
"backbone.blocks.6.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 700 |
+
"backbone.blocks.6.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 701 |
+
"backbone.blocks.6.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 702 |
+
"backbone.blocks.6.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 703 |
+
"backbone.blocks.6.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 704 |
+
"backbone.blocks.6.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 705 |
+
"backbone.blocks.6.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 706 |
+
"backbone.blocks.6.norm1.bias": "model-00001-of-00004.safetensors",
|
| 707 |
+
"backbone.blocks.6.norm1.weight": "model-00001-of-00004.safetensors",
|
| 708 |
+
"backbone.blocks.6.norm2.bias": "model-00001-of-00004.safetensors",
|
| 709 |
+
"backbone.blocks.6.norm2.weight": "model-00001-of-00004.safetensors",
|
| 710 |
+
"backbone.blocks.7.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 711 |
+
"backbone.blocks.7.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 712 |
+
"backbone.blocks.7.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 713 |
+
"backbone.blocks.7.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 714 |
+
"backbone.blocks.7.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 715 |
+
"backbone.blocks.7.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 716 |
+
"backbone.blocks.7.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 717 |
+
"backbone.blocks.7.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 718 |
+
"backbone.blocks.7.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 719 |
+
"backbone.blocks.7.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 720 |
+
"backbone.blocks.7.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 721 |
+
"backbone.blocks.7.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 722 |
+
"backbone.blocks.7.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 723 |
+
"backbone.blocks.7.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 724 |
+
"backbone.blocks.7.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 725 |
+
"backbone.blocks.7.norm1.bias": "model-00001-of-00004.safetensors",
|
| 726 |
+
"backbone.blocks.7.norm1.weight": "model-00001-of-00004.safetensors",
|
| 727 |
+
"backbone.blocks.7.norm2.bias": "model-00001-of-00004.safetensors",
|
| 728 |
+
"backbone.blocks.7.norm2.weight": "model-00001-of-00004.safetensors",
|
| 729 |
+
"backbone.blocks.8.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 730 |
+
"backbone.blocks.8.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 731 |
+
"backbone.blocks.8.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 732 |
+
"backbone.blocks.8.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 733 |
+
"backbone.blocks.8.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 734 |
+
"backbone.blocks.8.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 735 |
+
"backbone.blocks.8.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 736 |
+
"backbone.blocks.8.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 737 |
+
"backbone.blocks.8.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 738 |
+
"backbone.blocks.8.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 739 |
+
"backbone.blocks.8.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 740 |
+
"backbone.blocks.8.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 741 |
+
"backbone.blocks.8.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 742 |
+
"backbone.blocks.8.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 743 |
+
"backbone.blocks.8.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 744 |
+
"backbone.blocks.8.norm1.bias": "model-00001-of-00004.safetensors",
|
| 745 |
+
"backbone.blocks.8.norm1.weight": "model-00001-of-00004.safetensors",
|
| 746 |
+
"backbone.blocks.8.norm2.bias": "model-00001-of-00004.safetensors",
|
| 747 |
+
"backbone.blocks.8.norm2.weight": "model-00001-of-00004.safetensors",
|
| 748 |
+
"backbone.blocks.9.attn.gate_proj.bias": "model-00001-of-00004.safetensors",
|
| 749 |
+
"backbone.blocks.9.attn.gate_proj.weight": "model-00001-of-00004.safetensors",
|
| 750 |
+
"backbone.blocks.9.attn.k_norm.weight": "model-00001-of-00004.safetensors",
|
| 751 |
+
"backbone.blocks.9.attn.proj.bias": "model-00001-of-00004.safetensors",
|
| 752 |
+
"backbone.blocks.9.attn.proj.weight": "model-00001-of-00004.safetensors",
|
| 753 |
+
"backbone.blocks.9.attn.q_norm.weight": "model-00001-of-00004.safetensors",
|
| 754 |
+
"backbone.blocks.9.attn.qkv.weight": "model-00001-of-00004.safetensors",
|
| 755 |
+
"backbone.blocks.9.ls1.gamma": "model-00001-of-00004.safetensors",
|
| 756 |
+
"backbone.blocks.9.ls2.gamma": "model-00001-of-00004.safetensors",
|
| 757 |
+
"backbone.blocks.9.mlp.w1.bias": "model-00001-of-00004.safetensors",
|
| 758 |
+
"backbone.blocks.9.mlp.w1.weight": "model-00001-of-00004.safetensors",
|
| 759 |
+
"backbone.blocks.9.mlp.w2.bias": "model-00001-of-00004.safetensors",
|
| 760 |
+
"backbone.blocks.9.mlp.w2.weight": "model-00001-of-00004.safetensors",
|
| 761 |
+
"backbone.blocks.9.mlp.w3.bias": "model-00001-of-00004.safetensors",
|
| 762 |
+
"backbone.blocks.9.mlp.w3.weight": "model-00001-of-00004.safetensors",
|
| 763 |
+
"backbone.blocks.9.norm1.bias": "model-00001-of-00004.safetensors",
|
| 764 |
+
"backbone.blocks.9.norm1.weight": "model-00001-of-00004.safetensors",
|
| 765 |
+
"backbone.blocks.9.norm2.bias": "model-00001-of-00004.safetensors",
|
| 766 |
+
"backbone.blocks.9.norm2.weight": "model-00001-of-00004.safetensors",
|
| 767 |
+
"backbone.cls_token": "model-00001-of-00004.safetensors",
|
| 768 |
+
"backbone.local_cls_norm.bias": "model-00004-of-00004.safetensors",
|
| 769 |
+
"backbone.local_cls_norm.weight": "model-00004-of-00004.safetensors",
|
| 770 |
+
"backbone.mask_token": "model-00001-of-00004.safetensors",
|
| 771 |
+
"backbone.norm.bias": "model-00004-of-00004.safetensors",
|
| 772 |
+
"backbone.norm.weight": "model-00004-of-00004.safetensors",
|
| 773 |
+
"backbone.patch_embed.proj.bias": "model-00001-of-00004.safetensors",
|
| 774 |
+
"backbone.patch_embed.proj.weight": "model-00001-of-00004.safetensors",
|
| 775 |
+
"backbone.rope_embed.periods_h": "model-00001-of-00004.safetensors",
|
| 776 |
+
"backbone.rope_embed.periods_t": "model-00001-of-00004.safetensors",
|
| 777 |
+
"backbone.rope_embed.periods_w": "model-00001-of-00004.safetensors",
|
| 778 |
+
"backbone.storage_tokens": "model-00001-of-00004.safetensors"
|
| 779 |
+
}
|
| 780 |
+
}
|
modeling_motif.py
ADDED
|
@@ -0,0 +1,1671 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Copyright (c) Motif Technologies.
|
| 2 |
+
# Self-contained inference model for MOTIF Vision Encoder (image + video).
|
| 3 |
+
# Auto-assembled from the training repo's inference path; NO training code.
|
| 4 |
+
"""MOTIF Vision Encoder — unified image/video ViT backbone (inference-only).
|
| 5 |
+
|
| 6 |
+
Usage:
|
| 7 |
+
from transformers import AutoModel
|
| 8 |
+
import torch
|
| 9 |
+
model = AutoModel.from_pretrained("Motif-Technologies/motif-vision-encoder",
|
| 10 |
+
trust_remote_code=True).eval()
|
| 11 |
+
# image: (B, 3, H, W) video: (B, T, 3, H, W) (H,W multiples of 16)
|
| 12 |
+
out = model(pixel_values=torch.randn(1, 3, 224, 224))
|
| 13 |
+
out.last_hidden_state # (B, 1+num_register+N, D)
|
| 14 |
+
out.pooler_output # (B, D) CLS token
|
| 15 |
+
"""
|
| 16 |
+
import math
|
| 17 |
+
from typing import Callable, Literal
|
| 18 |
+
|
| 19 |
+
import torch
|
| 20 |
+
import torch.nn as nn
|
| 21 |
+
import torch.nn.functional as F
|
| 22 |
+
from torch import Tensor
|
| 23 |
+
|
| 24 |
+
from transformers import PreTrainedModel, PretrainedConfig
|
| 25 |
+
from transformers.modeling_outputs import BaseModelOutputWithPooling
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
|
| 29 |
+
# ---- utils ----
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
def cat_keep_shapes(x_list: list[Tensor]) -> tuple[Tensor, list[tuple[int]], list[int]]:
|
| 33 |
+
"""Concatenate list of tensors while preserving their shapes for later reconstruction."""
|
| 34 |
+
shapes = [x.shape for x in x_list]
|
| 35 |
+
num_tokens = [x.select(dim=-1, index=0).numel() for x in x_list]
|
| 36 |
+
flattened = torch.cat([x.flatten(0, -2) for x in x_list])
|
| 37 |
+
return flattened, shapes, num_tokens
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def uncat_with_shapes(flattened: Tensor, shapes: list[tuple[int]], num_tokens: list[int]) -> list[Tensor]:
|
| 42 |
+
"""Reverse of cat_keep_shapes: split and reshape flattened tensor back to original shapes."""
|
| 43 |
+
outputs_splitted = torch.split_with_sizes(flattened, num_tokens, dim=0)
|
| 44 |
+
shapes_adjusted = [shape[:-1] + torch.Size([flattened.shape[-1]]) for shape in shapes]
|
| 45 |
+
outputs_reshaped = [o.reshape(shape) for o, shape in zip(outputs_splitted, shapes_adjusted)]
|
| 46 |
+
return outputs_reshaped
|
| 47 |
+
|
| 48 |
+
|
| 49 |
+
|
| 50 |
+
def named_apply(
|
| 51 |
+
fn: Callable,
|
| 52 |
+
module: nn.Module,
|
| 53 |
+
name: str = "",
|
| 54 |
+
depth_first: bool = True,
|
| 55 |
+
include_root: bool = False,
|
| 56 |
+
) -> nn.Module:
|
| 57 |
+
"""Apply fn to all submodules (in-place, no replacement)."""
|
| 58 |
+
if not depth_first and include_root:
|
| 59 |
+
fn(module=module, name=name)
|
| 60 |
+
for child_name, child_module in module.named_children():
|
| 61 |
+
child_name = ".".join((name, child_name)) if name else child_name
|
| 62 |
+
named_apply(
|
| 63 |
+
fn=fn,
|
| 64 |
+
module=child_module,
|
| 65 |
+
name=child_name,
|
| 66 |
+
depth_first=depth_first,
|
| 67 |
+
include_root=True,
|
| 68 |
+
)
|
| 69 |
+
if depth_first and include_root:
|
| 70 |
+
fn(module=module, name=name)
|
| 71 |
+
return module
|
| 72 |
+
|
| 73 |
+
|
| 74 |
+
# ---- rms_norm ----
|
| 75 |
+
|
| 76 |
+
|
| 77 |
+
import torch
|
| 78 |
+
from torch import Tensor, nn
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
class RMSNorm(nn.Module):
|
| 82 |
+
"""Root Mean Square Layer Normalization.
|
| 83 |
+
|
| 84 |
+
A simpler alternative to LayerNorm that normalizes by RMS without centering.
|
| 85 |
+
|
| 86 |
+
Args:
|
| 87 |
+
dim: Number of features.
|
| 88 |
+
eps: Small constant for numerical stability.
|
| 89 |
+
"""
|
| 90 |
+
|
| 91 |
+
def __init__(
|
| 92 |
+
self,
|
| 93 |
+
dim: int,
|
| 94 |
+
eps: float = 1e-6,
|
| 95 |
+
device: torch.device | str | None = None,
|
| 96 |
+
) -> None:
|
| 97 |
+
super().__init__()
|
| 98 |
+
self.eps = eps
|
| 99 |
+
self.weight = nn.Parameter(torch.ones(dim, device=device))
|
| 100 |
+
|
| 101 |
+
def reset_parameters(self) -> None:
|
| 102 |
+
"""Reset weight to ones."""
|
| 103 |
+
nn.init.ones_(self.weight)
|
| 104 |
+
|
| 105 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 106 |
+
"""Apply RMS normalization."""
|
| 107 |
+
rms = torch.sqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
|
| 108 |
+
return x / rms * self.weight
|
| 109 |
+
|
| 110 |
+
|
| 111 |
+
# ---- layer_scale ----
|
| 112 |
+
|
| 113 |
+
|
| 114 |
+
import torch
|
| 115 |
+
from torch import Tensor, nn
|
| 116 |
+
|
| 117 |
+
|
| 118 |
+
class LayerScale(nn.Module):
|
| 119 |
+
"""Per-channel scaling that allows gradual incorporation of each layer's contribution.
|
| 120 |
+
|
| 121 |
+
Initializes to a small value (e.g., 1e-5) so that early in training, each layer's
|
| 122 |
+
contribution is nearly zero, stabilizing deep network training.
|
| 123 |
+
|
| 124 |
+
Args:
|
| 125 |
+
dim: Number of channels.
|
| 126 |
+
init_values: Initial value for all channels.
|
| 127 |
+
inplace: Whether to apply scaling in-place.
|
| 128 |
+
device: Device for parameter allocation.
|
| 129 |
+
"""
|
| 130 |
+
|
| 131 |
+
def __init__(
|
| 132 |
+
self,
|
| 133 |
+
dim: int,
|
| 134 |
+
init_values: float | Tensor = 1e-5,
|
| 135 |
+
inplace: bool = False,
|
| 136 |
+
device: torch.device | None = None,
|
| 137 |
+
) -> None:
|
| 138 |
+
super().__init__()
|
| 139 |
+
self.inplace = inplace
|
| 140 |
+
self.gamma = nn.Parameter(torch.empty(dim, device=device))
|
| 141 |
+
self.init_values = init_values
|
| 142 |
+
|
| 143 |
+
def reset_parameters(self) -> None:
|
| 144 |
+
"""Reset gamma to initial values."""
