Weidows commited on
Commit
6df33b0
·
verified ·
1 Parent(s): 247aebc

Upload folder using huggingface_hub

Browse files
.gitattributes CHANGED
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
 
 
33
  *.zip filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
+ tokenizer.json filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,435 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Tencent is pleased to support the open source community by making WeMM-Embedding-2B available.
2
+
3
+ Copyright (C) 2026 Tencent. All rights reserved.
4
+
5
+ The open-source software and/or models included in this distribution may have been modified by Tencent (“Tencent Modifications”). All Tencent Modifications are Copyright (C) Tencent.
6
+
7
+ WeMM-Embedding-2B is licensed under Apache-2.0, except for the third-party components listed below, which remain licensed under their respective original terms. WeMM-Embedding-2B does not impose any additional restrictions beyond those specified in the original licenses of these third-party components. Users are required to comply with all applicable terms and conditions of the original licenses and to ensure that the use of these third-party components conforms to all relevant laws and regulations.
8
+
9
+ For the avoidance of doubt, WeMM-Embedding-2B refers to code, parameters, and weights made publicly available by Tencent in accordance with Apache-2.0.
10
+
11
+ Terms of Apache-2.0:
12
+ --------------------------------------------------------------------
13
+ Apache License
14
+ Version 2.0, January 2004
15
+ http://www.apache.org/licenses/
16
+
17
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
18
+
19
+ 1. Definitions.
20
+
21
+ "License" shall mean the terms and conditions for use, reproduction,
22
+ and distribution as defined by Sections 1 through 9 of this document.
23
+
24
+ "Licensor" shall mean the copyright owner or entity authorized by
25
+ the copyright owner that is granting the License.
26
+
27
+ "Legal Entity" shall mean the union of the acting entity and all
28
+ other entities that control, are controlled by, or are under common
29
+ control with that entity. For the purposes of this definition,
30
+ "control" means (i) the power, direct or indirect, to cause the
31
+ direction or management of such entity, whether by contract or
32
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
33
+ outstanding shares, or (iii) beneficial ownership of such entity.
34
+
35
+ "You" (or "Your") shall mean an individual or Legal Entity
36
+ exercising permissions granted by this License.
37
+
38
+ "Source" form shall mean the preferred form for making modifications,
39
+ including but not limited to software source code, documentation
40
+ source, and configuration files.
41
+
42
+ "Object" form shall mean any form resulting from mechanical
43
+ transformation or translation of a Source form, including but
44
+ not limited to compiled object code, generated documentation,
45
+ and conversions to other media types.
46
+
47
+ "Work" shall mean the work of authorship, whether in Source or
48
+ Object form, made available under the License, as indicated by a
49
+ copyright notice that is included in or attached to the work
50
+ (an example is provided in the Appendix below).
51
+
52
+ "Derivative Works" shall mean any work, whether in Source or Object
53
+ form, that is based on (or derived from) the Work and for which the
54
+ editorial revisions, annotations, elaborations, or other modifications
55
+ represent, as a whole, an original work of authorship. For the purposes
56
+ of this License, Derivative Works shall not include works that remain
57
+ separable from, or merely link (or bind by name) to the interfaces of,
58
+ the Work and Derivative Works thereof.
59
+
60
+ "Contribution" shall mean any work of authorship, including
61
+ the original version of the Work and any modifications or additions
62
+ to that Work or Derivative Works thereof, that is intentionally
63
+ submitted to Licensor for inclusion in the Work by the copyright owner
64
+ or by an individual or Legal Entity authorized to submit on behalf of
65
+ the copyright owner. For the purposes of this definition, "submitted"
66
+ means any form of electronic, verbal, or written communication sent
67
+ to the Licensor or its representatives, including but not limited to
68
+ communication on electronic mailing lists, source code control systems,
69
+ and issue tracking systems that are managed by, or on behalf of, the
70
+ Licensor for the purpose of discussing and improving the Work, but
71
+ excluding communication that is conspicuously marked or otherwise
72
+ designated in writing by the copyright owner as "Not a Contribution."
73
+
74
+ "Contributor" shall mean Licensor and any individual or Legal Entity
75
+ on behalf of whom a Contribution has been received by Licensor and
76
+ subsequently incorporated within the Work.
77
+
78
+ 2. Grant of Copyright License. Subject to the terms and conditions of
79
+ this License, each Contributor hereby grants to You a perpetual,
80
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
81
+ copyright license to reproduce, prepare Derivative Works of,
82
+ publicly display, publicly perform, sublicense, and distribute the
83
+ Work and such Derivative Works in Source or Object form.
84
+
85
+ 3. Grant of Patent License. Subject to the terms and conditions of
86
+ this License, each Contributor hereby grants to You a perpetual,
87
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
88
+ (except as stated in this section) patent license to make, have made,
89
+ use, offer to sell, sell, import, and otherwise transfer the Work,
90
+ where such license applies only to those patent claims licensable
91
+ by such Contributor that are necessarily infringed by their
92
+ Contribution(s) alone or by combination of their Contribution(s)
93
+ with the Work to which such Contribution(s) was submitted. If You
94
+ institute patent litigation against any entity (including a
95
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
96
+ or a Contribution incorporated within the Work constitutes direct
97
+ or contributory patent infringement, then any patent licenses
98
+ granted to You under this License for that Work shall terminate
99
+ as of the date such litigation is filed.
100
+
101
+ 4. Redistribution. You may reproduce and distribute copies of the
102
+ Work or Derivative Works thereof in any medium, with or without
103
+ modifications, and in Source or Object form, provided that You
104
+ meet the following conditions:
105
+
106
+ (a) You must give any other recipients of the Work or
107
+ Derivative Works a copy of this License; and
108
+
109
+ (b) You must cause any modified files to carry prominent notices
110
+ stating that You changed the files; and
111
+
112
+ (c) You must retain, in the Source form of any Derivative Works
113
+ that You distribute, all copyright, patent, trademark, and
114
+ attribution notices from the Source form of the Work,
115
+ excluding those notices that do not pertain to any part of
116
+ the Derivative Works; and
117
+
118
+ (d) If the Work includes a "NOTICE" text file as part of its
119
+ distribution, then any Derivative Works that You distribute must
120
+ include a readable copy of the attribution notices contained
121
+ within such NOTICE file, excluding those notices that do not
122
+ pertain to any part of the Derivative Works, in at least one
123
+ of the following places: within a NOTICE text file distributed
124
+ as part of the Derivative Works; within the Source form or
125
+ documentation, if provided along with the Derivative Works; or,
126
+ within a display generated by the Derivative Works, if and
127
+ wherever such third-party notices normally appear. The contents
128
+ of the NOTICE file are for informational purposes only and
129
+ do not modify the License. You may add Your own attribution
130
+ notices within Derivative Works that You distribute, alongside
131
+ or as an addendum to the NOTICE text from the Work, provided
132
+ that such additional attribution notices cannot be construed
133
+ as modifying the License.
134
+
135
+ You may add Your own copyright statement to Your modifications and
136
+ may provide additional or different license terms and conditions
137
+ for use, reproduction, or distribution of Your modifications, or
138
+ for any such Derivative Works as a whole, provided Your use,
139
+ reproduction, and distribution of the Work otherwise complies with
140
+ the conditions stated in this License.
141
+
142
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
143
+ any Contribution intentionally submitted for inclusion in the Work
144
+ by You to the Licensor shall be under the terms and conditions of
145
+ this License, without any additional terms or conditions.
146
+ Notwithstanding the above, nothing herein shall supersede or modify
147
+ the terms of any separate license agreement you may have executed
148
+ with Licensor regarding such Contributions.
149
+
150
+ 6. Trademarks. This License does not grant permission to use the trade
151
+ names, trademarks, service marks, or product names of the Licensor,
152
+ except as required for reasonable and customary use in describing the
153
+ origin of the Work and reproducing the content of the NOTICE file.
154
+
155
+ 7. Disclaimer of Warranty. Unless required by applicable law or
156
+ agreed to in writing, Licensor provides the Work (and each
157
+ Contributor provides its Contributions) on an "AS IS" BASIS,
158
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
159
+ implied, including, without limitation, any warranties or conditions
160
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
161
+ PARTICULAR PURPOSE. You are solely responsible for determining the
162
+ appropriateness of using or redistributing the Work and assume any
163
+ risks associated with Your exercise of permissions under this License.
164
+
165
+ 8. Limitation of Liability. In no event and under no legal theory,
166
+ whether in tort (including negligence), contract, or otherwise,
167
+ unless required by applicable law (such as deliberate and grossly
168
+ negligent acts) or agreed to in writing, shall any Contributor be
169
+ liable to You for damages, including any direct, indirect, special,
170
+ incidental, or consequential damages of any character arising as a
171
+ result of this License or out of the use or inability to use the
172
+ Work (including but not limited to damages for loss of goodwill,
173
+ work stoppage, computer failure or malfunction, or any and all
174
+ other commercial damages or losses), even if such Contributor
175
+ has been advised of the possibility of such damages.
176
+
177
+ 9. Accepting Warranty or Additional Liability. While redistributing
178
+ the Work or Derivative Works thereof, You may choose to offer,
179
+ and charge a fee for, acceptance of support, warranty, indemnity,
180
+ or other liability obligations and/or rights consistent with this
181
+ License. However, in accepting such obligations, You may act only
182
+ on Your own behalf and on Your sole responsibility, not on behalf
183
+ of any other Contributor, and only if You agree to indemnify,
184
+ defend, and hold each Contributor harmless for any liability
185
+ incurred by, or claims asserted against, such Contributor by reason
186
+ of your accepting any such warranty or additional liability.
187
+
188
+ END OF TERMS AND CONDITIONS
189
+
190
+ APPENDIX: How to apply the Apache License to your work.
