yangapku commited on
Commit
580a183
1 Parent(s): e101d0b

upload resource files

Browse files
LICENSE ADDED
@@ -0,0 +1,54 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Tongyi Qianwen LICENSE AGREEMENT
2
+
3
+ Tongyi Qianwen Release Date: August 3, 2023
4
+
5
+ By clicking to agree or by using or distributing any portion or element of the Tongyi Qianwen Materials, you will be deemed to have recognized and accepted the content of this Agreement, which is effective immediately.
6
+
7
+ 1. Definitions
8
+ a. This Tongyi Qianwen LICENSE AGREEMENT (this "Agreement") shall mean the terms and conditions for use, reproduction, distribution and modification of the Materials as defined by this Agreement.
9
+ b. "We"(or "Us") shall mean Alibaba Cloud.
10
+ c. "You" (or "Your") shall mean a natural person or legal entity exercising the rights granted by this Agreement and/or using the Materials for any purpose and in any field of use.
11
+ d. "Third Parties" shall mean individuals or legal entities that are not under common control with Us or You.
12
+ e. "Tongyi Qianwen" shall mean the large language models (including Qwen-7b model and Qwen-7b-Chat model ), and software and
13
+ algorithms, consisting of trained model weights, parameters (including optimizer states), machine-learning model code, inference-enabling code, training-enabling code, fine-tuning enabling code and other elements of the foregoing distributed by Us.
14
+ f. "Materials" shall mean, collectively, Alibaba Cloud's proprietary Tongyi Qianwen and Documentation (and any portion thereof) made available under this Agreement.
15
+ g. "Source" form shall mean the preferred form for making modifications, including but not limited to model source code, documentation source, and configuration files.
16
+ h. "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,
17
+ and conversions to other media types.
18
+
19
+ 2. Grant of Rights
20
+ You are granted a non-exclusive, worldwide, non-transferable and royalty-free limited license under Alibaba Cloud's intellectual property or other rights owned by Us embodied in the Materials to use, reproduce, distribute, copy, create derivative works of, and make modifications to the Materials.
21
+
22
+ 3. Redistribution
23
+ You may reproduce and distribute copies of the Materials or derivative works thereof in any medium, with or without modifications, and in Source or Object form, provided that You meet the following conditions:
24
+ a. You shall give any other recipients of the Materials or derivative works a copy of this Agreement;
25
+ b. You shall cause any modified files to carry prominent notices stating that You changed the files;
26
+ c. You shall retain in all copies of the Materials that You distribute the following attribution notices within a "Notice" text file distributed as a part of such copies:"Tongyi Qianwen is licensed under the Tongyi Qianwen LICENSE AGREEMENT, Copyright (c) Alibaba Cloud. All Rights Reserved."; and
27
+ d. 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 terms and conditions of this Agreement.
28
+
29
+ 4. Restrictions
30
+ If you are commercially using the Materials, and your product or service has more than 100 million monthly active users, You shall request a license from Us. You cannot exercise your rights under this Agreement without our express authorization.
31
+
32
+ 5. Rules of use
33
+ a. The Materials may be subject to export controls or restrictions in China, the United States or other countries or regions. You shall comply with applicable laws and regulations in your use of the Materials.
34
+ b. You can not use the Materials or any output therefrom to improve any other large language model (excluding Tongyi Qianwen or derivative works thereof).
35
+
36
+ 6. Intellectual Property
37
+ a. We retain ownership of all intellectual property rights in and to the Materials and derivatives made by or for Us. Conditioned upon compliance with the terms and conditions of this Agreement, with respect to any derivative works and modifications of the Materials that are made by you, you are and will be the owner of such derivative works and modifications.
38
+ b. No trademark license is granted to use the trade names, trademarks, service marks, or product names of Us, except as required to fulfill notice requirements under this Agreement or as required for reasonable and customary use in describing and redistributing the Materials.
39
+ c. If you commence a lawsuit or other proceedings (including a cross-claim or counterclaim in a lawsuit) against Us or any entity alleging that the Materials or any output therefrom, or any part of the foregoing, infringe any intellectual property or other right owned or licensable by you, then all licences granted to you under this Agreement shall terminate as of the date such lawsuit or other proceeding is commenced or brought.
40
+
41
+ 7. Disclaimer of Warranty and Limitation of Liability
42
+
43
+ a. We are not obligated to support, update, provide training for, or develop any further version of the Tongyi Qianwen Model or to grant any license thereto.
44
+ b. THE MATERIALS ARE PROVIDED "AS IS" WITHOUT ANY EXPRESS OR IMPLIED WARRANTY OF ANY KIND INCLUDING WARRANTIES OF MERCHANTABILITY, NONINFRINGEMENT, OR FITNESS FOR A PARTICULAR PURPOSE. WE MAKE NO WARRANTY AND ASSUME NO RESPONSIBILITY FOR THE SAFETY OR STABILITY OF THE MATERIALS AND ANY OUTPUT THEREFROM.
45
+ c. IN NO EVENT SHALL WE BE LIABLE TO YOU FOR ANY DAMAGES, INCLUDING, BUT NOT LIMITED TO ANY DIRECT, OR INDIRECT, SPECIAL OR CONSEQUENTIAL DAMAGES ARISING FROM YOUR USE OR INABILITY TO USE THE MATERIALS OR ANY OUTPUT OF IT, NO MATTER HOW IT’S CAUSED.
46
+ d. You will indemnify and hold armless Us from and against any claim by any third party arising out of or related to your use or distribution of the Materials.
47
+
48
+ 8. Survival and Termination.
49
+ a. The term of this Agreement shall commence upon your acceptance of this Agreement or access to the Materials and will continue in full force and effect until terminated in accordance with the terms and conditions herein.
50
+ b. We may terminate this Agreement if you breach any of the terms or conditions of this Agreement. Upon termination of this Agreement, you must delete and cease use of the Materials. Sections 7 and 9 shall survive the termination of this Agreement.
51
+
52
+ 9. Governing Law and Jurisdiction.
53
+ a. This Agreement and any dispute arising out of or relating to it will be governed by the laws of China, without regard to conflict of law principles, and the UN Convention on Contracts for the International Sale of Goods does not apply to this Agreement.
54
+ b. The People's Courts in Hangzhou City shall have exclusive jurisdiction over any dispute arising out of this Agreement.
NOTICE ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ------------- LICENSE FOR NVIDIA Megatron-LM code --------------
2
+
3
+ Copyright (c) 2022, NVIDIA CORPORATION. All rights reserved.
4
+
5
+ Redistribution and use in source and binary forms, with or without
6
+ modification, are permitted provided that the following conditions
7
+ are met:
8
+ * Redistributions of source code must retain the above copyright
9
+ notice, this list of conditions and the following disclaimer.
10
+ * Redistributions in binary form must reproduce the above copyright
11
+ notice, this list of conditions and the following disclaimer in the
12
+ documentation and/or other materials provided with the distribution.
13
+ * Neither the name of NVIDIA CORPORATION nor the names of its
14
+ contributors may be used to endorse or promote products derived
15
+ from this software without specific prior written permission.
16
+
17
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS ``AS IS'' AND ANY
18
+ EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19
+ IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
20
+ PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR
21
+ CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL,
22
+ EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO,
23
+ PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR
24
+ PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY
25
+ OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
26
+ (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27
+ OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
README.md CHANGED
@@ -12,39 +12,40 @@ pipeline_tag: text-generation
12
  <p align="center">
13
  <img src="assets/logo.jpg" width="400"/>
14
  <p>
 