|
| 145 |
+
nn.init.constant_(self.gamma, self.init_values)
|
| 146 |
+
|
| 147 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 148 |
+
"""Apply per-channel scaling."""
|
| 149 |
+
return x.mul_(self.gamma) if self.inplace else x * self.gamma
|
| 150 |
+
|
| 151 |
+
|
| 152 |
+
# ---- patch_embed ----
|
| 153 |
+
|
| 154 |
+
|
| 155 |
+
import math
|
| 156 |
+
|
| 157 |
+
import torch.nn as nn
|
| 158 |
+
from torch import Tensor
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
class PatchEmbed(nn.Module):
|
| 162 |
+
"""Video (5D) or Image (4D) to Patch Embedding via 3D Convolution.
|
| 163 |
+
|
| 164 |
+
Handles both modalities through a single Conv3d projection:
|
| 165 |
+
- Image (B, C, H, W): unsqueeze temporal dim -> (B, C, 1, H, W) -> Conv3d
|
| 166 |
+
- Video (B, T, C, H, W): transpose -> (B, C, T, H, W) -> Conv3d
|
| 167 |
+
|
| 168 |
+
Output: (B, N_total, embed_dim)
|
| 169 |
+
N_total = (T // tubelet_size) * (H // patch_size) * (W // patch_size)
|
| 170 |
+
|
| 171 |
+
Args:
|
| 172 |
+
img_size: Input image size (used for reference only).
|
| 173 |
+
patch_size: Spatial patch size in pixels.
|
| 174 |
+
in_chans: Number of input channels.
|
| 175 |
+
embed_dim: Output embedding dimension.
|
| 176 |
+
tubelet_size: Temporal patch size (number of frames per temporal token).
|
| 177 |
+
flatten_embedding: Whether to flatten spatial dimensions.
|
| 178 |
+
"""
|
| 179 |
+
|
| 180 |
+
def __init__(
|
| 181 |
+
self,
|
| 182 |
+
img_size: int = 224,
|
| 183 |
+
patch_size: int = 16,
|
| 184 |
+
in_chans: int = 3,
|
| 185 |
+
embed_dim: int = 768,
|
| 186 |
+
tubelet_size: int = 1,
|
| 187 |
+
flatten_embedding: bool = True,
|
| 188 |
+
) -> None:
|
| 189 |
+
super().__init__()
|
| 190 |
+
self.img_size = img_size
|
| 191 |
+
self.patch_size = (patch_size, patch_size) if isinstance(patch_size, int) else patch_size
|
| 192 |
+
self.tubelet_size = tubelet_size
|
| 193 |
+
self.flatten_embedding = flatten_embedding
|
| 194 |
+
self.in_chans = in_chans
|
| 195 |
+
|
| 196 |
+
# 3D Convolution: kernel and stride = (tubelet_size, patch_h, patch_w)
|
| 197 |
+
self.proj = nn.Conv3d(
|
| 198 |
+
in_chans,
|
| 199 |
+
embed_dim,
|
| 200 |
+
kernel_size=(tubelet_size, self.patch_size[0], self.patch_size[1]),
|
| 201 |
+
stride=(tubelet_size, self.patch_size[0], self.patch_size[1]),
|
| 202 |
+
)
|
| 203 |
+
|
| 204 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 205 |
+
"""Tokenize input images or videos.
|
| 206 |
+
|
| 207 |
+
Args:
|
| 208 |
+
x: Input tensor.
|
| 209 |
+
Image: (B, C, H, W) or Video: (B, T, C, H, W)
|
| 210 |
+
|
| 211 |
+
Returns:
|
| 212 |
+
Patch tokens of shape (B, N_total, embed_dim).
|
| 213 |
+
"""
|
| 214 |
+
if x.ndim == 4:
|
| 215 |
+
# Image: (B, C, H, W) -> (B, C, 1, H, W)
|
| 216 |
+
x = x.unsqueeze(2)
|
| 217 |
+
# If tubelet_size > 1, repeat the single frame to match kernel size
|
| 218 |
+
if self.tubelet_size > 1:
|
| 219 |
+
x = x.expand(-1, -1, self.tubelet_size, -1, -1)
|
| 220 |
+
elif x.ndim == 5:
|
| 221 |
+
# Video: (B, T, C, H, W) -> (B, C, T, H, W)
|
| 222 |
+
x = x.transpose(1, 2)
|
| 223 |
+
|
| 224 |
+
# Conv3d Projection -> (B, embed_dim, T', H', W')
|
| 225 |
+
x = self.proj(x)
|
| 226 |
+
|
| 227 |
+
if self.flatten_embedding:
|
| 228 |
+
# Flatten spatial+temporal: (B, embed_dim, N) -> (B, N, embed_dim)
|
| 229 |
+
x = x.flatten(2).transpose(1, 2)
|
| 230 |
+
|
| 231 |
+
return x
|
| 232 |
+
|
| 233 |
+
def reset_parameters(self) -> None:
|
| 234 |
+
"""Reset Conv3d parameters with uniform initialization."""
|
| 235 |
+
k = 1 / (self.in_chans * self.tubelet_size * self.patch_size[0] * self.patch_size[1])
|
| 236 |
+
nn.init.uniform_(self.proj.weight, -math.sqrt(k), math.sqrt(k))
|
| 237 |
+
if self.proj.bias is not None:
|
| 238 |
+
nn.init.uniform_(self.proj.bias, -math.sqrt(k), math.sqrt(k))
|
| 239 |
+
|
| 240 |
+
|
| 241 |
+
# ---- rope ----
|
| 242 |
+
|
| 243 |
+
|
| 244 |
+
import math
|
| 245 |
+
from typing import Literal
|
| 246 |
+
|
| 247 |
+
import numpy as np
|
| 248 |
+
import torch
|
| 249 |
+
from torch import Tensor, nn
|
| 250 |
+
|
| 251 |
+
|
| 252 |
+
class RopePositionEmbedding3D(nn.Module):
|
| 253 |
+
"""Full 3D axial RoPE with independent T/H/W frequency bands.
|
| 254 |
+
|
| 255 |
+
Unlike the original MOTIF implementation which simply repeats 2D spatial angles
|
| 256 |
+
across temporal frames (making temporal positions indistinguishable), this
|
| 257 |
+
implementation partitions the head dimension into three axis groups:
|
| 258 |
+
|
| 259 |
+
D_head = D_T + D_H + D_W (no spare dimensions)
|
| 260 |
+
|
| 261 |
+
By default, the split is spatial-heavy for SSL (spatial quality is priority):
|
| 262 |
+
D_T = D_head // 4 (25% temporal)
|
| 263 |
+
D_H = (D_head - D_T) // 2 (37.5% height)
|
| 264 |
+
D_W = D_head - D_T - D_H (37.5% width)
|
| 265 |
+
e.g., D_head=64 → T=16, H=24, W=24
|
| 266 |
+
|
| 267 |
+
This can be overridden via ``fhw_dim=(D_T, D_H, D_W)`` for full control.
|
| 268 |
+
|
| 269 |
+
For images (T=1): t=0 for all tokens, making temporal angles constant
|
| 270 |
+
and the output is equivalent to spatial-only RoPE.
|
| 271 |
+
|
| 272 |
+
Args:
|
| 273 |
+
embed_dim: Total embedding dimension.
|
| 274 |
+
num_heads: Number of attention heads.
|
| 275 |
+
fhw_dim: Optional explicit (D_T, D_H, D_W) partition. Each must be even.
|
| 276 |
+
If None, uses the spatial-heavy default described above.
|
| 277 |
+
base: Frequency base (100.0 for spatial vision convention).
|
| 278 |
+
min_period: Minimum period (alternative to base).
|
| 279 |
+
max_period: Maximum period (alternative to base).
|
| 280 |
+
normalize_coords: How to normalize coordinates.
|
| 281 |
+
shift_coords: Random shift range during training.
|
| 282 |
+
jitter_coords: Random jitter multiplier during training.
|
| 283 |
+
rescale_coords: Random rescale multiplier during training.
|
| 284 |
+
dtype: Data type for computation.
|
| 285 |
+
device: Device for parameter allocation.
|
| 286 |
+
"""
|
| 287 |
+
|
| 288 |
+
def __init__(
|
| 289 |
+
self,
|
| 290 |
+
embed_dim: int,
|
| 291 |
+
*,
|
| 292 |
+
num_heads: int,
|
| 293 |
+
fhw_dim: tuple[int, int, int] | None = None,
|
| 294 |
+
base: float | None = 100.0,
|
| 295 |
+
min_period: float | None = None,
|
| 296 |
+
max_period: float | None = None,
|
| 297 |
+
normalize_coords: Literal["min", "max", "separate"] = "separate",
|
| 298 |
+
shift_coords: float | None = None,
|
| 299 |
+
jitter_coords: float | None = None,
|
| 300 |
+
rescale_coords: float | None = None,
|
| 301 |
+
dtype: torch.dtype | None = None,
|
| 302 |
+
device: torch.device | None = None,
|
| 303 |
+
) -> None:
|
| 304 |
+
super().__init__()
|
| 305 |
+
both_periods = min_period is not None and max_period is not None
|
| 306 |
+
if (base is None and not both_periods) or (base is not None and both_periods):
|
| 307 |
+
raise ValueError("Either `base` or `min_period`+`max_period` must be provided.")
|
| 308 |
+
|
| 309 |
+
D_head = embed_dim // num_heads
|
| 310 |
+
self.base = base
|
| 311 |
+
self.min_period = min_period
|
| 312 |
+
self.max_period = max_period
|
| 313 |
+
self.D_head = D_head
|
| 314 |
+
self.normalize_coords = normalize_coords
|
| 315 |
+
self.shift_coords = shift_coords
|
| 316 |
+
self.jitter_coords = jitter_coords
|
| 317 |
+
self.rescale_coords = rescale_coords
|
| 318 |
+
|
| 319 |
+
# Partition head dimension into 3 groups: T, H, W (no spare)
|
| 320 |
+
if fhw_dim is not None:
|
| 321 |
+
self.D_T, self.D_H, self.D_W = fhw_dim
|
| 322 |
+
assert self.D_T + self.D_H + self.D_W == D_head, (
|
| 323 |
+
f"fhw_dim must sum to D_head={D_head}, got {sum(fhw_dim)}"
|
| 324 |
+
)
|
| 325 |
+
else:
|
| 326 |
+
# Default: spatial-heavy split (SSL prioritizes spatial quality)
|
| 327 |
+
self.D_T = D_head // 4 # 25% temporal
|
| 328 |
+
self.D_H = (D_head - self.D_T) // 2 # 37.5% height
|
| 329 |
+
self.D_W = D_head - self.D_T - self.D_H # 37.5% width
|
| 330 |
+
assert self.D_T % 2 == 0 and self.D_H % 2 == 0 and self.D_W % 2 == 0, (
|
| 331 |
+
f"All axis dims must be even, got T={self.D_T}, H={self.D_H}, W={self.D_W}"
|
| 332 |
+
)
|
| 333 |
+
|
| 334 |
+
self.dtype = dtype
|
| 335 |
+
# Separate period buffers for each axis (n_freqs = D_axis // 2)
|
| 336 |
+
self.register_buffer(
|
| 337 |
+
"periods_t",
|
| 338 |
+
torch.empty(self.D_T // 2, device=device, dtype=dtype),
|
| 339 |
+
persistent=True,
|
| 340 |
+
)
|
| 341 |
+
self.register_buffer(
|
| 342 |
+
"periods_h",
|
| 343 |
+
torch.empty(self.D_H // 2, device=device, dtype=dtype),
|
| 344 |
+
persistent=True,
|
| 345 |
+
)
|
| 346 |
+
self.register_buffer(
|
| 347 |
+
"periods_w",
|
| 348 |
+
torch.empty(self.D_W // 2, device=device, dtype=dtype),
|
| 349 |
+
persistent=True,
|
| 350 |
+
)
|
| 351 |
+
self._init_weights()
|
| 352 |
+
|
| 353 |
+
def forward(self, *, T: int = 1, H: int, W: int) -> tuple[Tensor, Tensor]:
|
| 354 |
+
"""Compute 3D axial RoPE sin/cos for (T, H, W) grid.
|
| 355 |
+
|
| 356 |
+
The head dimension is partitioned as [D_T | D_H | D_W]:
|
| 357 |
+
- D_T: temporal frequency bands (angles vary with t)
|
| 358 |
+
- D_H: height frequency bands (angles vary with h)
|
| 359 |
+
- D_W: width frequency bands (angles vary with w)
|
| 360 |
+
|
| 361 |
+
For images (T=1), all tokens get t=0, so temporal angles are constant
|
| 362 |
+
and the output is equivalent to spatial-only RoPE.
|
| 363 |
+
|
| 364 |
+
Args:
|
| 365 |
+
T: Number of temporal positions (T_grid = num_frames // tubelet_size).
|
| 366 |
+
H: Height in patches.
|
| 367 |
+
W: Width in patches.
|
| 368 |
+
|
| 369 |
+
Returns:
|
| 370 |
+
Tuple of (sin, cos), each of shape (T*H*W, D_head).