191
+
192
+ To apply the Apache License to your work, attach the following
193
+ boilerplate notice, with the fields enclosed by brackets "[]"
194
+ replaced with your own identifying information. (Don't include
195
+ the brackets!) The text should be enclosed in the appropriate
196
+ comment syntax for the file format. We also recommend that a
197
+ file or class name and description of purpose be included on the
198
+ same "printed page" as the copyright notice for easier
199
+ identification within third-party archives.
200
+
201
+ Copyright [yyyy] [name of copyright owner]
202
+
203
+ Licensed under the Apache License, Version 2.0 (the "License");
204
+ you may not use this file except in compliance with the License.
205
+ You may obtain a copy of the License at
206
+
207
+ http://www.apache.org/licenses/LICENSE-2.0
208
+
209
+ Unless required by applicable law or agreed to in writing, software
210
+ distributed under the License is distributed on an "AS IS" BASIS,
211
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
212
+ See the License for the specific language governing permissions and
213
+ limitations under the License.
214
+
215
+
216
+ Dependencies and Licenses:
217
+
218
+ This open-source project, WeMM-Embedding-2B, builds upon the following open-source models and/or software components, each of which remains licensed under its original license. Certain models or software may include modifications made by Tencent (“Tencent Modifications”), which are Copyright (C) Tencent.
219
+
220
+ In case you believe there have been errors in the attribution below, you may submit the concerns to us for review and correction.
221
+
222
+
223
+ Open Source Model Licensed under the Apache-2.0:
224
+ --------------------------------------------------------------------
225
+ 1. Qwen3.5-2B
226
+ Copyright 2026 Alibaba Cloud
227
+ Please note this model has been modified by Tencent in this distribution.
228
+
229
+
230
+
231
+ Terms of the Apache-2.0:
232
+ --------------------------------------------------------------------
233
+ Apache License
234
+ Version 2.0, January 2004
235
+ http://www.apache.org/licenses/
236
+
237
+ TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
238
+
239
+ 1. Definitions.
240
+
241
+ "License" shall mean the terms and conditions for use, reproduction,
242
+ and distribution as defined by Sections 1 through 9 of this document.
243
+
244
+ "Licensor" shall mean the copyright owner or entity authorized by
245
+ the copyright owner that is granting the License.
246
+
247
+ "Legal Entity" shall mean the union of the acting entity and all
248
+ other entities that control, are controlled by, or are under common
249
+ control with that entity. For the purposes of this definition,
250
+ "control" means (i) the power, direct or indirect, to cause the
251
+ direction or management of such entity, whether by contract or
252
+ otherwise, or (ii) ownership of fifty percent (50%) or more of the
253
+ outstanding shares, or (iii) beneficial ownership of such entity.
254
+
255
+ "You" (or "Your") shall mean an individual or Legal Entity
256
+ exercising permissions granted by this License.
257
+
258
+ "Source" form shall mean the preferred form for making modifications,
259
+ including but not limited to software source code, documentation
260
+ source, and configuration files.
261
+
262
+ "Object" form shall mean any form resulting from mechanical
263
+ transformation or translation of a Source form, including but
264
+ not limited to compiled object code, generated documentation,
265
+ and conversions to other media types.
266
+
267
+ "Work" shall mean the work of authorship, whether in Source or
268
+ Object form, made available under the License, as indicated by a
269
+ copyright notice that is included in or attached to the work
270
+ (an example is provided in the Appendix below).
271
+
272
+ "Derivative Works" shall mean any work, whether in Source or Object
273
+ form, that is based on (or derived from) the Work and for which the
274
+ editorial revisions, annotations, elaborations, or other modifications
275
+ represent, as a whole, an original work of authorship. For the purposes
276
+ of this License, Derivative Works shall not include works that remain
277
+ separable from, or merely link (or bind by name) to the interfaces of,
278
+ the Work and Derivative Works thereof.
279
+
280
+ "Contribution" shall mean any work of authorship, including
281
+ the original version of the Work and any modifications or additions
282
+ to that Work or Derivative Works thereof, that is intentionally
283
+ submitted to Licensor for inclusion in the Work by the copyright owner
284
+ or by an individual or Legal Entity authorized to submit on behalf of
285
+ the copyright owner. For the purposes of this definition, "submitted"
286
+ means any form of electronic, verbal, or written communication sent
287
+ to the Licensor or its representatives, including but not limited to
288
+ communication on electronic mailing lists, source code control systems,
289
+ and issue tracking systems that are managed by, or on behalf of, the
290
+ Licensor for the purpose of discussing and improving the Work, but
291
+ excluding communication that is conspicuously marked or otherwise
292
+ designated in writing by the copyright owner as "Not a Contribution."
293
+
294
+ "Contributor" shall mean Licensor and any individual or Legal Entity
295
+ on behalf of whom a Contribution has been received by Licensor and
296
+ subsequently incorporated within the Work.
297
+
298
+ 2. Grant of Copyright License. Subject to the terms and conditions of
299
+ this License, each Contributor hereby grants to You a perpetual,
300
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
301
+ copyright license to reproduce, prepare Derivative Works of,
302
+ publicly display, publicly perform, sublicense, and distribute the
303
+ Work and such Derivative Works in Source or Object form.
304
+
305
+ 3. Grant of Patent License. Subject to the terms and conditions of
306
+ this License, each Contributor hereby grants to You a perpetual,
307
+ worldwide, non-exclusive, no-charge, royalty-free, irrevocable
308
+ (except as stated in this section) patent license to make, have made,
309
+ use, offer to sell, sell, import, and otherwise transfer the Work,
310
+ where such license applies only to those patent claims licensable
311
+ by such Contributor that are necessarily infringed by their
312
+ Contribution(s) alone or by combination of their Contribution(s)
313
+ with the Work to which such Contribution(s) was submitted. If You
314
+ institute patent litigation against any entity (including a
315
+ cross-claim or counterclaim in a lawsuit) alleging that the Work
316
+ or a Contribution incorporated within the Work constitutes direct
317
+ or contributory patent infringement, then any patent licenses
318
+ granted to You under this License for that Work shall terminate
319
+ as of the date such litigation is filed.
320
+
321
+ 4. Redistribution. You may reproduce and distribute copies of the
322
+ Work or Derivative Works thereof in any medium, with or without
323
+ modifications, and in Source or Object form, provided that You
324
+ meet the following conditions:
325
+
326
+ (a) You must give any other recipients of the Work or
327
+ Derivative Works a copy of this License; and
328
+
329
+ (b) You must cause any modified files to carry prominent notices
330
+ stating that You changed the files; and
331
+
332
+ (c) You must retain, in the Source form of any Derivative Works
333
+ that You distribute, all copyright, patent, trademark, and
334
+ attribution notices from the Source form of the Work,
335
+ excluding those notices that do not pertain to any part of
336
+ the Derivative Works; and
337
+
338
+ (d) If the Work includes a "NOTICE" text file as part of its
339
+ distribution, then any Derivative Works that You distribute must
340
+ include a readable copy of the attribution notices contained
341
+ within such NOTICE file, excluding those notices that do not
342
+ pertain to any part of the Derivative Works, in at least one
343
+ of the following places: within a NOTICE text file distributed
344
+ as part of the Derivative Works; within the Source form or
345
+ documentation, if provided along with the Derivative Works; or,
346
+ within a display generated by the Derivative Works, if and
347
+ wherever such third-party notices normally appear. The contents
348
+ of the NOTICE file are for informational purposes only and
349
+ do not modify the License. You may add Your own attribution
350
+ notices within Derivative Works that You distribute, alongside
351
+ or as an addendum to the NOTICE text from the Work, provided
352
+ that such additional attribution notices cannot be construed
353
+ as modifying the License.
354
+
355
+ You may add Your own copyright statement to Your modifications and
356
+ may provide additional or different license terms and conditions
357
+ for use, reproduction, or distribution of Your modifications, or
358
+ for any such Derivative Works as a whole, provided Your use,
359
+ reproduction, and distribution of the Work otherwise complies with
360
+ the conditions stated in this License.
361
+
362
+ 5. Submission of Contributions. Unless You explicitly state otherwise,
363
+ any Contribution intentionally submitted for inclusion in the Work
364
+ by You to the Licensor shall be under the terms and conditions of
365
+ this License, without any additional terms or conditions.
366
+ Notwithstanding the above, nothing herein shall supersede or modify
367
+ the terms of any separate license agreement you may have executed
368
+ with Licensor regarding such Contributions.
369
+
370
+ 6. Trademarks. This License does not grant permission to use the trade
371
+ names, trademarks, service marks, or product names of the Licensor,
372
+ except as required for reasonable and customary use in describing the
373
+ origin of the Work and reproducing the content of the NOTICE file.
374
+
375
+ 7. Disclaimer of Warranty. Unless required by applicable law or
376
+ agreed to in writing, Licensor provides the Work (and each
377
+ Contributor provides its Contributions) on an "AS IS" BASIS,
378
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
379
+ implied, including, without limitation, any warranties or conditions
380
+ of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
381
+ PARTICULAR PURPOSE. You are solely responsible for determining the
382
+ appropriateness of using or redistributing the Work and assume any
383
+ risks associated with Your exercise of permissions under this License.
384
+
385
+ 8. Limitation of Liability. In no event and under no legal theory,
386
+ whether in tort (including negligence), contract, or otherwise,
387
+ unless required by applicable law (such as deliberate and grossly
388
+ negligent acts) or agreed to in writing, shall any Contributor be
389
+ liable to You for damages, including any direct, indirect, special,
390
+ incidental, or consequential damages of any character arising as a
391
+ result of this License or out of the use or inability to use the
392
+ Work (including but not limited to damages for loss of goodwill,
393
+ work stoppage, computer failure or malfunction, or any and all
394
+ other commercial damages or losses), even if such Contributor
395
+ has been advised of the possibility of such damages.
