Add files using upload-large-folder tool

LICENSE

@@ -0,0 +1,202 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright (c) 2025 Baidu, Inc. All Rights Reserved.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.
+

added_tokens.json

@@ -0,0 +1 @@
+{"<|IMAGE_PLACEHOLDER|>": 100295, "<|AUDIO_PLACEHOLDER|>": 100296, "<|LOC_0|>": 100297, "<|LOC_1|>": 100298, "<|LOC_2|>": 100299, "<|LOC_3|>": 100300, "<|LOC_4|>": 100301, "<|LOC_5|>": 100302, "<|LOC_6|>": 100303, "<|LOC_7|>": 100304, "<|LOC_8|>": 100305, "<|LOC_9|>": 100306, "<|LOC_10|>": 100307, "<|LOC_11|>": 100308, "<|LOC_12|>": 100309, "<|LOC_13|>": 100310, "<|LOC_14|>": 100311, "<|LOC_15|>": 100312, "<|LOC_16|>": 100313, "<|LOC_17|>": 100314, "<|LOC_18|>": 100315, "<|LOC_19|>": 100316, "<|LOC_20|>": 100317, "<|LOC_21|>": 100318, "<|LOC_22|>": 100319, "<|LOC_23|>": 100320, "<|LOC_24|>": 100321, "<|LOC_25|>": 100322, "<|LOC_26|>": 100323, "<|LOC_27|>": 100324, "<|LOC_28|>": 100325, "<|LOC_29|>": 100326, "<|LOC_30|>": 100327, "<|LOC_31|>": 100328, "<|LOC_32|>": 100329, "<|LOC_33|>": 100330, "<|LOC_34|>": 100331, "<|LOC_35|>": 100332, "<|LOC_36|>": 100333, "<|LOC_37|>": 100334, "<|LOC_38|>": 100335, "<|LOC_39|>": 100336, "<|LOC_40|>": 100337, "<|LOC_41|>": 100338, "<|LOC_42|>": 100339, "<|LOC_43|>": 100340, "<|LOC_44|>": 100341, "<|LOC_45|>": 100342, "<|LOC_46|>": 100343, "<|LOC_47|>": 100344, "<|LOC_48|>": 100345, "<|LOC_49|>": 100346, "<|LOC_50|>": 100347, "<|LOC_51|>": 100348, "<|LOC_52|>": 100349, "<|LOC_53|>": 100350, "<|LOC_54|>": 100351, "<|LOC_55|>": 100352, "<|LOC_56|>": 100353, "<|LOC_57|>": 100354, "<|LOC_58|>": 100355, "<|LOC_59|>": 100356, "<|LOC_60|>": 100357, "<|LOC_61|>": 100358, "<|LOC_62|>": 100359, "<|LOC_63|>": 100360, "<|LOC_64|>": 100361, "<|LOC_65|>": 100362, "<|LOC_66|>": 100363, "<|LOC_67|>": 100364, "<|LOC_68|>": 100365, "<|LOC_69|>": 100366, "<|LOC_70|>": 100367, "<|LOC_71|>": 100368, "<|LOC_72|>": 100369, "<|LOC_73|>": 100370, "<|LOC_74|>": 100371, "<|LOC_75|>": 100372, "<|LOC_76|>": 100373, "<|LOC_77|>": 100374, "<|LOC_78|>": 100375, "<|LOC_79|>": 100376, "<|LOC_80|>": 100377, "<|LOC_81|>": 100378, "<|LOC_82|>": 100379, "<|LOC_83|>": 100380, "<|LOC_84|>": 100381, "<|LOC_85|>": 100382, "<|LOC_86|>": 100383, "<|LOC_87|>": 100384, "<|LOC_88|>": 100385, "<|LOC_89|>": 100386, "<|LOC_90|>": 100387, "<|LOC_91|>": 100388, "<|LOC_92|>": 100389, "<|LOC_93|>": 100390, "<|LOC_94|>": 100391, "<|LOC_95|>": 100392, "<|LOC_96|>": 100393, "<|LOC_97|>": 100394, "<|LOC_98|>": 100395, "<|LOC_99|>": 100396, "<|LOC_100|>": 100397, "<|LOC_101|>": 100398, "<|LOC_102|>": 100399, "<|LOC_103|>": 100400, "<|LOC_104|>": 100401, "<|LOC_105|>": 100402, "<|LOC_106|>": 100403, "<|LOC_107|>": 100404, "<|LOC_108|>": 100405, "<|LOC_109|>": 100406, "<|LOC_110|>": 100407, "<|LOC_111|>": 100408, "<|LOC_112|>": 100409, "<|LOC_113|>": 100410, "<|LOC_114|>": 100411, "<|LOC_115|>": 100412, "<|LOC_116|>": 100413, "<|LOC_117|>": 100414, "<|LOC_118|>": 100415, "<|LOC_119|>": 100416, "<|LOC_120|>": 100417, "<|LOC_121|>": 100418, "<|LOC_122|>": 100419, "<|LOC_123|>": 100420, "<|LOC_124|>": 100421, "<|LOC_125|>": 100422, "<|LOC_126|>": 100423, "<|LOC_127|>": 100424, "<|LOC_128|>": 100425, "<|LOC_129|>": 100426, "<|LOC_130|>": 100427, "<|LOC_131|>": 100428, "<|LOC_132|>": 100429, "<|LOC_133|>": 100430, "<|LOC_134|>": 100431, "<|LOC_135|>": 100432, "<|LOC_136|>": 100433, "<|LOC_137|>": 100434, "<|LOC_138|>": 100435, "<|LOC_139|>": 100436, "<|LOC_140|>": 100437, "<|LOC_141|>": 100438, "<|LOC_142|>": 100439, "<|LOC_143|>": 100440, "<|LOC_144|>": 100441, "<|LOC_145|>": 100442, "<|LOC_146|>": 100443, "<|LOC_147|>": 100444, "<|LOC_148|>": 100445, "<|LOC_149|>": 100446, "<|LOC_150|>": 100447, "<|LOC_151|>": 100448, "<|LOC_152|>": 100449, "<|LOC_153|>": 100450, "<|LOC_154|>": 100451, "<|LOC_155|>": 100452, "<|LOC_156|>": 100453, "<|LOC_157|>": 100454, "<|LOC_158|>": 100455, "<|LOC_159|>": 100456, "<|LOC_160|>": 100457, "<|LOC_161|>": 100458, "<|LOC_162|>": 100459, "<|LOC_163|>": 100460, "<|LOC_164|>": 100461, "<|LOC_165|>": 100462, "<|LOC_166|>": 100463, "<|LOC_167|>": 100464, "<|LOC_168|>": 100465, "<|LOC_169|>": 100466, "<|LOC_170|>": 100467, "<|LOC_171|>": 100468, "<|LOC_172|>": 100469, "<|LOC_173|>": 100470, "<|LOC_174|>": 100471, "<|LOC_175|>": 100472, "<|LOC_176|>": 100473, "<|LOC_177|>": 100474, "<|LOC_178|>": 100475, "<|LOC_179|>": 100476, "<|LOC_180|>": 100477, "<|LOC_181|>": 100478, "<|LOC_182|>": 100479, "<|LOC_183|>": 100480, "<|LOC_184|>": 100481, "<|LOC_185|>": 100482, "<|LOC_186|>": 100483, "<|LOC_187|>": 100484, "<|LOC_188|>": 100485, "<|LOC_189|>": 100486, "<|LOC_190|>": 100487, "<|LOC_191|>": 100488, "<|LOC_192|>": 100489, "<|LOC_193|>": 100490, "<|LOC_194|>": 100491, "<|LOC_195|>": 100492, "<|LOC_196|>": 100493, "<|LOC_197|>": 100494, "<|LOC_198|>": 100495, "<|LOC_199|>": 100496, "<|LOC_200|>": 100497, "<|LOC_201|>": 100498, "<|LOC_202|>": 100499, "<|LOC_203|>": 100500, "<|LOC_204|>": 100501, "<|LOC_205|>": 100502, "<|LOC_206|>": 100503, "<|LOC_207|>": 100504, "<|LOC_208|>": 100505, "<|LOC_209|>": 100506, "<|LOC_210|>": 100507, "<|LOC_211|>": 100508, "<|LOC_212|>": 100509, "<|LOC_213|>": 100510, "<|LOC_214|>": 100511, "<|LOC_215|>": 100512, "<|LOC_216|>": 100513, "<|LOC_217|>": 100514, "<|LOC_218|>": 100515, "<|LOC_219|>": 100516, "<|LOC_220|>": 100517, "<|LOC_221|>": 100518, "<|LOC_222|>": 100519, "<|LOC_223|>": 100520, "<|LOC_224|>": 100521, "<|LOC_225|>": 100522, "<|LOC_226|>": 100523, "<|LOC_227|>": 100524, "<|LOC_228|>": 100525, "<|LOC_229|>": 100526, "<|LOC_230|>": 100527, "<|LOC_231|>": 100528, "<|LOC_232|>": 100529, "<|LOC_233|>": 100530, "<|LOC_234|>": 100531, "<|LOC_235|>": 100532, "<|LOC_236|>": 100533, "<|LOC_237|>": 100534, "<|LOC_238|>": 100535, "<|LOC_239|>": 100536, "<|LOC_240|>": 100537, "<|LOC_241|>": 100538, "<|LOC_242|>": 100539, "<|LOC_243|>": 100540, "<|LOC_244|>": 100541, "<|LOC_245|>": 100542, "<|LOC_246|>": 100543, "<|LOC_247|>": 100544, "<|LOC_248|>": 100545, "<|LOC_249|>": 100546, "<|LOC_250|>": 100547, "<|LOC_251|>": 100548, "<|LOC_252|>": 100549, "<|LOC_253|>": 100550, "<|LOC_254|>": 100551, "<|LOC_255|>": 100552, "<|LOC_256|>": 100553, "<|LOC_257|>": 100554, "<|LOC_258|>": 100555, "<|LOC_259|>": 100556, "<|LOC_260|>": 100557, "<|LOC_261|>": 100558, "<|LOC_262|>": 100559, "<|LOC_263|>": 100560, "<|LOC_264|>": 100561, "<|LOC_265|>": 100562, "<|LOC_266|>": 100563, "<|LOC_267|>": 100564, "<|LOC_268|>": 100565, "<|LOC_269|>": 100566, "<|LOC_270|>": 100567, "<|LOC_271|>": 100568, "<|LOC_272|>": 100569, "<|LOC_273|>": 100570, "<|LOC_274|>": 100571, "<|LOC_275|>": 100572, "<|LOC_276|>": 100573, "<|LOC_277|>": 100574, "<|LOC_278|>": 100575, "<|LOC_279|>": 100576, "<|LOC_280|>": 100577, "<|LOC_281|>": 100578, "<|LOC_282|>": 100579, "<|LOC_283|>": 100580, "<|LOC_284|>": 100581, "<|LOC_285|>": 100582, "<|LOC_286|>": 100583, "<|LOC_287|>": 100584, "<|LOC_288|>": 100585, "<|LOC_289|>": 100586, "<|LOC_290|>": 100587, "<|LOC_291|>": 100588, "<|LOC_292|>": 100589, "<|LOC_293|>": 100590, "<|LOC_294|>": 100591, "<|LOC_295|>": 100592, "<|LOC_296|>": 100593, "<|LOC_297|>": 100594, "<|LOC_298|>": 100595, "<|LOC_299|>": 100596, "<|LOC_300|>": 100597, "<|LOC_301|>": 100598, "<|LOC_302|>": 100599, "<|LOC_303|>": 100600, "<|LOC_304|>": 100601, "<|LOC_305|>": 100602, "<|LOC_306|>": 100603, "<|LOC_307|>": 100604, "<|LOC_308|>": 100605, "<|LOC_309|>": 100606, "<|LOC_310|>": 100607, "<|LOC_311|>": 100608, "<|LOC_312|>": 100609, "<|LOC_313|>": 100610, "<|LOC_314|>": 100611, "<|LOC_315|>": 100612, "<|LOC_316|>": 100613, "<|LOC_317|>": 100614, "<|LOC_318|>": 100615, "<|LOC_319|>": 100616, "<|LOC_320|>": 100617, "<|LOC_321|>": 100618, "<|LOC_322|>": 100619, "<|LOC_323|>": 100620, "<|LOC_324|>": 100621, "<|LOC_325|>": 100622, "<|LOC_326|>": 100623, "<|LOC_327|>": 100624, "<|LOC_328|>": 100625, "<|LOC_329|>": 100626, "<|LOC_330|>": 100627, "<|LOC_331|>": 100628, "<|LOC_332|>": 100629, "<|LOC_333|>": 100630, "<|LOC_334|>": 100631, "<|LOC_335|>": 100632, "<|LOC_336|>": 100633, "<|LOC_337|>": 100634, "<|LOC_338|>": 100635, "<|LOC_339|>": 100636, "<|LOC_340|>": 100637, "<|LOC_341|>": 100638, "<|LOC_342|>": 100639, "<|LOC_343|>": 100640, "<|LOC_344|>": 100641, "<|LOC_345|>": 100642, "<|LOC_346|>": 100643, "<|LOC_347|>": 100644, "<|LOC_348|>": 100645, "<|LOC_349|>": 100646, "<|LOC_350|>": 100647, "<|LOC_351|>": 100648, "<|LOC_352|>": 100649, "<|LOC_353|>": 100650, "<|LOC_354|>": 100651, "<|LOC_355|>": 100652, "<|LOC_356|>": 100653, "<|LOC_357|>": 100654, "<|LOC_358|>": 100655, "<|LOC_359|>": 100656, "<|LOC_360|>": 100657, "<|LOC_361|>": 100658, "<|LOC_362|>": 100659, "<|LOC_363|>": 100660, "<|LOC_364|>": 100661, "<|LOC_365|>": 100662, "<|LOC_366|>": 100663, "<|LOC_367|>": 100664, "<|LOC_368|>": 100665, "<|LOC_369|>": 100666, "<|LOC_370|>": 100667, "<|LOC_371|>": 100668, "<|LOC_372|>": 100669, "<|LOC_373|>": 100670, "<|LOC_374|>": 100671, "<|LOC_375|>": 100672, "<|LOC_376|>": 100673, "<|LOC_377|>": 100674, "<|LOC_378|>": 100675, "<|LOC_379|>": 100676, "<|LOC_380|>": 100677, "<|LOC_381|>": 100678, "<|LOC_382|>": 100679, "<|LOC_383|>": 100680, "<|LOC_384|>": 100681, "<|LOC_385|>": 100682, "<|LOC_386|>": 100683, "<|LOC_387|>": 100684, "<|LOC_388|>": 100685, "<|LOC_389|>": 100686, "<|LOC_390|>": 100687, "<|LOC_391|>": 100688, "<|LOC_392|>": 100689, "<|LOC_393|>": 100690, "<|LOC_394|>": 100691, "<|LOC_395|>": 100692, "<|LOC_396|>": 100693, "<|LOC_397|>": 100694, "<|LOC_398|>": 100695, "<|LOC_399|>": 100696, "<|LOC_400|>": 100697, "<|LOC_401|>": 100698, "<|LOC_402|>": 100699, "<|LOC_403|>": 100700, "<|LOC_404|>": 100701, "<|LOC_405|>": 100702, "<|LOC_406|>": 100703, "<|LOC_407|>": 100704, "<|LOC_408|>": 100705, "<|LOC_409|>": 100706, "<|LOC_410|>": 100707, "<|LOC_411|>": 100708, "<|LOC_412|>": 100709, "<|LOC_413|>": 100710, "<|LOC_414|>": 100711, "<|LOC_415|>": 100712, "<|LOC_416|>": 100713, "<|LOC_417|>": 100714, "<|LOC_418|>": 100715, "<|LOC_419|>": 100716, "<|LOC_420|>": 100717, "<|LOC_421|>": 100718, "<|LOC_422|>": 100719, "<|LOC_423|>": 100720, "<|LOC_424|>": 100721, "<|LOC_425|>": 100722, "<|LOC_426|>": 100723, "<|LOC_427|>": 100724, "<|LOC_428|>": 100725, "<|LOC_429|>": 100726, "<|LOC_430|>": 100727, "<|LOC_431|>": 100728, "<|LOC_432|>": 100729, "<|LOC_433|>": 100730, "<|LOC_434|>": 100731, "<|LOC_435|>": 100732, "<|LOC_436|>": 100733, "<|LOC_437|>": 100734, "<|LOC_438|>": 100735, "<|LOC_439|>": 100736, "<|LOC_440|>": 100737, "<|LOC_441|>": 100738, "<|LOC_442|>": 100739, "<|LOC_443|>": 100740, "<|LOC_444|>": 100741, "<|LOC_445|>": 100742, "<|LOC_446|>": 100743, "<|LOC_447|>": 100744, "<|LOC_448|>": 100745, "<|LOC_449|>": 100746, "<|LOC_450|>": 100747, "<|LOC_451|>": 100748, "<|LOC_452|>": 100749, "<|LOC_453|>": 100750, "<|LOC_454|>": 100751, "<|LOC_455|>": 100752, "<|LOC_456|>": 100753, "<|LOC_457|>": 100754, "<|LOC_458|>": 100755, "<|LOC_459|>": 100756, "<|LOC_460|>": 100757, "<|LOC_461|>": 100758, "<|LOC_462|>": 100759, "<|LOC_463|>": 100760, "<|LOC_464|>": 100761, "<|LOC_465|>": 100762, "<|LOC_466|>": 100763, "<|LOC_467|>": 100764, "<|LOC_468|>": 100765, "<|LOC_469|>": 100766, "<|LOC_470|>": 100767, "<|LOC_471|>": 100768, "<|LOC_472|>": 100769, "<|LOC_473|>": 100770, "<|LOC_474|>": 100771, "<|LOC_475|>": 100772, "<|LOC_476|>": 100773, "<|LOC_477|>": 100774, "<|LOC_478|>": 100775, "<|LOC_479|>": 100776, "<|LOC_480|>": 100777, "<|LOC_481|>": 100778, "<|LOC_482|>": 100779, "<|LOC_483|>": 100780, "<|LOC_484|>": 100781, "<|LOC_485|>": 100782, "<|LOC_486|>": 100783, "<|LOC_487|>": 100784, "<|LOC_488|>": 100785, "<|LOC_489|>": 100786, "<|LOC_490|>": 100787, "<|LOC_491|>": 100788, "<|LOC_492|>": 100789, "<|LOC_493|>": 100790, "<|LOC_494|>": 100791, "<|LOC_495|>": 100792, "<|LOC_496|>": 100793, "<|LOC_497|>": 100794, "<|LOC_498|>": 100795, "<|LOC_499|>": 