|
| 371 |
+
"""
|
| 372 |
+
device = self.periods_t.device
|
| 373 |
+
dtype = self.dtype
|
| 374 |
+
dd = {"device": device, "dtype": dtype}
|
| 375 |
+
|
| 376 |
+
# 1. Compute normalized coordinates for each axis
|
| 377 |
+
if T > 1:
|
| 378 |
+
coords_t = torch.arange(0.5, T, **dd) / T # [T]
|
| 379 |
+
else:
|
| 380 |
+
coords_t = torch.tensor([0.5], **dd) # [1] - constant for images
|
| 381 |
+
|
| 382 |
+
coords_h, coords_w = self._compute_spatial_coords(H, W, **dd)
|
| 383 |
+
|
| 384 |
+
# Shift to [-1, +1] range
|
| 385 |
+
coords_t = 2.0 * coords_t - 1.0 # [T]
|
| 386 |
+
coords_h = 2.0 * coords_h - 1.0 # [H]
|
| 387 |
+
coords_w = 2.0 * coords_w - 1.0 # [W]
|
| 388 |
+
|
| 389 |
+
# Apply training-time augmentations to spatial coords only
|
| 390 |
+
if self.training:
|
| 391 |
+
coords_h, coords_w = self._augment_spatial_coords(coords_h, coords_w, dd)
|
| 392 |
+
|
| 393 |
+
# 2. Compute raw angles for each axis (n_freqs = D_axis // 2)
|
| 394 |
+
angles_t = 2 * math.pi * coords_t[:, None] / self.periods_t[None, :] # [T, D_T//2]
|
| 395 |
+
angles_h = 2 * math.pi * coords_h[:, None] / self.periods_h[None, :] # [H, D_H//2]
|
| 396 |
+
angles_w = 2 * math.pi * coords_w[:, None] / self.periods_w[None, :] # [W, D_W//2]
|
| 397 |
+
|
| 398 |
+
# 3. Build full 3D grid: create (T*H*W, D_head) angle tensor
|
| 399 |
+
t_idx, h_idx, w_idx = torch.meshgrid(
|
| 400 |
+
torch.arange(T, device=device),
|
| 401 |
+
torch.arange(H, device=device),
|
| 402 |
+
torch.arange(W, device=device),
|
| 403 |
+
indexing="ij",
|
| 404 |
+
)
|
| 405 |
+
t_idx = t_idx.flatten() # [T*H*W]
|
| 406 |
+
h_idx = h_idx.flatten() # [T*H*W]
|
| 407 |
+
w_idx = w_idx.flatten() # [T*H*W]
|
| 408 |
+
|
| 409 |
+
# Gather per-token raw angles and concatenate to D_head//2
|
| 410 |
+
token_angles_t = angles_t[t_idx] # [T*H*W, D_T//2]
|
| 411 |
+
token_angles_h = angles_h[h_idx] # [T*H*W, D_H//2]
|
| 412 |
+
token_angles_w = angles_w[w_idx] # [T*H*W, D_W//2]
|
| 413 |
+
angles_half = torch.cat([token_angles_t, token_angles_h, token_angles_w], dim=-1) # [T*H*W, D_head//2]
|
| 414 |
+
|
| 415 |
+
# tile(2) on full concat — matches MOTIF 2D RoPE pattern
|
| 416 |
+
# This ensures rotate_half pairs (dim i ↔ dim i+D//2) have identical angles,
|
| 417 |
+
# making the rotation orthogonal (preserves dot products in attention).
|
| 418 |
+
angles = angles_half.tile(2) # [T*H*W, D_head]
|
| 419 |
+
|
| 420 |
+
cos = torch.cos(angles)
|
| 421 |
+
sin = torch.sin(angles)
|
| 422 |
+
|
| 423 |
+
return (sin, cos)
|
| 424 |
+
|
| 425 |
+
def _compute_spatial_coords(self, H: int, W: int, **dd) -> tuple[Tensor, Tensor]:
|
| 426 |
+
"""Compute normalized spatial coordinates."""
|
| 427 |
+
if self.normalize_coords == "max":
|
| 428 |
+
max_HW = max(H, W)
|
| 429 |
+
coords_h = torch.arange(0.5, H, **dd) / max_HW
|
| 430 |
+
coords_w = torch.arange(0.5, W, **dd) / max_HW
|
| 431 |
+
elif self.normalize_coords == "min":
|
| 432 |
+
min_HW = min(H, W)
|
| 433 |
+
coords_h = torch.arange(0.5, H, **dd) / min_HW
|
| 434 |
+
coords_w = torch.arange(0.5, W, **dd) / min_HW
|
| 435 |
+
elif self.normalize_coords == "separate":
|
| 436 |
+
coords_h = torch.arange(0.5, H, **dd) / H
|
| 437 |
+
coords_w = torch.arange(0.5, W, **dd) / W
|
| 438 |
+
else:
|
| 439 |
+
raise ValueError(f"Unknown normalize_coords: {self.normalize_coords}")
|
| 440 |
+
return coords_h, coords_w
|
| 441 |
+
|
| 442 |
+
def _augment_spatial_coords(
|
| 443 |
+
self,
|
| 444 |
+
coords_h: Tensor,
|
| 445 |
+
coords_w: Tensor,
|
| 446 |
+
dd: dict,
|
| 447 |
+
) -> tuple[Tensor, Tensor]:
|
| 448 |
+
"""Apply training-time coordinate augmentations to spatial coords."""
|
| 449 |
+
if self.shift_coords is not None:
|
| 450 |
+
shift = torch.empty(2, **dd).uniform_(-self.shift_coords, self.shift_coords)
|
| 451 |
+
coords_h = coords_h + shift[0]
|
| 452 |
+
coords_w = coords_w + shift[1]
|
| 453 |
+
if self.jitter_coords is not None:
|
| 454 |
+
jitter_max = np.log(self.jitter_coords)
|
| 455 |
+
jitter = torch.empty(2, **dd).uniform_(-jitter_max, jitter_max).exp()
|
| 456 |
+
coords_h = coords_h * jitter[0]
|
| 457 |
+
coords_w = coords_w * jitter[1]
|
| 458 |
+
if self.rescale_coords is not None:
|
| 459 |
+
rescale_max = np.log(self.rescale_coords)
|
| 460 |
+
rescale = torch.empty(1, **dd).uniform_(-rescale_max, rescale_max).exp()
|
| 461 |
+
coords_h = coords_h * rescale
|
| 462 |
+
coords_w = coords_w * rescale
|
| 463 |
+
return coords_h, coords_w
|
| 464 |
+
|
| 465 |
+
def _compute_periods(self, n_freqs: int, device: torch.device, dtype: torch.dtype | None) -> Tensor:
|
| 466 |
+
"""Compute frequency periods for a single axis.
|
| 467 |
+
|
| 468 |
+
Args:
|
| 469 |
+
n_freqs: Number of frequency bands (D_axis // 2).
|
| 470 |
+
device: Device for tensor allocation.
|
| 471 |
+
dtype: Data type for computation.
|
| 472 |
+
|
| 473 |
+
Returns:
|
| 474 |
+
Tensor of shape (n_freqs,) with logarithmically spaced periods.
|
| 475 |
+
"""
|
| 476 |
+
if self.base is not None:
|
| 477 |
+
return self.base ** (
|
| 478 |
+
2 * torch.arange(n_freqs, device=device, dtype=dtype) / (2 * n_freqs)
|
| 479 |
+
)
|
| 480 |
+
else:
|
| 481 |
+
base = self.max_period / self.min_period
|
| 482 |
+
exponents = torch.linspace(0, 1, n_freqs, device=device, dtype=dtype)
|
| 483 |
+
periods = base**exponents
|
| 484 |
+
periods = periods / base
|
| 485 |
+
return periods * self.max_period
|
| 486 |
+
|
| 487 |
+
def _init_weights(self) -> None:
|
| 488 |
+
"""Initialize frequency periods for all three axes.
|
| 489 |
+
|
| 490 |
+
Each axis gets its own frequency schedule based on its dimension size:
|
| 491 |
+
periods[i] = base^(2i / D_axis)
|
| 492 |
+
|
| 493 |
+
This produces logarithmically spaced periods from 1.0 to base,
|
| 494 |
+
with more frequencies for axes with more allocated dimensions.
|
| 495 |
+
"""
|
| 496 |
+
device = self.periods_t.device
|
| 497 |
+
dtype = self.dtype
|
| 498 |
+
|
| 499 |
+
self.periods_t.data = self._compute_periods(self.D_T // 2, device, dtype)
|
| 500 |
+
self.periods_h.data = self._compute_periods(self.D_H // 2, device, dtype)
|
| 501 |
+
self.periods_w.data = self._compute_periods(self.D_W // 2, device, dtype)
|
| 502 |
+
|
| 503 |
+
|
| 504 |
+
# ---- attention ----
|
| 505 |
+
|
| 506 |
+
|
| 507 |
+
import math
|
| 508 |
+
|
| 509 |
+
import torch
|
| 510 |
+
import torch.nn.functional as F
|
| 511 |
+
from torch import Tensor, nn
|
| 512 |
+
|
| 513 |
+
|
| 514 |
+
|
| 515 |
+
def rope_rotate_half(x: Tensor) -> Tensor:
|
| 516 |
+
"""Rotate half of the dimensions: [-x2, x1] from [x1, x2].
|
| 517 |
+
|
| 518 |
+
Args:
|
| 519 |
+
x: Input tensor of shape (..., D).
|
| 520 |
+
|
| 521 |
+
Returns:
|
| 522 |
+
Rotated tensor of shape (..., D).
|
| 523 |
+
"""
|
| 524 |
+
x1, x2 = x.chunk(2, dim=-1)
|
| 525 |
+
return torch.cat([-x2, x1], dim=-1)
|
| 526 |
+
|
| 527 |
+
|
| 528 |
+
def rope_apply(x: Tensor, sin: Tensor, cos: Tensor) -> Tensor:
|
| 529 |
+
"""Apply rotary position embedding to input tensor.
|
| 530 |
+
|
| 531 |
+
Args:
|
| 532 |
+
x: Input tensor of shape (..., D).
|
| 533 |
+
sin: Sine angles of shape (..., D).
|
| 534 |
+
cos: Cosine angles of shape (..., D).
|
| 535 |
+
|
| 536 |
+
Returns:
|
| 537 |
+
Rotated tensor of shape (..., D).
|
| 538 |
+
"""
|
| 539 |
+
return (x * cos) + (rope_rotate_half(x) * sin)
|
| 540 |
+
|
| 541 |
+
|
| 542 |
+
class LinearKMaskedBias(nn.Linear):
|
| 543 |
+
"""Linear layer with masked bias for the K component of QKV.
|
| 544 |
+
|
| 545 |
+
Zeroes out the bias for the K component (middle third of output)
|
| 546 |
+
to avoid interference with RoPE positional encoding.
|
| 547 |
+
"""
|
| 548 |
+
|
| 549 |
+
def __init__(self, *args, **kwargs) -> None:
|
| 550 |
+
super().__init__(*args, **kwargs)
|
| 551 |
+
o = self.out_features
|
| 552 |
+
assert o % 3 == 0
|
| 553 |
+
if self.bias is not None:
|
| 554 |
+
self.register_buffer("bias_mask", torch.full_like(self.bias, fill_value=math.nan))
|
| 555 |
+
|
| 556 |
+
def forward(self, input: Tensor) -> Tensor:
|
| 557 |
+
"""Forward pass with masked bias."""
|
| 558 |
+
masked_bias = self.bias * self.bias_mask.to(self.bias.dtype) if self.bias is not None else None
|
| 559 |
+
return F.linear(input, self.weight, masked_bias)
|
| 560 |
+
|
| 561 |
+
|
| 562 |
+
class SelfAttention(nn.Module):
|
| 563 |
+
"""Multi-head self-attention with RoPE support.
|
| 564 |
+
|
| 565 |
+
Uses torch.nn.functional.scaled_dot_product_attention for FlashAttention
|
| 566 |
+
compatibility. RoPE is applied to Q and K on patch tokens only (not CLS/register).
|
| 567 |
+
|
| 568 |
+
Args:
|
| 569 |
+
dim: Model dimension.
|
| 570 |
+
num_heads: Number of attention heads.
|
| 571 |
+
qkv_bias: Whether to use bias in QKV projection.
|
| 572 |
+
proj_bias: Whether to use bias in output projection.
|
| 573 |
+
attn_drop: Attention dropout probability.
|
| 574 |
+
proj_drop: Output projection dropout probability.
|
| 575 |
+
mask_k_bias: Whether to mask K bias (for RoPE compatibility).
|
| 576 |
+
device: Device for parameter allocation.
|
| 577 |
+
gated_attention: Gated attention variant. None disables gating,
|
| 578 |
+
"headwise" applies a per-head scalar gate, "elementwise" applies
|
| 579 |
+
a per-element gate. Gate scores are query-dependent (derived from
|
| 580 |
+
input) and applied as sigmoid after SDPA.