15
 
16
  <p align="center">
17
- ModelScope[Base|Chat]&nbsp | &nbspHuggingface[Base|Chat]&nbsp | &nbspDemo&nbsp | &nbspReport
18
  </p>
19
- <br><br>
20
 
21
  ## 介绍 (Introduction)
22
 
23
- 通义千问-7B(`Qwen-7B`) 是阿里云研发的通义千问大模型系列的70亿参数规模的模型。`Qwen-7B`是基于Transformer的大语言模型, 在超大规模的预训练数据上进行训练得到。预训练数据类型多样,覆盖广泛,包括大量网络文本、专业书籍、代码等。同时,在`Qwen-7B`的基础上,我们使用对齐机制打造了基于大语言模型的AI助手`Qwen-7B-Chat`。本仓库为`Qwen-7B`的仓库。
24
 
25
  通义千问-7B(Qwen-7B)主要有以下特点:
26
 
27
  1. **大规模高质量训练语料**:使用超过2.2万亿tokens的数据进行预训练,包含高质量中、英、多语言、代码、数学等数据,涵盖通用及专业领域的训练语料。通过大量对比实验对预训练语料分布进行了优化。
28
  2. **强大的性能**:Qwen-7B在多个中英文下游评测任务上(涵盖常识推理、代码、数学、翻译等),效果显著超越现有的相近规模开源模型,甚至在部分指标上相比更大尺寸模型也有较强竞争力。具体评测结果请详见下文。
29
- 3. **覆盖更全面的词表**:相比目前以中英词表为主的开源模型,`Qwen-7B`使用了约15万大小的词表。该词表对多语言更加友好,方便用户在不扩展词表的情况下对部分语种进行能力增强和扩展。
30
 