100796, "<|LOC_500|>": 100797, "<|LOC_501|>": 100798, "<|LOC_502|>": 100799, "<|LOC_503|>": 100800, "<|LOC_504|>": 100801, "<|LOC_505|>": 100802, "<|LOC_506|>": 100803, "<|LOC_507|>": 100804, "<|LOC_508|>": 100805, "<|LOC_509|>": 100806, "<|LOC_510|>": 100807, "<|LOC_511|>": 100808, "<|LOC_512|>": 100809, "<|LOC_513|>": 100810, "<|LOC_514|>": 100811, "<|LOC_515|>": 100812, "<|LOC_516|>": 100813, "<|LOC_517|>": 100814, "<|LOC_518|>": 100815, "<|LOC_519|>": 100816, "<|LOC_520|>": 100817, "<|LOC_521|>": 100818, "<|LOC_522|>": 100819, "<|LOC_523|>": 100820, "<|LOC_524|>": 100821, "<|LOC_525|>": 100822, "<|LOC_526|>": 100823, "<|LOC_527|>": 100824, "<|LOC_528|>": 100825, "<|LOC_529|>": 100826, "<|LOC_530|>": 100827, "<|LOC_531|>": 100828, "<|LOC_532|>": 100829, "<|LOC_533|>": 100830, "<|LOC_534|>": 100831, "<|LOC_535|>": 100832, "<|LOC_536|>": 100833, "<|LOC_537|>": 100834, "<|LOC_538|>": 100835, "<|LOC_539|>": 100836, "<|LOC_540|>": 100837, "<|LOC_541|>": 100838, "<|LOC_542|>": 100839, "<|LOC_543|>": 100840, "<|LOC_544|>": 100841, "<|LOC_545|>": 100842, "<|LOC_546|>": 100843, "<|LOC_547|>": 100844, "<|LOC_548|>": 100845, "<|LOC_549|>": 100846, "<|LOC_550|>": 100847, "<|LOC_551|>": 100848, "<|LOC_552|>": 100849, "<|LOC_553|>": 100850, "<|LOC_554|>": 100851, "<|LOC_555|>": 100852, "<|LOC_556|>": 100853, "<|LOC_557|>": 100854, "<|LOC_558|>": 100855, "<|LOC_559|>": 100856, "<|LOC_560|>": 100857, "<|LOC_561|>": 100858, "<|LOC_562|>": 100859, "<|LOC_563|>": 100860, "<|LOC_564|>": 100861, "<|LOC_565|>": 100862, "<|LOC_566|>": 100863, "<|LOC_567|>": 100864, "<|LOC_568|>": 100865, "<|LOC_569|>": 100866, "<|LOC_570|>": 100867, "<|LOC_571|>": 100868, "<|LOC_572|>": 100869, "<|LOC_573|>": 100870, "<|LOC_574|>": 100871, "<|LOC_575|>": 100872, "<|LOC_576|>": 100873, "<|LOC_577|>": 100874, "<|LOC_578|>": 100875, "<|LOC_579|>": 100876, "<|LOC_580|>": 100877, "<|LOC_581|>": 100878, "<|LOC_582|>": 100879, "<|LOC_583|>": 100880, "<|LOC_584|>": 100881, "<|LOC_585|>": 100882, "<|LOC_586|>": 100883, "<|LOC_587|>": 100884, "<|LOC_588|>": 100885, "<|LOC_589|>": 100886, "<|LOC_590|>": 100887, "<|LOC_591|>": 100888, "<|LOC_592|>": 100889, "<|LOC_593|>": 100890, "<|LOC_594|>": 100891, "<|LOC_595|>": 100892, "<|LOC_596|>": 100893, "<|LOC_597|>": 100894, "<|LOC_598|>": 100895, "<|LOC_599|>": 100896, "<|LOC_600|>": 100897, "<|LOC_601|>": 100898, "<|LOC_602|>": 100899, "<|LOC_603|>": 100900, "<|LOC_604|>": 100901, "<|LOC_605|>": 100902, "<|LOC_606|>": 100903, "<|LOC_607|>": 100904, "<|LOC_608|>": 100905, "<|LOC_609|>": 100906, "<|LOC_610|>": 100907, "<|LOC_611|>": 100908, "<|LOC_612|>": 100909, "<|LOC_613|>": 100910, "<|LOC_614|>": 100911, "<|LOC_615|>": 100912, "<|LOC_616|>": 100913, "<|LOC_617|>": 100914, "<|LOC_618|>": 100915, "<|LOC_619|>": 100916, "<|LOC_620|>": 100917, "<|LOC_621|>": 100918, "<|LOC_622|>": 100919, "<|LOC_623|>": 100920, "<|LOC_624|>": 100921, "<|LOC_625|>": 100922, "<|LOC_626|>": 100923, "<|LOC_627|>": 100924, "<|LOC_628|>": 100925, "<|LOC_629|>": 100926, "<|LOC_630|>": 100927, "<|LOC_631|>": 100928, "<|LOC_632|>": 100929, "<|LOC_633|>": 100930, "<|LOC_634|>": 100931, "<|LOC_635|>": 100932, "<|LOC_636|>": 100933, "<|LOC_637|>": 100934, "<|LOC_638|>": 100935, "<|LOC_639|>": 100936, "<|LOC_640|>": 100937, "<|LOC_641|>": 100938, "<|LOC_642|>": 100939, "<|LOC_643|>": 100940, "<|LOC_644|>": 100941, "<|LOC_645|>": 100942, "<|LOC_646|>": 100943, "<|LOC_647|>": 100944, "<|LOC_648|>": 100945, "<|LOC_649|>": 100946, "<|LOC_650|>": 100947, "<|LOC_651|>": 100948, "<|LOC_652|>": 100949, "<|LOC_653|>": 100950, "<|LOC_654|>": 100951, "<|LOC_655|>": 100952, "<|LOC_656|>": 100953, "<|LOC_657|>": 100954, "<|LOC_658|>": 100955, "<|LOC_659|>": 100956, "<|LOC_660|>": 100957, "<|LOC_661|>": 100958, "<|LOC_662|>": 100959, "<|LOC_663|>": 100960, "<|LOC_664|>": 100961, "<|LOC_665|>": 100962, "<|LOC_666|>": 100963, "<|LOC_667|>": 100964, "<|LOC_668|>": 100965, "<|LOC_669|>": 100966, "<|LOC_670|>": 100967, "<|LOC_671|>": 100968, "<|LOC_672|>": 100969, "<|LOC_673|>": 100970, "<|LOC_674|>": 100971, "<|LOC_675|>": 100972, "<|LOC_676|>": 100973, "<|LOC_677|>": 100974, "<|LOC_678|>": 100975, "<|LOC_679|>": 100976, "<|LOC_680|>": 100977, "<|LOC_681|>": 100978, "<|LOC_682|>": 100979, "<|LOC_683|>": 100980, "<|LOC_684|>": 100981, "<|LOC_685|>": 100982, "<|LOC_686|>": 100983, "<|LOC_687|>": 100984, "<|LOC_688|>": 100985, "<|LOC_689|>": 100986, "<|LOC_690|>": 100987, "<|LOC_691|>": 100988, "<|LOC_692|>": 100989, "<|LOC_693|>": 100990, "<|LOC_694|>": 100991, "<|LOC_695|>": 100992, "<|LOC_696|>": 100993, "<|LOC_697|>": 100994, "<|LOC_698|>": 100995, "<|LOC_699|>": 100996, "<|LOC_700|>": 100997, "<|LOC_701|>": 100998, "<|LOC_702|>": 100999, "<|LOC_703|>": 101000, "<|LOC_704|>": 101001, "<|LOC_705|>": 101002, "<|LOC_706|>": 101003, "<|LOC_707|>": 101004, "<|LOC_708|>": 101005, "<|LOC_709|>": 101006, "<|LOC_710|>": 101007, "<|LOC_711|>": 101008, "<|LOC_712|>": 101009, "<|LOC_713|>": 101010, "<|LOC_714|>": 101011, "<|LOC_715|>": 101012, "<|LOC_716|>": 101013, "<|LOC_717|>": 101014, "<|LOC_718|>": 101015, "<|LOC_719|>": 101016, "<|LOC_720|>": 101017, "<|LOC_721|>": 101018, "<|LOC_722|>": 101019, "<|LOC_723|>": 101020, "<|LOC_724|>": 101021, "<|LOC_725|>": 101022, "<|LOC_726|>": 101023, "<|LOC_727|>": 101024, "<|LOC_728|>": 101025, "<|LOC_729|>": 101026, "<|LOC_730|>": 101027, "<|LOC_731|>": 101028, "<|LOC_732|>": 101029, "<|LOC_733|>": 101030, "<|LOC_734|>": 101031, "<|LOC_735|>": 101032, "<|LOC_736|>": 101033, "<|LOC_737|>": 101034, "<|LOC_738|>": 101035, "<|LOC_739|>": 101036, "<|LOC_740|>": 101037, "<|LOC_741|>": 101038, "<|LOC_742|>": 101039, "<|LOC_743|>": 101040, "<|LOC_744|>": 101041, "<|LOC_745|>": 101042, "<|LOC_746|>": 101043, "<|LOC_747|>": 101044, "<|LOC_748|>": 101045, "<|LOC_749|>": 101046, "<|LOC_750|>": 101047, "<|LOC_751|>": 101048, "<|LOC_752|>": 101049, "<|LOC_753|>": 101050, "<|LOC_754|>": 101051, "<|LOC_755|>": 101052, "<|LOC_756|>": 101053, "<|LOC_757|>": 101054, "<|LOC_758|>": 101055, "<|LOC_759|>": 101056, "<|LOC_760|>": 101057, "<|LOC_761|>": 101058, "<|LOC_762|>": 101059, "<|LOC_763|>": 101060, "<|LOC_764|>": 101061, "<|LOC_765|>": 101062, "<|LOC_766|>": 101063, "<|LOC_767|>": 101064, "<|LOC_768|>": 101065, "<|LOC_769|>": 101066, "<|LOC_770|>": 101067, "<|LOC_771|>": 101068, "<|LOC_772|>": 101069, "<|LOC_773|>": 101070, "<|LOC_774|>": 101071, "<|LOC_775|>": 101072, "<|LOC_776|>": 101073, "<|LOC_777|>": 101074, "<|LOC_778|>": 101075, "<|LOC_779|>": 101076, "<|LOC_780|>": 101077, "<|LOC_781|>": 101078, "<|LOC_782|>": 101079, "<|LOC_783|>": 101080, "<|LOC_784|>": 101081, "<|LOC_785|>": 101082, "<|LOC_786|>": 101083, "<|LOC_787|>": 101084, "<|LOC_788|>": 101085, "<|LOC_789|>": 101086, "<|LOC_790|>": 101087, "<|LOC_791|>": 101088, "<|LOC_792|>": 101089, "<|LOC_793|>": 101090, "<|LOC_794|>": 101091, "<|LOC_795|>": 101092, "<|LOC_796|>": 101093, "<|LOC_797|>": 101094, "<|LOC_798|>": 101095, "<|LOC_799|>": 101096, "<|LOC_800|>": 101097, "<|LOC_801|>": 101098, "<|LOC_802|>": 101099, "<|LOC_803|>": 101100, "<|LOC_804|>": 101101, "<|LOC_805|>": 101102, "<|LOC_806|>": 101103, "<|LOC_807|>": 101104, "<|LOC_808|>": 101105, "<|LOC_809|>": 101106, "<|LOC_810|>": 101107, "<|LOC_811|>": 101108, "<|LOC_812|>": 101109, "<|LOC_813|>": 101110, "<|LOC_814|>": 101111, "<|LOC_815|>": 101112, "<|LOC_816|>": 101113, "<|LOC_817|>": 101114, "<|LOC_818|>": 101115, "<|LOC_819|>": 101116, "<|LOC_820|>": 101117, "<|LOC_821|>": 101118, "<|LOC_822|>": 101119, "<|LOC_823|>": 101120, "<|LOC_824|>": 101121, "<|LOC_825|>": 101122, "<|LOC_826|>": 101123, "<|LOC_827|>": 101124, "<|LOC_828|>": 101125, "<|LOC_829|>": 101126, "<|LOC_830|>": 101127, "<|LOC_831|>": 101128, "<|LOC_832|>": 101129, "<|LOC_833|>": 101130, "<|LOC_834|>": 101131, "<|LOC_835|>": 101132, "<|LOC_836|>": 101133, "<|LOC_837|>": 101134, "<|LOC_838|>": 101135, "<|LOC_839|>": 101136, "<|LOC_840|>": 101137, "<|LOC_841|>": 101138, "<|LOC_842|>": 101139, "<|LOC_843|>": 101140, "<|LOC_844|>": 101141, "<|LOC_845|>": 101142, "<|LOC_846|>": 101143, "<|LOC_847|>": 101144, "<|LOC_848|>": 101145, "<|LOC_849|>": 101146, "<|LOC_850|>": 101147, "<|LOC_851|>": 101148, "<|LOC_852|>": 101149, "<|LOC_853|>": 101150, "<|LOC_854|>": 101151, "<|LOC_855|>": 101152, "<|LOC_856|>": 101153, "<|LOC_857|>": 101154, "<|LOC_858|>": 101155, "<|LOC_859|>": 101156, "<|LOC_860|>": 101157, "<|LOC_861|>": 101158, "<|LOC_862|>": 101159, "<|LOC_863|>": 101160, "<|LOC_864|>": 101161, "<|LOC_865|>": 101162, "<|LOC_866|>": 101163, "<|LOC_867|>": 101164, "<|LOC_868|>": 101165, "<|LOC_869|>": 101166, "<|LOC_870|>": 101167, "<|LOC_871|>": 101168, "<|LOC_872|>": 101169, "<|LOC_873|>": 101170, "<|LOC_874|>": 101171, "<|LOC_875|>": 101172, "<|LOC_876|>": 101173, "<|LOC_877|>": 101174, "<|LOC_878|>": 101175, "<|LOC_879|>": 101176, "<|LOC_880|>": 101177, "<|LOC_881|>": 101178, "<|LOC_882|>": 101179, "<|LOC_883|>": 101180, "<|LOC_884|>": 101181, "<|LOC_885|>": 101182, "<|LOC_886|>": 101183, "<|LOC_887|>": 101184, "<|LOC_888|>": 101185, "<|LOC_889|>": 101186, "<|LOC_890|>": 101187, "<|LOC_891|>": 101188, "<|LOC_892|>": 101189, "<|LOC_893|>": 101190, "<|LOC_894|>": 101191, "<|LOC_895|>": 101192, "<|LOC_896|>": 101193, "<|LOC_897|>": 101194, "<|LOC_898|>": 101195, "<|LOC_899|>": 101196, "<|LOC_900|>": 101197, "<|LOC_901|>": 101198, "<|LOC_902|>": 101199, "<|LOC_903|>": 101200, "<|LOC_904|>": 101201, "<|LOC_905|>": 101202, "<|LOC_906|>": 101203, "<|LOC_907|>": 101204, "<|LOC_908|>": 101205, "<|LOC_909|>": 101206, "<|LOC_910|>": 101207, "<|LOC_911|>": 101208, "<|LOC_912|>": 101209, "<|LOC_913|>": 101210, "<|LOC_914|>": 101211, "<|LOC_915|>": 101212, "<|LOC_916|>": 101213, "<|LOC_917|>": 101214, "<|LOC_918|>": 101215, "<|LOC_919|>": 101216, "<|LOC_920|>": 101217, "<|LOC_921|>": 101218, "<|LOC_922|>": 101219, "<|LOC_923|>": 101220, "<|LOC_924|>": 101221, "<|LOC_925|>": 101222, "<|LOC_926|>": 101223, "<|LOC_927|>": 101224, "<|LOC_928|>": 101225, "<|LOC_929|>": 101226, "<|LOC_930|>": 101227, "<|LOC_931|>": 101228, "<|LOC_932|>": 101229, "<|LOC_933|>": 101230, "<|LOC_934|>": 101231, "<|LOC_935|>": 101232, "<|LOC_936|>": 101233, "<|LOC_937|>": 101234, "<|LOC_938|>": 101235, "<|LOC_939|>": 101236, "<|LOC_940|>": 101237, "<|LOC_941|>": 101238, "<|LOC_942|>": 101239, "<|LOC_943|>": 101240, "<|LOC_944|>": 101241, "<|LOC_945|>": 101242, "<|LOC_946|>": 101243, "<|LOC_947|>": 101244, "<|LOC_948|>": 101245, "<|LOC_949|>": 101246, "<|LOC_950|>": 101247, "<|LOC_951|>": 101248, "<|LOC_952|>": 101249, "<|LOC_953|>": 101250, "<|LOC_954|>": 101251, "<|LOC_955|>": 101252, "<|LOC_956|>": 101253, "<|LOC_957|>": 101254, "<|LOC_958|>": 101255, "<|LOC_959|>": 101256, "<|LOC_960|>": 101257, "<|LOC_961|>": 101258, "<|LOC_962|>": 101259, "<|LOC_963|>": 101260, "<|LOC_964|>": 101261, "<|LOC_965|>": 101262, "<|LOC_966|>": 101263, "<|LOC_967|>": 101264, "<|LOC_968|>": 101265, "<|LOC_969|>": 101266, "<|LOC_970|>": 101267, "<|LOC_971|>": 101268, "<|LOC_972|>": 101269, "<|LOC_973|>": 101270, "<|LOC_974|>": 101271, "<|LOC_975|>": 101272, "<|LOC_976|>": 101273, "<|LOC_977|>": 101274, "<|LOC_978|>": 101275, "<|LOC_979|>": 101276, "<|LOC_980|>": 101277, "<|LOC_981|>": 101278, "<|LOC_982|>": 101279, "<|LOC_983|>": 101280, "<|LOC_984|>": 101281, "<|LOC_985|>": 101282, "<|LOC_986|>": 101283, "<|LOC_987|>": 101284, "<|LOC_988|>": 101285, "<|LOC_989|>": 101286, "<|LOC_990|>": 101287, "<|LOC_991|>": 101288, "<|LOC_992|>": 101289, "<|LOC_993|>": 101290, "<|LOC_994|>": 101291, "<|LOC_995|>": 101292, "<|LOC_996|>": 101293, "<|LOC_997|>": 101294, "<|LOC_998|>": 101295, "<|LOC_999|>": 101296, "<|LOC_1000|>": 101297, "<|LOC_BEGIN|>": 101298, "<|LOC_END|>": 101299, "<|LOC_SEP|>": 101300, "<|CROP_COL_SEP|>": 101301, "<|CROP_ROW_SEP|>": 101302, "<|IMAGE_SEP|>": 101303}