|
| 581 |
+
Reference: https://arxiv.org/abs/2505.06708
|
| 582 |
+
"""
|
| 583 |
+
|
| 584 |
+
def __init__(
|
| 585 |
+
self,
|
| 586 |
+
dim: int,
|
| 587 |
+
num_heads: int = 8,
|
| 588 |
+
qkv_bias: bool = False,
|
| 589 |
+
proj_bias: bool = True,
|
| 590 |
+
attn_drop: float = 0.0,
|
| 591 |
+
proj_drop: float = 0.0,
|
| 592 |
+
mask_k_bias: bool = False,
|
| 593 |
+
device: str | None = None,
|
| 594 |
+
gated_attention: str | None = None,
|
| 595 |
+
qk_norm: bool = False,
|
| 596 |
+
) -> None:
|
| 597 |
+
super().__init__()
|
| 598 |
+
self.num_heads = num_heads
|
| 599 |
+
self.head_dim = dim // num_heads
|
| 600 |
+
self.scale = self.head_dim**-0.5
|
| 601 |
+
|
| 602 |
+
linear_class = LinearKMaskedBias if mask_k_bias else nn.Linear
|
| 603 |
+
self.qkv = linear_class(dim, dim * 3, bias=qkv_bias, device=device)
|
| 604 |
+
self.attn_drop = nn.Dropout(attn_drop)
|
| 605 |
+
self.proj = nn.Linear(dim, dim, bias=proj_bias, device=device)
|
| 606 |
+
self.proj_drop = nn.Dropout(proj_drop)
|
| 607 |
+
|
| 608 |
+
self.qk_norm = qk_norm
|
| 609 |
+
if qk_norm:
|
| 610 |
+
self.q_norm = RMSNorm(self.head_dim, device=device)
|
| 611 |
+
self.k_norm = RMSNorm(self.head_dim, device=device)
|
| 612 |
+
|
| 613 |
+
self.gated_attention = gated_attention
|
| 614 |
+
if gated_attention == "headwise":
|
| 615 |
+
self.gate_proj = nn.Linear(dim, num_heads, bias=True, device=device)
|
| 616 |
+
elif gated_attention == "elementwise":
|
| 617 |
+
self.gate_proj = nn.Linear(dim, dim, bias=True, device=device)
|
| 618 |
+
elif gated_attention is not None:
|
| 619 |
+
raise ValueError(f"Unknown gated_attention mode: {gated_attention!r}. Use 'headwise' or 'elementwise'.")
|
| 620 |
+
|
| 621 |
+
def apply_rope(
|
| 622 |
+
self,
|
| 623 |
+
q: Tensor,
|
| 624 |
+
k: Tensor,
|
| 625 |
+
rope: tuple[Tensor, Tensor],
|
| 626 |
+
) -> tuple[Tensor, Tensor]:
|
| 627 |
+
"""Apply RoPE to query and key tensors.
|
| 628 |
+
|
| 629 |
+
RoPE is applied only to patch tokens (prefix tokens like CLS and register
|
| 630 |
+
are excluded based on the difference between sequence length and rope length).
|
| 631 |
+
|
| 632 |
+
Args:
|
| 633 |
+
q: Query tensor of shape (B, heads, N, D_head).
|
| 634 |
+
k: Key tensor of shape (B, heads, N, D_head).
|
| 635 |
+
rope: Tuple of (sin, cos), each of shape (N_patches, D_head).
|
| 636 |
+
|
| 637 |
+
Returns:
|
| 638 |
+
Tuple of rotated (q, k) tensors.
|
| 639 |
+
"""
|
| 640 |
+
q_dtype = q.dtype
|
| 641 |
+
k_dtype = k.dtype
|
| 642 |
+
sin, cos = rope
|
| 643 |
+
rope_dtype = sin.dtype
|
| 644 |
+
q = q.to(dtype=rope_dtype)
|
| 645 |
+
k = k.to(dtype=rope_dtype)
|
| 646 |
+
N = q.shape[-2]
|
| 647 |
+
prefix = N - sin.shape[-2]
|
| 648 |
+
assert prefix >= 0
|
| 649 |
+
q_prefix = q[:, :, :prefix, :]
|
| 650 |
+
q = rope_apply(q[:, :, prefix:, :], sin, cos)
|
| 651 |
+
q = torch.cat((q_prefix, q), dim=-2)
|
| 652 |
+
k_prefix = k[:, :, :prefix, :]
|
| 653 |
+
k = rope_apply(k[:, :, prefix:, :], sin, cos)
|
| 654 |
+
k = torch.cat((k_prefix, k), dim=-2)
|
| 655 |
+
q = q.to(dtype=q_dtype)
|
| 656 |
+
k = k.to(dtype=k_dtype)
|
| 657 |
+
return q, k
|
| 658 |
+
|
| 659 |
+
def forward(self, x: Tensor, attn_bias: Tensor | None = None, rope: Tensor | None = None) -> Tensor:
|
| 660 |
+
"""Forward pass for single tensor input.
|
| 661 |
+
|
| 662 |
+
Args:
|
| 663 |
+
x: Input tensor of shape (B, N, D).
|
| 664 |
+
attn_bias: Unused (kept for interface compatibility).
|
| 665 |
+
rope: Optional RoPE (sin, cos) tuple.
|
| 666 |
+
|
| 667 |
+
Returns:
|
| 668 |
+
Output tensor of shape (B, N, D).
|
| 669 |
+
"""
|
| 670 |
+
gate_score = self._compute_gate(x) if self.gated_attention else None
|
| 671 |
+
qkv = self.qkv(x)
|
| 672 |
+
attn_v = self.compute_attention(qkv=qkv, attn_bias=attn_bias, rope=rope, gate_score=gate_score)
|
| 673 |
+
x = self.proj(attn_v)
|
| 674 |
+
x = self.proj_drop(x)
|
| 675 |
+
return x
|
| 676 |
+
|
| 677 |
+
def forward_list(
|
| 678 |
+
self,
|
| 679 |
+
x_list: list[Tensor],
|
| 680 |
+
attn_bias: Tensor | None = None,
|
| 681 |
+
rope_list: list[tuple[Tensor, Tensor]] | None = None,
|
| 682 |
+
) -> list[Tensor]:
|
| 683 |
+
"""Forward pass for list of tensors (multi-crop efficiency).
|
| 684 |
+
|
| 685 |
+
Concatenates inputs for a single QKV projection, then splits for per-crop
|
| 686 |
+
attention computation (needed because different crops have different RoPE).
|
| 687 |
+
|
| 688 |
+
Args:
|
| 689 |
+
x_list: List of input tensors.
|
| 690 |
+
attn_bias: Unused.
|
| 691 |
+
rope_list: List of RoPE (sin, cos) tuples, one per input.
|
| 692 |
+
|
| 693 |
+
Returns:
|
| 694 |
+
List of output tensors.
|
| 695 |
+
"""
|
| 696 |
+
assert len(x_list) == len(rope_list)
|
| 697 |
+
x_flat, shapes, num_tokens = cat_keep_shapes(x_list)
|
| 698 |
+
qkv_flat = self.qkv(x_flat)
|
| 699 |
+
qkv_list = uncat_with_shapes(qkv_flat, shapes, num_tokens)
|
| 700 |
+
|
| 701 |
+
if self.gated_attention:
|
| 702 |
+
gate_flat = self._compute_gate(x_flat)
|
| 703 |
+
gate_list = uncat_with_shapes(gate_flat, shapes, num_tokens)
|
| 704 |
+
else:
|
| 705 |
+
gate_list = [None] * len(x_list)
|
| 706 |
+
|
| 707 |
+
att_out = []
|
| 708 |
+
for qkv, _, rope, gate_score in zip(qkv_list, shapes, rope_list, gate_list):
|
| 709 |
+
att_out.append(self.compute_attention(qkv, attn_bias=attn_bias, rope=rope, gate_score=gate_score))
|
| 710 |
+
x_flat, shapes, num_tokens = cat_keep_shapes(att_out)
|
| 711 |
+
x_flat = self.proj(x_flat)
|
| 712 |
+
return uncat_with_shapes(x_flat, shapes, num_tokens)
|
| 713 |
+
|
| 714 |
+
def _compute_gate(self, x: Tensor) -> Tensor:
|
| 715 |
+
"""Compute raw gate scores from input.
|
| 716 |
+
|
| 717 |
+
Returns the raw projection without reshaping so that the output keeps
|
| 718 |
+
the same number of leading dimensions as ``x``. This is critical for
|
| 719 |
+
``forward_list`` where ``uncat_with_shapes`` must split a 2-D flat
|
| 720 |
+
tensor back to per-crop 3-D tensors — adding extra dims here would
|
| 721 |
+
break that reshape. The per-head unflatten happens later inside
|
| 722 |
+
``compute_attention`` where B and N are known.
|
| 723 |
+
|
| 724 |
+
Args:
|
| 725 |
+
x: Input tensor of shape (..., D). Supports both 2D (flat) and 3D (batched).
|
| 726 |
+
|
| 727 |
+
Returns:
|
| 728 |
+
Raw gate projection. Headwise: (..., num_heads). Elementwise: (..., D).
|
| 729 |
+
"""
|
| 730 |
+
return self.gate_proj(x)
|
| 731 |
+
|
| 732 |
+
def compute_attention(
|
| 733 |
+
self,
|
| 734 |
+
qkv: Tensor,
|
| 735 |
+
attn_bias: Tensor | None = None,
|
| 736 |
+
rope: tuple[Tensor, Tensor] | None = None,
|
| 737 |
+
gate_score: Tensor | None = None,
|
| 738 |
+
) -> Tensor:
|
| 739 |
+
"""Compute scaled dot-product attention.
|
| 740 |
+
|
| 741 |
+
Args:
|
| 742 |
+
qkv: Combined QKV tensor of shape (B, N, 3*D).
|
| 743 |
+
attn_bias: Unused.
|
| 744 |
+
rope: Optional RoPE (sin, cos) tuple.
|
| 745 |
+
gate_score: Optional gate tensor from _compute_gate.
|
| 746 |
+
|
| 747 |
+
Returns:
|
| 748 |
+
Attention output of shape (B, N, D).
|
| 749 |
+
"""
|
| 750 |
+
assert attn_bias is None
|
| 751 |
+
B, N, _ = qkv.shape
|
| 752 |
+
C = self.qkv.in_features
|
| 753 |
+
|
| 754 |
+
qkv = qkv.reshape(B, N, 3, self.num_heads, self.head_dim)
|
| 755 |
+
q, k, v = torch.unbind(qkv, 2)
|
| 756 |
+
q, k, v = [t.transpose(1, 2) for t in [q, k, v]]
|
| 757 |
+
if self.qk_norm:
|
| 758 |
+
q = self.q_norm(q)
|
| 759 |
+
k = self.k_norm(k)
|
| 760 |
+
if rope is not None:
|
| 761 |
+
q, k = self.apply_rope(q, k, rope)
|
| 762 |
+
x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
|
| 763 |
+
x = x.transpose(1, 2) # (B, N, num_heads, head_dim)
|
| 764 |
+
if gate_score is not None:
|
| 765 |
+
# _compute_gate returns raw projection: (..., num_heads) or (..., D).
|
| 766 |
+
# Reshape to (B, N, num_heads, 1) or (B, N, num_heads, head_dim) here.
|
| 767 |
+
if self.gated_attention == "headwise":
|
| 768 |
+
gate_score = gate_score.unflatten(-1, (self.num_heads, 1))
|
| 769 |
+
else: # elementwise
|
| 770 |
+
gate_score = gate_score.unflatten(-1, (self.num_heads, self.head_dim))
|
| 771 |
+
x = x * torch.sigmoid(gate_score)
|
| 772 |
+
return x.reshape([B, N, C])
|
| 773 |
+
|
| 774 |
+
|
| 775 |
+
# ---- ffn ----
|
| 776 |
+
|
| 777 |
+
|
| 778 |
+
from typing import Callable
|
| 779 |
+
|
| 780 |
+
import torch.nn.functional as F
|
| 781 |
+
from torch import Tensor, nn
|
| 782 |
+
|
| 783 |
+
|
| 784 |
+
|
| 785 |
+
class ListForwardMixin:
|
| 786 |
+
"""Mixin providing forward_list for efficient multi-crop processing."""
|
| 787 |
+
|
| 788 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 789 |
+
"""Forward pass for a single tensor."""
|
| 790 |
+
raise NotImplementedError
|
| 791 |
+
|
| 792 |
+
def forward_list(self, x_list: list[Tensor]) -> list[Tensor]:
|
| 793 |
+
"""Forward pass for a list of tensors, concatenated for efficiency."""
|
| 794 |
+
x_flat, shapes, num_tokens = cat_keep_shapes(x_list)
|
| 795 |
+
x_flat = self.forward(x_flat)
|
| 796 |
+
return uncat_with_shapes(x_flat, shapes, num_tokens)
|
| 797 |
+
|
| 798 |
+
|
| 799 |
+
class SwiGLUFFN(nn.Module, ListForwardMixin):
|
| 800 |
+
"""SwiGLU Feed-Forward Network: w3(silu(w1(x)) * w2(x)).
|
| 801 |
+
|
| 802 |
+
Used for larger ViT models (SO400M+) due to better gradient flow.
|
| 803 |
+
Hidden dimension is aligned to a multiple of `align_to` for GPU efficiency.
|
| 804 |
+
|
| 805 |
+
Args:
|
| 806 |
+
in_features: Input dimension.
|
| 807 |
+
hidden_features: Hidden dimension before alignment.
|
| 808 |
+
out_features: Output dimension (default: same as in_features).
|
| 809 |
+
act_layer: Unused (SwiGLU has built-in SiLU activation).
|
| 810 |
+
drop: Unused (no dropout in SwiGLU).
|
| 811 |
+
bias: Whether to use bias in linear layers.
|
| 812 |
+
align_to: Align hidden dimension to this multiple.
|
| 813 |
+
device: Device for parameter allocation.