31
  如果您想了解更多关于通义千问7B开源模型的细节,我们建议您参阅Github代码库。
32
 
33
- `Qwen-7B` is the 7B-parameter version of the large language model series, Qwen (abbr. of Tongyi Qianwen), proposed by Aibaba Cloud. `Qwen-7B` is a Transformer-based large language model, which is pretrained on a large volume of data, including web texts, books, codes, etc. Additionally, based on the pretrained `Qwen-7B`, we release `Qwen-7B-Chat`, a large-model-based AI assistant, which is trained with alignment techniques. This repository is the one for `Qwen-7B`.
34
 
35
- The features of `Qwen-7B` include:
36
 
37
  1. **Large-scale high-quality training corpora**: It is pretrained on over 2.2 trillion tokens, including Chinese, English, multilingual texts, code, and mathematics, covering general and professional fields. The distribution of the pre-training corpus has been optimized through a large number of ablation experiments.
38
  2. **Competitive performance**: It significantly surpasses existing open-source models of similar scale on multiple Chinese and English downstream evaluation tasks (including commonsense, reasoning, code, mathematics, etc.), and even surpasses some larger-scale models in several benchmarks. See below for specific evaluation results.
39
- 3. **More comprehensive vocabulary coverage**: Compared with other open-source models based on Chinese and English vocabularies, `Qwen-7B` uses a vocabulary of over 150K tokens. This vocabulary is more friendly to multiple languages, enabling users to directly further enhance the capability for certain languages without expanding the vocabulary.
40
 
41
- For more details about the open-source model of `Qwen-7B`, please refer to the Github code repository.
42
 
43
  ## 依赖项 (Dependency)
44
 
45
- 运行`Qwen-7B`,请确保机器环境torch版本不低于1.12,再执行以下pip命令安装依赖库
46
 
47
- To run Qwen-7B, please make sure that pytorch version is not lower than 1.12, and then execute the following pip commands to install the dependency libraries.
48
 