config.json

@@ -0,0 +1,41 @@
+{
+    "_attn_implementation": "eager",
+    "architectures": [
+        "Ernie4_5_ForCausalLM"
+    ],
+    "auto_map": {
+        "AutoConfig": "configuration_ernie4_5_moe.Ernie4_5_MoeConfig",
+        "AutoModel": "modeling_ernie4_5_moe.Ernie4_5_Model",
+        "AutoModelForCausalLM": "modeling_ernie4_5_moe.Ernie4_5_MoeForCausalLM"
+    },
+    "bos_token_id": 1,
+    "eos_token_id": 2,
+    "hidden_act": "silu",
+    "hidden_size": 8192,
+    "intermediate_size": 28672,
+    "max_position_embeddings": 131072,
+    "model_type": "ernie4_5_moe",
+    "moe_capacity": [
+        64,
+        64,
+        64
+    ],
+    "moe_gate": "topk",
+    "moe_intermediate_size": 3584,
+    "moe_k": 8,
+    "moe_layer_interval": 1,
+    "moe_layer_start_index": 3,
+    "moe_num_experts": 64,
+    "moe_use_aux_free": true,
+    "num_attention_heads": 64,
+    "num_hidden_layers": 54,
+    "num_key_value_heads": 8,
+    "pad_token_id": 0,
+    "rms_norm_eps": 1e-05,
+    "rope_theta": 500000,
+    "tie_word_embeddings": false,
+    "torch_dtype": "bfloat16",
+    "use_bias": false,
+    "use_cache": true,
+    "vocab_size": 103424
+}

configuration_ernie4_5_moe.py

@@ -0,0 +1,192 @@
+# Copyright (c) 2025 Baidu, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Ernie4_5_Moe model configuration"""
+
+from transformers import PretrainedConfig
+
+
+
+class Ernie4_5_MoeConfig(PretrainedConfig):
+    r"""
+    This is the configuration class to store the configuration of a [`Ernie4_5_Model`].
+    It is used to instantiate an ERNIE-4.5 model according to the specified arguments,
+    defining the model architecture. Instantiating a configuration with the defaults
+    will yield a similar configuration to that of ERNIE-4.5-300B-A47B-Base-PT [baidu/ERNIE-4.5-300B-A47B-Base-PT].
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+
+    Args:
+        vocab_size (int): Size of the vocabulary (number of unique tokens)
+        hidden_size (int): Dimensionality of the encoder layers and the pooler layer
+        intermediate_size (int): Dimensionality of the "intermediate" (feed-forward) layer
+        max_position_embeddings (int): Maximum sequence length the model can handle
+        num_hidden_layers (int): Number of hidden layers in the Transformer encoder
+        num_attention_heads (int): Number of attention heads for each attention layer
+        rms_norm_eps (float): The epsilon used by the RMS normalization layers
+        use_cache (bool): Whether to use caching for faster generation (decoding)
+        use_flash_attention (bool): Whether to use FlashAttention for optimized attention computation
+        pad_token_id (int): Token ID used for padding sequences
+        bos_token_id (int): Token ID used for beginning-of-sequence
+        eos_token_id (int): Token ID used for end-of-sequence
+        use_bias (bool): Whether to use bias terms in linear layers
+        rope_theta (float): The base period of the RoPE embeddings
+        weight_share_add_bias (bool): Whether to share bias weights in certain layers
+        ignored_index (int): Target value that is ignored during loss computation
+        attention_probs_dropout_prob (float): Dropout probability for attention weights
+        hidden_dropout_prob (float): Dropout probability for hidden layers
+        num_key_value_heads (int): Number of key/value heads (for Grouped Query Attention)
+        max_sequence_length (int): Maximum sequence length for positional embeddings
+        moe_num_experts: Number of experts in MoE layers
+        moe_capacity: Capacity configuration for MoE layers
+        moe_layer_interval: Interval between MoE layers
+        moe_layer_start_index: Starting layer index for MoE
+        moe_layer_end_index: Ending layer index for MoE (-1 means last layer)
+        sinkhorn_2gate: Whether to use sinkhorn 2-gate routing
+        sinkhorn_temp: Temperature for sinkhorn routing
+        moe_dropout_prob: Dropout probability for MoE layers
+        moe_gate: Type of gating mechanism ('top2', etc.)
+        moe_intermediate_size: Intermediate size for MoE layers
+        moe_gate_act: Activation function for gating
+        moe_k: Number of experts to route to
+        **kwargs: Additional base model configuration parameters
+    """
+
+    model_type = "ernie4_5_moe"
+    use_keep_in_fp32_modules = True
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    attribute_map = {
+        "n_positions": "max_position_embeddings",
+        "n_embd": "hidden_size",
+        "n_layer": "num_hidden_layers",
+        "n_head": "num_attention_heads",
+        "n_inner": "intermediate_size",
+        "activation_function": "hidden_act",
+    }
+
+    # Default tensor parallel plan for base model `ernie_4_5_moe`
+    base_model_tp_plan = {
+        "model.layers.*.self_attn.q_proj": "colwise_rep",
+        "model.layers.*.self_attn.k_proj": "colwise_rep",
+        "model.layers.*.self_attn.v_proj": "colwise_rep",
+        "model.layers.*.self_attn.o_proj": "rowwise_rep",
+        "model.layers.*.mlp.experts.*.gate_proj": "colwise",
+        "model.layers.*.mlp.experts.*.up_proj": "colwise",
+        "model.layers.*.mlp.experts.*.down_proj": "rowwise",
+        "model.layers.*.mlp.gate_proj": "colwise",
+        "model.layers.*.mlp.up_proj": "colwise",
+        "model.layers.*.mlp.down_proj": "rowwise",
+    }
+    base_model_pp_plan = {
+        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
+        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
+        "norm": (["hidden_states"], ["hidden_states"]),
+    }
+
+    def __init__(
+        self,
+        vocab_size=32000,
+        hidden_size=768,
+        intermediate_size=11008,
+        num_hidden_layers=2,
+        num_attention_heads=2,
+        num_key_value_heads=None,
+        max_position_embeddings=32768,
+        use_sliding_window=None,
+        sliding_window=None,
+        rms_norm_eps=1e-6,
+        use_cache=False,
+        pad_token_id=0,
+        bos_token_id=1,
+        eos_token_id=2,
+        attention_probs_dropout_prob=0.0,
+        hidden_dropout_prob=0.0,
+        rope_theta=10000.0,
+        use_flash_attention=False,
+        use_rmsnorm=True,
+        use_bias=False,
+        weight_share_add_bias=True,
+        max_sequence_length=None,
+        ignored_index=-100,
+        use_moe=True,
+        moe_num_experts=64,
+        moe_capacity=(64, 64, 64),
+        moe_layer_interval=2,
+        moe_layer_start_index=0,
+        moe_layer_end_index=-1,
+        sinkhorn_2gate=True,
+        sinkhorn_temp=3e-2,
+        moe_dropout_prob=0.0,
+        moe_gate="top2",
+        moe_intermediate_size=3584,
+        moe_k=2,
+        moe_gate_act="softmax",
+        moe_use_aux_free=False,
+        **kwargs
+    ):
+        self.vocab_size = vocab_size
+        self.max_position_embeddings = max_position_embeddings
+        self.use_sliding_window = use_sliding_window
+        self.sliding_window = sliding_window
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+
+        if num_key_value_heads is None:
+            num_key_value_heads = num_attention_heads
+
+        self.num_key_value_heads = num_key_value_heads
+        self.use_rmsnorm = use_rmsnorm
+        self.rms_norm_eps = rms_norm_eps
+        self.rope_theta = rope_theta
+        self.max_sequence_length = max_sequence_length
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.eos_token_id = eos_token_id
+        self.ignored_index = ignored_index
+        self.use_cache = use_cache
+        self.use_bias = use_bias
+        self.weight_share_add_bias = weight_share_add_bias
+        self.use_flash_attention = use_flash_attention
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.hidden_dropout_prob = hidden_dropout_prob
+
+        self.use_moe = moe_num_experts > 0 and use_moe
+        self.moe_num_experts = moe_num_experts
+        self.moe_capacity = moe_capacity
+        self.sinkhorn_2gate = sinkhorn_2gate
+        self.sinkhorn_temp = sinkhorn_temp
+        self.moe_layer_interval = moe_layer_interval
+        self.moe_dropout_prob = moe_dropout_prob
+        self.moe_gate = moe_gate
+        self.moe_intermediate_size = moe_intermediate_size
+        self.moe_k = moe_k
+        self.moe_layer_start_index = moe_layer_start_index
+        self.moe_layer_end_index = self.num_hidden_layers - 1 if moe_layer_end_index == -1 else moe_layer_end_index
+        self.moe_gate_act = moe_gate_act
+        self.moe_use_aux_free = moe_use_aux_free
+
+        # Set default for tied embeddings if not specified.
+        if "tie_word_embeddings" not in kwargs:
+            kwargs["tie_word_embeddings"] = False
+        
+        super().__init__(
+            pad_token_id=pad_token_id,
+            bos_token_id=bos_token_id,
+            eos_token_id=eos_token_id,
+            **kwargs,
+        )

generation_config.json

@@ -0,0 +1,13 @@
+{
+  "do_sample": true,
+  "top_p": 0.8,
+  "temperature": 0.8,
+  "repetition_penalty": 1.0,
+  "frequency_penalty": 0.0,
+  "presence_penalty": 0.0,
+  "bos_token_id": 1,
+  "eos_token_id": 2,
+  "pad_token_id": 0,
+  "transformers_version": "4.52.4",
+  "use_cache": true
+}