|
| 814 |
+
"""
|
| 815 |
+
|
| 816 |
+
def __init__(
|
| 817 |
+
self,
|
| 818 |
+
in_features: int,
|
| 819 |
+
hidden_features: int | None = None,
|
| 820 |
+
out_features: int | None = None,
|
| 821 |
+
act_layer: Callable[..., nn.Module] | None = None,
|
| 822 |
+
drop: float = 0.0,
|
| 823 |
+
bias: bool = True,
|
| 824 |
+
align_to: int = 8,
|
| 825 |
+
device: str | None = None,
|
| 826 |
+
) -> None:
|
| 827 |
+
super().__init__()
|
| 828 |
+
out_features = out_features or in_features
|
| 829 |
+
hidden_features = hidden_features or in_features
|
| 830 |
+
d = int(hidden_features * 2 / 3)
|
| 831 |
+
swiglu_hidden_features = d + (-d % align_to)
|
| 832 |
+
self.w1 = nn.Linear(in_features, swiglu_hidden_features, bias=bias, device=device)
|
| 833 |
+
self.w2 = nn.Linear(in_features, swiglu_hidden_features, bias=bias, device=device)
|
| 834 |
+
self.w3 = nn.Linear(swiglu_hidden_features, out_features, bias=bias, device=device)
|
| 835 |
+
|
| 836 |
+
def forward(self, x: Tensor) -> Tensor:
|
| 837 |
+
"""Forward pass: w3(silu(w1(x)) * w2(x))."""
|
| 838 |
+
x1 = self.w1(x)
|
| 839 |
+
x2 = self.w2(x)
|
| 840 |
+
hidden = F.silu(x1) * x2
|
| 841 |
+
return self.w3(hidden)
|
| 842 |
+
|
| 843 |
+
|
| 844 |
+
# ---- block ----
|
| 845 |
+
|
| 846 |
+
|
| 847 |
+
from typing import Callable
|
| 848 |
+
|
| 849 |
+
import torch
|
| 850 |
+
from torch import Tensor, nn
|
| 851 |
+
|
| 852 |
+
|
| 853 |
+
torch._dynamo.config.automatic_dynamic_shapes = False
|
| 854 |
+
torch._dynamo.config.accumulated_cache_size_limit = 1024
|
| 855 |
+
# Per-code-object recompile cap (default 8). With dynamic=False + fullgraph=True,
|
| 856 |
+
# every distinct input shape forces a static recompile of the block. Gram adds the
|
| 857 |
+
# Gram-teacher forward at gram_teacher_crops_size (e.g. 768/1152) on top of the
|
| 858 |
+
# 512/768 image + video + local-crop shapes, pushing the distinct-shape count past
|
| 859 |
+
# 8 → FailOnRecompileLimitHit (hard crash, not the eager-fallback warning seen when
|
| 860 |
+
# fullgraph=False). Raise the cap so all Gram shapes compile; the accumulated cap
|
| 861 |
+
# (1024) still bounds total graphs. Only the shapes that actually occur are compiled,
|
| 862 |
+
# so a high cap costs nothing when Gram is off.
|
| 863 |
+
torch._dynamo.config.recompile_limit = 64
|
| 864 |
+
|
| 865 |
+
|
| 866 |
+
|
| 867 |
+
class SelfAttentionBlock(nn.Module):
|
| 868 |
+
"""Pre-norm transformer block: Norm -> Attention -> LayerScale -> DropPath -> Residual.
|
| 869 |
+
|
| 870 |
+
Supports both single-tensor and list-of-tensors forward for efficient multi-crop
|
| 871 |
+
processing during MOTIF training.
|
| 872 |
+
|
| 873 |
+
Args:
|
| 874 |
+
dim: Model dimension.
|
| 875 |
+
num_heads: Number of attention heads.
|
| 876 |
+
ffn_ratio: FFN hidden dimension ratio.
|
| 877 |
+
qkv_bias: Whether to use bias in QKV projection.
|
| 878 |
+
proj_bias: Whether to use bias in output projection.
|
| 879 |
+
ffn_bias: Whether to use bias in FFN layers.
|
| 880 |
+
drop: Dropout probability.
|
| 881 |
+
attn_drop: Attention dropout probability.
|
| 882 |
+
init_values: LayerScale initial values (None disables LayerScale).
|
| 883 |
+
drop_path: Stochastic depth drop probability.
|
| 884 |
+
act_layer: Activation function class.
|
| 885 |
+
norm_layer: Normalization layer class.
|
| 886 |
+
attn_class: Attention class.
|
| 887 |
+
ffn_layer: FFN class.
|
| 888 |
+
mask_k_bias: Whether to mask K bias.
|
| 889 |
+
device: Device for parameter allocation.
|
| 890 |
+
"""
|
| 891 |
+
|
| 892 |
+
def __init__(
|
| 893 |
+
self,
|
| 894 |
+
dim: int,
|
| 895 |
+
num_heads: int,
|
| 896 |
+
ffn_ratio: float = 4.0,
|
| 897 |
+
qkv_bias: bool = False,
|
| 898 |
+
proj_bias: bool = True,
|
| 899 |
+
ffn_bias: bool = True,
|
| 900 |
+
drop: float = 0.0,
|
| 901 |
+
attn_drop: float = 0.0,
|
| 902 |
+
init_values: float | None = None,
|
| 903 |
+
drop_path: float = 0.0,
|
| 904 |
+
act_layer: Callable[..., nn.Module] = nn.GELU,
|
| 905 |
+
norm_layer: Callable[..., nn.Module] = nn.LayerNorm,
|
| 906 |
+
attn_class: Callable[..., nn.Module] = SelfAttention,
|
| 907 |
+
ffn_layer: Callable[..., nn.Module] = SwiGLUFFN,
|
| 908 |
+
mask_k_bias: bool = False,
|
| 909 |
+
device: str | None = None,
|
| 910 |
+
gated_attention: str | None = None,
|
| 911 |
+
qk_norm: bool = False,
|
| 912 |
+
) -> None:
|
| 913 |
+
super().__init__()
|
| 914 |
+
self.norm1 = norm_layer(dim)
|
| 915 |
+
self.attn = attn_class(
|
| 916 |
+
dim,
|
| 917 |
+
num_heads=num_heads,
|
| 918 |
+
qkv_bias=qkv_bias,
|
| 919 |
+
proj_bias=proj_bias,
|
| 920 |
+
attn_drop=attn_drop,
|
| 921 |
+
proj_drop=drop,
|
| 922 |
+
mask_k_bias=mask_k_bias,
|
| 923 |
+
device=device,
|
| 924 |
+
gated_attention=gated_attention,
|
| 925 |
+
qk_norm=qk_norm,
|
| 926 |
+
)
|
| 927 |
+
self.ls1 = LayerScale(dim, init_values=init_values, device=device) if init_values else nn.Identity()
|
| 928 |
+
|
| 929 |
+
self.norm2 = norm_layer(dim)
|
| 930 |
+
mlp_hidden_dim = int(dim * ffn_ratio)
|
| 931 |
+
self.mlp = ffn_layer(
|
| 932 |
+
in_features=dim,
|
| 933 |
+
hidden_features=mlp_hidden_dim,
|
| 934 |
+
act_layer=act_layer,
|
| 935 |
+
drop=drop,
|
| 936 |
+
bias=ffn_bias,
|
| 937 |
+
device=device,
|
| 938 |
+
)
|
| 939 |
+
self.ls2 = LayerScale(dim, init_values=init_values, device=device) if init_values else nn.Identity()
|
| 940 |
+
|
| 941 |
+
self.sample_drop_ratio = drop_path
|
| 942 |
+
|
| 943 |
+
@staticmethod
|
| 944 |
+
def _maybe_index_rope(
|
| 945 |
+
rope: tuple[Tensor, Tensor] | None,
|
| 946 |
+
indices: Tensor,
|
| 947 |
+
) -> tuple[Tensor, Tensor] | None:
|
| 948 |
+
"""Optionally index into RoPE embeddings for stochastic depth."""
|
| 949 |
+
if rope is None:
|
| 950 |
+
return None
|
| 951 |
+
sin, cos = rope
|
| 952 |
+
assert sin.ndim == cos.ndim
|
| 953 |
+
if sin.ndim == 4:
|
| 954 |
+
return sin[indices], cos[indices]
|
| 955 |
+
else:
|
| 956 |
+
return sin, cos
|
| 957 |
+
|
| 958 |
+
def _forward_list(self, x_list: list[Tensor], rope_list: list | None = None) -> list[Tensor]:
|
| 959 |
+
"""Forward pass for list of tensors with stochastic depth support.
|
| 960 |
+
|
| 961 |
+
Concatenates tokens from multiple inputs for efficient elementwise operations,
|
| 962 |
+
then splits for per-input attention computation (different RoPE per crop).
|
| 963 |
+
"""
|
| 964 |
+
b_list = [x.shape[0] for x in x_list]
|
| 965 |
+
sample_subset_sizes = [max(int(b * (1 - self.sample_drop_ratio)), 1) for b in b_list]
|
| 966 |
+
residual_scale_factors = [b / s for b, s in zip(b_list, sample_subset_sizes)]
|
| 967 |
+
|
| 968 |
+
if self.training and self.sample_drop_ratio > 0.0:
|
| 969 |
+
indices_1_list = [
|
| 970 |
+
(torch.randperm(b, device=x.device))[:s]
|
| 971 |
+
for x, b, s in zip(x_list, b_list, sample_subset_sizes)
|
| 972 |
+
]
|
| 973 |
+
x_subset_1_list = [x[idx] for x, idx in zip(x_list, indices_1_list)]
|
| 974 |
+
|
| 975 |
+
if rope_list is not None:
|
| 976 |
+
rope_subset_list = [
|
| 977 |
+
self._maybe_index_rope(rope, idx) for rope, idx in zip(rope_list, indices_1_list)
|
| 978 |
+
]
|
| 979 |
+
else:
|
| 980 |
+
rope_subset_list = rope_list
|
| 981 |
+
|
| 982 |
+
norm1 = [self.norm1(x) for x in x_subset_1_list]
|
| 983 |
+
residual_1_list = self.attn.forward_list(norm1, rope_list=rope_subset_list)
|
| 984 |
+
|
| 985 |
+
x_attn_list = [
|
| 986 |
+
torch.index_add(
|
| 987 |
+
x, dim=0, source=self.ls1(r), index=idx, alpha=scale,
|
| 988 |
+
)
|
| 989 |
+
for x, r, idx, scale in zip(x_list, residual_1_list, indices_1_list, residual_scale_factors)
|
| 990 |
+
]
|
| 991 |
+
|
| 992 |
+
indices_2_list = [
|
| 993 |
+
(torch.randperm(b, device=x.device))[:s]
|
| 994 |
+
for x, b, s in zip(x_list, b_list, sample_subset_sizes)
|
| 995 |
+
]
|
| 996 |
+
x_subset_2_list = [x[idx] for x, idx in zip(x_attn_list, indices_2_list)]
|
| 997 |
+
norm2_list = [self.norm2(x) for x in x_subset_2_list]
|
| 998 |
+
|
| 999 |
+
residual_2_list = self.mlp.forward_list(norm2_list)
|
| 1000 |
+
|
| 1001 |
+
x_ffn = [
|
| 1002 |
+
torch.index_add(
|
| 1003 |
+
xa, dim=0, source=self.ls2(r), index=idx, alpha=scale,
|
| 1004 |
+
)
|
| 1005 |
+
for xa, r, idx, scale in zip(x_attn_list, residual_2_list, indices_2_list, residual_scale_factors)
|
| 1006 |
+
]
|
| 1007 |
+
else:
|
| 1008 |
+
x_out = []
|
| 1009 |
+
for x, rope in zip(x_list, rope_list):
|
| 1010 |
+
x_attn = x + self.ls1(self.attn(self.norm1(x), rope=rope))
|
| 1011 |
+
x_ffn_item = x_attn + self.ls2(self.mlp(self.norm2(x_attn)))
|
| 1012 |
+
x_out.append(x_ffn_item)
|
| 1013 |
+
x_ffn = x_out
|
| 1014 |
+
|
| 1015 |
+
return x_ffn
|
| 1016 |
+
|
| 1017 |
+
def forward(
|
| 1018 |
+
self,
|
| 1019 |
+
x_or_x_list: Tensor | list[Tensor],
|
| 1020 |
+
rope_or_rope_list: tuple | list | None = None,
|
| 1021 |
+
) -> Tensor | list[Tensor]:
|
| 1022 |
+
"""Forward pass accepting either a single tensor or list of tensors.
|
| 1023 |
+
|
| 1024 |
+
Args:
|
| 1025 |
+
x_or_x_list: Single tensor (B, N, D) or list of tensors.
|
| 1026 |
+
rope_or_rope_list: Single RoPE tuple or list of RoPE tuples.
|
| 1027 |
+
|
| 1028 |
+
Returns:
|
| 1029 |
+
Output tensor(s) matching input format.