396
+
397
+ 9. Accepting Warranty or Additional Liability. While redistributing
398
+ the Work or Derivative Works thereof, You may choose to offer,
399
+ and charge a fee for, acceptance of support, warranty, indemnity,
400
+ or other liability obligations and/or rights consistent with this
401
+ License. However, in accepting such obligations, You may act only
402
+ on Your own behalf and on Your sole responsibility, not on behalf
403
+ of any other Contributor, and only if You agree to indemnify,
404
+ defend, and hold each Contributor harmless for any liability
405
+ incurred by, or claims asserted against, such Contributor by reason
406
+ of your accepting any such warranty or additional liability.
407
+
408
+ END OF TERMS AND CONDITIONS
409
+
410
+ APPENDIX: How to apply the Apache License to your work.
411
+
412
+ To apply the Apache License to your work, attach the following
413
+ boilerplate notice, with the fields enclosed by brackets "[]"
414
+ replaced with your own identifying information. (Don't include
415
+ the brackets!) The text should be enclosed in the appropriate
416
+ comment syntax for the file format. We also recommend that a
417
+ file or class name and description of purpose be included on the
418
+ same "printed page" as the copyright notice for easier
419
+ identification within third-party archives.
420
+
421
+ Copyright [yyyy] [name of copyright owner]
422
+
423
+ Licensed under the Apache License, Version 2.0 (the "License");
424
+ you may not use this file except in compliance with the License.
425
+ You may obtain a copy of the License at
426
+
427
+ http://www.apache.org/licenses/LICENSE-2.0
428
+
429
+ Unless required by applicable law or agreed to in writing, software
430
+ distributed under the License is distributed on an "AS IS" BASIS,
431
+ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
432
+ See the License for the specific language governing permissions and
433
+ limitations under the License.
434
+
435
+
README.md ADDED
@@ -0,0 +1,66 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ base_model: tencent/WeMM-Embedding-2B
3
+ license: apache-2.0
4
+ tags:
5
+ - embedding
6
+ - multimodal
7
+ - fp8
8
+ - qwen3_5
9
+ - image-text-to-text
10
+ - video-text-to-text
11
+ ---
12
+
13
+ # WeMM-Embedding-2B — FP8 Quantization
14
+
15
+ FP8 (8-bit float, E4M3) quantization of [tencent/WeMM-Embedding-2B](https://huggingface.co/tencent/WeMM-Embedding-2B), intended for vLLM / SGLang-class backends with FP8 support (RTX 4090 / Ada and newer have native FP8 tensor cores).
16
+
17
+ This is a **separate repo from the GGUF build** — GGUF targets llama.cpp; this FP8 build targets GPU inference servers.
18
+
19
+ ## Files
20
+
21
+ - `model.fp8.safetensors` — weights stored as `float8_e4m3fn` (per-tensor scale).
22
+ - `fp8_scales.json` — per-layer dequant scale (layer name -> scalar).
23
+ - Supporting files (config, tokenizer, custom `WeMMEmbedding` modeling, chat templates) are mirrored from the base model so `AutoModel` can load it.
24
+
25
+ > Note: the saved weights are raw fp8 + scale. To run, dequantize at load time (fp8 -> bf16) or serve through a backend that natively consumes fp8. See Usage below.
26
+
27
+ ## Evaluation (STS-B, same engine)
28
+
29
+ All numbers use the **same engine** (transformers / torch `AutoModel`) for both the BF16 baseline and the FP8 model, so Δρ is pure FP8 rounding error.
30
+
31
+ | Model | Bits/Weight | STS-B Spearman ρ | Δρ vs BF16 | Emb Cosine vs BF16 |
32
+ |---|---|---|---|---|
33
+ | BF16 | 16.00 | 0.8124 | — | — |
34
+ | FP8 (manual per-tensor E4M3) | 8.00 | 0.8114 | +0.12% | 0.9987 |
35
+
36
+ Metrics:
37
+ - **STS-B Spearman ρ**: rank correlation between model cosine similarities and human similarity scores (0-5). Higher is better.
38
+ - **Δρ vs BF16**: relative drop of ρ against the BF16 baseline. Negative = scored slightly above baseline (within noise).
39
+ - **Emb Cosine vs BF16**: mean cosine between each sentence embedding and its BF16 counterpart (space fidelity). 1.0 = identical.
40
+
41
+ Conclusion: FP8 causes negligible quality loss (Δρ = +0.12%, Emb Cosine = 0.999) on STS-B. This is the recommended format when serving on FP8-capable GPUs.
42
+
43
+ ## Usage
44
+
45
+ ### Dequantize to bf16 at load time (transformers)
46
+
47
+ ```python
48
+ import torch, json, safetensors.torch as st
49
+ from transformers import AutoModel, AutoProcessor
50
+
51
+ sd = st.load_file("model.fp8.safetensors")
52
+ scales = json.load(open("fp8_scales.json"))
53
+ for k, s in scales.items(): # k like 'model.xxx.weight'
54
+ sd[k] = (sd[k].to(torch.float32) * s).to(torch.bfloat16) # dequant
55
+ # save a runnable bf16 copy, or load directly:
56
+ model = AutoModel.from_pretrained(".", trust_remote_code=True, dtype=torch.bfloat16)
57
+ ```
58
+
59
+ ### vLLM / SGLang
60
+
61
+ These backends expect a compressed-tensors / native fp8 checkpoint. The raw fp8+scale layout here is not yet wrapped for direct vLLM loading; repackaging into the backend's fp8 format (or quantizing the base model with the backend's own fp8 path) is required before serving. The quality numbers above already demonstrate the FP8 format itself is near-lossless.
62
+
63
+ ## Notes
64
+
65
+ - FP8 was produced by **manual per-tensor fp8 (E4M3) quantization of all 285 Linear layers** (including the vision tower). `llmcompressor` oneshot did not actually quantize the qwen3_5 custom Linear layers, so the manual path is used for the reported numbers.
66
+ - The GGUF variants (Q8_0 / Q6_K / Q5_K_M / IQ4_XS / IQ3_M) live in a separate repo.
chat_template.jinja ADDED
@@ -0,0 +1,154 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {%- set image_count = namespace(value=0) %}
2
+ {%- set video_count = namespace(value=0) %}
3
+ {%- macro render_content(content, do_vision_count, is_system_content=false) %}
4
+ {%- if content is string %}
5
+ {{- content }}
6
+ {%- elif content is iterable and content is not mapping %}
7
+ {%- for item in content %}
8
+ {%- if 'image' in item or 'image_url' in item or item.type == 'image' %}
9
+ {%- if is_system_content %}
10
+ {{- raise_exception('System message cannot contain images.') }}
11
+ {%- endif %}
12
+ {%- if do_vision_count %}
13
+ {%- set image_count.value = image_count.value + 1 %}
14
+ {%- endif %}
15
+ {%- if add_vision_id %}
16
+ {{- 'Picture ' ~ image_count.value ~ ': ' }}
17
+ {%- endif %}
18
+ {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
19
+ {%- elif 'video' in item or item.type == 'video' %}
20
+ {%- if is_system_content %}
21
+ {{- raise_exception('System message cannot contain videos.') }}
22
+ {%- endif %}
23
+ {%- if do_vision_count %}
24
+ {%- set video_count.value = video_count.value + 1 %}
25
+ {%- endif %}
26
+ {%- if add_vision_id %}
27
+ {{- 'Video ' ~ video_count.value ~ ': ' }}
28
+ {%- endif %}
29
+ {{- '<|vision_start|><|video_pad|><|vision_end|>' }}
30
+ {%- elif 'text' in item %}
31
+ {{- item.text }}
32
+ {%- else %}
33
+ {{- raise_exception('Unexpected item type in content.') }}
34
+ {%- endif %}
35
+ {%- endfor %}
36
+ {%- elif content is none or content is undefined %}
37
+ {{- '' }}
38
+ {%- else %}
39
+ {{- raise_exception('Unexpected content type.') }}
40
+ {%- endif %}
41
+ {%- endmacro %}
42
+ {%- if not messages %}
43
+ {{- raise_exception('No messages provided.') }}
44
+ {%- endif %}
45
+ {%- if tools and tools is iterable and tools is not mapping %}
46
+ {{- '<|im_start|>system\n' }}
47
+ {{- "# Tools\n\nYou have access to the following functions:\n\n<tools>" }}
48
+ {%- for tool in tools %}
49
+ {{- "\n" }}
50
+ {{- tool | tojson }}
51
+ {%- endfor %}
52
+ {{- "\n</tools>" }}
53