modeling_ernie4_5_moe.py

@@ -0,0 +1,1590 @@
+# Copyright (c) 2025 Baidu, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Ernie4_5_Moe model """
+
+from copy import deepcopy
+from dataclasses import dataclass
+from functools import partial
+from typing import Callable, Optional, Tuple, Union
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+
+from transformers.cache_utils import (
+    Cache,
+    DynamicCache,
+    SlidingWindowCache,
+    StaticCache,
+)
+from transformers.generation import GenerationMixin
+from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
+from transformers.modeling_outputs import ModelOutput, MoeCausalLMOutputWithPast
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+from transformers.processing_utils import Unpack
+from transformers.utils import (
+    LossKwargs,
+    auto_docstring,
+    can_return_tuple,
+    logging,
+    is_torch_flex_attn_available,
+)
+
+from .configuration_ernie4_5_moe import Ernie4_5_MoeConfig
+
+
+if is_torch_flex_attn_available():
+    from torch.nn.attention.flex_attention import BlockMask
+
+    from transformers.integrations.flex_attention import make_flex_block_causal_mask
+
+logger = logging.get_logger(__name__)
+
+
+class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs):
+    """Kwargs class used during autoregressive generation"""
+
+    ...
+
+
+@dataclass
+class Erine4_5_MoeModelOutputWithPast(ModelOutput):
+    """Class for Ernie4_5_Moe model outputs with past keys."""
+
+    last_hidden_state: Optional[torch.FloatTensor] = None
+    past_key_values: Optional[Cache] = None
+    hidden_states: Optional[tuple[torch.FloatTensor, ...]] = None
+    attentions: Optional[tuple[torch.FloatTensor, ...]] = None
+    router_loss: Optional[torch.FloatTensor] = None
+    gate_logits: Optional[tuple[torch.FloatTensor, ...]] = None
+
+
+@dataclass
+class Ernie4_5_MoeCausalLMOutputWithPast(MoeCausalLMOutputWithPast):
+    """Class for  Ernie4_5_Moe causal LM output with past keys"""
+
+    router_loss: Optional[torch.FloatTensor] = None
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+
+    x1 = x[..., 0::2]
+    x2 = x[..., 1::2]
+    return torch.stack((-x2, x1), dim=-1).reshape(x.shape)
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(
+        batch, num_key_value_heads, n_rep, slen, head_dim
+    )
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1):
+    """Applies Rotary Position Embedding to the query and key tensors.
+
+    Args:
+        q (`torch.Tensor`): The query tensor.
+        k (`torch.Tensor`): The key tensor.
+        cos (`torch.Tensor`): The cosine part of the rotary embedding.
+        sin (`torch.Tensor`): The sine part of the rotary embedding.
+        position_ids (`torch.Tensor`, *optional*):
+            Deprecated and unused.
+        unsqueeze_dim (`int`, *optional*, defaults to 1):
+            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+    Returns:
+        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+    """
+    orig_dtype = q.dtype
+    sin_pos = torch.stack([sin, sin], dim=-1).reshape(*sin.shape[:-1], -1)
+    cos_pos = torch.stack([cos, cos], dim=-1).reshape(*sin.shape[:-1], -1)
+    q_embed = (q.float() * cos_pos) + (rotate_half(q).float() * sin_pos)
+    k_embed = (k.float() * cos_pos) + (rotate_half(k).float() * sin_pos)
+    return q_embed.to(orig_dtype), k_embed.to(orig_dtype)
+
+
+def eager_attention_forward(
+    module: nn.Module,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    value: torch.Tensor,
+    attention_mask: Optional[torch.Tensor],
+    scaling: float,
+    dropout: float = 0.0,
+    **kwargs,
+):
+    """
+    Eager attention for Ernie4_5_Attention forward function.
+    """
+    key_states = repeat_kv(key, module.num_key_value_groups)
+    value_states = repeat_kv(value, module.num_key_value_groups)
+
+    attn_weights = torch.matmul(query, key_states.transpose(2, 3)) * scaling
+    if attention_mask is not None:
+        causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+        attn_weights = attn_weights + causal_mask.to(attn_weights.device)
+
+    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(
+        query.dtype
+    )
+    attn_weights = nn.functional.dropout(
+        attn_weights, p=dropout, training=module.training
+    )
+    attn_output = torch.matmul(attn_weights, value_states)
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    return attn_output, attn_weights
+
+
+def topk_gate_func(
+    module: nn.Module,
+    hidden_states: torch.Tensor,
+):
+    """
+    Topk gate function for Ernie4_5_MoEMlp
+    """
+    capacity = module.get_capacity(hidden_states.shape[0])
+    with torch.autocast(device_type="cuda", dtype=torch.float32):
+        logits = module.gate(hidden_states.float())
+    router_loss = torch.zeros([1], dtype=torch.float32, device=hidden_states.device)
+    router_loss.detach()
+    return logits, capacity, router_loss
+
+
+class Ernie4_5_ResidualWithDropout(nn.Module):
+    """
+    Fused dropout implementation with residual connection support.
+
+    This layer combines dropout and residual addition in a single operation for better performance,
+    particularly on GPU devices. The dropout is conditionally applied based on the probability.
+
+    Args:
+        prob (float): Dropout probability (between 0 and 1)
+
+    Attributes:
+        prob (float): Stores the dropout probability
+        dropout (nn.Dropout): The actual dropout layer instance
+    """
+
+    def __init__(self, prob):
+        """
+        Initialize the fused dropout layer.
+
+        Args:
+            prob (float): Dropout probability (0 means no dropout)
+        """
+        super().__init__()
+        self.prob = prob
+        self.dropout = nn.Dropout(p=prob)
+
+    def forward(self, x, y):
+        """
+        Forward pass of the fused dropout layer.
+
+        Args:
+            x (torch.Tensor): Input tensor to potentially apply dropout on
+            y (torch.Tensor): Residual tensor to add to the (possibly dropped out) x
+
+        Returns:
+            torch.Tensor: Result of x (with optional dropout) + y
+        """
+        if self.prob > 0:
+            x = self.dropout(x)
+        output = x + y
+
+        return output
+
+
+class Ernie4_5_Attention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(self, config, layer_idx=0):
+        """
+        Args:
+            config (ErnieConfig): Model configuration.
+            layer_idx (int, optional): Index in transformer stack. Defaults to 0.
+        """
+        super().__init__()
+        self.layer_idx = layer_idx
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.num_key_value_heads = (
+            config.num_key_value_heads
+            if config.num_key_value_heads is not None
+            else self.nums_head
+        )
+        self.num_key_value_groups = (
+            config.num_attention_heads // config.num_key_value_heads
+        )
+        self.head_dim = self.hidden_size // self.num_heads
+        self.freq_allocation = (
+            config.freq_allocation if hasattr(config, "freq_allocation") else 0
+        )
+        self.scaling = self.head_dim**-0.5
+        self.attention_dropout = getattr(config, "attention_probs_dropout_prob", 0.0)
+        self.is_causal = True
+
+        self.q_proj = nn.Linear(
+            self.hidden_size,
+            self.num_heads * self.head_dim,
+            bias=config.use_bias,
+        )
+
+        self.k_proj = nn.Linear(
+            self.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=config.use_bias,
+        )
+
+        self.v_proj = nn.Linear(
+            self.hidden_size,
+            self.num_key_value_heads * self.head_dim,
+            bias=config.use_bias,
+        )
+
+        self.o_proj = nn.Linear(
+            self.hidden_size,
+            self.hidden_size,
+            bias=config.use_bias,
+        )
+
+        self.config = config
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Cache] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] = None,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> Tuple[
+        torch.Tensor,
+        Optional[torch.Tensor],
+        Optional[Tuple[torch.Tensor, torch.Tensor]],
+    ]:
+        """ 
+            Ernie4_5_Attention forward function
+        """
+        B, L = hidden_states.shape[:-1]
+
+        query_states = (
+            self.q_proj(hidden_states).view(B, L, self.num_heads, -1).transpose(1, 2)
+        )
+        key_states = (
+            self.k_proj(hidden_states)
+            .view(B, L, self.num_key_value_heads, -1)
+            .transpose(1, 2)
+        )
+        value_states = (
+            self.v_proj(hidden_states)
+            .view(B, L, self.num_key_value_heads, -1)
+            .transpose(1, 2)
+        )
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb(
+            query_states, key_states, cos, sin
+        )
+
+        if past_key_value is not None:
+            # sin and cos are specific to RoPE models; cache_position needed for the static cache
+            cache_kwargs = {"cache_position": cache_position}
+            key_states, value_states = past_key_value.update(
+                key_states, value_states, self.layer_idx, cache_kwargs
+            )
+
+        attention_interface: Callable = eager_attention_forward
+        if self.config._attn_implementation != "eager":
+            attention_interface = ALL_ATTENTION_FUNCTIONS[
+                self.config._attn_implementation
+            ]
+
+        attn_output, attn_weights = attention_interface(
+            self,
+            query_states,
+            key_states,
+            value_states,
+            attention_mask,
+            dropout=0.0 if not self.training else self.attention_dropout,
+            scaling=self.scaling,
+            **kwargs,
+        )
+        attn_output = attn_output.reshape(B, L, -1).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output, attn_weights
+
+
+class Ernie4_5_MLP(nn.Module):
+    """
+    Ernie4_5_MLP - Gated Multi-Layer Perceptron module used in Ernie model.
+    """
+
+    def __init__(self, config, intermediate_size=None):
+        """
+        Initialize the MLP module with configuration options.
+
+        Args:
+            config: Model configuration object with attributes:
+                - hidden_size: int
+                - intermediate_size: int
+                - use_bias: bool
+            layer_idx (int): Index of current layer (default: 0)
+        """
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = (
+            intermediate_size
+            if intermediate_size is not None
+            else config.intermediate_size
+        )
+        self.gate_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=config.use_bias
+        )
+        self.up_proj = nn.Linear(
+            self.hidden_size, self.intermediate_size, bias=config.use_bias
+        )
+        self.down_proj = nn.Linear(
+            self.intermediate_size, self.hidden_size, bias=config.use_bias
+        )
+
+    def forward(self, x):
+        """
+        Args:
+            x (Tensor): shape [batch_size, seq_len, hidden_size]
+
+        Returns:
+            Tensor: shape [batch_size, seq_len, hidden_size]
+        """
+        down_proj = self.down_proj(F.silu(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class Ernie4_5_MoeStatics(nn.Module):
+    """
+    Stores MoE (Mixture of Experts) statistics
+    and expert usage information.
+    """
+
+    def __init__(self, config):
+        """
+        Initialize MoE statistics tracking.
+
+        Args:
+            config: Model configuration containing MoE parameters
+        """
+        super().__init__()
+
+        num_experts = config.moe_num_experts
+        num_experts_groups = 1
+
+        self.e_score_correction_bias = nn.Parameter(
+            torch.zeros(num_experts_groups, num_experts, dtype=torch.float32),
+            requires_grad=False,
+        )
+
+
+class Ernie4_5_MoeMLP(nn.Module):
+    """Mixture of Experts (MoE) variant of ERNIE's MLP layer."""
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.k = config.moe_k
+        self.sinkhorn_2gate = config.sinkhorn_2gate
+        self.sinkhorn_temp = config.sinkhorn_temp
+
+        moe_intermediate_size = (
+            config.moe_intermediate_size
+            if config.moe_intermediate_size
+            else config.intermediate_size
+        )
+        self.gate = nn.Linear(
+            config.hidden_size, config.moe_num_experts, bias=False, dtype=torch.float32
+        )
+        if config.moe_gate_act == "softmax":
+            self.gate_act = partial(F.softmax, dim=-1)
+        elif config.moe_gate_act == "sigmoid":
+            self.gate_act = F.sigmoid
+        else:
+            raise ValueError(f"{config.moe_gate_act} is not supported.")
+
+        self.experts = nn.ModuleList(
+            [
+                Ernie4_5_MLP(config, moe_intermediate_size)
+                for i in range(config.moe_num_experts)
+            ]
+        )
+
+        if config.moe_use_aux_free:
+            self.moe_statics = Ernie4_5_MoeStatics(config)
+
+        self.use_correction_bias = config.moe_use_aux_free
+        self.num_local_experts = len(self.experts)
+
+        self.shared_experts = self._init_shared_experts()
+
+    def _init_shared_experts(self):
+        """
+        Initialize the shared expert module.
+
+        Returns:
+            shared_experts: Shared expert module, returns None if no shared experts are needed.
+
+        """
+        cfg = deepcopy(self.config)
+        if getattr(cfg, "moe_num_shared_experts", 0) > 0:
+            if getattr(cfg, "moe_intermediate_size", None):
+                cfg.intermediate_size = (
+                    cfg.moe_intermediate_size * cfg.moe_num_shared_experts
+                )
+            else:
+                cfg.intermediate_size = (
+                    cfg.intermediate_size * cfg.moe_num_shared_experts
+                )
+            shared_experts = Ernie4_5_MLP(cfg, cfg.intermediate_size)
+        else:
+            shared_experts = None
+        return shared_experts
+
+    def forward(
+        self,
+        input: torch.Tensor,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Forward pass through MoE layer.
+
+        Args:
+            input (Tensor): Input tensor of shape [s, d].
+            token_type_ids: Optional tensor for token types.
+
+        Returns:
+            tuple: (output, combine_weights, router_loss, gate_logits)
+        """
+
+        if input.dim() == 3:
+            orig_shape = input.shape
+            input = input.reshape(-1, input.shape[-1])
+        else:
+            orig_shape = None
+        assert (
+            input.dim() == 2
+        ), f"input Tensor must have dimensions: (s)equence, (d)im, got:{input.shape}"
+
+        assert self.gate is not None
+
+        gate_input = input
+
+        (
+            dispatched_input,
+            combine_weights,
+            dispatch_mask,
+            scatter_index,
+            router_loss,
+            gate_logits,
+            gate_prob,
+        ) = self.gate_and_dispatch(gate_input)
+
+        expert_out = self.forward_experts(dispatched_input)
+
+        combined_output = self.combine_expert_output(
+            expert_out, combine_weights, scatter_index
+        )
+
+        if self.shared_experts is not None:
+            shared_expert_out = self.shared_experts(gate_input)
+            combined_output += shared_expert_out
+
+        if orig_shape:
+            combined_output = combined_output.reshape(
+                orig_shape[:-1] + (combined_output.shape[-1],)
+            )
+
+        return combined_output, combine_weights, router_loss, gate_logits
+
+    def forward_experts(self, dispatched_input: torch.Tensor) -> torch.Tensor:
+        """
+        Forward pass through experts sequentially.
+
+        Args:
+            dispatched_input (Tensor): Input tensor of shape [num_experts, capacity, dim].
+
+        Returns:
+            Tensor: Expert outputs of shape [num_experts, capacity, dim].
+        """
+        true_experts = self.experts
+        dispatched_input = dispatched_input.reshape(
+            1, self.num_local_experts, -1, dispatched_input.shape[-1]
+        )
+        expert_outputs = []
+        if isinstance(self.experts, nn.ModuleList):
+            chunks = dispatched_input.permute(1, 0, 2, 3).contiguous().unbind(0)
+            assert len(chunks) == len(
+                true_experts
+            ), f"{len(chunks)}, {len(true_experts)}"
+            for chunk, expert in zip(chunks, true_experts):
+                expert_outputs.append(expert(chunk))
+        else:
+            dispatched_input = dispatched_input.permute(1, 0, 2, 3).contiguous()
+            orig_shape = dispatched_input.shape
+            chunks = dispatched_input.reshape(orig_shape[0], -1, orig_shape[-1])
+            chunks = self.experts(chunks)
+            chunks = chunks.reshape(orig_shape[:-1] + (chunks.shape[-1],)).unbind(0)
+            expert_outputs.extend(chunks)
+
+        expert_output = torch.stack(expert_outputs, dim=1)
+        return expert_output
+
+    def moe_gate_dispatch(
+        self,
+        x: torch.Tensor,
+        gate_logits: torch.Tensor,
+        k: int,
+        capacity: Optional[int],
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+            Dispatch inputs to experts based on their routing probabilities.
+        """
+        S, H = x.shape
+        E = gate_logits.shape[1]
+        device = x.device