49
  ```bash
50
  pip install transformers==4.31.0 accelerate tiktoken einops
@@ -88,9 +89,9 @@ For more information, please refer to our Github repo for more information.
88
 
89
  ## 模型细节 (Model)
90
 
91
- `Qwen-7B`模型规模基本情况如下所示:
92
 
93
- The details of the model architecture of `Qwen-7B` are listed as follows:
94
 
95
  | Hyperparameter | Value |
96
  |:---------------:|-------:|
@@ -100,19 +101,17 @@ The details of the model architecture of `Qwen-7B` are listed as follows:
100
  | vocab size | 151851 |
101
  | sequence length | 2048 |
102
 
103
-
104
  在位置编码、FFN激活函数和normalization的实现方式上,我们也采用了目前最流行的做法,
105
  即RoPE相对位置编码、SwiGLU激活函数、RMSNorm(可选安装flash-attention加速)。
106
 
107
- 在分词器方面,相比目前主流开源模型以中英词表为主,`Qwen-7B`使用了超过15万token大小的词表。
108
- 该词表在GPT-4使用的BPE词表`cl100k_base`基础上,对中文、多语言进行了优化,在对中、英、代码数据的高效编解码的基础上,对部分多语言更加友好,方便用户在不扩展词表的情况下对部分语种进行能力增强。
109
  词表对数字按单个数字位切分。调用较为高效的[tiktoken分词库](https://github.com/openai/tiktoken)进行分词。
110
 
111
  我们从部分语种各随机抽取100万个文档语料,以对比不同模型的编码压缩率(以支持100语种的XLM-R为基准值1,越低越好),具体性能见图。
112
 
113
- 可以看到`Qwen-7B`在保持中英代码高效解码的前提下,对部分使用人群较多的语种(泰语th、希伯来语he、阿拉伯语ar、韩语ko、越南语vi、日语ja、土耳其语tr、印尼语id、波兰语pl、俄语ru、荷兰语nl、葡萄牙语pt、意大利语it、德语de、西班牙语es、法语fr等)上也实现了较高的压缩率,使得模型在这些语种上也具备较强的可扩展性和较高的训练和推理效率。
114
 
115
- 在预训练数据方面,`Qwen-7B`模型一方面利用了部分开源通用语料,
116
  另一方面也积累了海量全网语料以及高质量文本内容,去重及过滤后的语料超过2.2T tokens。
117
  囊括全网文本、百科、书籍、代码、数学及各个领域垂类。
118
 
@@ -122,13 +121,13 @@ The details of the model architecture of `Qwen-7B` are listed as follows:
122
 
123
  For position encoding, FFN activation function, and normalization methods, we adopt the prevalent practices, i.e., RoPE relative position encoding, SwiGLU for activation function, and RMSNorm for normalization (optional installation of flash-attention for acceleration).
124
 
125
- For tokenization, compared to the current mainstream open-source models based on Chinese and English vocabularies, `Qwen-7B` uses a vocabulary of over 150K tokens. It first considers efficient encoding of Chinese, English, and code data, and is also more friendly to multilingual languages, enabling users to directly enhance the capability of some languages without expanding the vocabulary. It segments numbers by single digit, and calls the [tiktoken](https://github.com/openai/tiktoken) tokenizer library for efficient tokenization.
126
 
127
  We randomly selected 1 million document corpus of each language to test and compare the encoding compression rates of different models (with XLM-R, which supports 100 languages, as the base value 1). The specific performance is shown in the figure above.
128
 
129
- As can be seen, while ensuring the efficient decoding of Chinese, English, and code, `Qwen-7B` also achieves a high compression rate for many other languages (such as th, he, ar, ko, vi, ja, tr, id, pl, ru, nl, pt, it, de, es, fr etc.), equipping the model with strong scalability as well as high training and inference efficiency in these languages.
130
 
131
- For pre-training data, on the one hand, `Qwen-7B` uses part of the open-source generic corpus. On the other hand, it uses a massive amount of accumulated web corpus and high-quality text content. The scale of corpus reaches over 2.2T tokens after deduplication and filtration, encompassing web text, encyclopedias, books, code, mathematics, and various domain.
132
 
133
  ## 评测效果(Evaluation)
134
 
@@ -137,13 +136,13 @@ For pre-training data, on the one hand, `Qwen-7B` uses part of the open-source g
137
  #### C-Eval
138
 
139
  [C-Eval](https://arxiv.org/abs/2305.08322)是评测预训练模型中文常识能力的常用测评框架,覆盖人文、社科、理工、其他专业四个大方向共52个学科。
140
- 我们按照标准做法,以开发集样本作为few-shot来源,评价`Qwen-7B`预训练模型的5-shot验证集与测试集准确率。
141
 
142
- [C-Eval](https://arxiv.org/abs/2305.08322) is a common evaluation benchmark for testing the common sense capability of pre-trained models in Chinese. It covers 52 subjects in four major directions: humanities, social sciences, STEM, and other specialties. According to the standard practice, we use the development set samples as the source of few-shot, to evaluate the 5-shot validation set and test set accuracy of the `Qwen-7B` pre-trained model.
143
 
144
- 在C-Eval验证集上,`Qwen-7B`模型和其他模型的准确率对比如下:
145
 
146
- The accuracy comparison of `Qwen-7B` and the other models on the C-Eval validation set is shown as follows:
147
 
148
  | Model | Avg. |
149