+ {{- '\n\nIf you choose to call a function ONLY reply in the following format with NO suffix:\n\n<tool_call>\n<function=example_function_name>\n<parameter=example_parameter_1>\nvalue_1\n</parameter>\n<parameter=example_parameter_2>\nThis is the value for the second parameter\nthat can span\nmultiple lines\n</parameter>\n</function>\n</tool_call>\n\n<IMPORTANT>\nReminder:\n- Function calls MUST follow the specified format: an inner <function=...></function> block must be nested within <tool_call></tool_call> XML tags\n- Required parameters MUST be specified\n- You may provide optional reasoning for your function call in natural language BEFORE the function call, but NOT after\n- If there is no function call available, answer the question like normal with your current knowledge and do not tell the user about function calls\n</IMPORTANT>' }}
54
+ {%- if messages[0].role == 'system' %}
55
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
56
+ {%- if content %}
57
+ {{- '\n\n' + content }}
58
+ {%- endif %}
59
+ {%- endif %}
60
+ {{- '<|im_end|>\n' }}
61
+ {%- else %}
62
+ {%- if messages[0].role == 'system' %}
63
+ {%- set content = render_content(messages[0].content, false, true)|trim %}
64
+ {{- '<|im_start|>system\n' + content + '<|im_end|>\n' }}
65
+ {%- endif %}
66
+ {%- endif %}
67
+ {%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
68
+ {%- for message in messages[::-1] %}
69
+ {%- set index = (messages|length - 1) - loop.index0 %}
70
+ {%- if ns.multi_step_tool and message.role == "user" %}
71
+ {%- set content = render_content(message.content, false)|trim %}
72
+ {%- if not(content.startswith('<tool_response>') and content.endswith('</tool_response>')) %}
73
+ {%- set ns.multi_step_tool = false %}
74
+ {%- set ns.last_query_index = index %}
75
+ {%- endif %}
76
+ {%- endif %}
77
+ {%- endfor %}
78
+ {%- if ns.multi_step_tool %}
79
+ {{- raise_exception('No user query found in messages.') }}
80
+ {%- endif %}
81
+ {%- for message in messages %}
82
+ {%- set content = render_content(message.content, true)|trim %}
83
+ {%- if message.role == "system" %}
84
+ {%- if not loop.first %}
85
+ {{- raise_exception('System message must be at the beginning.') }}
86
+ {%- endif %}
87
+ {%- elif message.role == "user" %}
88
+ {{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
89
+ {%- elif message.role == "assistant" %}
90
+ {%- set reasoning_content = '' %}
91
+ {%- if message.reasoning_content is string %}
92
+ {%- set reasoning_content = message.reasoning_content %}
93
+ {%- else %}
94
+ {%- if '</think>' in content %}
95
+ {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
96
+ {%- set content = content.split('</think>')[-1].lstrip('\n') %}
97
+ {%- endif %}
98
+ {%- endif %}
99
+ {%- set reasoning_content = reasoning_content|trim %}
100
+ {%- if loop.index0 > ns.last_query_index %}
101
+ {{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content + '\n</think>\n\n' + content }}
102
+ {%- else %}
103
+ {{- '<|im_start|>' + message.role + '\n' + content }}
104
+ {%- endif %}
105
+ {%- if message.tool_calls and message.tool_calls is iterable and message.tool_calls is not mapping %}
106
+ {%- for tool_call in message.tool_calls %}
107
+ {%- if tool_call.function is defined %}
108
+ {%- set tool_call = tool_call.function %}
109
+ {%- endif %}
110
+ {%- if loop.first %}
111
+ {%- if content|trim %}
112
+ {{- '\n\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
113
+ {%- else %}
114
+ {{- '<tool_call>\n<function=' + tool_call.name + '>\n' }}
115
+ {%- endif %}
116
+ {%- else %}
117
+ {{- '\n<tool_call>\n<function=' + tool_call.name + '>\n' }}
118
+ {%- endif %}
119
+ {%- if tool_call.arguments is defined %}
120
+ {%- for args_name, args_value in tool_call.arguments|items %}
121
+ {{- '<parameter=' + args_name + '>\n' }}
122
+ {%- set args_value = args_value | tojson | safe if args_value is mapping or (args_value is sequence and args_value is not string) else args_value | string %}
123
+ {{- args_value }}
124
+ {{- '\n</parameter>\n' }}
125
+ {%- endfor %}
126
+ {%- endif %}
127
+ {{- '</function>\n</tool_call>' }}
128
+ {%- endfor %}
129
+ {%- endif %}
130
+ {{- '<|im_end|>\n' }}
131
+ {%- elif message.role == "tool" %}
132
+ {%- if loop.previtem and loop.previtem.role != "tool" %}
133
+ {{- '<|im_start|>user' }}
134
+ {%- endif %}
135
+ {{- '\n<tool_response>\n' }}
136
+ {{- content }}
137
+ {{- '\n</tool_response>' }}
138
+ {%- if not loop.last and loop.nextitem.role != "tool" %}
139
+ {{- '<|im_end|>\n' }}
140
+ {%- elif loop.last %}
141
+ {{- '<|im_end|>\n' }}
142
+ {%- endif %}
143
+ {%- else %}
144
+ {{- raise_exception('Unexpected message role.') }}
145
+ {%- endif %}
146
+ {%- endfor %}
147
+ {%- if add_generation_prompt %}
148
+ {{- '<|im_start|>assistant\n' }}
149
+ {%- if enable_thinking is defined and enable_thinking is true %}
150
+ {{- '<think>\n' }}
151
+ {%- else %}
152
+ {{- '<think>\n\n</think>\n\n' }}
153
+ {%- endif %}
154
+ {%- endif %}
config.json ADDED
@@ -0,0 +1,113 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "Qwen3_5ForConditionalGeneration"
4
+ ],
5
+ "dtype": "bfloat16",
6
+ "image_token_id": 248056,
7
+ "model_type": "qwen3_5",
8
+ "text_config": {
9
+ "attention_bias": false,
10
+ "attention_dropout": 0.0,
11
+ "attn_output_gate": true,
12
+ "bos_token_id": null,
13
+ "dtype": "bfloat16",
14
+ "eos_token_id": 248044,
15
+ "full_attention_interval": 4,
16
+ "head_dim": 256,
17
+ "hidden_act": "silu",
18
+ "hidden_size": 2048,
19
+ "initializer_range": 0.02,
20
+ "intermediate_size": 6144,
21
+ "layer_types": [
22
+ "linear_attention",
23
+ "linear_attention",
24
+ "linear_attention",
25
+ "full_attention",
26
+ "linear_attention",
27
+ "linear_attention",
28
+ "linear_attention",
29
+ "full_attention",
30
+ "linear_attention",
31
+ "linear_attention",
32
+ "linear_attention",
33
+ "full_attention",
34
+ "linear_attention",
35
+ "linear_attention",
36
+ "linear_attention",
37
+ "full_attention",
38
+ "linear_attention",
39
+ "linear_attention",
40
+ "linear_attention",
41
+ "full_attention",
42
+ "linear_attention",
43
+ "linear_attention",
44
+ "linear_attention",
45
+ "full_attention"
46
+ ],
47
+ "linear_conv_kernel_dim": 4,
48
+ "linear_key_head_dim": 128,
49
+ "linear_num_key_heads": 16,
50
+ "linear_num_value_heads": 16,
51
+ "linear_value_head_dim": 128,
52
+ "mamba_ssm_dtype": "float32",
53
+ "max_position_embeddings": 262144,
54
+ "mlp_only_layers": [],
55
+ "model_type": "qwen3_5_text",
56
+ "mtp_num_hidden_layers": 1,
57
+ "mtp_use_dedicated_embeddings": false,
58
+ "num_attention_heads": 8,
59
+ "num_hidden_layers": 24,
60
+ "num_key_value_heads": 2,
61
+ "pad_token_id": null,
62
+ "partial_rotary_factor": 0.25,
63
+ "rms_norm_eps": 1e-06,
64
+ "rope_parameters": {
65
+ "mrope_interleaved": true,
66
+ "mrope_section": [
67
+ 11,
68
+ 11,
69
+ 10
70
+ ],
71
+ "partial_rotary_factor": 0.25,
72
+ "rope_theta": 10000000,
73
+ "rope_type": "default"
74
+ },
75
+ "tie_word_embeddings": false,
76
+ "use_cache": true,
77
+ "vocab_size": 248078
78
+ },
79
+ "tie_word_embeddings": false,
80
+ "transformers_version": "5.2.0",
81
+ "video_token_id": 248057,
82
+ "vision_config": {
83
+ "deepstack_visual_indexes": [],
84
+ "depth": 24,
85
+ "dtype": "bfloat16",
86
+ "hidden_act": "gelu_pytorch_tanh",
87
+ "hidden_size": 1024,
88
+ "in_channels": 3,
89
+ "initializer_range": 0.02,
90
+ "intermediate_size": 4096,
91
+ "model_type": "qwen3_5",
92
+ "num_heads": 16,
93
+ "num_position_embeddings": 2304,
94
+ "out_hidden_size": 2048,
95
+ "patch_size": 16,
96
+ "spatial_merge_size": 2,
97
+ "temporal_patch_size": 2
98
+ },
99
+ "vision_end_token_id": 248054,
100
+ "vision_start_token_id": 248053,
101
+ "auto_map": {
102
+ "AutoModel": "modeling_wemm_embedding.WeMMEmbedding",
103
+ "AutoModelForCausalLM": "modeling_wemm_embedding.WeMMEmbedding"
104
+ },
105
+ "matryoshka_dimensions": [
106
+ 64,
107
+ 128,
108
+ 256,
109
+ 512,
110
+ 1024,
111
+ 2048
112
+ ]
113
+ }
config_sentence_transformers.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "__version__": {
3
+ "pytorch": "2.11.0+cu128",
4
+ "sentence_transformers": "5.7.0",
5
+ "transformers": "5.2.0"
6
+ },
7
+ "default_prompt_name": null,
8
+ "model_type": "SentenceTransformer",
9
+ "prompts": {},
10
+ "similarity_fn_name": "cosine"
11
+ }
embedding_chat_template.jinja ADDED
@@ -0,0 +1,28 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {# Direct-tokenization template for vLLM 0.27 embedding requests.