+        topk_prob, topk_idx = torch.topk(gate_logits, k, dim=-1)
+        combine_weights = topk_prob
+        expert_id = topk_idx
+        y = x.new_zeros((E, capacity, H))
+        scatter_index = x.new_full((k, S), -1, dtype=torch.int32)
+
+        # per-expert slot counters
+        slot_counter = torch.zeros(E, dtype=torch.int32, device=device)
+
+        for tok in range(S):
+            for route in range(k):
+                e = expert_id[tok, route].item()
+                slot = slot_counter[e].item()
+                if slot >= capacity:
+                    combine_weights[tok, route] = 0.0
+                    continue
+
+                # record mapping & dispatch activation
+                scatter_index[route, tok] = e * capacity + slot
+                y[e, slot] = x[tok]
+                slot_counter[e] += 1
+
+        expert_offset = torch.cumsum(slot_counter, 0, dtype=torch.int64)
+
+        return y, combine_weights, scatter_index, expert_offset, expert_id
+
+    def combine_expert_output(
+        self,
+        expert_output: torch.Tensor,
+        combine_weights: torch.Tensor,
+        scatter_index: torch.Tensor,
+    ) -> torch.Tensor:
+        """
+        Combine expert outputs using combination weights.
+
+        Args:
+            expert_output (Tensor): Expert outputs [num_experts, capacity, dim].
+            combine_weights (Tensor): Combination weights.
+            scatter_index (Tensor): Scatter indices.
+
+        Returns:
+            Tensor: Combined output [seqlen, dim].
+        """
+        expert_output = expert_output.reshape(-1, expert_output.shape[-1])
+        combined_output = self.combining(expert_output, combine_weights, scatter_index)
+        return combined_output
+
+    def combining(self, x, combine_weights, scatter_index):
+        """
+        Combines and aggregates input matrix using combination weights.
+
+        Args:
+            x (Tensor): Input tensor of shape [num_experts * capacity, dim]
+            combine_weights (Tensor): Combination weights of shape [seq, 2]
+            scatter_index (Tensor): Scatter indices of shape [seq, 2]
+
+        Returns:
+            Tensor: Combined output tensor of shape [seq, dim]
+        """
+        dim = x.shape[-1]
+
+        scatter_index = scatter_index.reshape([-1])
+        num_k = combine_weights.shape[-1]
+
+        combine_weights = combine_weights.unsqueeze(1)
+
+        x = x[scatter_index].reshape([-1, num_k, dim])
+
+        return torch.matmul(combine_weights, x).squeeze(1)
+
+    def gate_and_dispatch(self, input):
+        """
+        Calculate gate and dispatch inputs.
+
+        Args:
+            input: Input tensor of shape [seq, dim]
+
+        Returns:
+            tuple: (dispatched_input, combine_weights, dispatch_mask,
+            scatter_index, router_loss, gate_logits, gate_prob)
+        """
+        gate_logits, capacity, router_loss = topk_gate_func(self, input)
+
+        # capacity no use
+        prob = self.gate_act(gate_logits)
+        (
+            dispatched_input,
+            combine_weights_unnorm,
+            scatter_index,
+            dispatch_mask,
+            _,
+        ) = self.moe_gate_dispatch(input, prob, k=self.k, capacity=capacity)
+        dispatch_mask = torch.diff(F.pad(dispatch_mask, (1, 0)))
+
+        scatter_index.detach()
+        dispatch_mask.detach()
+
+        scatter_index = scatter_index.transpose(0, 1)  # [k, s] -> [s, k]
+        combine_weights = combine_weights_unnorm / torch.clamp(
+            combine_weights_unnorm.sum(dim=-1, keepdim=True), min=1e-12
+        )
+        combine_weights = combine_weights.to(dtype=dispatched_input.dtype)
+
+        return (
+            dispatched_input,
+            combine_weights,
+            dispatch_mask,
+            scatter_index,
+            router_loss,
+            gate_logits,
+            prob,
+        )
+
+    def get_capacity(self, num_tokens, cap_factor=None):
+        """
+        Calculate capacity based on number of tokens.
+
+        Args:
+            num_tokens: Number of input tokens
+            cap_factor: Optional capacity factor override
+
+        Returns:
+            int: Calculated capacity
+        """
+        num_experts = self.config.moe_num_experts
+        if cap_factor is not None:
+            cap = cap_factor
+        else:
+            if self.training:
+                cap = self.config.moe_capacity[0]
+            elif num_tokens < num_experts:
+                cap = self.config.moe_capacity[2]
+            else:
+                cap = self.config.moe_capacity[1]
+
+        capacity = int(cap * num_tokens // num_experts)
+        assert (
+            capacity > 0
+        ), f"requires capacity to >= 0. cap={cap}, num_tokens={num_tokens}"
+        return capacity
+
+
+class Ernie4_5_RMSNorm(nn.Module):
+    """
+    Ernie Root Mean Square Layer Normalization (Ernie4_5_RMSNorm) implementation.
+
+    Ernie4_5_RMSNorm is a simplified version of LayerNorm that focuses on the root mean square of inputs,
+    omitting the mean-centering operation. This provides computational efficiency while maintaining
+    good performance.
+
+    """
+
+    def __init__(self, config):
+        """
+        Initialize RMSNorm layer.
+
+        Args:
+            config (ErnieConfig): Model configuration.
+        """
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.weight = nn.Parameter(torch.ones(config.hidden_size))
+        self.variance_epsilon = config.rms_norm_eps
+
+    def forward(self, hidden_states):
+        """
+        Apply RMS normalization to input hidden states.
+
+        Args:
+            hidden_states (Tensor): Input tensor of shape [batch_size, seq_len, hidden_size]
+
+        Returns:
+            Tensor: Normalized output tensor of same shape as input
+        """
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(dim=-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+
+        return self.weight * hidden_states.to(input_dtype)
+
+
+class Ernie4_5_RopeEmbedding(nn.Module):
+    """
+    Implements Rotary Position Embedding (RoPE) for Ernie4_5_MoeModel.
+    """
+
+    def __init__(self, config: Ernie4_5_MoeConfig, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type")
+            )
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        inv_freq_expanded = self.inv_freq[None, None, :].float()
+        position_ids_expanded = position_ids[..., None].float()
+        freqs = inv_freq_expanded.float() * position_ids_expanded.float()
+        cos = torch.cos(freqs) * self.attention_scaling
+        sin = torch.sin(freqs) * self.attention_scaling
+        return cos, sin
+        # return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+class Ernie4_5_DecoderLayer(nn.Module):
+    """A single transformer decoder layer in ERNIE-MoE model.
+
+    Contains self-attention and feed-forward components with optional MoE (Mixture of Experts)
+    support, residual connections, and layer normalization.
+    """
+
+    def __init__(self, config, layer_idx):
+        """Initialize the decoder layer.
+
+        Args:
+            config (ErnieMoEConfig): Model configuration.
+            layer_idx (int): Index of this layer in the transformer stack
+        """
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.layer_idx = layer_idx
+        self.config = config
+        self.use_moe = config.use_moe
+        self.self_attn = Ernie4_5_Attention(config, layer_idx)
+
+        moe_layer_start_index = (
+            min(config.moe_layer_start_index)
+            if isinstance(config.moe_layer_start_index, (tuple, list))
+            else config.moe_layer_start_index
+        )
+        moe_layer_end_index = (
+            max(config.moe_layer_end_index)
+            if isinstance(config.moe_layer_end_index, (tuple, list))
+            else config.moe_layer_end_index
+        )
+
+        if (
+            self.use_moe
+            and ((layer_idx + 1) % config.moe_layer_interval == 0)
+            and layer_idx >= moe_layer_start_index
+            and layer_idx <= moe_layer_end_index
+        ):
+            self.mlp = Ernie4_5_MoeMLP(config)
+        else:
+            self.mlp = Ernie4_5_MLP(config)
+
+        self.input_layernorm = Ernie4_5_RMSNorm(config)
+        self.post_attention_layernorm = Ernie4_5_RMSNorm(config)
+
+        self.residual_add1 = Ernie4_5_ResidualWithDropout(config.hidden_dropout_prob)
+        self.residual_add2 = Ernie4_5_ResidualWithDropout(config.hidden_dropout_prob)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.Tensor] = None,
+        past_key_value: Optional[Tuple[torch.Tensor]] = None,
+        output_attentions: Optional[bool] = False,
+        use_cache: Optional[bool] = False,
+        cache_position: Optional[torch.LongTensor] = None,
+        position_embeddings: Optional[
+            tuple[torch.Tensor, torch.Tensor]
+        ] = None,  # necessary, but kept here for BC
+        output_router_loss: bool = True,
+        output_gate_logits: bool = True,
+        **kwargs: Unpack[FlashAttentionKwargs],
+    ) -> tuple[
+        torch.FloatTensor, Optional[tuple[torch.FloatTensor, torch.FloatTensor]]
+    ]:
+        """Forward pass through the decoder layer.
+
+        Args:
+            hidden_states (torch.Tensor): Input tensor [batch_size, seq_len, hidden_size]
+            attention_mask (Optional[torch.Tensor]): Attention mask tensor
+            position_ids (Optional[torch.Tensor]): Position indices for rotary embeddings
+            past_key_value (Optional[Tuple[torch.Tensor]]): Cached key/value states
+            output_attentions (Optional[bool]): Whether to return attention weights
+            use_cache (Optional[bool]): Whether to cache key/value states
+            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            position_embeddings (`tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+                with `head_dim` being the embedding dimension of each attention head.
+            output_router_loss (bool): Whether to return MoE router loss
+            output_gate_logits (bool): Whether to return MoE gate logits
+
+        Returns:
+            Union: Various output combinations depending on arguments:
+                - Base case: Hidden states tensor
+                - With attention: Tuple of (hidden_states, attention_weights)
+                - With router loss: May include gate logits in output tuple
+                - With MoE gate logits: May include gate logits in output tuple
+        """
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states, self_attn_weights = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            past_key_value=past_key_value,
+            position_ids=position_ids,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            position_embeddings=position_embeddings,
+            **kwargs,
+        )
+
+        hidden_states = self.residual_add1(hidden_states, residual)
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+
+        router_loss = None
+        gate_logits = None
+
+        if isinstance(self.mlp, Ernie4_5_MoeMLP):
+            hidden_states, _, router_loss, gate_logits = self.mlp(hidden_states)
+        else:
+            hidden_states = self.mlp(hidden_states)
+
+        hidden_states = self.residual_add2(hidden_states, residual)
+
+        outputs = (hidden_states,)
+
+        if output_attentions:
+            outputs += (self_attn_weights,)
+
+        if output_router_loss:
+            outputs += (router_loss,)
+
+        if output_gate_logits:
+            outputs += (gate_logits,)
+
+        return outputs
+
+
+@auto_docstring
+class Ernie4_5_PretrainedModel(PreTrainedModel):
+    """Base class for ERNIE pretrained models."""
+
+    config_class = Ernie4_5_MoeConfig
+    base_model_prefix = "model"
+    supports_gradient_checkpointing = True
+    _no_split_modules = ["Ernie4_5_DecoderLayer"]
+    _skip_keys_device_placement = ["past_key_values"]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+    _supports_flex_attn = True
+    _supports_cache_class = True
+    _supports_quantized_cache = True
+    _supports_static_cache = False  # MoE models don't work with torch.compile (`torch.where(condition)` not supported)
+
+
+def subbatch(f, arg_idx, axis, bs, out_idx, same_arg_idx={}):
+    """
+    Converts a function to one that applies to subbatch of an input dimension.
+    Useful for processing large tensors in smaller chunks to reduce memory usage.
+
+    Args:
+        f (Callable): Function to be subbatched.
+        arg_idx ([int]): Indices of the inputs to be subbatched.
+        axis ([int]): Indices of the dimensions to be subbatched for each input.
+        bs (int): Subbatch size.
+        out_idx (int): Dimension to concatenate outputs along.
+        same_arg_idx (dict): Mapping of argument indices that share the same tensor.
+
+    Returns:
+        Callable: New function that processes inputs in subbatches.
+    """
+
+    @functools.wraps(f)
+    def wrapper(*args, **kwargs):
+
+        assert len(arg_idx) == len(
+            axis
+        ), "Number of batching args and number of batching dims should match."
+
+        inps = [args[i] for i in arg_idx]
+        axis_width = [inp.shape[d] for inp, d in zip(inps, axis)]
+        assert len(set(axis_width)) == 1, "Batch sizes should be kept equal."
+
+        inp_axis = {idx: d for idx, d in zip(arg_idx, axis)}
+
+        axis_width = axis_width[0]
+        if axis_width < bs:
+            return f(*args, **kwargs)
+
+        outs = []
+        for slice_at in range(0, axis_width, bs):
+            _args = []
+            for i, inp in enumerate(args):
+                if i in same_arg_idx:
+                    assert (
+                        i > same_arg_idx[i]
+                    ), f"expect i > same_arg_idx[i], but got i: {i} and same_arg_idx[i]: {same_arg_idx[i]}"
+                    _args.append(_args[same_arg_idx[i]])
+                elif i in arg_idx:
+                    d = inp_axis[i]
+                    start = slice_at
+                    end = min(inp.shape[d], slice_at + bs)
+                    # Build slice for all dims, only slice along axis d
+                    slices = [slice(None)] * inp.ndim
+                    slices[d] = slice(start, end)
+                    _args.append(inp[tuple(slices)])
+                else:
+                    _args.append(inp)
+
+            out = f(*_args, **kwargs)
+            outs.append(out)
+
+        return torch.cat(outs, dim=out_idx)
+
+    return wrapper
+
+
+class ErniePretrainingCriterion(nn.Module):
+    """Criterion for ERNIE pretraining task."""
+
+    def __init__(self, config, return_tuple=True):
+        """Initialize the pretraining criterion.
+
+        Args:
+            config (ErnieConfig): Model configuration.
+            return_tuple (bool): Whether to return loss as tuple (loss, loss_sum). Defaults to True.
+        """
+        super().__init__()
+        self.ignored_index = getattr(config, "ignored_index", -100)
+        self.config = config
+        self.return_tuple = return_tuple
+
+        self.loss_func = nn.CrossEntropyLoss(reduction="none")
+
+    def forward(self, prediction_scores, masked_lm_labels, loss_mask, router_loss=None):
+        """Compute the combined pretraining loss.
+
+        Args:
+            prediction_scores: Prediction scores tensor, [batch_size, seq_len, vocab_size]
+            masked_lm_labels: Target labels tensor [batch_size, seq_len]
+            loss_mask: Optional mask for valid tokens
+            router_loss: Optional MoE router loss tensor
+
+        Returns:
+            Union:
+                - If return_tuple=True: Tuple of (combined_loss, mlm_loss_sum)
+                - If return_tuple=False: Combined loss tensor
+        """
+        res = self.forward_impl(prediction_scores, masked_lm_labels, loss_mask)
+
+        if self.return_tuple:
+            loss, loss_sum = res
+        else:
+            loss, loss_sum = res, None
+
+        if router_loss is not None and isinstance(router_loss, torch.Tensor):
+            loss = loss + router_loss - router_loss.detach()
+
+        return loss, loss_sum
+
+    def loss_impl(
+        self, prediction_scores: torch.Tensor, masked_lm_labels: torch.Tensor
+    ) -> torch.Tensor:
+        """
+        Core loss computation without reduction (but per-token).
+
+        Args:
+            prediction_scores (torch.Tensor): Logits tensor [batch_size, seq_len, vocab_size].
+            masked_lm_labels (torch.Tensor): Target labels tensor [batch_size, seq_len].
+
+        Returns:
+            torch.Tensor: Unreduced loss tensor of shape [batch_size, seq_len].
+                          Losses are calculated in float32.
+        """
+        scores_float32 = prediction_scores.to(torch.float32)
+        # prediction_scores: [batch_size, seq_len, vocab_size]
+        # masked_lm_labels: [batch_size, seq_len]
+        # Transpose prediction_scores to [batch_size, vocab_size, seq_len]
+        unreduced_loss = self.loss_func(
+            scores_float32.transpose(1, 2),  # Shape: [batch_size, vocab_size, seq_len]
+            masked_lm_labels.long(),  # Shape: [batch_size, seq_len], ensure long type
+        )
+        # unreduced_loss will be of shape [batch_size, seq_len] and dtype float32
+        return unreduced_loss
+
+    def forward_impl(self, prediction_scores, masked_lm_labels, loss_mask=None):
+        """
+            Loss function forward pass implementation.