  |:---------------:|---------:|
@@ -159,7 +158,7 @@ The accuracy comparison of `Qwen-7B` and the other models on the C-Eval validati
159
 
160
  在C-Eval测试集上,Qwen-7B预训练模型与其他模型的效果对比如下表所示:
161
 
162
- The performance comparison of `Qwen-7B` and other models on the C-Eval test set is shown in the following table:
163
 
164
  | Model | Avg. | Avg. (Hard) | STEM | Social Sciences | Humanities | Others |
165
  |:--------------:|------:|------:|------:|------:|------:|------:|
@@ -176,9 +175,9 @@ The performance comparison of `Qwen-7B` and other models on the C-Eval test set
176
  | ChatGPT | 54.4 | 41.4 | 52.9 | 61.8 | 50.9 | 53.6 |
177
  | **Qwen-7B** | **59.6** | 41.0 | 52.8 | 74.1 | 63.1 | 55.2 |
178
 
179
- 可以看到,`Qwen-7B`在同等规模现有模型中取得了最高的分数,甚至相比更大规模模��也具有较强竞争力。
180
 
181
- As can be seen, `Qwen-7B` achieves the best performance out of all existing models with similar scale and even surpasses larger-scale models.
182
 
183
  ### 英文评测(English Evaluation)
184
 
@@ -186,9 +185,9 @@ As can be seen, `Qwen-7B` achieves the best performance out of all existing mode
186
 
187
  [MMLU](https://arxiv.org/abs/2009.03300)是目前评测英文综合能力最权威的基准评测之一,同样覆盖了不同学科领域、不同难度层级的57个子任务。
188
 
189
- `Qwen-7B`在MMLU 5-shot准确率表现如下表:
190
 
191
- [MMLU](https://arxiv.org/abs/2009.03300) is currently one of the most recognized benchmarks for evaluating English comprehension abilities, covering 57 subtasks across different academic fields and difficulty levels. The MMLU 5-shot accuracy performance of `Qwen-7B` is shown in the following table:
192
 
193
  | Model | Avg. | STEM | Social Sciences | Humanities | Others |
194
  |:--------------:|------:|------:|------:|------:|------:|
@@ -203,9 +202,9 @@ As can be seen, `Qwen-7B` achieves the best performance out of all existing mode
203
  | ChatGLM2-12B | 56.2 | 48.2 | 65.1 | 52.6 | 60.9 |
204
  | **Qwen-7B** | **56.7** | 47.6 | 65.9 | 51.5 | 64.7 |
205
 
206
- 在英文方面,`Qwen-7B`的效果同样超过了目前国内外其他同类开源预训练模型,同样对比更大规模版本的模型也具有较强竞争力。
207
 
208
- In terms of English, `Qwen-7B` also surpasses other similar open-source pre-trained models, and is competitive when compared to larger versions of other models.
209
 
210
  ### 代码评测(Coding Evaluation)
211
 
@@ -253,24 +252,28 @@ We compared the math capabilities of pre-trained models on [GSM8K](https://githu
253
 
254
  We compared the translation capabilities of pre-trained models on [WMT22](https://www.statmt.org/wmt22/translation-task.html) zh-en and en-zh (5-shot BLEU), and the results are as follows:
255
 
256
- | Model | Avg. | zh-en | en-zh |
257
- |:--------------:|------:|------:|------:|
258
- | InternLM-7B | 11.8 | 9.0 | 14.5
259
- | LLaMA-7B | 12.7 | 16.7 | 8.7 |
260
- | LLaMA-13B | 15.8 | 19.5 | 12.0 |
261
- | LLaMA2-7B | 19.9 | 21.9 | 17.9 |
262
- | Bloom-7B | 20.3 | 19.1 | 21.4 |
263
- | LLaMA2-13B | 23.3 | 22.4 | 24.2 |
264
- | PolyLM-13B | 23.6 | 20.2 | 27.0 |
265
- | Baichuan-7B | 24.6 | 22.6 | 26.6 |
266
  | **Qwen-7B** | **27.5** | **24.3** | **30.6** |
267
 
268
  ### 长序列评测(Long-Context Evaluation)
269
 
270
- 我们引入NTK插值,LogN注意力缩放,窗口注意力等技巧,将模型的上下文长度扩展到8K以上。在arXiv数据上使用PPL指标测试`Qwen-7B`在不同长度下的表现,结果如下:
 
 
271
 
272
  We introduce NTK-aware interpolation, LogN attention scaling, Window attention, etc. to extend the context length to over 8K tokens. We conduct language modeling experiments on the arXiv dataset with the PPL evaluation. Results are demonstrated below:
273
 
 
 
274
  <table>
275
  <tr>
276
  <th rowspan="2">Model</th><th colspan="5" align="center">序列长度 Sequence Length</th>
@@ -288,7 +291,7 @@ We introduce NTK-aware interpolation, LogN attention scaling, Window attention,
288
  <td>+ dynamic_ntk + logn</td><td align="center"><b>4.23</b></td><td align="center"><b>3.78</b></td><td align="center"><b>3.58</b></td><td align="center">3.56</td><td align="center">4.62</td>
289
  </tr>
290
  <tr>
291
- <td>+ dynamic_ntk + logn + local_attn</td><td align="center"><b>4.23</b></td><td align="center"><b>3.78</b></td><td align="center"><b>3.58</b></td><td align="center"><b>3.49</b></td><td align="center"><b>4.32</b></td>
292
  </tr>
293
  </table>
294
 
@@ -322,7 +325,7 @@ model = AutoModelForCausalLM.from_pretrained(
322
 
323
  上述方法可以让我们将模型量化成`NF4`和`Int8`精度的模型进行读取,帮助我们节省显存开销。我们也提供了相关性能数据。我们发现尽管模型在效果上存在损失,但模型的显存开销大幅降低。
324
 
325