2
+ Qwen's processor post-processor removes the role newline before images,
3
+ keeps it before text/video, and appends <embedding> without a preceding
4
+ newline. vLLM 0.27 adds each frame's vision boundaries itself, so video
5
+ uses a bare <video_pad> placeholder. #}
6
+ {%- for message in messages %}
7
+ {{- '<|im_start|>' + message.role }}
8
+ {%- if message.content is string %}
9
+ {{- '\n' + message.content }}
10
+ {%- else %}
11
+ {%- for item in message.content %}
12
+ {%- if item.type in ['image', 'image_url'] or 'image' in item or 'image_url' in item %}
13
+ {{- '<|vision_start|><|image_pad|><|vision_end|>' }}
14
+ {%- elif item.type in ['video', 'video_url'] or 'video' in item or 'video_url' in item %}
15
+ {{- '\n<|video_pad|>' }}
16
+ {%- elif item.type == 'text' or 'text' in item %}
17
+ {{- ('\n' if loop.first else '') + item.text }}
18
+ {%- else %}
19
+ {{- raise_exception('Unsupported embedding content type') }}
20
+ {%- endif %}
21
+ {%- endfor %}
22
+ {%- endif %}
23
+ {{- '<|im_end|>' }}
24
+ {%- if not loop.last %}
25
+ {{- '\n' }}
26
+ {%- endif %}
27
+ {%- endfor %}
28
+ {{- '<embedding>' }}
fp8_scales.json ADDED
@@ -0,0 +1 @@
 
 
1
+ {"model.visual.blocks.0.attn.qkv.weight": 0.0014299665344879031, "model.visual.blocks.0.attn.proj.weight": 0.0007890973938629031, "model.visual.blocks.0.mlp.linear_fc1.weight": 0.0006452288362197578, "model.visual.blocks.0.mlp.linear_fc2.weight": 0.0005841936799697578, "model.visual.blocks.1.attn.qkv.weight": 0.0011247907532379031, "model.visual.blocks.1.attn.proj.weight": 0.0007934570894576609, "model.visual.blocks.1.mlp.linear_fc1.weight": 0.0008588518830947578, "model.visual.blocks.1.mlp.linear_fc2.weight": 0.0013253348879516125, "model.visual.blocks.2.attn.qkv.weight": 0.0006757464143447578, "model.visual.blocks.2.attn.proj.weight": 0.0005711147096008062, "model.visual.blocks.2.mlp.linear_fc1.weight": 0.0007934570894576609, "model.visual.blocks.2.mlp.linear_fc2.weight": 0.0012991769472137094, "model.visual.blocks.3.attn.qkv.weight": 0.0005841936799697578, "model.visual.blocks.3.attn.proj.weight": 0.0006844656891189516, "model.visual.blocks.3.mlp.linear_fc1.weight": 0.0008283343049697578, "model.visual.blocks.3.mlp.linear_fc2.weight": 0.0012294225161895156, "model.visual.blocks.4.attn.qkv.weight": 0.00047302249004133046, "model.visual.blocks.4.attn.proj.weight": 0.0006670270813629031, "model.visual.blocks.4.mlp.linear_fc1.weight": 0.0006321498658508062, "model.visual.blocks.4.mlp.linear_fc2.weight": 0.0012991769472137094, "model.visual.blocks.5.attn.qkv.weight": 0.00054931640625, "model.visual.blocks.5.attn.proj.weight": 0.0004512242157943547, "model.visual.blocks.5.mlp.linear_fc1.weight": 0.00044250491191633046, "model.visual.blocks.5.mlp.linear_fc2.weight": 0.0010157994693145156, "model.visual.blocks.6.attn.qkv.weight": 0.0004141671524848789, "model.visual.blocks.6.attn.proj.weight": 0.00041634697117842734, "model.visual.blocks.6.mlp.linear_fc1.weight": 0.000579833984375, "model.visual.blocks.6.mlp.linear_fc2.weight": 0.0013253348879516125, "model.visual.blocks.7.attn.qkv.weight": 0.0005623953766189516, "model.visual.blocks.7.attn.proj.weight": 0.00026484898990020156, "model.visual.blocks.7.mlp.linear_fc1.weight": 0.0006277902284637094, "model.visual.blocks.7.mlp.linear_fc2.weight": 0.0005711147096008062, "model.visual.blocks.8.attn.qkv.weight": 0.0005078997346572578, "model.visual.blocks.8.attn.proj.weight": 0.0006321498658508062, "model.visual.blocks.8.mlp.linear_fc1.weight": 0.0006365095032379031, "model.visual.blocks.8.mlp.linear_fc2.weight": 0.0005514962831512094, "model.visual.blocks.9.attn.qkv.weight": 0.00047302249004133046, "model.visual.blocks.9.attn.proj.weight": 0.0002506801101844758, "model.visual.blocks.9.mlp.linear_fc1.weight": 0.0006103515625, "model.visual.blocks.9.mlp.linear_fc2.weight": 0.0006539481109939516, "model.visual.blocks.10.attn.qkv.weight": 0.0005209787050262094, "model.visual.blocks.10.attn.proj.weight": 0.0002659388992469758, "model.visual.blocks.10.mlp.linear_fc1.weight": 0.0006277902284637094, "model.visual.blocks.10.mlp.linear_fc2.weight": 0.0013863700442016125, "model.visual.blocks.11.attn.qkv.weight": 0.00039672854472883046, "model.visual.blocks.11.attn.proj.weight": 0.00022016253205947578, "model.visual.blocks.11.mlp.linear_fc1.weight": 0.0007324219332076609, "model.visual.blocks.11.mlp.linear_fc2.weight": 0.0012032645754516125, "model.visual.blocks.12.attn.qkv.weight": 0.0003574916336219758, "model.visual.blocks.12.attn.proj.weight": 0.00032479423680342734, "model.visual.blocks.12.mlp.linear_fc1.weight": 0.0009111677063629031, "model.visual.blocks.12.mlp.linear_fc2.weight": 0.0008980887942016125, "model.visual.blocks.13.attn.qkv.weight": 0.0003749302704818547, "model.visual.blocks.13.attn.proj.weight": 0.00023651124502066523, "model.visual.blocks.13.mlp.linear_fc1.weight": 0.0009286063723266125, "model.visual.blocks.13.mlp.linear_fc2.weight": 0.0007629395113326609, "model.visual.blocks.14.attn.qkv.weight": 0.0003444126923568547, "model.visual.blocks.14.attn.proj.weight": 0.0003596714814193547, "model.visual.blocks.14.mlp.linear_fc1.weight": 0.0003291539032943547, "model.visual.blocks.14.mlp.linear_fc2.weight": 0.0010027204407379031, "model.visual.blocks.15.attn.qkv.weight": 0.0003073556290473789, "model.visual.blocks.15.attn.proj.weight": 0.0003923688782379031, "model.visual.blocks.15.mlp.linear_fc1.weight": 0.00033569338847883046, "model.visual.blocks.15.mlp.linear_fc2.weight": 0.0009765625, "model.visual.blocks.16.attn.qkv.weight": 0.00042724612285383046, "model.visual.blocks.16.attn.proj.weight": 0.0002593994140625, "model.visual.blocks.16.mlp.linear_fc1.weight": 0.00032479423680342734, "model.visual.blocks.16.mlp.linear_fc2.weight": 0.0007716587861068547, "model.visual.blocks.17.attn.qkv.weight": 0.00026375908055342734, "model.visual.blocks.17.attn.proj.weight": 0.00044904439710080624, "model.visual.blocks.17.mlp.linear_fc1.weight": 0.0003378732071723789, "model.visual.blocks.17.mlp.linear_fc2.weight": 0.0006452288362197578, "model.visual.blocks.18.attn.qkv.weight": 0.0004926409455947578, "model.visual.blocks.18.attn.proj.weight": 0.00035095217754133046, "model.visual.blocks.18.mlp.linear_fc1.weight": 0.00033351354068145156, "model.visual.blocks.18.mlp.linear_fc2.weight": 0.0007106236298568547, "model.visual.blocks.19.attn.qkv.weight": 0.00029427665867842734, "model.visual.blocks.19.attn.proj.weight": 0.0002964564773719758, "model.visual.blocks.19.mlp.linear_fc1.weight": 0.0003749302704818547, "model.visual.blocks.19.mlp.linear_fc2.weight": 0.00036621096660383046, "model.visual.blocks.20.attn.qkv.weight": 0.0003008161438629031, "model.visual.blocks.20.attn.proj.weight": 0.00031825475161895156, "model.visual.blocks.20.mlp.linear_fc1.weight": 0.00033569338847883046, "model.visual.blocks.20.mlp.linear_fc2.weight": 0.0003422328445594758, "model.visual.blocks.21.attn.qkv.weight": 0.0003880092117469758, "model.visual.blocks.21.attn.proj.weight": 0.0003073556290473789, "model.visual.blocks.21.mlp.linear_fc1.weight": 0.0005754743469879031, "model.visual.blocks.21.mlp.linear_fc2.weight": 0.00046212333836592734, "model.visual.blocks.22.attn.qkv.weight": 0.00037057060399092734, "model.visual.blocks.22.attn.proj.weight": 0.00040980748599395156, "model.visual.blocks.22.mlp.linear_fc1.weight": 0.0005296979798004031, "model.visual.blocks.22.mlp.linear_fc2.weight": 0.00034005302586592734, "model.visual.blocks.23.attn.qkv.weight": 0.0003313337219879031, "model.visual.blocks.23.attn.proj.weight": 0.0005449567688629031, "model.visual.blocks.23.mlp.linear_fc1.weight": 0.0018920899601653218, "model.visual.blocks.23.mlp.linear_fc2.weight": 0.0015607562381774187, "model.visual.merger.linear_fc1.weight": 0.00021362306142691523, "model.visual.merger.linear_fc2.weight": 0.00044686454930342734, "model.language_model.layers.0.linear_attn.out_proj.weight": 0.0008762904908508062, "model.language_model.layers.0.linear_attn.in_proj_qkv.weight": 0.0008850098238326609, "model.language_model.layers.0.linear_attn.in_proj_z.weight": 0.0008283343049697578, "model.language_model.layers.0.linear_attn.in_proj_b.weight": 0.00027138847508467734, "model.language_model.layers.0.linear_attn.in_proj_a.weight": 0.0004926409455947578, "model.language_model.layers.0.mlp.gate_proj.weight": 0.0004228864563629031, "model.language_model.layers.0.mlp.up_proj.weight": 0.0002626691712066531, "model.language_model.layers.0.mlp.down_proj.weight": 