+        """
+        prediction_scores_dims = len(prediction_scores.shape)
+
+        loss_subbatch_seqlen_config_key = "loss_subbatch_seqlen"
+        default_loss_subbatch_seqlen = 32768
+
+        current_loss_subbatch_seqlen = self.config.get(
+            loss_subbatch_seqlen_config_key, default_loss_subbatch_seqlen
+        )
+
+        if (
+            prediction_scores_dims == 2
+            and prediction_scores.shape[0] > current_loss_subbatch_seqlen
+        ):
+            sb_loss_func = subbatch(
+                self.loss_impl, [0, 1], [0, 0], current_loss_subbatch_seqlen, 0
+            )
+            masked_lm_loss = sb_loss_func(prediction_scores, masked_lm_labels)
+        elif (
+            prediction_scores_dims == 3
+            and prediction_scores.shape[1] > current_loss_subbatch_seqlen
+        ):
+            sb_loss_func = subbatch(
+                self.loss_impl, [0, 1], [1, 1], current_loss_subbatch_seqlen, 1
+            )
+            masked_lm_loss = sb_loss_func(prediction_scores, masked_lm_labels)
+        else:
+            masked_lm_loss = self.loss_impl(prediction_scores, masked_lm_labels)
+
+        if loss_mask is None:
+            loss_mask = masked_lm_labels != self.ignored_index
+
+        loss_mask = loss_mask.reshape(-1).to(torch.float32)
+
+        masked_lm_loss = torch.sum(
+            masked_lm_loss.to(torch.float32).reshape(-1) * loss_mask
+        )
+
+        # The division will be in float32
+        loss = masked_lm_loss / loss_mask.sum()
+
+        loss_sum = masked_lm_loss.sum().detach()
+
+        if not self.return_tuple:
+            if self.training:
+                return loss
+            return loss_sum
+        return loss, loss_sum
+
+
+@auto_docstring
+class Ernie4_5_Model(Ernie4_5_PretrainedModel):
+    """The core ERNIE transformer model with MoE (Mixture of Experts) support."""
+
+    _keep_in_fp32_modules = ["gate"]
+
+    def __init__(self, config: Ernie4_5_MoeConfig):
+        """Initialize the ERNIE model architecture."""
+        super().__init__(config)
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.hidden_size = config.hidden_size
+        self.config = config
+
+        self.embed_tokens = nn.Embedding(
+            self.vocab_size,
+            self.hidden_size,
+        )
+
+        self.layers = nn.ModuleList(
+            [Ernie4_5_DecoderLayer(config, i) for i in range(config.num_hidden_layers)]
+        )
+        self.norm = Ernie4_5_RMSNorm(config)
+        self.rotary_emb = Ernie4_5_RopeEmbedding(config=config)
+
+        self.gradient_checkpointing = False
+
+        self.post_init()
+
+    def get_input_embeddings(self):
+        """Get the input embedding layer."""
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        """Set new input embeddings."""
+        self.embed_tokens = value
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Cache] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+    ):
+        """Forward pass through the ERNIE model."""
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+
+        if (input_ids is None) ^ (inputs_embeds is not None):
+            raise ValueError(
+                "You must specify exactly one of input_ids or inputs_embeds"
+            )
+
+        if self.gradient_checkpointing and self.training:
+            if use_cache:
+                logger.warning_once(
+                    "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+                )
+                use_cache = False
+
+        if use_cache and past_key_values is None:
+            past_key_values = DynamicCache()
+
+        if inputs_embeds is None:
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        inputs_embeds = inputs_embeds.to(self.embed_tokens.weight.dtype)
+
+        if cache_position is None:
+            past_seen_tokens = (
+                past_key_values.get_seq_length() if past_key_values is not None else 0
+            )
+            cache_position = torch.arange(
+                past_seen_tokens,
+                past_seen_tokens + inputs_embeds.shape[1],
+                device=inputs_embeds.device,
+            )
+        if position_ids is None:
+            position_ids = cache_position.unsqueeze(0)
+
+        causal_mask = self._update_causal_mask(
+            attention_mask,
+            inputs_embeds,
+            cache_position,
+            past_key_values,
+            output_attentions,
+        )
+
+        hidden_states = inputs_embeds
+
+        # create position embeddings to be shared across the decoder layers
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_router_loss = (
+            torch.tensor(0.0, device=inputs_embeds.device)
+            if self.config.use_moe
+            else None
+        )
+        all_gate_logits = ()
+
+        for decoder_layer in self.layers:
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+
+            if self.gradient_checkpointing and self.training:
+                layer_outputs = self._gradient_checkpointing_func(
+                    partial(decoder_layer.__call__, **flash_attn_kwargs),
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                )
+            else:
+                layer_outputs = decoder_layer(
+                    hidden_states,
+                    causal_mask,
+                    position_ids,
+                    past_key_values,
+                    output_attentions,
+                    use_cache,
+                    cache_position,
+                    position_embeddings,
+                    **flash_attn_kwargs,
+                )
+
+            hidden_states = layer_outputs[0]
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+            if self.config.use_moe:
+                layer_outputs, gate_logits = layer_outputs[:-1], layer_outputs[-1]
+                all_gate_logits = all_gate_logits + (gate_logits,)
+
+        hidden_states = self.norm(hidden_states)
+
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        # assert all_router_loss is None, f'moe not support `return-dict`'
+        return Erine4_5_MoeModelOutputWithPast(
+            last_hidden_state=hidden_states,
+            past_key_values=past_key_values,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            router_loss=all_router_loss,
+            gate_logits=all_gate_logits,
+        )
+
+    def _update_causal_mask(
+        self,
+        attention_mask: Union[torch.Tensor, "BlockMask"],
+        input_tensor: torch.Tensor,
+        cache_position: torch.Tensor,
+        past_key_values: Cache,
+        output_attentions: bool = False,
+    ):
+        if self.config._attn_implementation == "flash_attention_2":
+            if attention_mask is not None and past_key_values is not None:
+                is_padding_right = (
+                    attention_mask[:, -1].sum().item() != input_tensor.size()[0]
+                )
+                if is_padding_right:
+                    raise ValueError(
+                        "You are attempting to perform batched generation with padding_side='right'"
+                        " this may lead to unexpected behaviour for Flash Attention version of Qwen3. Make sure to "
+                        " call `tokenizer.padding_side  = 'left'` before tokenizing the input. "
+                    )
+            if attention_mask is not None and 0.0 in attention_mask:
+                return attention_mask
+            return None
+        if self.config._attn_implementation == "flex_attention":
+            if isinstance(attention_mask, torch.Tensor):
+                attention_mask = make_flex_block_causal_mask(attention_mask)
+            return attention_mask
+
+        # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+        # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+        # to infer the attention mask.
+        past_seen_tokens = (
+            past_key_values.get_seq_length() if past_key_values is not None else 0
+        )
+        using_static_cache = isinstance(past_key_values, StaticCache)
+        using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
+
+        # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+        if (
+            self.config._attn_implementation == "sdpa"
+            and not (using_static_cache or using_sliding_window_cache)
+            and not output_attentions
+        ):
+            if AttentionMaskConverter._ignore_causal_mask_sdpa(
+                attention_mask,
+                inputs_embeds=input_tensor,
+                past_key_values_length=past_seen_tokens,
+                sliding_window=self.config.sliding_window,
+                is_training=self.training,
+            ):
+                return None
+
+        dtype = input_tensor.dtype
+        min_dtype = torch.finfo(dtype).min
+        sequence_length = input_tensor.shape[1]
+        # SlidingWindowCache or StaticCache
+        if using_sliding_window_cache or using_static_cache:
+            target_length = past_key_values.get_max_cache_shape()
+        # DynamicCache or no cache
+        else:
+            target_length = (
+                attention_mask.shape[-1]
+                if isinstance(attention_mask, torch.Tensor)
+                else past_seen_tokens + sequence_length + 1
+            )
+
+        # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+        causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+            attention_mask,
+            sequence_length=sequence_length,
+            target_length=target_length,
+            dtype=dtype,
+            cache_position=cache_position,
+            batch_size=input_tensor.shape[0],
+            config=self.config,
+            past_key_values=past_key_values,
+        )
+
+        if (
+            self.config._attn_implementation == "sdpa"
+            and attention_mask is not None
+            and attention_mask.device.type in ["cuda", "xpu", "npu"]
+            and not output_attentions
+        ):
+            # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+            # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+            # Details: https://github.com/pytorch/pytorch/issues/110213
+            causal_mask = AttentionMaskConverter._unmask_unattended(
+                causal_mask, min_dtype
+            )
+
+        return causal_mask
+
+    @staticmethod
+    def _prepare_4d_causal_attention_mask_with_cache_position(
+        attention_mask: torch.Tensor,
+        sequence_length: int,
+        target_length: int,
+        dtype: torch.dtype,
+        cache_position: torch.Tensor,
+        batch_size: int,
+        config: Ernie4_5_MoeConfig,
+        past_key_values: Cache,
+    ):
+        """
+        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+        Args:
+            attention_mask (`torch.Tensor`):
+                A 2D attention mask of shape `(batch_size, key_value_length)`, 
+                or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+            sequence_length (`int`):
+                The sequence length being processed.
+            target_length (`int`):
+                The target length: when generating with static cache, the mask should be as long as the static cache,
+                 to account for the 0 padding, the part of the cache that is not filled yet.
+            dtype (`torch.dtype`):
+                The dtype to use for the 4D attention mask.
+            cache_position (`torch.Tensor`):
+                Indices depicting the position of the input sequence tokens in the sequence.
+            batch_size (`torch.Tensor`):
+                Batch size.
+            config (`Ernie4_5_MoeConfig`):
+                The model's configuration class
+            past_key_values (`Cache`):
+                The cache class that is being used currently to generate
+        """
+        if attention_mask is not None and attention_mask.dim() == 4:
+            # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+            causal_mask = attention_mask
+        else:
+            min_dtype = torch.finfo(dtype).min
+            causal_mask = torch.full(
+                (sequence_length, target_length),
+                fill_value=min_dtype,
+                dtype=dtype,
+                device=cache_position.device,
+            )
+            diagonal_attend_mask = torch.arange(
+                target_length, device=cache_position.device
+            ) > cache_position.reshape(-1, 1)
+            text_config = config.get_text_config()
+            if (
+                getattr(text_config, "use_sliding_window", True)
+                and text_config.sliding_window is not None
+            ):
+                if (
+                    not isinstance(past_key_values, SlidingWindowCache)
+                    or sequence_length > target_length
+                ):
+                    sliding_attend_mask = torch.arange(
+                        target_length, device=cache_position.device
+                    ) <= (cache_position.reshape(-1, 1) - text_config.sliding_window)
+                    diagonal_attend_mask.bitwise_or_(sliding_attend_mask)
+            causal_mask *= diagonal_attend_mask
+            causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+            if attention_mask is not None:
+                causal_mask = (
+                    causal_mask.clone()
+                )  # copy to contiguous memory for in-place edit
+                if attention_mask.shape[-1] > target_length:
+                    attention_mask = attention_mask[:, :target_length]
+                mask_length = attention_mask.shape[-1]
+                padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[
+                    :, None, None, :
+                ].to(causal_mask.device)
+                padding_mask = padding_mask == 0
+                causal_mask[:, :, :, :mask_length] = causal_mask[
+                    :, :, :, :mask_length
+                ].masked_fill(padding_mask, min_dtype)
+        return causal_mask
+
+
+@auto_docstring
+class Ernie4_5_MoeForCausalLM(Ernie4_5_PretrainedModel, GenerationMixin):
+    """ERNIE Mixture of Experts (MoE) model for causal language modeling."""
+
+    _tied_weights_keys = ["lm_head.weight"]
+    _tp_plan = {"lm_head": "colwise_rep"}
+    _pp_plan = {"lm_head": (["hidden_states"], ["logits"])}
+
+    def __init__(self, config):
+        """
+        Initializes the ERNIE MoE model for causal language modeling.
+
+        Args:
+            config (dict): Model configuration.
+        """
+        super().__init__(config)
+        self.config = config
+        self.model = Ernie4_5_Model(config)
+        self.lm_head = nn.Linear(
+            config.hidden_size,
+            config.vocab_size,
+            bias=config.weight_share_add_bias and config.use_bias,
+        )  # TODO
+        self.loss_function = ErniePretrainingCriterion(config)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+
+    def get_input_embeddings(self):
+        """Returns the input embeddings layer."""
+        return self.model.embed_tokens
+
+    def set_input_embeddings(self, value):
+        """Sets the input embeddings layer."""
+        self.ernie.embed_tokens = value
+
+    def get_output_embeddings(self):
+        """Returns the output embeddings (LM head)."""
+        return self.lm_head
+
+    def set_output_embeddings(self, new_embeddings):
+        """Sets the output embeddings layer."""
+        self.lm_head = new_embeddings
+
+    def set_decoder(self, decoder):
+        """Sets the ERNIE decoder model."""
+        self.model = decoder
+
+    def get_decoder(self):
+        """Get the transformer decoder."""
+        return self.model
+
+    @can_return_tuple
+    def forward(
+        self,
+        input_ids,
+        attention_mask=None,
+        position_ids=None,
+        past_key_values: Optional[list[torch.FloatTensor]] = None,
+        inputs_embeds=None,
+        labels=None,
+        loss_mask=None,
+        use_cache=False,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        **kwargs: Unpack[KwargsForCausalLM],
+    ):
+        """
+        Forward pass for causal language modeling.
+        """
+        output_attentions = (
+            output_attentions
+            if output_attentions is not None
+            else self.config.output_attentions
+        )
+        output_hidden_states = (
+            output_hidden_states
+            if output_hidden_states is not None
+            else self.config.output_hidden_states
+        )
+
+        outputs = self.model(
+            input_ids,
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            past_key_values=past_key_values,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            **kwargs,
+        )
+
+        hidden_states = outputs.last_hidden_state
+        logits = self.lm_head(hidden_states)
+
+        loss, router_loss = None, None
+        if getattr(self.config, "use_moe", False):
+            router_loss = outputs.router_loss
+
+        if labels is not None:
+            loss, _ = self.loss_function(logits, labels, loss_mask, router_loss)
+
+        return Ernie4_5_MoeCausalLMOutputWithPast(
+            loss=loss,
+            logits=logits,
+            past_key_values=outputs.past_key_values,
+            hidden_states=outputs.hidden_states,
+            attentions=outputs.attentions,
+            router_loss=router_loss,
+        )
+
+
+__all__ = ["Ernie4_5_Model", "Ernie4_5_MoeForCausalLM", "Ernie4_5_PretrainedModel"]