- With this method, it is available to load `Qwen-7B` in `NF4`and`Int8`, which saves you memory usage. We provide related statistics of model performance below. We find that the quantization downgrades the effectiveness slightly but significantly increases inference efficiency and reduces memory costs.
326
 
327
  | Precision | MMLU | Memory |
328
  | :---------: | -------: | -----: |
@@ -330,8 +333,6 @@ With this method, it is available to load `Qwen-7B` in `NF4`and`Int8`, which sav
330
  | Int8 | 52.8 | 10.1G |
331
  | NF4 | 48.9 | 7.4G |
332
 
333
-
334
-
335
  ## 评测复现(Reproduction)
336
 
337
  我们提供了评测脚本,方便大家复现模型效果,详见[链接](eval/EVALUATION.md)。提示:由于硬件和框架造成的舍入误差,复现结果如有小幅波动属于正常现象。
@@ -342,12 +343,11 @@ We have provided evaluation scripts to reproduce the performance of our model, d
342
 
343
  我们的代码和模型权重对学术研究完全开放,并支持商用。请查看LICENSE了解具体的开源协议细节。
344
 
345
- Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check LICENSE.txt for more details about the license.
346
 
347
  ## 联系我们(Contact Us)
348
 
349
  如果你想给我们的研发团队和产品团队留言,请通过邮件(qianwen_opensource@alibabacloud.com)联系我们。
350
 
351
-
352
  If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.
353
 
 
12
  <p align="center">
13
  <img src="assets/logo.jpg" width="400"/>
14
  <p>
15
+ <br>
16
 
17
  <p align="center">
18
+ Qwen-7B <a href="https://modelscope.cn/models/qwen/Qwen-7B/summary">🤖 <a> | <a href="https://huggingface.co/Qwen/Qwen-7B">🤗</a>&nbsp | Qwen-7B-Chat <a href="https://modelscope.cn/models/qwen/Qwen-7B-Chat/summary">🤖 <a>| <a href="https://huggingface.co/Qwen/Qwen-7B-Chat">🤗</a>&nbsp | &nbspDemo&nbsp | &nbsp<a href="https://github.com/QwenLM/Qwen-7B/tech_memo.md">Report</a>
19
  </p>
20
+ <br>
21
 
22
  ## 介绍 (Introduction)
23
 
24
+ **通义千问-7BQwen-7B)**是阿里云研发的通义千问大模型系列的70亿参数规模的模型。Qwen-7B是基于Transformer的大语言模型, 在超大规模的预训练数据上进行训练得到。预训练数据类型多样,覆盖广泛,包括大量网络文本、专业书籍、代码等。同时,在Qwen-7B的基础上,我们使用对齐机制打造了基于大语言模型的AI助手Qwen-7B-Chat。本仓库为Qwen-7B的仓库。
25
 
26
  通义千问-7B(Qwen-7B)主要有以下特点:
27
 
28
  1. **大规模高质量训练语料**:使用超过2.2万亿tokens的数据进行预训练,包含高质量中、英、多语言、代码、数学等数据,涵盖通用及专业领域的训练语料。通过大量对比实验对预训练语料分布进行了优化。
29
  2. **强大的性能**:Qwen-7B在多个中英文下游评测任务上(涵盖常识推理、代码、数学、翻译等),效果显著超越现有的相近规模开源模型,甚至在部分指标上相比更大尺寸模型也有较强竞争力。具体评测结果请详见下文。
30
+ 3. **覆盖更全面的词表**:相比目前以中英词表为主的开源模型,Qwen-7B使用了约15万大小的词表。该词表对多语言更加友好,方便用户在不扩展词表的情况下对部分语种进行能力增强和扩展。
31
 
32
  如果您想了解更多关于通义千问7B开源模型的细节,我们建议您参阅Github代码库。
33
 
34
+ **Qwen-7B** is the 7B-parameter version of the large language model series, Qwen (abbr. Tongyi Qianwen), proposed by Aibaba Cloud. Qwen-7B is a Transformer-based large language model, which is pretrained on a large volume of data, including web texts, books, codes, etc. Additionally, based on the pretrained Qwen-7B, we release Qwen-7B-Chat, a large-model-based AI assistant, which is trained with alignment techniques. This repository is the one for Qwen-7B.
35
 
36
+ The features of Qwen-7B include:
37
 
38
  1. **Large-scale high-quality training corpora**: It is pretrained on over 2.2 trillion tokens, including Chinese, English, multilingual texts, code, and mathematics, covering general and professional fields. The distribution of the pre-training corpus has been optimized through a large number of ablation experiments.
39
  2. **Competitive performance**: It significantly surpasses existing open-source models of similar scale on multiple Chinese and English downstream evaluation tasks (including commonsense, reasoning, code, mathematics, etc.), and even surpasses some larger-scale models in several benchmarks. See below for specific evaluation results.
40
+ 3. **More comprehensive vocabulary coverage**: Compared with other open-source models based on Chinese and English vocabularies, Qwen-7B uses a vocabulary of over 150K tokens. This vocabulary is more friendly to multiple languages, enabling users to directly further enhance the capability for certain languages without expanding the vocabulary.
41
 
42
+ For more details about the open-source model of Qwen-7B, please refer to the Github code repository.
43
 
44
  ## 依赖项 (Dependency)
45
 
46
+ 运行Qwen-7B,请确保机器环境torch版本不低于1.12,再执行以下pip命令安装依赖库
47
 
48
+ To run Qwen-7B, please make sure that pytorch version is not lower than 1.12, and then execute the following pip commands to install the dependent libraries.
49
 
50
  ```bash
51
  pip install transformers==4.31.0 accelerate tiktoken einops
 