0.0007324219332076609, "model.language_model.layers.1.linear_attn.out_proj.weight": 0.0010157994693145156, "model.language_model.layers.1.linear_attn.in_proj_qkv.weight": 0.0005711147096008062, "model.language_model.layers.1.linear_attn.in_proj_z.weight": 0.00044904439710080624, "model.language_model.layers.1.linear_attn.in_proj_b.weight": 0.0003422328445594758, "model.language_model.layers.1.linear_attn.in_proj_a.weight": 0.0005057199159637094, "model.language_model.layers.1.mlp.gate_proj.weight": 0.00031825475161895156, "model.language_model.layers.1.mlp.up_proj.weight": 0.0002692086563911289, "model.language_model.layers.1.mlp.down_proj.weight": 0.0006103515625, "model.language_model.layers.2.linear_attn.out_proj.weight": 0.0009111677063629031, "model.language_model.layers.2.linear_attn.in_proj_qkv.weight": 0.0005405971314758062, "model.language_model.layers.2.linear_attn.in_proj_z.weight": 0.00039454869693145156, "model.language_model.layers.2.linear_attn.in_proj_b.weight": 0.00033569338847883046, "model.language_model.layers.2.linear_attn.in_proj_a.weight": 0.0004359654267318547, "model.language_model.layers.2.mlp.gate_proj.weight": 0.00039454869693145156, "model.language_model.layers.2.mlp.up_proj.weight": 0.0003226144181098789, "model.language_model.layers.2.mlp.down_proj.weight": 0.0008196149719879031, "model.language_model.layers.3.self_attn.q_proj.weight": 0.0007280622376129031, "model.language_model.layers.3.self_attn.k_proj.weight": 0.00032479423680342734, "model.language_model.layers.3.self_attn.v_proj.weight": 0.0003073556290473789, "model.language_model.layers.3.self_attn.o_proj.weight": 0.0008588518830947578, "model.language_model.layers.3.mlp.gate_proj.weight": 0.0003531319962348789, "model.language_model.layers.3.mlp.up_proj.weight": 0.0002593994140625, "model.language_model.layers.3.mlp.down_proj.weight": 0.0007237026002258062, "model.language_model.layers.4.linear_attn.out_proj.weight": 0.0008850098238326609, "model.language_model.layers.4.linear_attn.in_proj_qkv.weight": 0.0006452288362197578, "model.language_model.layers.4.linear_attn.in_proj_z.weight": 0.00036403111880645156, "model.language_model.layers.4.linear_attn.in_proj_b.weight": 0.00028773717349395156, "model.language_model.layers.4.linear_attn.in_proj_a.weight": 0.0003880092117469758, "model.language_model.layers.4.mlp.gate_proj.weight": 0.0004599435196723789, "model.language_model.layers.4.mlp.up_proj.weight": 0.00022125245595816523, "model.language_model.layers.4.mlp.down_proj.weight": 0.0005558559205383062, "model.language_model.layers.5.linear_attn.out_proj.weight": 0.0009024484315887094, "model.language_model.layers.5.linear_attn.in_proj_qkv.weight": 0.0006888253847137094, "model.language_model.layers.5.linear_attn.in_proj_z.weight": 0.0003422328445594758, "model.language_model.layers.5.linear_attn.in_proj_b.weight": 0.00022997175983618945, "model.language_model.layers.5.linear_attn.in_proj_a.weight": 0.00033569338847883046, "model.language_model.layers.5.mlp.gate_proj.weight": 0.0002964564773719758, "model.language_model.layers.5.mlp.up_proj.weight": 0.00019073487783316523, "model.language_model.layers.5.mlp.down_proj.weight": 0.0005667550722137094, "model.language_model.layers.6.linear_attn.out_proj.weight": 0.0006844656891189516, "model.language_model.layers.6.linear_attn.in_proj_qkv.weight": 0.00047956197522580624, "model.language_model.layers.6.linear_attn.in_proj_z.weight": 0.0004359654267318547, "model.language_model.layers.6.linear_attn.in_proj_b.weight": 0.0003269740554969758, "model.language_model.layers.6.linear_attn.in_proj_a.weight": 0.00030953544774092734, "model.language_model.layers.6.mlp.gate_proj.weight": 0.0005209787050262094, "model.language_model.layers.6.mlp.up_proj.weight": 0.00021798271336592734, "model.language_model.layers.6.mlp.down_proj.weight": 0.0005340576171875, "model.language_model.layers.7.self_attn.q_proj.weight": 0.0005841936799697578, "model.language_model.layers.7.self_attn.k_proj.weight": 0.0002681187470443547, "model.language_model.layers.7.self_attn.v_proj.weight": 0.0003138951142318547, "model.language_model.layers.7.self_attn.o_proj.weight": 0.0005209787050262094, "model.language_model.layers.7.mlp.gate_proj.weight": 0.00041634697117842734, "model.language_model.layers.7.mlp.up_proj.weight": 0.00024087089695967734, "model.language_model.layers.7.mlp.down_proj.weight": 0.0006844656891189516, "model.language_model.layers.8.linear_attn.out_proj.weight": 0.0007978167268447578, "model.language_model.layers.8.linear_attn.in_proj_qkv.weight": 0.0005035400390625, "model.language_model.layers.8.linear_attn.in_proj_z.weight": 0.0004991804016754031, "model.language_model.layers.8.linear_attn.in_proj_b.weight": 0.00018419539264868945, "model.language_model.layers.8.linear_attn.in_proj_a.weight": 0.0002811976883094758, "model.language_model.layers.8.mlp.gate_proj.weight": 0.00034877232974395156, "model.language_model.layers.8.mlp.up_proj.weight": 0.00020381383365020156, "model.language_model.layers.8.mlp.down_proj.weight": 0.0006539481109939516, "model.language_model.layers.9.linear_attn.out_proj.weight": 0.0006757464143447578, "model.language_model.layers.9.linear_attn.in_proj_qkv.weight": 0.0006801060517318547, "model.language_model.layers.9.linear_attn.in_proj_z.weight": 0.00047084264224395156, "model.language_model.layers.9.linear_attn.in_proj_b.weight": 0.00022779192659072578, "model.language_model.layers.9.linear_attn.in_proj_a.weight": 0.0005929129547439516, "model.language_model.layers.9.mlp.gate_proj.weight": 0.0003749302704818547, "model.language_model.layers.9.mlp.up_proj.weight": 0.00021035331883467734, "model.language_model.layers.9.mlp.down_proj.weight": 0.0004032680008094758, "model.language_model.layers.10.linear_attn.out_proj.weight": 0.0006931850221008062, "model.language_model.layers.10.linear_attn.in_proj_qkv.weight": 0.0005471365875564516, "model.language_model.layers.10.linear_attn.in_proj_z.weight": 0.00033351354068145156, "model.language_model.layers.10.linear_attn.in_proj_b.weight": 0.00020926340948790312, "model.language_model.layers.10.linear_attn.in_proj_a.weight": 0.0002615792618598789, "model.language_model.layers.10.mlp.gate_proj.weight": 0.0003117152664344758, "model.language_model.layers.10.mlp.up_proj.weight": 0.0002768380509223789, "model.language_model.layers.10.mlp.down_proj.weight": 0.00046430318616330624, "model.language_model.layers.11.self_attn.q_proj.weight": 0.0005580357392318547, "model.language_model.layers.11.self_attn.k_proj.weight": 0.0003138951142318547, "model.language_model.layers.11.self_attn.v_proj.weight": 0.0003531319962348789, "model.language_model.layers.11.self_attn.o_proj.weight": 0.0005427769501693547, "model.language_model.layers.11.mlp.gate_proj.weight": 0.00037928990786895156, "model.language_model.layers.11.mlp.up_proj.weight": 0.0003073556290473789, "model.language_model.layers.11.mlp.down_proj.weight": 0.0005427769501693547, "model.language_model.layers.12.linear_attn.out_proj.weight": 0.0006670270813629031, "model.language_model.layers.12.linear_attn.in_proj_qkv.weight": 0.0006103515625, "model.language_model.layers.12.linear_attn.in_proj_z.weight": 0.0004032680008094758, "model.language_model.layers.12.linear_attn.in_proj_b.weight": 0.00017983574070967734, "model.language_model.layers.12.linear_attn.in_proj_a.weight": 0.00033351354068145156, "model.language_model.layers.12.mlp.gate_proj.weight": 0.0005318777984939516, "model.language_model.layers.12.mlp.up_proj.weight": 0.00025612968602217734, "model.language_model.layers.12.mlp.down_proj.weight": 0.0004207066376693547, "model.language_model.layers.13.linear_attn.out_proj.weight": 0.0007019043550826609, "model.language_model.layers.13.linear_attn.in_proj_qkv.weight": 0.0005057199159637094, "model.language_model.layers.13.linear_attn.in_proj_z.weight": 0.00040108818211592734, "model.language_model.layers.13.linear_attn.in_proj_b.weight": 0.00020708357624243945, "model.language_model.layers.13.linear_attn.in_proj_a.weight": 0.00033351354068145156, "model.language_model.layers.13.mlp.gate_proj.weight": 0.0004294259415473789, "model.language_model.layers.13.mlp.up_proj.weight": 0.0002692086563911289, "model.language_model.layers.13.mlp.down_proj.weight": 0.0005929129547439516, "model.language_model.layers.14.linear_attn.out_proj.weight": 0.0009329660097137094, "model.language_model.layers.14.linear_attn.in_proj_qkv.weight": 0.0007106236298568547, "model.language_model.layers.14.linear_attn.in_proj_z.weight": 