special_tokens_map.json

@@ -0,0 +1 @@
+{"bos_token": "<s>", "eos_token": "</s>", "pad_token": "<unk>", "unk_token": "<unk>", "cls_token": "<|begin_of_sentence|>", "sep_token": "<|end_of_sentence|>", "mask_token": "<mask:1>", "sys_start_token": "<mask:4>", "sys_end_token": "<mask:5>", "header_start_token": "<mask:6>", "header_end_token": "<mask:7>",  "additional_special_tokens": ["<|IMAGE_PLACEHOLDER|>", "<|AUDIO_PLACEHOLDER|>", "<|LOC_0|>", "<|LOC_1|>", "<|LOC_2|>", "<|LOC_3|>", "<|LOC_4|>", "<|LOC_5|>", "<|LOC_6|>", "<|LOC_7|>", "<|LOC_8|>", "<|LOC_9|>", "<|LOC_10|>", "<|LOC_11|>", "<|LOC_12|>", "<|LOC_13|>", "<|LOC_14|>", "<|LOC_15|>", "<|LOC_16|>", "<|LOC_17|>", "<|LOC_18|>", "<|LOC_19|>", "<|LOC_20|>", "<|LOC_21|>", "<|LOC_22|>", "<|LOC_23|>", "<|LOC_24|>", "<|LOC_25|>", "<|LOC_26|>", "<|LOC_27|>", "<|LOC_28|>", "<|LOC_29|>", "<|LOC_30|>", "<|LOC_31|>", "<|LOC_32|>", "<|LOC_33|>", "<|LOC_34|>", "<|LOC_35|>", "<|LOC_36|>", "<|LOC_37|>", "<|LOC_38|>", "<|LOC_39|>", "<|LOC_40|>", "<|LOC_41|>", "<|LOC_42|>", "<|LOC_43|>", "<|LOC_44|>", "<|LOC_45|>", "<|LOC_46|>", "<|LOC_47|>", "<|LOC_48|>", "<|LOC_49|>", "<|LOC_50|>", "<|LOC_51|>", "<|LOC_52|>", "<|LOC_53|>", "<|LOC_54|>", "<|LOC_55|>", "<|LOC_56|>", "<|LOC_57|>", "<|LOC_58|>", "<|LOC_59|>", "<|LOC_60|>", "<|LOC_61|>", "<|LOC_62|>", "<|LOC_63|>", "<|LOC_64|>", "<|LOC_65|>", "<|LOC_66|>", "<|LOC_67|>", "<|LOC_68|>", "<|LOC_69|>", "<|LOC_70|>", "<|LOC_71|>", "<|LOC_72|>", "<|LOC_73|>", "<|LOC_74|>", "<|LOC_75|>", "<|LOC_76|>", "<|LOC_77|>", "<|LOC_78|>", "<|LOC_79|>", "<|LOC_80|>", "<|LOC_81|>", "<|LOC_82|>", "<|LOC_83|>", "<|LOC_84|>", "<|LOC_85|>", "<|LOC_86|>", "<|LOC_87|>", "<|LOC_88|>", "<|LOC_89|>", "<|LOC_90|>", "<|LOC_91|>", "<|LOC_92|>", "<|LOC_93|>", "<|LOC_94|>", "<|LOC_95|>", "<|LOC_96|>", "<|LOC_97|>", "<|LOC_98|>", "<|LOC_99|>", "<|LOC_100|>", "<|LOC_101|>", "<|LOC_102|>", "<|LOC_103|>", "<|LOC_104|>", "<|LOC_105|>", "<|LOC_106|>", "<|LOC_107|>", "<|LOC_108|>", "<|LOC_109|>", "<|LOC_110|>", "<|LOC_111|>", "<|LOC_112|>", "<|LOC_113|>", "<|LOC_114|>", "<|LOC_115|>", "<|LOC_116|>", "<|LOC_117|>", "<|LOC_118|>", "<|LOC_119|>", "<|LOC_120|>", "<|LOC_121|>", "<|LOC_122|>", "<|LOC_123|>", "<|LOC_124|>", "<|LOC_125|>", "<|LOC_126|>", "<|LOC_127|>", "<|LOC_128|>", "<|LOC_129|>", "<|LOC_130|>", "<|LOC_131|>", "<|LOC_132|>", "<|LOC_133|>", "<|LOC_134|>", "<|LOC_135|>", "<|LOC_136|>", "<|LOC_137|>", "<|LOC_138|>", "<|LOC_139|>", "<|LOC_140|>", "<|LOC_141|>", "<|LOC_142|>", "<|LOC_143|>", "<|LOC_144|>", "<|LOC_145|>", "<|LOC_146|>", "<|LOC_147|>", "<|LOC_148|>", "<|LOC_149|>", "<|LOC_150|>", "<|LOC_151|>", "<|LOC_152|>", "<|LOC_153|>", "<|LOC_154|>", "<|LOC_155|>", "<|LOC_156|>", "<|LOC_157|>", "<|LOC_158|>", "<|LOC_159|>", "<|LOC_160|>", "<|LOC_161|>", "<|LOC_162|>", "<|LOC_163|>", "<|LOC_164|>", "<|LOC_165|>", "<|LOC_166|>", "<|LOC_167|>", "<|LOC_168|>", "<|LOC_169|>", "<|LOC_170|>", "<|LOC_171|>", "<|LOC_172|>", "<|LOC_173|>", "<|LOC_174|>", "<|LOC_175|>", "<|LOC_176|>", "<|LOC_177|>", "<|LOC_178|>", "<|LOC_179|>", "<|LOC_180|>", "<|LOC_181|>", "<|LOC_182|>", "<|LOC_183|>", "<|LOC_184|>", "<|LOC_185|>", "<|LOC_186|>", "<|LOC_187|>", "<|LOC_188|>", "<|LOC_189|>", "<|LOC_190|>", "<|LOC_191|>", "<|LOC_192|>", "<|LOC_193|>", "<|LOC_194|>", "<|LOC_195|>", "<|LOC_196|>", "<|LOC_197|>", "<|LOC_198|>", "<|LOC_199|>", "<|LOC_200|>", "<|LOC_201|>", "<|LOC_202|>", "<|LOC_203|>", "<|LOC_204|>", "<|LOC_205|>", "<|LOC_206|>", "<|LOC_207|>", "<|LOC_208|>", "<|LOC_209|>", "<|LOC_210|>", "<|LOC_211|>", "<|LOC_212|>", "<|LOC_213|>", "<|LOC_214|>", "<|LOC_215|>", "<|LOC_216|>", "<|LOC_217|>", "<|LOC_218|>", "<|LOC_219|>", "<|LOC_220|>", "<|LOC_221|>", "<|LOC_222|>", "<|LOC_223|>", "<|LOC_224|>", "<|LOC_225|>", "<|LOC_226|>", "<|LOC_227|>", "<|LOC_228|>", "<|LOC_229|>", "<|LOC_230|>", "<|LOC_231|>", "<|LOC_232|>", "<|LOC_233|>", "<|LOC_234|>", "<|LOC_235|>", "<|LOC_236|>", "<|LOC_237|>", "<|LOC_238|>", "<|LOC_239|>", "<|LOC_240|>", "<|LOC_241|>", "<|LOC_242|>", "<|LOC_243|>", "<|LOC_244|>", "<|LOC_245|>", "<|LOC_246|>", "<|LOC_247|>", "<|LOC_248|>", "<|LOC_249|>", "<|LOC_250|>", "<|LOC_251|>", "<|LOC_252|>", "<|LOC_253|>", "<|LOC_254|>", "<|LOC_255|>", "<|LOC_256|>", "<|LOC_257|>", "<|LOC_258|>", "<|LOC_259|>", "<|LOC_260|>", "<|LOC_261|>", "<|LOC_262|>", "<|LOC_263|>", "<|LOC_264|>", "<|LOC_265|>", "<|LOC_266|>", "<|LOC_267|>", "<|LOC_268|>", "<|LOC_269|>", "<|LOC_270|>", "<|LOC_271|>", "<|LOC_272|>", "<|LOC_273|>", "<|LOC_274|>", "<|LOC_275|>", "<|LOC_276|>", "<|LOC_277|>", "<|LOC_278|>", "<|LOC_279|>", "<|LOC_280|>", "<|LOC_281|>", "<|LOC_282|>", "<|LOC_283|>", "<|LOC_284|>", "<|LOC_285|>", "<|LOC_286|>", "<|LOC_287|>", "<|LOC_288|>", "<|LOC_289|>", "<|LOC_290|>", "<|LOC_291|>", "<|LOC_292|>", "<|LOC_293|>", "<|LOC_294|>", "<|LOC_295|>", "<|LOC_296|>", "<|LOC_297|>", "<|LOC_298|>", "<|LOC_299|>", "<|LOC_300|>", "<|LOC_301|>", "<|LOC_302|>", "<|LOC_303|>", "<|LOC_304|>", "<|LOC_305|>", "<|LOC_306|>", "<|LOC_307|>", "<|LOC_308|>", "<|LOC_309|>", "<|LOC_310|>", "<|LOC_311|>", "<|LOC_312|>", "<|LOC_313|>", "<|LOC_314|>", "<|LOC_315|>", "<|LOC_316|>", "<|LOC_317|>", "<|LOC_318|>", "<|LOC_319|>", "<|LOC_320|>", "<|LOC_321|>", "<|LOC_322|>", "<|LOC_323|>", "<|LOC_324|>", "<|LOC_325|>", "<|LOC_326|>", "<|LOC_327|>", "<|LOC_328|>", "<|LOC_329|>", "<|LOC_330|>", "<|LOC_331|>", "<|LOC_332|>", "<|LOC_333|>", "<|LOC_334|>", "<|LOC_335|>", "<|LOC_336|>", "<|LOC_337|>", "<|LOC_338|>", "<|LOC_339|>", "<|LOC_340|>", "<|LOC_341|>", "<|LOC_342|>", "<|LOC_343|>", "<|LOC_344|>", "<|LOC_345|>", "<|LOC_346|>", "<|LOC_347|>", "<|LOC_348|>", "<|LOC_349|>", "<|LOC_350|>", "<|LOC_351|>", "<|LOC_352|>", "<|LOC_353|>", "<|LOC_354|>", "<|LOC_355|>", "<|LOC_356|>", "<|LOC_357|>", "<|LOC_358|>", "<|LOC_359|>", "<|LOC_360|>", "<|LOC_361|>", "<|LOC_362|>", "<|LOC_363|>", "<|LOC_364|>", "<|LOC_365|>", "<|LOC_366|>", "<|LOC_367|>", "<|LOC_368|>", "<|LOC_369|>", "<|LOC_370|>", "<|LOC_371|>", "<|LOC_372|>", "<|LOC_373|>", "<|LOC_374|>", "<|LOC_375|>", "<|LOC_376|>", "<|LOC_377|>", "<|LOC_378|>", "<|LOC_379|>", "<|LOC_380|>", "<|LOC_381|>", "<|LOC_382|>", "<|LOC_383|>", "<|LOC_384|>", "<|LOC_385|>", "<|LOC_386|>", "<|LOC_387|>", "<|LOC_388|>", "<|LOC_389|>", "<|LOC_390|>", "<|LOC_391|>", "<|LOC_392|>", "<|LOC_393|>", "<|LOC_394|>", "<|LOC_395|>", "<|LOC_396|>", "<|LOC_397|>", "<|LOC_398|>", "<|LOC_399|>", "<|LOC_400|>", "<|LOC_401|>", "<|LOC_402|>", "<|LOC_403|>", "<|LOC_404|>", "<|LOC_405|>", "<|LOC_406|>", "<|LOC_407|>", "<|LOC_408|>", "<|LOC_409|>", "<|LOC_410|>", "<|LOC_411|>", "<|LOC_412|>", "<|LOC_413|>", "<|LOC_414|>", "<|LOC_415|>", "<|LOC_416|>", "<|LOC_417|>", "<|LOC_418|>", "<|LOC_419|>", "<|LOC_420|>", "<|LOC_421|>", "<|LOC_422|>", "<|LOC_423|>", "<|LOC_424|>", "<|LOC_425|>", "<|LOC_426|>", "<|LOC_427|>", "<|LOC_428|>", "<|LOC_429|>", "<|LOC_430|>", "<|LOC_431|>", "<|LOC_432|>", "<|LOC_433|>", "<|LOC_434|>", "<|LOC_435|>", "<|LOC_436|>", "<|LOC_437|>", "<|LOC_438|>", "<|LOC_439|>", "<|LOC_440|>", "<|LOC_441|>", "<|LOC_442|>", "<|LOC_443|>", "<|LOC_444|>", "<|LOC_445|>", "<|LOC_446|>", "<|LOC_447|>", "<|LOC_448|>", "<|LOC_449|>", "<|LOC_450|>", "<|LOC_451|>", "<|LOC_452|>", "<|LOC_453|>", "<|LOC_454|>", "<|LOC_455|>", "<|LOC_456|>", "<|LOC_457|>", "<|LOC_458|>", "<|LOC_459|>", "<|LOC_460|>", "<|LOC_461|>", "<|LOC_462|>", "<|LOC_463|>", "<|LOC_464|>", "<|LOC_465|>", "<|LOC_466|>", "<|LOC_467|>", "<|LOC_468|>", "<|LOC_469|>", "<|LOC_470|>", "<|LOC_471|>", "<|LOC_472|>", "<|LOC_473|>", "<|LOC_474|>", "<|LOC_475|>", "<|LOC_476|>", "<|LOC_477|>", "<|LOC_478|>", "<|LOC_479|>", "<|LOC_480|>", "<|LOC_481|>", "<|LOC_482|>", "<|LOC_483|>", "<|LOC_484|>", "<|LOC_485|>", "<|LOC_486|>", "<|LOC_487|>", "<|LOC_488|>", "<|LOC_489|>", "<|LOC_490|>", "<|LOC_491|>", "<|LOC_492|>", "<|LOC_493|>", "<|LOC_494|>", "<|LOC_495|>", "<|LOC_496|>", "<|LOC_497|>", "<|LOC_498|>", "<|LOC_499|>", "<|LOC_500|>", "<|LOC_501|>", "<|LOC_502|>", "<|LOC_503|>", "<|LOC_504|>", "<|LOC_505|>", "<|LOC_506|>", "<|LOC_507|>", "<|LOC_508|>", "<|LOC_509|>", "<|LOC_510|>", "<|LOC_511|>", "<|LOC_512|>", "<|LOC_513|>", "<|LOC_514|>", "<|LOC_515|>", "<|LOC_516|>", "<|LOC_517|>", "<|LOC_518|>", "<|LOC_519|>", "<|LOC_520|>", "<|LOC_521|>", "<|LOC_522|>", "<|LOC_523|>", "<|LOC_524|>", "<|LOC_525|>", "<|LOC_526|>", "<|LOC_527|>", "<|LOC_528|>", "<|LOC_529|>", "<|LOC_530|>", "<|LOC_531|>", "<|LOC_532|>", "<|LOC_533|>", "<|LOC_534|>", "<|LOC_535|>", "<|LOC_536|>", "<|LOC_537|>", "<|LOC_538|>", "<|LOC_539|>", "<|LOC_540|>", "<|LOC_541|>", "<|LOC_542|>", "<|LOC_543|>", "<|LOC_544|>", "<|LOC_545|>", "<|LOC_546|>", "<|LOC_547|>", "<|LOC_548|>", "<|LOC_549|>", "<|LOC_550|>", "<|LOC_551|>", "<|LOC_552|>", "<|LOC_553|>", "<|LOC_554|>", "<|LOC_555|>", "<|LOC_556|>", "<|LOC_557|>", "<|LOC_558|>", "<|LOC_559|>", "<|LOC_560|>", "<|LOC_561|>", "<|LOC_562|>", "<|LOC_563|>", "<|LOC_564|>", "<|LOC_565|>", "<|LOC_566|>", "<|LOC_567|>", "<|LOC_568|>", "<|LOC_569|>", "<|LOC_570|>", "<|LOC_571|>", "<|LOC_572|>", "<|LOC_573|>", "<|LOC_574|>", "<|LOC_575|>", "<|LOC_576|>", "<|LOC_577|>", "<|LOC_578|>", "<|LOC_579|>", "<|LOC_580|>", "<|LOC_581|>", "<|LOC_582|>", "<|LOC_583|>", "<|LOC_584|>", "<|LOC_585|>", "<|LOC_586|>", "<|LOC_587|>", "<|LOC_588|>", "<|LOC_589|>", "<|LOC_590|>", "<|LOC_591|>", "<|LOC_592|>", "<|LOC_593|>", "<|LOC_594|>", "<|LOC_595|>", "<|LOC_596|>", "<|LOC_597|>", "<|LOC_598|>", "<|LOC_599|>", "<|LOC_600|>", "<|LOC_601|>", "<|LOC_602|>", "<|LOC_603|>", "<|LOC_604|>", "<|LOC_605|>", "<|LOC_606|>", "<|LOC_607|>", "<|LOC_608|>", "<|LOC_609|>", "<|LOC_610|>", "<|LOC_611|>", "<|LOC_612|>", "<|LOC_613|>", "<|LOC_614|>", "<|LOC_615|>", "<|LOC_616|>", "<|LOC_617|>", "<|LOC_618|>", "<|LOC_619|>", "<|LOC_620|>", "<|LOC_621|>", "<|LOC_622|>", "<|LOC_623|>", "<|LOC_624|>", "<|LOC_625|>", "<|LOC_626|>", "<|LOC_627|>", "<|LOC_628|>", "<|LOC_629|>", "<|LOC_630|>", "<|LOC_631|>", "<|LOC_632|>", "<|LOC_633|>", "<|LOC_634|>", "<|LOC_635|>", "<|LOC_636|>", "<|LOC_637|>", "<|LOC_638|>", "<|LOC_639|>", "<|LOC_640|>", "<|LOC_641|>", "<|LOC_642|>", "<|LOC_643|>", "<|LOC_644|>", "<|LOC_645|>", "<|LOC_646|>", "<|LOC_647|>", "<|LOC_648|>", "<|LOC_649|>", "<|LOC_650|>", "<|LOC_651|>", "<|LOC_652|>", "<|LOC_653|>", "<|LOC_654|>", "<|LOC_655|>", "<|LOC_656|>", "<|LOC_657|>", "<|LOC_658|>", "<|LOC_659|>", "<|LOC_660|>", "<|LOC_661|>", "<|LOC_662|>", "<|LOC_663|>", "<|LOC_664|>", "<|LOC_665|>", "<|LOC_666|>", "<|LOC_667|>", "<|LOC_668|>", "<|LOC_669|>", "<|LOC_670|>", "<|LOC_671|>", "<|LOC_672|>", "<|LOC_673|>", "<|LOC_674|>", "<|LOC_675|>", "<|LOC_676|>", "<|LOC_677|>", "<|LOC_678|>", "<|LOC_679|>", "<|LOC_680|>", "<|LOC_681|>", "<|LOC_682|>", "<|LOC_683|>", "<|LOC_684|>", "<|LOC_685|>", "<|LOC_686|>", "<|LOC_687|>", "<|LOC_688|>", "<|LOC_689|>", "<|LOC_690|>", "<|LOC_691|>", "<|LOC_692|>", "<|LOC_693|>", "<|LOC_694|>", "<|LOC_695|>", "<|LOC_696|>", "<|LOC_697|>", "<|LOC_698|>", "<|LOC_699|>", "<|LOC_700|>", "<|LOC_701|>", "<|LOC_702|>", "<|LOC_703|>", "<|LOC_704|>", "<|LOC_705|>", "<|LOC_706|>", "<|LOC_707|>", "<|LOC_708|>", "<|LOC_709|>", "<|LOC_710|>", "<|LOC_711|>", "<|LOC_712|>", "<|LOC_713|>", "<|LOC_714|>", "<|LOC_715|>", "<|LOC_716|>", "<|LOC_717|>", "<|LOC_718|>", "<|LOC_719|>", "<|LOC_720|>", "<|LOC_721|>", "<|LOC_722|>", "<|LOC_723|>", "<|LOC_724|>", "<|LOC_725|>", "<|LOC_726|>", "<|LOC_727|>", "<|LOC_728|>", "<|LOC_729|>", "<|LOC_730|>", "<|LOC_731|>", "<|LOC_732|>", "<|LOC_733|>", "<|LOC_734|>", "<|LOC_735|>", "<|LOC_736|>", "<|LOC_737|>", "<|LOC_738|>", "<|LOC_739|>", "<|LOC_740|>", "<|LOC_741|>", "<|LOC_742|>", "<|LOC_743|>", "<|LOC_744|>", "<|LOC_745|>", "<|LOC_746|>", "<|LOC_747|>", "<|LOC_748|>", "<|LOC_749|>", "<|LOC_750|>", "<|LOC_751|>", "<|LOC_752|>", "<|LOC_753|>", "<|LOC_754|>", "<|LOC_755|>", "<|LOC_756|>", "<|LOC_757|>", "<|LOC_758|>", "<|LOC_759|>", "<|LOC_760|>", "<|LOC_761|>", "<|LOC_762|>", "<|LOC_763|>", "<|LOC_764|>", "<|LOC_765|>", "<|LOC_766|>", "<|LOC_767|>", "<|LOC_768|>", "<|LOC_769|>", "<|LOC_770|>", "<|LOC_771|>", "<|LOC_772|>", "<|LOC_773|>", "<|LOC_774|>", "<|LOC_775|>", "<|LOC_776|>", "<|LOC_777|>", "<|LOC_778|>", "<|LOC_779|>", "<|LOC_780|>", "<|LOC_781|>", "<|LOC_782|>", "<|LOC_783|>", "<|LOC_784|>", "<|LOC_785|>", "<|LOC_786|>", "<|LOC_787|>", "<|LOC_788|>", "<|LOC_789|>", "<|LOC_790|>", "<|LOC_791|>", "<|LOC_792|>", "<|LOC_793|>", "<|LOC_794|>", "<|LOC_795|>", "<|LOC_796|>", "<|LOC_797|>", "<|LOC_798|>", "<|LOC_799|>", "<|LOC_800|>", "<|LOC_801|>", "<|LOC_802|>", "<|LOC_803|>", "<|LOC_804|>", "<|LOC_805|>", "<|LOC_806|>", "<|LOC_807|>", "<|LOC_808|>", "<|LOC_809|>", "<|LOC_810|>", "<|LOC_811|>", "<|LOC_812|>", "<|LOC_813|>", "<|LOC_814|>", "<|LOC_815|>", "<|LOC_816|>", "<|LOC_817|>", "<|LOC_818|>", "<|LOC_819|>", "<|LOC_820|>", "<|LOC_821|>", "<|LOC_822|>", "<|LOC_823|>", "<|LOC_824|>", "<|LOC_825|>", "<|LOC_826|>", "<|LOC_827|>", "<|LOC_828|>", "<|LOC_829|>", "<|LOC_830|>", "<|LOC_831|>", "<|LOC_832|>", "<|LOC_833|>", "<|LOC_834|>", "<|LOC_835|>", "<|LOC_836|>", "<|LOC_837|>", "<|LOC_838|>", "<|LOC_839|>", "<|LOC_840|>", "<|LOC_841|>", "<|LOC_842|>", "<|LOC_843|>", "<|LOC_844|>", "<|LOC_845|>", "<|LOC_846|>", "<|LOC_847|>", "<|LOC_848|>", "<|LOC_849|>", "<|LOC_850|>", "<|LOC_851|>", "<|LOC_852|>", "<|LOC_853|>", "<|LOC_854|>", "<|LOC_855|>", "<|LOC_856|>", "<|LOC_857|>", "<|LOC_858|>", "<|LOC_859|>", "<|LOC_860|>", "<|LOC_861|>", "<|LOC_862|>", "<|LOC_863|>", "<|LOC_864|>", "<|LOC_865|>", "<|LOC_866|>", "<|LOC_867|>", "<|LOC_868|>", "<|LOC_869|>", "<|LOC_870|>", "<|LOC_871|>", "<|LOC_872|>", "<|LOC_873|>", "<|LOC_874|>", "<|LOC_875|>", "<|LOC_876|>", "<|LOC_877|>", "<|LOC_878|>", "<|LOC_879|>", "<|LOC_880|>", "<|LOC_881|>", "<|LOC_882|>", "<|LOC_883|>", "<|LOC_884|>", "<|LOC_885|>", "<|LOC_886|>", "<|LOC_887|>", "<|LOC_888|>", "<|LOC_889|>", "<|LOC_890|>", "<|LOC_891|>", "<|LOC_892|>", "<|LOC_893|>", "<|LOC_894|>", "<|LOC_895|>", "<|LOC_896|>", "<|LOC_897|>", "<|LOC_898|>", "<|LOC_899|>", "<|LOC_900|>", "<|LOC_901|>", "<|LOC_902|>", "<|LOC_903|>", "<|LOC_904|>", "<|LOC_905|>", "<|LOC_906|>", "<|LOC_907|>", "<|LOC_908|>", "<|LOC_909|>", "<|LOC_910|>", "<|LOC_911|>", "<|LOC_912|>", "<|LOC_913|>", "<|LOC_914|>", "<|LOC_915|>", "<|LOC_916|>", "<|LOC_917|>", "<|LOC_918|>", "<|LOC_919|>", "<|LOC_920|>", "<|LOC_921|>", "<|LOC_922|>", "<|LOC_923|>", "<|LOC_924|>", "<|LOC_925|>", "<|LOC_926|>", "<|LOC_927|>", "<|LOC_928|>", "<|LOC_929|>", "<|LOC_930|>", "<|LOC_931|>", "<|LOC_932|>", "<|LOC_933|>", "<|LOC_934|>", "<|LOC_935|>", "<|LOC_936|>", "<|LOC_937|>", "<|LOC_938|>", "<|LOC_939|>", "<|LOC_940|>", "<|LOC_941|>", "<|LOC_942|>", "<|LOC_943|>", "<|LOC_944|>", "<|LOC_945|>", "<|LOC_946|>", "<|LOC_947|>", "<|LOC_948|>", "<|LOC_949|>", "<|LOC_950|>", "<|LOC_951|>", "<|LOC_952|>", "<|LOC_953|>", "<|LOC_954|>", "<|LOC_955|>", "<|LOC_956|>", "<|LOC_957|>", "<|LOC_958|>", "<|LOC_959|>", "<|LOC_960|>", "<|LOC_961|>", "<|LOC_962|>", "<|LOC_963|>", "<|LOC_964|>", "<|LOC_965|>", "<|LOC_966|>", "<|LOC_967|>", "<|LOC_968|>", "<|LOC_969|>", "<|LOC_970|>", "<|LOC_971|>", "<|LOC_972|>", "<|LOC_973|>", "<|LOC_974|>", "<|LOC_975|>", "<|LOC_976|>", "<|LOC_977|>", "<|LOC_978|>", "<|LOC_979|>", "<|LOC_980|>", "<|LOC_981|>", "<|LOC_982|>", "<|LOC_983|>", "<|LOC_984|>", "<|LOC_985|>", "<|LOC_986|>", "<|LOC_987|>", "<|LOC_988|>", "<|LOC_989|>", "<|LOC_990|>", "<|LOC_991|>", "<|LOC_992|>", "<|LOC_993|>", "<|LOC_994|>", "<|LOC_995|>", "<|LOC_996|>", "<|LOC_997|>", "<|LOC_998|>", "<|LOC_999|>", "<|LOC_1000|>", "<|LOC_BEGIN|>", "<|LOC_END|>", "<|LOC_SEP|>", "<|CROP_COL_SEP|>", "<|CROP_ROW_SEP|>", "<|IMAGE_SEP|>"]}