89
 
90
  ## 模型细节 (Model)
91
 
92
+ Qwen-7B模型规模基本情况如下所示:
93
 
94
+ The details of the model architecture of Qwen-7B are listed as follows:
95
 
96
  | Hyperparameter | Value |
97
  |:---------------:|-------:|
 
101
  | vocab size | 151851 |
102
  | sequence length | 2048 |
103
 
 
104
  在位置编码、FFN激活函数和normalization的实现方式上,我们也采用了目前最流行的做法,
105
  即RoPE相对位置编码、SwiGLU激活函数、RMSNorm(可选安装flash-attention加速)。
106
 
107
+ 在分词器方面,相比目前主流开源模型以中英词表为主,Qwen-7B使用了超过15万token大小的词表。 该词表在GPT-4使用的BPE词表`cl100k_base`基础上,对中文、多语言进行了优化,在对中、英、代码数据的高效编解码的基础上,对部分多语言更加友好,方便用户在不扩展词表的情况下对部分语种进行能力增强。
 
108
  词表对数字按单个数字位切分。调用较为高效的[tiktoken分词库](https://github.com/openai/tiktoken)进行分词。
109
 
110
  我们从部分语种各随机抽取100万个文档语料,以对比不同模型的编码压缩率(以支持100语种的XLM-R为基准值1,越低越好),具体性能见图。
111
 
112
+ 可以看到Qwen-7B在保持中英代码高效解码的前提下,对部分使用人群较多的语种(泰语th、希伯来语he、阿拉伯语ar、韩语ko、越南语vi、日语ja、土耳其语tr、印尼语id、波兰语pl、俄语ru、荷兰语nl、葡萄牙语pt、意大利语it、德语de、西班牙语es、法语fr等)上也实现了较高的压缩率,使得模型在这些语种上也具备较强的可扩展性和较高的训练和推理效率。
113
 
114
+ 在预训练数据方面,Qwen-7B模型一方面利用了部分开源通用语料,
115
  另一方面也积累了海量全网语料以及高质量文本内容,去重及过滤后的语料超过2.2T tokens。
116
  囊括全网文本、百科、书籍、代码、数学及各个领域垂类。
117
 
 
121
 
122
  For position encoding, FFN activation function, and normalization methods, we adopt the prevalent practices, i.e., RoPE relative position encoding, SwiGLU for activation function, and RMSNorm for normalization (optional installation of flash-attention for acceleration).
123
 
124
+ For tokenization, compared to the current mainstream open-source models based on Chinese and English vocabularies, Qwen-7B uses a vocabulary of over 150K tokens. It first considers efficient encoding of Chinese, English, and code data, and is also more friendly to multilingual languages, enabling users to directly enhance the capability of some languages without expanding the vocabulary. It segments numbers by single digit, and calls the [tiktoken](https://github.com/openai/tiktoken) tokenizer library for efficient tokenization.
125
 
126
  We randomly selected 1 million document corpus of each language to test and compare the encoding compression rates of different models (with XLM-R, which supports 100 languages, as the base value 1). The specific performance is shown in the figure above.
127
 
128
+ As can be seen, while ensuring the efficient decoding of Chinese, English, and code, Qwen-7B also achieves a high compression rate for many other languages (such as th, he, ar, ko, vi, ja, tr, id, pl, ru, nl, pt, it, de, es, fr etc.), equipping the model with strong scalability as well as high training and inference efficiency in these languages.
129
 
130
+ For pre-training data, on the one hand, Qwen-7B uses part of the open-source generic corpus. On the other hand, it uses a massive amount of accumulated web corpus and high-quality text content. The scale of corpus reaches over 2.2T tokens after deduplication and filtration, encompassing web text, encyclopedias, books, code, mathematics, and various domain.
131
 