|
| 1030 |
+
"""
|
| 1031 |
+
if isinstance(x_or_x_list, Tensor):
|
| 1032 |
+
return self._forward_list([x_or_x_list], rope_list=[rope_or_rope_list])[0]
|
| 1033 |
+
elif isinstance(x_or_x_list, list):
|
| 1034 |
+
if rope_or_rope_list is None:
|
| 1035 |
+
rope_or_rope_list = [None for _ in x_or_x_list]
|
| 1036 |
+
return self._forward_list(x_or_x_list, rope_list=rope_or_rope_list)
|
| 1037 |
+
else:
|
| 1038 |
+
raise AssertionError(f"Unexpected input type: {type(x_or_x_list)}")
|
| 1039 |
+
|
| 1040 |
+
|
| 1041 |
+
# ---- vision_transformer ----
|
| 1042 |
+
|
| 1043 |
+
|
| 1044 |
+
import logging
|
| 1045 |
+
from functools import partial
|
| 1046 |
+
from typing import Any
|
| 1047 |
+
|
| 1048 |
+
import torch
|
| 1049 |
+
import torch.nn.init
|
| 1050 |
+
from torch import Tensor, nn
|
| 1051 |
+
|
| 1052 |
+
|
| 1053 |
+
logger = logging.getLogger("motif")
|
| 1054 |
+
|
| 1055 |
+
ffn_layer_dict: dict[str, type] = {
|
| 1056 |
+
"swiglu": SwiGLUFFN,
|
| 1057 |
+
"swiglu32": partial(SwiGLUFFN, align_to=32),
|
| 1058 |
+
"swiglu64": partial(SwiGLUFFN, align_to=64),
|
| 1059 |
+
"swiglu128": partial(SwiGLUFFN, align_to=128),
|
| 1060 |
+
}
|
| 1061 |
+
|
| 1062 |
+
norm_layer_dict: dict[str, type] = {
|
| 1063 |
+
"layernorm": partial(nn.LayerNorm, eps=1e-6),
|
| 1064 |
+
"layernormbf16": partial(nn.LayerNorm, eps=1e-5),
|
| 1065 |
+
"rmsnorm": RMSNorm,
|
| 1066 |
+
}
|
| 1067 |
+
|
| 1068 |
+
dtype_dict: dict[str, torch.dtype] = {
|
| 1069 |
+
"fp32": torch.float32,
|
| 1070 |
+
"fp16": torch.float16,
|
| 1071 |
+
"bf16": torch.bfloat16,
|
| 1072 |
+
}
|
| 1073 |
+
|
| 1074 |
+
|
| 1075 |
+
def init_weights_vit(module: nn.Module, name: str = "") -> None:
|
| 1076 |
+
"""Initialize weights for ViT components.
|
| 1077 |
+
|
| 1078 |
+
Applied recursively via named_apply to all submodules.
|
| 1079 |
+
|
| 1080 |
+
Args:
|
| 1081 |
+
module: Module to initialize.
|
| 1082 |
+
name: Module name (for logging, unused).
|
| 1083 |
+
"""
|
| 1084 |
+
if isinstance(module, nn.Linear):
|
| 1085 |
+
torch.nn.init.trunc_normal_(module.weight, std=0.02)
|
| 1086 |
+
if module.bias is not None:
|
| 1087 |
+
nn.init.zeros_(module.bias)
|
| 1088 |
+
if hasattr(module, "bias_mask") and module.bias_mask is not None:
|
| 1089 |
+
o = module.out_features
|
| 1090 |
+
module.bias_mask.fill_(1)
|
| 1091 |
+
module.bias_mask[o // 3 : 2 * o // 3].fill_(0)
|
| 1092 |
+
if isinstance(module, nn.LayerNorm):
|
| 1093 |
+
module.reset_parameters()
|
| 1094 |
+
if isinstance(module, LayerScale):
|
| 1095 |
+
module.reset_parameters()
|
| 1096 |
+
if isinstance(module, PatchEmbed):
|
| 1097 |
+
module.reset_parameters()
|
| 1098 |
+
if isinstance(module, RMSNorm):
|
| 1099 |
+
module.reset_parameters()
|
| 1100 |
+
|
| 1101 |
+
|
| 1102 |
+
class MotifVisionTransformer(nn.Module):
|
| 1103 |
+
"""Vision Transformer backbone with 3D RoPE for unified image/video processing.
|
| 1104 |
+
|
| 1105 |
+
Key features:
|
| 1106 |
+
- PatchEmbed (Conv3d) for unified image/video tokenization
|
| 1107 |
+
- Full 3D axial RoPE positional encoding (T/H/W)
|
| 1108 |
+
- CLS token + register (storage) tokens
|
| 1109 |
+
- LayerScale and stochastic depth (DropPath)
|
| 1110 |
+
- MLP or SwiGLU FFN variants
|
| 1111 |
+
- FSDP-compatible (SHARD_GRAD_OP strategy)
|
| 1112 |
+
|
| 1113 |
+
Token sequence layout: [CLS] + [Register x n_storage_tokens] + [Patch x N_total]
|
| 1114 |
+
|
| 1115 |
+
Args:
|
| 1116 |
+
img_size: Input image size.
|
| 1117 |
+
patch_size: Spatial patch size.
|
| 1118 |
+
in_chans: Number of input channels.
|
| 1119 |
+
embed_dim: Embedding dimension.
|
| 1120 |
+
depth: Number of transformer blocks.
|
| 1121 |
+
num_heads: Number of attention heads.
|
| 1122 |
+
ffn_ratio: FFN hidden dimension ratio.
|
| 1123 |
+
qkv_bias: Whether to use bias in QKV.
|
| 1124 |
+
drop_path_rate: Stochastic depth rate.
|
| 1125 |
+
layerscale_init: LayerScale initial value (None to disable).
|
| 1126 |
+
norm_layer: Normalization layer name.
|
| 1127 |
+
ffn_layer: FFN layer name.
|
| 1128 |
+
ffn_bias: Whether to use bias in FFN.
|
| 1129 |
+
proj_bias: Whether to use bias in attention output projection.
|
| 1130 |
+
n_storage_tokens: Number of register tokens.
|
| 1131 |
+
mask_k_bias: Whether to mask K bias in attention.
|
| 1132 |
+
untie_cls_and_patch_norms: Use separate norms for CLS and patch tokens.
|
| 1133 |
+
untie_global_and_local_cls_norm: Use separate norm for local CLS tokens.
|
| 1134 |
+
device: Device for parameter allocation.
|
| 1135 |
+
num_frames: Number of input video frames.
|
| 1136 |
+
tubelet_size: Temporal patch size for Conv3d.
|
| 1137 |
+
pos_embed_rope_base: RoPE frequency base (100.0 for vision).
|
| 1138 |
+
gated_attention: Gated attention variant (None, "headwise", "elementwise").
|
| 1139 |
+
See https://arxiv.org/abs/2505.06708.
|
| 1140 |
+
"""
|
| 1141 |
+
|
| 1142 |
+
def __init__(
|
| 1143 |
+
self,
|
| 1144 |
+
*,
|
| 1145 |
+
img_size: int = 224,
|
| 1146 |
+
patch_size: int = 16,
|
| 1147 |
+
in_chans: int = 3,
|
| 1148 |
+
embed_dim: int = 768,
|
| 1149 |
+
depth: int = 12,
|
| 1150 |
+
num_heads: int = 12,
|
| 1151 |
+
ffn_ratio: float = 4.0,
|
| 1152 |
+
qkv_bias: bool = True,
|
| 1153 |
+
drop_path_rate: float = 0.0,
|
| 1154 |
+
layerscale_init: float | None = None,
|
| 1155 |
+
norm_layer: str = "layernorm",
|
| 1156 |
+
ffn_layer: str = "mlp",
|
| 1157 |
+
ffn_bias: bool = True,
|
| 1158 |
+
proj_bias: bool = True,
|
| 1159 |
+
n_storage_tokens: int = 0,
|
| 1160 |
+
mask_k_bias: bool = False,
|
| 1161 |
+
untie_cls_and_patch_norms: bool = False,
|
| 1162 |
+
untie_global_and_local_cls_norm: bool = False,
|
| 1163 |
+
device: Any | None = None,
|
| 1164 |
+
num_frames: int = 1,
|
| 1165 |
+
tubelet_size: int = 1,
|
| 1166 |
+
pos_embed_rope_base: float = 100.0,
|
| 1167 |
+
pos_embed_rope_rescale_coords: float | None = None,
|
| 1168 |
+
pos_embed_rope_shift_coords: float | None = None,
|
| 1169 |
+
pos_embed_rope_jitter_coords: float | None = None,
|
| 1170 |
+
pos_embed_rope_fhw_dim: tuple[int, int, int] | None = None,
|
| 1171 |
+
gated_attention: str | None = None,
|
| 1172 |
+
qk_norm: bool = False,
|
| 1173 |
+
**ignored_kwargs,
|
| 1174 |
+
) -> None:
|
| 1175 |
+
super().__init__()
|
| 1176 |
+
if len(ignored_kwargs) > 0:
|
| 1177 |
+
logger.warning(f"Ignored kwargs: {ignored_kwargs}")
|
| 1178 |
+
|
| 1179 |
+
norm_layer_cls = norm_layer_dict[norm_layer]
|
| 1180 |
+
|
| 1181 |
+
self.num_features = self.embed_dim = embed_dim
|
| 1182 |
+
self.n_blocks = depth
|
| 1183 |
+
self.num_heads = num_heads
|
| 1184 |
+
self.patch_size = patch_size
|
| 1185 |
+
|
| 1186 |
+
self.patch_embed = PatchEmbed(
|
| 1187 |
+
img_size=img_size,
|
| 1188 |
+
patch_size=patch_size,
|
| 1189 |
+
in_chans=in_chans,
|
| 1190 |
+
embed_dim=embed_dim,
|
| 1191 |
+
tubelet_size=tubelet_size,
|
| 1192 |
+
flatten_embedding=True,
|
| 1193 |
+
)
|
| 1194 |
+
|
| 1195 |
+
self.cls_token = nn.Parameter(torch.empty(1, 1, embed_dim, device=device))
|
| 1196 |
+
self.n_storage_tokens = n_storage_tokens
|
| 1197 |
+
if self.n_storage_tokens > 0:
|
| 1198 |
+
self.storage_tokens = nn.Parameter(torch.empty(1, n_storage_tokens, embed_dim, device=device))
|
| 1199 |
+
|
| 1200 |
+
# Convert 0.0 to None for backward compat (0.0 means disabled)
|
| 1201 |
+
_rescale = pos_embed_rope_rescale_coords if pos_embed_rope_rescale_coords else None
|
| 1202 |
+
_shift = pos_embed_rope_shift_coords if pos_embed_rope_shift_coords else None
|
| 1203 |
+
_jitter = pos_embed_rope_jitter_coords if pos_embed_rope_jitter_coords else None
|
| 1204 |
+
self.rope_embed = RopePositionEmbedding3D(
|
| 1205 |
+
embed_dim=embed_dim,
|
| 1206 |
+
num_heads=num_heads,
|
| 1207 |
+
fhw_dim=pos_embed_rope_fhw_dim,
|
| 1208 |
+
base=pos_embed_rope_base,
|
| 1209 |
+
rescale_coords=_rescale,
|
| 1210 |
+
shift_coords=_shift,
|
| 1211 |
+
jitter_coords=_jitter,
|
| 1212 |
+
)
|
| 1213 |
+
|
| 1214 |
+
logger.info(f"using {ffn_layer} layer as FFN")
|
| 1215 |
+
ffn_layer_cls = ffn_layer_dict[ffn_layer]
|
| 1216 |
+
ffn_ratio_sequence = [ffn_ratio] * depth
|
| 1217 |
+
|
| 1218 |
+
blocks_list = [
|
| 1219 |
+
SelfAttentionBlock(
|
| 1220 |
+
dim=embed_dim,
|
| 1221 |
+
num_heads=num_heads,
|
| 1222 |
+
ffn_ratio=ffn_ratio_sequence[i],
|
| 1223 |
+
qkv_bias=qkv_bias,
|
| 1224 |
+
proj_bias=proj_bias,
|
| 1225 |
+
ffn_bias=ffn_bias,
|
| 1226 |
+
drop_path=drop_path_rate,
|
| 1227 |
+
norm_layer=norm_layer_cls,
|
| 1228 |
+
act_layer=nn.GELU,
|
| 1229 |
+
ffn_layer=ffn_layer_cls,
|
| 1230 |
+
init_values=layerscale_init,
|
| 1231 |
+
mask_k_bias=mask_k_bias,
|
| 1232 |
+
device=device,
|
| 1233 |
+
gated_attention=gated_attention,
|
| 1234 |
+
qk_norm=qk_norm,
|
| 1235 |
+
)
|
| 1236 |
+
for i in range(depth)
|
| 1237 |
+
]
|
| 1238 |
+
|
| 1239 |
+
self.chunked_blocks = False
|
| 1240 |
+
self.blocks = nn.ModuleList(blocks_list)
|
| 1241 |
+
|
| 1242 |
+
self.norm = norm_layer_cls(embed_dim)
|
| 1243 |
+
|
| 1244 |
+
self.untie_cls_and_patch_norms = untie_cls_and_patch_norms
|
| 1245 |
+
if untie_cls_and_patch_norms:
|
| 1246 |
+
self.cls_norm = norm_layer_cls(embed_dim)
|
| 1247 |
+
else:
|
| 1248 |
+
self.cls_norm = None
|
| 1249 |
+
|
| 1250 |
+
self.untie_global_and_local_cls_norm = untie_global_and_local_cls_norm
|
| 1251 |
+
if untie_global_and_local_cls_norm:
|
| 1252 |
+
self.local_cls_norm = norm_layer_cls(embed_dim)
|
| 1253 |
+
else:
|
| 1254 |
+
self.local_cls_norm = None
|
| 1255 |
+
self.head = nn.Identity()
|
| 1256 |
+
self.mask_token = nn.Parameter(torch.empty(1, embed_dim, device=device))
|
| 1257 |
+
|
| 1258 |
+
def init_weights(self) -> None:
|
| 1259 |
+
"""Initialize all model weights."""