tokenization_ernie4_5.py

@@ -0,0 +1,353 @@
+# Copyright (c) 2025 Baidu, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Ernie4_5_Tokenizer"""
+
+import os
+from shutil import copyfile
+from typing import Dict, List, Optional, Tuple, Union
+import torch
+import numpy as np
+import sentencepiece as spm
+
+from transformers.tokenization_utils import PreTrainedTokenizer
+from transformers.tokenization_utils_base import (
+    PaddingStrategy,
+)
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+
+class Ernie4_5_Tokenizer(PreTrainedTokenizer):
+    """
+    Ernie4_5_Tokenizer
+        vocab_files_names (dict): Mapping vocabulary-related config name to actual filename.
+        model_input_names (List): Model input names expected by the tokenizer
+        padding_side (str): Padding side (where to add padding tokens)
+    """
+    vocab_files_names = {
+        "vocab_file": "tokenizer.model",
+    }
+    model_input_names = ["input_ids", "position_ids", "attention_mask", "labels"]
+    padding_side = "right"
+
+    def __init__(
+        self,
+        vocab_file,
+        bos_token="<s>",
+        cls_token="<cls>",
+        eos_token="</s>",
+        mask_token="<mask:0>",
+        pad_token="<pad>",
+        sep_token="<sep>",
+        unk_token="<unk>",
+        additional_special_tokens=None,
+        split_special_tokens=False,
+        tokenizer_alpha=None,
+        **kwargs
+    ):
+        """
+        Initialize the ERNIE tokenizer.
+
+        Args:
+            vocab_file (str): Path to the SentencePiece model file.
+            bos_token (str, optional): Beginning of sentence token. Defaults to "<s>".
+            cls_token (str, optional): Classification token. Defaults to "<cls>".
+            eos_token (str, optional): End of sentence token. Defaults to "</s>".
+            mask_token (str, optional): Mask token. Defaults to "<mask:0>".
+            pad_token (str, optional): Padding token. Defaults to "<pad>".
+            sep_token (str, optional): Separator token. Defaults to "<sep>".
+            unk_token (str, optional): Unknown token. Defaults to "<unk>".
+            additional_special_tokens (List[str], optional): Additional special tokens.
+                Defaults to ["<mask:1>", "<mask:7>"].
+            split_special_tokens (bool, optional): Whether to split special tokens. Defaults to False.
+            tokenizer_alpha (float, optional): Alpha parameter for SentencePiece sampling.
+            **kwargs: Additional keyword arguments passed to the parent class.
+        """
+
+        self.vocab_file = vocab_file
+        self.sp_model = spm.SentencePieceProcessor()
+        self.sp_model.Load(vocab_file)
+        self.pad_id = self._convert_token_to_id(pad_token)
+        self.tokenizer_alpha = tokenizer_alpha
+
+        if additional_special_tokens is None:
+            additional_special_tokens = ["<mask:1>", "<mask:7>"]
+        super().__init__(
+            bos_token=bos_token,
+            cls_token=cls_token,
+            eos_token=eos_token,
+            mask_token=mask_token,
+            pad_token=pad_token,
+            sep_token=sep_token,
+            unk_token=unk_token,
+            additional_special_tokens=additional_special_tokens,
+            split_special_tokens=split_special_tokens,
+            **kwargs,
+        )
+
+    @property
+    def vocab_size(self):
+        """Returns the size of the vocabulary.
+
+        Returns:
+            int: The number of tokens in the vocabulary.
+        """
+        return self.sp_model.vocab_size()
+
+    def get_vocab(self):
+        """Get the vocabulary as a dictionary mapping tokens to their IDs.
+
+        Returns:
+            dict: A dictionary mapping tokens to their corresponding IDs.
+        """
+        vocab = {self.convert_ids_to_tokens(i): i for i in range(self.vocab_size)}
+        vocab.update(self.added_tokens_encoder)
+        return vocab
+
+    def _tokenize(self, text):
+        """Tokenize text using SentencePiece.
+
+        Args:
+            text (str): The text to tokenize.
+
+        Returns:
+            list: A list of tokens.
+        """
+        if self.tokenizer_alpha is not None:
+            return self.sp_model.encode_as_pieces(
+                text,
+                enable_sampling=True,
+                nbest_size=-1,
+                alpha=self.tokenizer_alpha,
+            )
+        else:
+            return self.sp_model.encode_as_pieces(text)
+
+    def _convert_token_to_id(self, token):
+        """Convert a token (str) to an ID using the vocabulary.
+
+        Args:
+            token (str): The token to convert.
+
+        Returns:
+            int: The corresponding token ID.
+        """
+        return self.sp_model.piece_to_id(token)
+
+    def _convert_id_to_token(self, id):
+        """Convert an ID to a token (str) using the vocabulary.
+
+        Args:
+            id (int): The token ID to convert.
+
+        Returns:
+            str: The corresponding token.
+        """
+        if id >= self.vocab_size:
+            return self.unk_token
+        else:
+            return self.sp_model.id_to_piece(id)
+
+    def convert_tokens_to_string(self, tokens):
+        """Convert a sequence of tokens back to a single string.
+
+        Args:
+            tokens (List[str]): A list of tokens to convert.
+
+        Returns:
+            str: The reconstructed string.
+        """
+        current_sub_tokens = []
+        out_string = ""
+        prev_is_special = False
+        for token in tokens:
+            # make sure that special tokens are not decoded using sentencepiece model
+            if token in self.all_special_tokens:
+                if not prev_is_special:
+                    out_string += " "
+                out_string += self.sp_model.decode(current_sub_tokens) + token
+                prev_is_special = True
+                current_sub_tokens = []
+            else:
+                current_sub_tokens.append(token)
+                prev_is_special = False
+        out_string += self.sp_model.decode(current_sub_tokens)
+        return out_string
+
+    def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
+        """Build model inputs by adding special tokens to sequences.
+
+        Args:
+            token_ids_0 (List[int]): List of token IDs for the first sequence.
+            token_ids_1 (List[int], optional): List of token IDs for the second sequence.
+
+        Returns:
+            List[int]: List of token IDs with special tokens added.
+        """
+        output = token_ids_0
+        last_cls_index = -1
+        last_sep_index = -1
+        if self.cls_token_id in output:
+            last_cls_index = len(output) - output[::-1].index(self.cls_token_id) - 1
+        if self.sep_token_id in output:
+            last_sep_index = len(output) - output[::-1].index(self.sep_token_id) - 1
+
+        if last_cls_index > last_sep_index:
+            next_token_id = self.sep_token_id
+        elif last_sep_index > last_cls_index:
+            next_token_id = self.cls_token_id
+        else:
+            output = [self.cls_token_id] + token_ids_0 + [self.sep_token_id]
+            next_token_id = self.cls_token_id
+
+        output = [self.bos_token_id] + output
+        # Assume no markup in text if token_ids_1 is given.
+        if token_ids_1 is not None:
+            output = output + token_ids_1 + [next_token_id]
+        return output
+
+    def get_special_tokens_mask(self, token_ids_0, token_ids_1=None, already_has_special_tokens=False):
+        """Get a mask showing which tokens are special tokens.
+
+        Args:
+            token_ids_0 (List[int]): List of token IDs for the first sequence.
+            token_ids_1 (List[int], optional): List of token IDs for the second sequence.
+            already_has_special_tokens (bool): Whether the tokens already include special tokens.
+
+        Returns:
+            List[int]: A mask where 1 indicates special tokens and 0 indicates regular tokens.
+        """
+        if already_has_special_tokens:
+            return super().get_special_tokens_mask(token_ids_0, token_ids_1, already_has_special_tokens=True)
+
+        # [bos_token, cls_token, tokens_0, sep_token]
+        if token_ids_1 is None:
+            return [1, 1] + ([0] * len(token_ids_0)) + [1]
+        # [bos_token, cls_token, tokens_0, sep_token, tokens_1, cls_token]
+        return [1, 1] + ([0] * len(token_ids_0)) + [1] + ([0] * len(token_ids_1)) + [1]
+
+    def save_vocabulary(self, save_directory, filename_prefix: Optional[str] = None) -> Tuple[str]:
+        """
+        Save the vocabulary and special tokens file to a directory.
+
+        Args:
+            save_directory (str): The directory in which to save the vocabulary.
+            filename_prefix (Optional[str]): Optional prefix for the saved filename.
+
+        Returns:
+            Tuple[str]: Paths to the files saved.
+
+        Raises:
+            ValueError: If the save_directory is not a valid directory.
+        """
+        if not os.path.isdir(save_directory):
+            logger.error(f"Vocabulary path ({save_directory}) should be a directory")
+            return
+        out_vocab_file = os.path.join(
+            save_directory,
+            (filename_prefix + "-" if filename_prefix else "") + self.resource_files_names["vocab_file"],
+        )
+
+        if os.path.abspath(self.vocab_file) != os.path.abspath(out_vocab_file) and os.path.isfile(self.vocab_file):
+            copyfile(self.vocab_file, out_vocab_file)
+        elif not os.path.isfile(self.vocab_file):
+            with open(out_vocab_file, "wb") as fi:
+                content_spiece_model = self.sp_model.serialized_model_proto()
+                fi.write(content_spiece_model)
+
+        return (out_vocab_file,)
+
+    def _pad(
+        self,
+        encoded_inputs: Union[Dict],
+        max_length: Optional[int] = None,
+        padding_strategy: PaddingStrategy = PaddingStrategy.DO_NOT_PAD,
+        pad_to_multiple_of: Optional[int] = None,
+        padding_side: Optional[str] = None,
+        return_attention_mask: Optional[bool] = None,
+    ) -> dict:
+        """
+        Pad encoded inputs according to specified strategy.
+
+        Args:
+            encoded_inputs (Union[Dict]): Dictionary of encoded inputs.
+            max_length (Optional[int]): Maximum length to pad to.
+            padding_strategy (PaddingStrategy): Strategy for padding.
+            pad_to_multiple_of (Optional[int]): Pad to a multiple of this value.
+            return_attention_mask (Optional[bool]): Whether to return attention mask.
+
+        Returns:
+            dict: Dictionary with padded inputs and optional attention mask.
+
+        Raises:
+            ValueError: If attention_mask has unexpected type or invalid padding strategy.
+        """
+        if return_attention_mask is None:
+            return_attention_mask = "attention_mask" in self.model_input_names
+        if return_attention_mask:
+            required_input = encoded_inputs[self.model_input_names[0]]
+            if padding_strategy == PaddingStrategy.LONGEST:
+                max_length = len(required_input)
+            if max_length is not None and pad_to_multiple_of is not None and (max_length % pad_to_multiple_of != 0):
+                max_length = ((max_length // pad_to_multiple_of) + 1) * pad_to_multiple_of
+            needs_to_be_padded = padding_strategy != PaddingStrategy.DO_NOT_PAD and len(required_input) != max_length
+
+            if "attention_mask" in encoded_inputs and encoded_inputs["attention_mask"] is not None:
+                attention_mask = encoded_inputs.pop("attention_mask")
+                if isinstance(attention_mask, torch.Tensor):
+                    attention_mask = attention_mask.numpy()
+                elif isinstance(attention_mask, list):
+                    attention_mask = np.array(attention_mask)
+                elif not isinstance(attention_mask, np.ndarray):
+                    raise ValueError(f"Unexpected type {type(attention_mask)} of attention_mask, ")
+            else:
+                # Create default attention mask if none provided
+                attention_mask = np.tril(np.ones((len(required_input), len(required_input)), dtype=np.int64))
+                attention_mask = np.expand_dims(attention_mask, axis=0)
+
+            if needs_to_be_padded:
+                difference = max_length - len(required_input)
+                if self.padding_side == "right":
+                    if attention_mask.ndim == 1:
+                        pad_width = [(0, difference)]
+                    else:
+                        pad_width = [(0, 0), (0, difference), (0, difference)]
+                elif self.padding_side == "left":
+                    if attention_mask.ndim == 1:
+                        pad_width = [(difference, 0)]
+                    else:
+                        pad_width = [(0, 0), (difference, 0), (difference, 0)]
+                else:
+                    raise ValueError("Invalid padding strategy:" + str(self.padding_side))
+                attention_mask = np.pad(
+                    attention_mask,
+                    pad_width=pad_width,
+                    mode="constant",
+                    constant_values=0,
+                )
+
+        encoded_inputs = super()._pad(
+            encoded_inputs,
+            max_length,
+            padding_strategy=padding_strategy,
+            pad_to_multiple_of=pad_to_multiple_of,
+            return_attention_mask=False,
+        )
+        if return_attention_mask:
+            encoded_inputs["attention_mask"] = attention_mask.tolist()
+        return encoded_inputs
+

tokenizer.model

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:34ef7db83df785924fb83d7b887b6e822a031c56e15cff40aaf9b982988180df
+size 1614363

tokenizer_config.json

@@ -0,0 +1,22 @@
+{
+    "bos_token": "<s>",
+    "eos_token": "</s>",
+    "pad_token": "<unk>",
+    "unk_token": "<unk>",
+    "cls_token": "<|begin_of_sentence|>",
+    "sep_token": "<|end_of_sentence|>",
+    "mask_token": "<mask:1>",
+    "sys_start_token": "<mask:4>",
+    "sys_end_token": "<mask:5>",
+    "header_start_token": "<mask:6>",
+    "header_end_token": "<mask:7>",
+    "additional_special_tokens": null,
+    "chat_template": "{%- if not add_generation_prompt is defined -%}\n    {%- set add_generation_prompt = true -%}\n{%- endif -%}\n{%- if not cls_token is defined -%}\n    {%- set cls_token = \"<|begin_of_sentence|>\" -%}\n{%- endif -%}\n{%- if not sep_token is defined -%}\n    {%- set sep_token = \"<|end_of_sentence|>\" -%}\n{%- endif -%}\n{{- cls_token -}}\n{%- for message in messages -%}\n    {%- if message[\"role\"] == \"user\" -%}\n        {{- \"User: \" + message[\"content\"] + \"\n\" -}}\n    {%- elif message[\"role\"] == \"assistant\" -%}\n        {{- \"Assistant: \" + message[\"content\"] + sep_token -}}\n    {%- elif message[\"role\"] == \"system\" -%}\n        {{- message[\"content\"] + \"\n\" -}}\n    {%- endif -%}\n{%- endfor -%}\n{%- if add_generation_prompt -%}\n    {{- \"Assistant: \" -}}\n{%- endif -%}",
+    "tokenizer_class": "Ernie4_5_Tokenizer",
+    "auto_map": {
+        "AutoTokenizer": [
+            "tokenization_ernie4_5.Ernie4_5_Tokenizer",
+            "tokenization_ernie4_5.Ernie4_5_Tokenizer"
+        ]
+    }
+}