0.00038582939305342734, "model.language_model.layers.14.linear_attn.in_proj_b.weight": 0.00024087089695967734, "model.language_model.layers.14.linear_attn.in_proj_a.weight": 0.00036403111880645156, "model.language_model.layers.14.mlp.gate_proj.weight": 0.0005929129547439516, "model.language_model.layers.14.mlp.up_proj.weight": 0.00030953544774092734, "model.language_model.layers.14.mlp.down_proj.weight": 0.0006234305328689516, "model.language_model.layers.15.self_attn.q_proj.weight": 0.0005929129547439516, "model.language_model.layers.15.self_attn.k_proj.weight": 0.0003836495743598789, "model.language_model.layers.15.self_attn.v_proj.weight": 0.0008719308534637094, "model.language_model.layers.15.self_attn.o_proj.weight": 0.00046430318616330624, "model.language_model.layers.15.mlp.gate_proj.weight": 0.0005667550722137094, "model.language_model.layers.15.mlp.up_proj.weight": 0.00040980748599395156, "model.language_model.layers.15.mlp.down_proj.weight": 0.00044904439710080624, "model.language_model.layers.16.linear_attn.out_proj.weight": 0.0007367815705947578, "model.language_model.layers.16.linear_attn.in_proj_qkv.weight": 0.0006234305328689516, "model.language_model.layers.16.linear_attn.in_proj_z.weight": 0.00035531181492842734, "model.language_model.layers.16.linear_attn.in_proj_b.weight": 0.0004512242157943547, "model.language_model.layers.16.linear_attn.in_proj_a.weight": 0.00033351354068145156, "model.language_model.layers.16.mlp.gate_proj.weight": 0.0005667550722137094, "model.language_model.layers.16.mlp.up_proj.weight": 0.00024850029149092734, "model.language_model.layers.16.mlp.down_proj.weight": 0.0007934570894576609, "model.language_model.layers.17.linear_attn.out_proj.weight": 0.0005885533173568547, "model.language_model.layers.17.linear_attn.in_proj_qkv.weight": 0.0008152553346008062, "model.language_model.layers.17.linear_attn.in_proj_z.weight": 0.0003008161438629031, "model.language_model.layers.17.linear_attn.in_proj_b.weight": 0.00025177001953125, "model.language_model.layers.17.linear_attn.in_proj_a.weight": 0.0004926409455947578, "model.language_model.layers.17.mlp.gate_proj.weight": 0.0008806501282379031, "model.language_model.layers.17.mlp.up_proj.weight": 0.0003073556290473789, "model.language_model.layers.17.mlp.down_proj.weight": 0.0005972726503387094, "model.language_model.layers.18.linear_attn.out_proj.weight": 0.0008893694612197578, "model.language_model.layers.18.linear_attn.in_proj_qkv.weight": 0.0006626674439758062, "model.language_model.layers.18.linear_attn.in_proj_z.weight": 0.0004686628235504031, "model.language_model.layers.18.linear_attn.in_proj_b.weight": 0.0003291539032943547, "model.language_model.layers.18.linear_attn.in_proj_a.weight": 0.0004228864563629031, "model.language_model.layers.18.mlp.gate_proj.weight": 0.0006147112580947578, "model.language_model.layers.18.mlp.up_proj.weight": 0.00022888185048941523, "model.language_model.layers.18.mlp.down_proj.weight": 0.0008457729127258062, "model.language_model.layers.19.self_attn.q_proj.weight": 0.00036403111880645156, "model.language_model.layers.19.self_attn.k_proj.weight": 0.00041198733379133046, "model.language_model.layers.19.self_attn.v_proj.weight": 0.001220703125, "model.language_model.layers.19.self_attn.o_proj.weight": 0.0007672991487197578, "model.language_model.layers.19.mlp.gate_proj.weight": 0.0007716587861068547, "model.language_model.layers.19.mlp.up_proj.weight": 0.0002768380509223789, "model.language_model.layers.19.mlp.down_proj.weight": 0.0008675712160766125, "model.language_model.layers.20.linear_attn.out_proj.weight": 0.0008937290986068547, "model.language_model.layers.20.linear_attn.in_proj_qkv.weight": 0.0008850098238326609, "model.language_model.layers.20.linear_attn.in_proj_z.weight": 0.0006103515625, "model.language_model.layers.20.linear_attn.in_proj_b.weight": 0.00036621096660383046, "model.language_model.layers.20.linear_attn.in_proj_a.weight": 0.0005449567688629031, "model.language_model.layers.20.mlp.gate_proj.weight": 0.0005929129547439516, "model.language_model.layers.20.mlp.up_proj.weight": 0.00032479423680342734, "model.language_model.layers.20.mlp.down_proj.weight": 0.0008283343049697578, "model.language_model.layers.21.linear_attn.out_proj.weight": 0.0007716587861068547, "model.language_model.layers.21.linear_attn.in_proj_qkv.weight": 0.0008675712160766125, "model.language_model.layers.21.linear_attn.in_proj_z.weight": 0.00047956197522580624, "model.language_model.layers.21.linear_attn.in_proj_b.weight": 0.0002692086563911289, "model.language_model.layers.21.linear_attn.in_proj_a.weight": 0.0005449567688629031, "model.language_model.layers.21.mlp.gate_proj.weight": 0.0008457729127258062, "model.language_model.layers.21.mlp.up_proj.weight": 0.0004512242157943547, "model.language_model.layers.21.mlp.down_proj.weight": 0.0006495884736068547, "model.language_model.layers.22.linear_attn.out_proj.weight": 0.0012468610657379031, "model.language_model.layers.22.linear_attn.in_proj_qkv.weight": 0.0009678432252258062, "model.language_model.layers.22.linear_attn.in_proj_z.weight": 0.0005667550722137094, "model.language_model.layers.22.linear_attn.in_proj_b.weight": 0.0004359654267318547, "model.language_model.layers.22.linear_attn.in_proj_a.weight": 0.0005623953766189516, "model.language_model.layers.22.mlp.gate_proj.weight": 0.0008893694612197578, "model.language_model.layers.22.mlp.up_proj.weight": 0.0003422328445594758, "model.language_model.layers.22.mlp.down_proj.weight": 0.0008588518830947578, "model.language_model.layers.23.self_attn.q_proj.weight": 0.0006757464143447578, "model.language_model.layers.23.self_attn.k_proj.weight": 0.00044250491191633046, "model.language_model.layers.23.self_attn.v_proj.weight": 0.0008196149719879031, "model.language_model.layers.23.self_attn.o_proj.weight": 0.0008283343049697578, "model.language_model.layers.23.mlp.gate_proj.weight": 0.0009547642548568547, "model.language_model.layers.23.mlp.up_proj.weight": 0.0005667550722137094, "model.language_model.layers.23.mlp.down_proj.weight": 0.0020490374881774187, "lm_head.weight": 0.00044032506411895156}
model.fp8.safetensors ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:750b5b58f160fb405ead4fb21df9774a492aec5a4367c0d7484e8db2a0b7ade2
3
+ size 3233793676
modeling_st_wemm.py ADDED
@@ -0,0 +1,70 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """Sentence Transformers module for WeMM-Embedding.
2
+
3
+ Reproduces the `transformers` usage from the model card inside a Sentence Transformers
4
+ pipeline: vision inputs are prepared with `qwen_vl_utils.process_vision_info` and the
5
+ embedding is read from `WeMMEmbedding.embedding`, which pools the `<embedding>` position and
6
+ L2-normalizes. Everything else (batching, prompts, truncation, `encode_query` /
7
+ `encode_document`, similarity) comes from the stock `Transformer` module.
8
+ """
9
+
10
+ from __future__ import annotations
11
+
12
+ import inspect
13
+ from typing import Any
14
+
15
+ from sentence_transformers.models import Transformer
16
+
17
+
18
+ class WeMMTransformer(Transformer):
19
+ """`Transformer` that prepares images and videos the way the model card's snippet does."""
20
+
21
+ def __init__(self, model_name_or_path: str, **kwargs: Any) -> None:
22
+ super().__init__(model_name_or_path, **kwargs)
23
+ vision_config = getattr(self.config, "vision_config", None)
24
+ self.image_patch_size = int(getattr(vision_config, "patch_size", 16))
25
+
26
+ # `embedding` hands its **kwargs to the inner model, so filtering on its own signature
27
+ # would drop `pixel_values`. Filter on the inner model's parameters instead, plus the
28
+ # processor's input names for anything the model only accepts as **kwargs.
29
+ inner_model = getattr(self.model, "model", self.model)
30
+ signature = set(inspect.signature(inner_model.forward).parameters)
31
+ signature |= set(getattr(self.processor, "model_input_names", ()))
32
+ for modality_params in self.modality_config.values():
33
+ method_name = modality_params["method"]
34
+ if method_name != "forward":
35
+ self._method_signature_cache.setdefault(method_name, signature)
36
+
37
+ def _apply_chat_template(
38
+ self,
39
+ messages: list[list[dict[str, Any]]],
40
+ modality_kwargs: dict[str, dict[str, Any]],
41
+ common_kwargs: dict[str, Any],
42
+ chat_template_kwargs: dict[str, Any],
43
+ ) -> dict[str, Any]:
44
+ """Render the chat template and prepare images / videos exactly as the model card does."""