|
| 1260 |
+
self.rope_embed._init_weights()
|
| 1261 |
+
nn.init.normal_(self.cls_token, std=0.02)
|
| 1262 |
+
if self.n_storage_tokens > 0:
|
| 1263 |
+
nn.init.normal_(self.storage_tokens, std=0.02)
|
| 1264 |
+
nn.init.zeros_(self.mask_token)
|
| 1265 |
+
named_apply(init_weights_vit, self)
|
| 1266 |
+
|
| 1267 |
+
def prepare_tokens_with_masks(
|
| 1268 |
+
self,
|
| 1269 |
+
x: Tensor,
|
| 1270 |
+
masks: Tensor | None = None,
|
| 1271 |
+
) -> tuple[Tensor, tuple[int, int, int]]:
|
| 1272 |
+
"""Tokenize input and assemble token sequence with CLS + register + patches.
|
| 1273 |
+
|
| 1274 |
+
Args:
|
| 1275 |
+
x: Input tensor. Image: (B, C, H, W) or Video: (B, T, C, H, W).
|
| 1276 |
+
masks: Boolean mask of shape (B, N_spatial) indicating which patches to mask.
|
| 1277 |
+
|
| 1278 |
+
Returns:
|
| 1279 |
+
Tuple of:
|
| 1280 |
+
- Token sequence: (B, 1 + n_storage + N_total, embed_dim)
|
| 1281 |
+
- Grid dimensions: (T_grid, H_grid, W_grid)
|
| 1282 |
+
"""
|
| 1283 |
+
if x.ndim == 5:
|
| 1284 |
+
B, T, C, H, W = x.shape
|
| 1285 |
+
# Video: Conv3d kernel=stride=tubelet downsamples raw T to T // tubelet.
|
| 1286 |
+
T_grid = T // self.patch_embed.tubelet_size
|
| 1287 |
+
else:
|
| 1288 |
+
B, C, H, W = x.shape
|
| 1289 |
+
# Image (4D): PatchEmbed expands raw T=1 to tubelet then Conv3d(stride=tubelet)
|
| 1290 |
+
# produces a single temporal token (output T_out = (tubelet - tubelet)/tubelet + 1 = 1).
|
| 1291 |
+
# vjepa2 vision_transformer.py:171-173 passes T=1 (no division) for the same reason.
|
| 1292 |
+
T_grid = 1
|
| 1293 |
+
|
| 1294 |
+
x = self.patch_embed(x) # (B, N_total, D)
|
| 1295 |
+
|
| 1296 |
+
# Grid dimensions for RoPE computation
|
| 1297 |
+
H_grid = H // self.patch_embed.patch_size[0]
|
| 1298 |
+
W_grid = W // self.patch_embed.patch_size[1]
|
| 1299 |
+
|
| 1300 |
+
if masks is not None:
|
| 1301 |
+
# Expand spatial mask to spatio-temporal if needed (tube masking)
|
| 1302 |
+
if masks.shape[1] != x.shape[1]:
|
| 1303 |
+
ratio = x.shape[1] // masks.shape[1]
|
| 1304 |
+
masks = masks.unsqueeze(1).repeat(1, ratio, 1).flatten(1)
|
| 1305 |
+
|
| 1306 |
+
# Replace masked positions with mask_token
|
| 1307 |
+
x = torch.where(masks.unsqueeze(-1), self.mask_token.to(x.dtype).unsqueeze(0), x)
|
| 1308 |
+
cls_token = self.cls_token
|
| 1309 |
+
else:
|
| 1310 |
+
# Include mask_token in computation graph even when not masking
|
| 1311 |
+
cls_token = self.cls_token + 0 * self.mask_token
|
| 1312 |
+
|
| 1313 |
+
if self.n_storage_tokens > 0:
|
| 1314 |
+
storage_tokens = self.storage_tokens
|
| 1315 |
+
else:
|
| 1316 |
+
storage_tokens = torch.empty(
|
| 1317 |
+
1, 0, cls_token.shape[-1],
|
| 1318 |
+
dtype=cls_token.dtype, device=cls_token.device,
|
| 1319 |
+
)
|
| 1320 |
+
|
| 1321 |
+
x = torch.cat(
|
| 1322 |
+
[
|
| 1323 |
+
cls_token.expand(B, -1, -1),
|
| 1324 |
+
storage_tokens.expand(B, -1, -1),
|
| 1325 |
+
x,
|
| 1326 |
+
],
|
| 1327 |
+
dim=1,
|
| 1328 |
+
)
|
| 1329 |
+
|
| 1330 |
+
return x, (T_grid, H_grid, W_grid)
|
| 1331 |
+
|
| 1332 |
+
def forward_features_list(
|
| 1333 |
+
self,
|
| 1334 |
+
x_list: list[Tensor],
|
| 1335 |
+
masks_list: list[Tensor | None],
|
| 1336 |
+
) -> list[dict[str, Tensor]]:
|
| 1337 |
+
"""Forward pass for a list of inputs (multi-crop).
|
| 1338 |
+
|
| 1339 |
+
Args:
|
| 1340 |
+
x_list: List of input tensors (global crops, local crops).
|
| 1341 |
+
masks_list: List of corresponding masks (None for unmasked).
|
| 1342 |
+
|
| 1343 |
+
Returns:
|
| 1344 |
+
List of output dictionaries, one per input, containing:
|
| 1345 |
+
- x_norm_clstoken: Normalized CLS token (B, D)
|
| 1346 |
+
- x_storage_tokens: Normalized register tokens (B, n_storage, D)
|
| 1347 |
+
- x_norm_patchtokens: Normalized patch tokens (B, N, D)
|
| 1348 |
+
- x_prenorm: Pre-normalization features (B, 1+n_storage+N, D)
|
| 1349 |
+
- masks: Original masks
|
| 1350 |
+
"""
|
| 1351 |
+
x = []
|
| 1352 |
+
rope_params = []
|
| 1353 |
+
for t_x, t_masks in zip(x_list, masks_list):
|
| 1354 |
+
t2_x, grid_tuple = self.prepare_tokens_with_masks(t_x, t_masks)
|
| 1355 |
+
x.append(t2_x)
|
| 1356 |
+
rope_params.append(grid_tuple)
|
| 1357 |
+
|
| 1358 |
+
# Pre-compute RoPE sin/cos once — identical across all blocks.
|
| 1359 |
+
# Hoisting this out of the loop avoids breaking FSDP2's forward prefetch
|
| 1360 |
+
# chain (rope_embed is part of the outer FSDP unit, calling it between
|
| 1361 |
+
# block forwards disrupts the prefetch scheduling).
|
| 1362 |
+
if self.rope_embed is not None:
|
| 1363 |
+
rope_sincos = [self.rope_embed(T=t, H=h, W=w) for t, h, w in rope_params]
|
| 1364 |
+
else:
|
| 1365 |
+
rope_sincos = [None for _ in rope_params]
|
| 1366 |
+
|
| 1367 |
+
for _, blk in enumerate(self.blocks):
|
| 1368 |
+
x = blk(x, rope_sincos)
|
| 1369 |
+
|
| 1370 |
+
all_x = x
|
| 1371 |
+
output = []
|
| 1372 |
+
for idx, (x, masks) in enumerate(zip(all_x, masks_list)):
|
| 1373 |
+
if self.untie_cls_and_patch_norms or self.untie_global_and_local_cls_norm:
|
| 1374 |
+
if self.untie_global_and_local_cls_norm and self.training and idx == 1:
|
| 1375 |
+
x_norm_cls_reg = self.local_cls_norm(x[:, : self.n_storage_tokens + 1])
|
| 1376 |
+
elif self.untie_cls_and_patch_norms:
|
| 1377 |
+
x_norm_cls_reg = self.cls_norm(x[:, : self.n_storage_tokens + 1])
|
| 1378 |
+
else:
|
| 1379 |
+
x_norm_cls_reg = self.norm(x[:, : self.n_storage_tokens + 1])
|
| 1380 |
+
x_norm_patch = self.norm(x[:, self.n_storage_tokens + 1 :])
|
| 1381 |
+
else:
|
| 1382 |
+
x_norm = self.norm(x)
|
| 1383 |
+
x_norm_cls_reg = x_norm[:, : self.n_storage_tokens + 1]
|
| 1384 |
+
x_norm_patch = x_norm[:, self.n_storage_tokens + 1 :]
|
| 1385 |
+
output.append(
|
| 1386 |
+
{
|
| 1387 |
+
"x_norm_clstoken": x_norm_cls_reg[:, 0],
|
| 1388 |
+
"x_storage_tokens": x_norm_cls_reg[:, 1:],
|
| 1389 |
+
"x_norm_patchtokens": x_norm_patch,
|
| 1390 |
+
"x_prenorm": x,
|
| 1391 |
+
"masks": masks,
|
| 1392 |
+
}
|
| 1393 |
+
)
|
| 1394 |
+
return output
|
| 1395 |
+
|
| 1396 |
+
def forward_features(
|
| 1397 |
+
self,
|
| 1398 |
+
x: Tensor | list[Tensor],
|
| 1399 |
+
masks: Tensor | list[Tensor | None] | None = None,
|
| 1400 |
+
) -> dict[str, Tensor] | list[dict[str, Tensor]]:
|
| 1401 |
+
"""Forward pass for single or multiple inputs.
|
| 1402 |
+
|
| 1403 |
+
Args:
|
| 1404 |
+
x: Single tensor or list of tensors.
|
| 1405 |
+
masks: Single mask or list of masks.
|
| 1406 |
+
|
| 1407 |
+
Returns:
|
| 1408 |
+
Output dict (single input) or list of output dicts (multiple inputs).
|
| 1409 |
+
"""
|
| 1410 |
+
if isinstance(x, torch.Tensor):
|
| 1411 |
+
return self.forward_features_list([x], [masks])[0]
|
| 1412 |
+
else:
|
| 1413 |
+
return self.forward_features_list(x, masks)
|
| 1414 |
+
|
| 1415 |
+
def _get_intermediate_layers_not_chunked(
|
| 1416 |
+
self,
|
| 1417 |
+
x: Tensor,
|
| 1418 |
+
n: int | list[int] = 1,
|
| 1419 |
+
) -> list[Tensor]:
|
| 1420 |
+
"""Run forward pass and collect intermediate block outputs.
|
| 1421 |
+
|
| 1422 |
+
Args:
|
| 1423 |
+
x: Input tensor (B, C, H, W) or (B, T, C, H, W).
|
| 1424 |
+
n: If int, return last n layers. If list, return specific layer indices.
|
| 1425 |
+
|
| 1426 |
+
Returns:
|
| 1427 |
+
List of intermediate outputs, each (B, 1+n_storage+N, D).
|
| 1428 |
+
"""
|
| 1429 |
+
x, grid_tuple = self.prepare_tokens_with_masks(x, masks=None)
|
| 1430 |
+
T, H, W = grid_tuple
|
| 1431 |
+
|
| 1432 |
+
output, total_block_len = [], len(self.blocks)
|
| 1433 |
+
blocks_to_take = range(total_block_len - n, total_block_len) if isinstance(n, int) else n
|
| 1434 |
+
|
| 1435 |
+
if self.rope_embed is not None:
|
| 1436 |
+
rope_sincos = self.rope_embed(T=T, H=H, W=W)
|
| 1437 |
+
else:
|
| 1438 |
+
rope_sincos = None
|
| 1439 |
+
|
| 1440 |
+
for i, blk in enumerate(self.blocks):
|
| 1441 |
+
x = blk([x], [rope_sincos])[0]
|
| 1442 |
+
if i in blocks_to_take:
|
| 1443 |
+
output.append(x)
|
| 1444 |
+
|
| 1445 |
+
assert len(output) == len(blocks_to_take), (
|
| 1446 |
+
f"only {len(output)} / {len(blocks_to_take)} blocks found"
|
| 1447 |
+
)
|
| 1448 |
+
return output
|
| 1449 |
+
|
| 1450 |
+
def get_intermediate_layers(
|
| 1451 |
+
self,
|
| 1452 |
+
x: Tensor,
|
| 1453 |
+
n: int | list[int] = 1,
|
| 1454 |
+
reshape: bool = False,
|
| 1455 |
+
return_class_token: bool = False,
|
| 1456 |
+
norm: bool = True,
|
| 1457 |
+
) -> tuple[Tensor, ...]:
|
| 1458 |
+
"""Extract intermediate layer outputs for downstream evaluation.
|
| 1459 |
+
|
| 1460 |
+
This method is critical for dense prediction tasks (segmentation, depth)
|
| 1461 |
+
that need multi-scale features from different transformer blocks.
|
| 1462 |
+
|
| 1463 |
+
Args:
|
| 1464 |
+
x: Input image (B, C, H, W) or video (B, T, C, H, W).
|
| 1465 |
+
n: If int, return outputs from last n layers.
|
| 1466 |
+
If list[int], return outputs from specific layer indices.
|
| 1467 |
+
reshape: If True, reshape patch tokens to spatial form (B, D, H_grid, W_grid).
|
| 1468 |
+
return_class_token: If True, return (patch_tokens, cls_token) tuples.
|
| 1469 |
+
norm: If True, apply final LayerNorm to outputs.
|
| 1470 |
+
|
| 1471 |
+
Returns:
|
| 1472 |
+
If return_class_token is False:
|
| 1473 |
+
Tuple of patch token tensors, one per requested layer.
|
| 1474 |
+
Each tensor is (B, N, D) or (B, D, H_grid, W_grid) if reshape=True.
|
| 1475 |
+
If return_class_token is True:
|
| 1476 |
+
Tuple of (patch_tokens, cls_token) pairs.
|
| 1477 |
+
"""
|
| 1478 |
+
# Determine spatial dims for reshape
|
| 1479 |
+
if x.ndim == 5:
|
| 1480 |
+
B, T_in, C, H, W = x.shape
|
| 1481 |
+
else:
|
| 1482 |
+
B, C, H, W = x.shape
|
| 1483 |
+
T_in = 1
|
| 1484 |
+
|
| 1485 |
+
outputs = self._get_intermediate_layers_not_chunked(x, n)
|
| 1486 |
+
|
| 1487 |
+
if norm:
|
| 1488 |
+
outputs_normed = []
|
| 1489 |
+
for out in outputs:
|
| 1490 |
+
if self.untie_cls_and_patch_norms:
|
| 1491 |
+
x_norm_cls_reg = self.cls_norm(out[:, : self.n_storage_tokens + 1])
|
| 1492 |
+
x_norm_patch = self.norm(out[:, self.n_storage_tokens + 1 :])
|
| 1493 |
+
outputs_normed.append(torch.cat((x_norm_cls_reg, x_norm_patch), dim=1))
|
| 1494 |
+
else:
|
| 1495 |
+
outputs_normed.append(self.norm(out))
|
| 1496 |
+
outputs = outputs_normed
|
| 1497 |
+
|
| 1498 |
+
class_tokens = [out[:, 0] for out in outputs]
|
| 1499 |
+
outputs = [out[:, self.n_storage_tokens + 1 :] for out in outputs]
|
| 1500 |
+
|
| 1501 |
+
if reshape:
|
| 1502 |
+
# Image (T_in=1): PatchEmbed expands to tubelet then Conv3d(stride=tubelet) → T_out=1.
|
| 1503 |
+
# Video: Conv3d downsamples T_in → T_in // tubelet. Matches prepare_tokens_with_masks
|
| 1504 |
+
# and vjepa2 vision_transformer.py:171-177.
|
| 1505 |
+
if x.ndim == 5:
|
| 1506 |
+
T_grid = T_in // self.patch_embed.tubelet_size
|
| 1507 |
+
else:
|
| 1508 |
+
T_grid = 1
|
| 1509 |
+
H_grid = H // self.patch_size
|
| 1510 |
+
W_grid = W // self.patch_size
|
| 1511 |
+
if T_grid > 1:
|
| 1512 |
+
# Video: reshape to (B, D, T_grid, H_grid, W_grid)
|
| 1513 |
+
outputs = [
|
| 1514 |
+
out.reshape(B, T_grid, H_grid, W_grid, -1).permute(0, 4, 1, 2, 3).contiguous()
|
| 1515 |
+
for out in outputs
|
| 1516 |
+
]
|
| 1517 |
+
else:
|
| 1518 |
+
# Image: reshape to (B, D, H_grid, W_grid)
|
| 1519 |
+
outputs = [
|
| 1520 |
+
out.reshape(B, H_grid, W_grid, -1).permute(0, 3, 1, 2).contiguous()
|
| 1521 |
+
for out in outputs
|
| 1522 |
+
]
|
| 1523 |
+
|
| 1524 |
+
if return_class_token:
|
| 1525 |
+
return tuple(zip(outputs, class_tokens))
|
| 1526 |
+
return tuple(outputs)
|
| 1527 |
+
|
| 1528 |
+
def forward(
|
| 1529 |
+
self,
|
| 1530 |
+
*args,
|
| 1531 |
+
is_training: bool = False,
|
| 1532 |
+
**kwargs,
|
| 1533 |
+
) -> dict[str, Tensor] | list[dict[str, Tensor]] | Tensor:
|
| 1534 |
+
"""High-level forward: training returns feature dict, inference returns CLS logits.
|
| 1535 |
+
|
| 1536 |
+
Args:
|
| 1537 |
+
is_training: If True, return full feature dictionary.
|
| 1538 |
+
|
| 1539 |
+
Returns:
|
| 1540 |
+
Feature dict(s) if training, CLS token logits if inference.
|
| 1541 |
+
"""
|
| 1542 |
+
ret = self.forward_features(*args, **kwargs)
|
| 1543 |
+
if is_training:
|
| 1544 |
+
return ret
|
| 1545 |
+
else:
|
| 1546 |
+
return self.head(ret["x_norm_clstoken"])
|
| 1547 |
+
|
| 1548 |
+
|
| 1549 |
+
# ============================================================================
|
| 1550 |
+
# HuggingFace transformers wrapper (inference-only)
|
| 1551 |
+
# ============================================================================
|
| 1552 |
+
class MotifVisionConfig(PretrainedConfig):
|
| 1553 |
+
"""Config for the MOTIF Vision Encoder backbone (image + video)."""
|
| 1554 |
+
|
| 1555 |
+
model_type = "motif_vision"
|
| 1556 |
+
|
| 1557 |
+
def __init__(
|
| 1558 |
+
self,
|
| 1559 |
+
img_size: int = 224,
|
| 1560 |
+
patch_size: int = 16,
|
| 1561 |
+
in_chans: int = 3,
|
| 1562 |
+
embed_dim: int = 4096,
|
| 1563 |
+
depth: int = 40,
|
| 1564 |
+
num_heads: int = 32,
|
| 1565 |
+
ffn_ratio: float = 3.0,
|
| 1566 |
+
qkv_bias: bool = False,
|
| 1567 |
+
drop_path_rate: float = 0.0,
|
| 1568 |
+
layerscale_init: float | None = 1.0e-5,
|
| 1569 |
+
norm_layer: str = "layernormbf16",
|
| 1570 |
+
ffn_layer: str = "swiglu64",
|
| 1571 |
+
ffn_bias: bool = True,
|
| 1572 |
+
proj_bias: bool = True,
|
| 1573 |
+
n_storage_tokens: int = 4,
|
| 1574 |
+
mask_k_bias: bool = True,
|
| 1575 |
+
untie_cls_and_patch_norms: bool = False,
|
| 1576 |
+
untie_global_and_local_cls_norm: bool = True,
|
| 1577 |
+
num_frames: int = 1,
|
| 1578 |
+
tubelet_size: int = 2,
|
| 1579 |
+
pos_embed_rope_base: float = 100.0,
|
| 1580 |
+
pos_embed_rope_rescale_coords: float | None = 2.0,
|
| 1581 |
+
gated_attention: str | None = "elementwise",
|
| 1582 |
+
qk_norm: bool = True,
|
| 1583 |
+
**kwargs,
|
| 1584 |
+
):
|
| 1585 |
+
self.img_size = img_size
|
| 1586 |
+
self.patch_size = patch_size
|
| 1587 |
+
self.in_chans = in_chans
|
| 1588 |
+
self.embed_dim = embed_dim
|
| 1589 |
+
self.depth = depth
|
| 1590 |
+
self.num_heads = num_heads
|
| 1591 |
+
self.ffn_ratio = ffn_ratio
|
| 1592 |
+
self.qkv_bias = qkv_bias
|
| 1593 |
+
self.drop_path_rate = drop_path_rate
|
| 1594 |
+
self.layerscale_init = layerscale_init
|
| 1595 |
+
self.norm_layer = norm_layer
|
| 1596 |
+
self.ffn_layer = ffn_layer
|
| 1597 |
+
self.ffn_bias = ffn_bias
|
| 1598 |
+
self.proj_bias = proj_bias
|
| 1599 |
+
self.n_storage_tokens = n_storage_tokens
|
| 1600 |
+
self.mask_k_bias = mask_k_bias
|
| 1601 |
+
self.untie_cls_and_patch_norms = untie_cls_and_patch_norms
|
| 1602 |
+
self.untie_global_and_local_cls_norm = untie_global_and_local_cls_norm
|
| 1603 |
+
self.num_frames = num_frames
|
| 1604 |
+
self.tubelet_size = tubelet_size
|
| 1605 |
+
self.pos_embed_rope_base = pos_embed_rope_base
|
| 1606 |
+
self.pos_embed_rope_rescale_coords = pos_embed_rope_rescale_coords
|
| 1607 |
+
self.gated_attention = gated_attention
|
| 1608 |
+
self.qk_norm = qk_norm
|
| 1609 |
+
super().__init__(**kwargs)
|
| 1610 |
+
|
| 1611 |
+
|
| 1612 |
+
class MotifVisionModel(PreTrainedModel):
|
| 1613 |
+
"""MOTIF Vision Encoder for HF `AutoModel` (inference). Returns dense + CLS features."""
|
| 1614 |
+
|
| 1615 |
+
config_class = MotifVisionConfig
|
| 1616 |
+
base_model_prefix = "motif"
|
| 1617 |
+
main_input_name = "pixel_values"
|
| 1618 |
+
_no_split_modules = ["SelfAttentionBlock"]
|
| 1619 |
+
supports_gradient_checkpointing = False
|
| 1620 |
+
|
| 1621 |
+
def __init__(self, config: MotifVisionConfig):
|
| 1622 |
+
super().__init__(config)
|
| 1623 |
+
self.backbone = MotifVisionTransformer(
|
| 1624 |
+
img_size=config.img_size,
|
| 1625 |
+
patch_size=config.patch_size,
|
| 1626 |
+
in_chans=config.in_chans,
|
| 1627 |
+
embed_dim=config.embed_dim,
|
| 1628 |
+
depth=config.depth,
|
| 1629 |
+
num_heads=config.num_heads,
|
| 1630 |
+
ffn_ratio=config.ffn_ratio,
|
| 1631 |
+
qkv_bias=config.qkv_bias,
|
| 1632 |
+
drop_path_rate=config.drop_path_rate,
|
| 1633 |
+
layerscale_init=config.layerscale_init,
|
| 1634 |
+
norm_layer=config.norm_layer,
|
| 1635 |
+
ffn_layer=config.ffn_layer,
|
| 1636 |
+
ffn_bias=config.ffn_bias,
|
| 1637 |
+
proj_bias=config.proj_bias,
|
| 1638 |
+
n_storage_tokens=config.n_storage_tokens,
|
| 1639 |
+
mask_k_bias=config.mask_k_bias,
|
| 1640 |
+
untie_cls_and_patch_norms=config.untie_cls_and_patch_norms,
|
| 1641 |
+
untie_global_and_local_cls_norm=config.untie_global_and_local_cls_norm,
|
| 1642 |
+
num_frames=config.num_frames,
|
| 1643 |
+
tubelet_size=config.tubelet_size,
|
| 1644 |
+
pos_embed_rope_base=config.pos_embed_rope_base,
|
| 1645 |
+
pos_embed_rope_rescale_coords=config.pos_embed_rope_rescale_coords,
|
| 1646 |
+
gated_attention=config.gated_attention,
|
| 1647 |
+
qk_norm=config.qk_norm,
|
| 1648 |
+
)
|
| 1649 |
+
self.post_init()
|
| 1650 |
+
|
| 1651 |
+
@torch.no_grad()
|
| 1652 |
+
def forward(self, pixel_values: Tensor, return_dict: bool = True, **kwargs):
|
| 1653 |
+
"""pixel_values: image (B,3,H,W) or video (B,T,3,H,W). H,W multiples of patch_size."""
|
| 1654 |
+
out = self.backbone.forward_features(pixel_values)
|
| 1655 |
+
cls = out["x_norm_clstoken"]
|
| 1656 |
+
reg = out["x_storage_tokens"]
|
| 1657 |
+
patch = out["x_norm_patchtokens"]
|
| 1658 |
+
last_hidden = torch.cat([cls.unsqueeze(1), reg, patch], dim=1)
|
| 1659 |
+
if not return_dict:
|
| 1660 |
+
return (last_hidden, cls)
|
| 1661 |
+
return BaseModelOutputWithPooling(last_hidden_state=last_hidden, pooler_output=cls)
|
| 1662 |
+
|
| 1663 |
+
|
| 1664 |
+
AutoConfig_registered = False
|
| 1665 |
+
try:
|
| 1666 |
+
from transformers import AutoConfig, AutoModel
|
| 1667 |
+
AutoConfig.register("motif_vision", MotifVisionConfig)
|
| 1668 |
+
AutoModel.register(MotifVisionConfig, MotifVisionModel)
|
| 1669 |
+
AutoConfig_registered = True
|
| 1670 |
+
except Exception:
|
| 1671 |
+
pass
|
preprocessor_config.json
ADDED
|
@@ -0,0 +1,27 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"image_processor_type": "BitImageProcessor",
|
| 3 |
+
"do_resize": true,
|
| 4 |
+
"size": {
|
| 5 |
+
"shortest_edge": 512
|
| 6 |
+
},
|
| 7 |
+
"do_center_crop": true,
|
| 8 |
+
"crop_size": {
|
| 9 |
+
"height": 512,
|
| 10 |
+
"width": 512
|
| 11 |
+
},
|
| 12 |
+
"do_rescale": true,
|
| 13 |
+
"rescale_factor": 0.00392156862745098,
|
| 14 |
+
"do_normalize": true,
|
| 15 |
+
"image_mean": [
|
| 16 |
+
0.485,
|
| 17 |
+
0.456,
|
| 18 |
+
0.406
|
| 19 |
+
],
|
| 20 |
+
"image_std": [
|
| 21 |
+
0.229,
|
| 22 |
+
0.224,
|
| 23 |
+
0.225
|
| 24 |
+
],
|
| 25 |
+
"resample": 3,
|
| 26 |
+
"do_convert_rgb": true
|
| 27 |
+
}
|