45
+ from qwen_vl_utils import process_vision_info
46
+
47
+ chat_template_kwargs = {"add_generation_prompt": False, **chat_template_kwargs}
48
+ texts = [
49
+ self.processor.apply_chat_template(conversation, tokenize=False, **chat_template_kwargs)
50
+ for conversation in messages
51
+ ]
52
+ images, videos, video_kwargs = process_vision_info(
53
+ [list(conversation) for conversation in messages],
54
+ image_patch_size=self.image_patch_size,
55
+ return_video_kwargs=True,
56
+ return_video_metadata=True,
57
+ )
58
+ if videos is not None:
59
+ videos, video_metadata = (list(part) for part in zip(*videos))
60
+ video_kwargs = {**video_kwargs, "video_metadata": video_metadata}
61
+
62
+ return self.processor(
63
+ text=texts,
64
+ images=images,
65
+ videos=videos,
66
+ text_kwargs=modality_kwargs["text"],
67
+ images_kwargs=modality_kwargs["image"],
68
+ videos_kwargs={**modality_kwargs["video"], **video_kwargs},
69
+ common_kwargs=common_kwargs,
70
+ )
modeling_wemm_embedding.py ADDED
@@ -0,0 +1,33 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import torch
2
+ import torch.nn.functional as F
3
+ from transformers import Qwen3_5ForConditionalGeneration
4
+
5
+
6
+ class WeMMEmbedding(Qwen3_5ForConditionalGeneration):
7
+ def embedding(self, input_ids=None, attention_mask=None, **kwargs):
8
+ # transformers < 5.15 reuses the rope_deltas cached by the previous multimodal
9
+ # forward for a text-only one, which shifts its position ids.
10
+ self.model.rope_deltas = None
11
+
12
+ outputs = self.model(
13
+ input_ids=input_ids,
14
+ attention_mask=attention_mask,
15
+ **kwargs
16
+ )
17
+
18
+ last_hidden_state = outputs.last_hidden_state
19
+
20
+ if attention_mask is not None:
21
+ eos_positions = attention_mask.sum(dim=1) - 1
22
+ else:
23
+ eos_positions = torch.full((last_hidden_state.shape[0],), last_hidden_state.shape[1] - 1, device=last_hidden_state.device)
24
+
25
+ eos_positions = eos_positions.clamp(min=0)
26
+
27
+ batch_indices = torch.arange(last_hidden_state.size(0), device=last_hidden_state.device)
28
+
29
+ embeddings = last_hidden_state[batch_indices, eos_positions]
30
+
31
+ embeddings = F.normalize(embeddings, dim=-1)
32
+
33
+ return embeddings
modules.json ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ [
2
+ {
3
+ "idx": 0,
4
+ "name": "0",
5
+ "path": "",
6
+ "type": "modeling_st_wemm.WeMMTransformer"
7
+ }
8
+ ]
patch_sglang_video.py ADDED
@@ -0,0 +1,73 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ #!/usr/bin/env python3
2
+ """Align SGLang 0.5.9 video preprocessing with WeMM-Embedding."""
3
+
4
+ import importlib.metadata
5
+ import py_compile
6
+ import shutil
7
+ from pathlib import Path
8
+
9
+ import sglang
10
+
11
+
12
+ EXPECTED_VERSION = "0.5.9"
13
+ REPLACEMENTS = (
14
+ (
15
+ "IMAGE_FACTOR = 28",
16
+ "IMAGE_FACTOR = 32 # patch_size=16 * spatial_merge_size=2",
17
+ ),
18
+ (
19
+ " idx = np.linspace(0, total_frames - 1, num=nframes, dtype=np.int64)",
20
+ " idx = torch.linspace(0, total_frames - 1, nframes).round().long().cpu().numpy()",
21
+ ),
22
+ (
23
+ """ video = torchvision.transforms.functional.resize(
24
+ video,
25
+ [resized_height, resized_width],
26
+ interpolation=InterpolationMode.BILINEAR,
27
+ )
28
+ """,
29
+ """ video = torchvision.transforms.functional.resize(
30
+ video,
31
+ [resized_height, resized_width],
32
+ interpolation=InterpolationMode.BICUBIC,
33
+ antialias=True,
34
+ ).float()
35
+ """,
36
+ ),
37
+ )
38
+
39
+
40
+ def main() -> None:
41
+ version = importlib.metadata.version("sglang")
42
+ if version != EXPECTED_VERSION:
43
+ raise RuntimeError(f"Expected sglang=={EXPECTED_VERSION}, found {version}")
44
+
45
+ path = (
46
+ Path(sglang.__file__).resolve().parent
47
+ / "srt"
48
+ / "multimodal"
49
+ / "processors"
50
+ / "qwen_vl.py"
51
+ )
52
+ text = path.read_text(encoding="utf-8")
53
+ if all(new in text for _, new in REPLACEMENTS):
54
+ print("SGLang video preprocessing is ready.")
55
+ return
56
+
57
+ updated = text
58
+ for old, new in REPLACEMENTS:
59
+ if updated.count(old) != 1:
60
+ raise RuntimeError(f"Unexpected SGLang source: {old.splitlines()[0]}")
61
+ updated = updated.replace(old, new, 1)
62
+
63
+ backup = path.with_suffix(path.suffix + ".original")
64
+ if not backup.exists():
65
+ shutil.copy2(path, backup)
66
+ path.write_text(updated, encoding="utf-8")
67
+ py_compile.compile(str(path), doraise=True)
68
+ print("SGLang video preprocessing is ready.")
69
+
70
+
71
+ if __name__ == "__main__":
72
+ main()
73
+
processor_config.json ADDED
@@ -0,0 +1,60 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "image_processor": {
3
+ "do_convert_rgb": true,
4
+ "do_normalize": true,
5
+ "do_rescale": true,
6
+ "do_resize": true,
7
+ "image_mean": [
8
+ 0.5,
9
+ 0.5,
10
+ 0.5
11
+ ],
12
+ "image_processor_type": "Qwen2VLImageProcessor",
13
+ "image_std": [
14
+ 0.5,
15
+ 0.5,
16
+ 0.5
17
+ ],
18
+ "merge_size": 2,
19
+ "patch_size": 16,
20
+ "resample": 3,
21
+ "rescale_factor": 0.00392156862745098,
22
+ "size": {
23
+ "longest_edge": 16777216,
24
+ "shortest_edge": 65536
25
+ },
26
+ "temporal_patch_size": 2
27
+ },
28
+ "processor_class": "Qwen3VLProcessor",
29
+ "video_processor": {
30
+ "do_convert_rgb": true,
31
+ "do_normalize": true,
32
+ "do_rescale": true,
33
+ "do_resize": true,
34
+ "do_sample_frames": true,
35
+ "fps": 2,
36
+ "image_mean": [
37
+ 0.5,
38
+ 0.5,
39
+ 0.5
40
+ ],
41
+ "image_std": [
42
+ 0.5,
43
+ 0.5,
44
+ 0.5
45
+ ],
46
+ "max_frames": 768,
47
+ "merge_size": 2,
48
+ "min_frames": 4,
49
+ "patch_size": 16,
50
+ "resample": 3,
51
+ "rescale_factor": 0.00392156862745098,
52
+ "return_metadata": false,
53
+ "size": {
54
+ "longest_edge": 234881024,
55
+ "shortest_edge": 4096
56
+ },
57
+ "temporal_patch_size": 2,
58
+ "video_processor_type": "Qwen3VLVideoProcessor"
59
+ }
60
+ }
sentence_bert_config.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "transformer_task": "feature-extraction",
3
+ "modality_config": {
4
+ "text": {
5
+ "method": "embedding",
6
+ "method_output_name": null
7
+ },
8
+ "image": {
9
+ "method": "embedding",
10
+ "method_output_name": null
11
+ },
12
+ "video": {
13
+ "method": "embedding",
14
+ "method_output_name": null
15
+ },
16
+ "image+text": {
17
+ "method": "embedding",
18
+ "method_output_name": null
19
+ },
20
+ "text+video": {
21
+ "method": "embedding",
22
+ "method_output_name": null
23
+ },
24
+ "message": {
25
+ "method": "embedding",
26
+ "method_output_name": null,
27
+ "format": "structured"
28
+ }
29
+ },
30
+ "module_output_name": "sentence_embedding",
31
+ "processing_kwargs": {
32
+ "chat_template": {
33
+ "chat_template": "sentence_transformers"
34
+ }
35
+ },
36
+ "unpad_inputs": false
37
+ }
tokenizer.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:40e444c744512f423da4c8443c47c21e22ff76056ba4e9796a81c04c13a9daf0
3
+ size 19990378
tokenizer_config.json ADDED
@@ -0,0 +1,32 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "add_prefix_space": false,
3
+ "audio_bos_token": "<|audio_start|>",
4
+ "audio_eos_token": "<|audio_end|>",
5
+ "audio_token": "<|audio_pad|>",
6
+ "backend": "tokenizers",
7
+ "bos_token": null,
8
+ "clean_up_tokenization_spaces": false,
9
+ "eos_token": "<|endoftext|>",
10
+ "errors": "replace",
11
+ "image_token": "<|image_pad|>",
12
+ "is_local": true,
13
+ "model_max_length": 262144,
14
+ "model_specific_special_tokens": {
15
+ "audio_bos_token": "<|audio_start|>",
16
+ "audio_eos_token": "<|audio_end|>",
17
+ "audio_token": "<|audio_pad|>",
18
+ "image_token": "<|image_pad|>",
19
+ "video_token": "<|video_pad|>",
20
+ "vision_bos_token": "<|vision_start|>",
21
+ "vision_eos_token": "<|vision_end|>"
22
+ },
23
+ "pad_token": "<|endoftext|>",
24
+ "pretokenize_regex": "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?[\\p{L}\\p{M}]+|\\p{N}| ?[^\\s\\p{L}\\p{M}\\p{N}]+[\\r\\n]*|\\s*[\\r\\n]+|\\s+(?!\\S)|\\s+",
25
+ "processor_class": "Qwen3VLProcessor",
26
+ "split_special_tokens": false,
27
+ "tokenizer_class": "PreTrainedTokenizerFast",
28
+ "unk_token": null,
29
+ "video_token": "<|video_pad|>",
30
+ "vision_bos_token": "<|vision_start|>",
31
+ "vision_eos_token": "<|vision_end|>"
32
+ }