132
  ## 评测效果(Evaluation)
133
 
 
136
  #### C-Eval
137
 
138
  [C-Eval](https://arxiv.org/abs/2305.08322)是评测预训练模型中文常识能力的常用测评框架,覆盖人文、社科、理工、其他专业四个大方向共52个学科。
139
+ 我们按照标准做法,以开发集样本作为few-shot来源,评价Qwen-7B预训练模型的5-shot验证集与测试集准确率。
140
 
141
+ [C-Eval](https://arxiv.org/abs/2305.08322) is a common evaluation benchmark for testing the common sense capability of pre-trained models in Chinese. It covers 52 subjects in four major directions: humanities, social sciences, STEM, and other specialties. According to the standard practice, we use the development set samples as the source of few-shot, to evaluate the 5-shot validation set and test set accuracy of the Qwen-7B pre-trained model.
142
 
143
+ 在C-Eval验证集上,Qwen-7B模型和其他模型的准确率对比如下:
144
 
145
+ The accuracy comparison of Qwen-7B and the other models on the C-Eval validation set is shown as follows:
146
 
147
  | Model | Avg. |
148
  |:---------------:|---------:|
 
158
 
159
  在C-Eval测试集上,Qwen-7B预训练模型与其他模型的效果对比如下表所示:
160
 
161
+ The performance comparison of Qwen-7B and other models on the C-Eval test set is shown in the following table:
162
 
163
  | Model | Avg. | Avg. (Hard) | STEM | Social Sciences | Humanities | Others |
164
  |:--------------:|------:|------:|------:|------:|------:|------:|
 
175
  | ChatGPT | 54.4 | 41.4 | 52.9 | 61.8 | 50.9 | 53.6 |
176
  | **Qwen-7B** | **59.6** | 41.0 | 52.8 | 74.1 | 63.1 | 55.2 |
177
 
178
+ 可以看到,Qwen-7B在同等规模现有模型中取得了最高的分数,甚至相比更大规模模型也具有较强竞争力。
179
 
180
+ As can be seen, Qwen-7B achieves the best performance out of all existing models with similar scale and even surpasses larger-scale models.
181
 
182
  ### 英文评测(English Evaluation)
183
 
 
185
 
186
  [MMLU](https://arxiv.org/abs/2009.03300)是目前评测英文综合能力最权威的基准评测之一,同样覆盖了不同学科领域、不同难度层级的57个子任务。
187
 
188
+ Qwen-7BMMLU 5-shot准确率表现如下表:
189
 
190
+ [MMLU](https://arxiv.org/abs/2009.03300) is currently one of the most recognized benchmarks for evaluating English comprehension abilities, covering 57 subtasks across different academic fields and difficulty levels. The MMLU 5-shot accuracy performance of Qwen-7B is shown in the following table:
191
 
192
  | Model | Avg. | STEM | Social Sciences | Humanities | Others |
193
  |:--------------:|------:|------:|------:|------:|------:|
 
202
  | ChatGLM2-12B | 56.2 | 48.2 | 65.1 | 52.6 | 60.9 |
203
  | **Qwen-7B** | **56.7** | 47.6 | 65.9 | 51.5 | 64.7 |
204
 
205
+ 在英文方面,Qwen-7B的效果同样超过了目前国内外其他同类开源预训练模型,同样对比更大规模版本的模型也具有较强竞争力。
206
 
207
+ In terms of English, Qwen-7B also surpasses other similar open-source pre-trained models, and is competitive when compared to larger versions of other models.
208
 
209
  ### 代码评测(Coding Evaluation)
210
 
 
252
 
253
  We compared the translation capabilities of pre-trained models on [WMT22](https://www.statmt.org/wmt22/translation-task.html) zh-en and en-zh (5-shot BLEU), and the results are as follows:
254
 
255
+ | Model | Avg. | zh-en | en-zh |
256
+ |:-----------:|---------:|---------:|---------:|
257
+ | InternLM-7B | 11.8 | 9.0 | 14.5 |
258
+ | LLaMA-7B | 12.7 | 16.7 | 8.7 |
259
+ | LLaMA-13B | 15.8 | 19.5 | 12.0 |
260
+ | LLaMA2-7B | 19.9 | 21.9 | 17.9 |
261
+ | Bloom-7B | 20.3 | 19.1 | 21.4 |
262
+ | LLaMA2-13B | 23.3 | 22.4 | 24.2 |
263
+ | PolyLM-13B | 23.6 | 20.2 | 27.0 |
264
+ | Baichuan-7B | 24.6 | 22.6 | 26.6 |
265
  | **Qwen-7B** | **27.5** | **24.3** | **30.6** |
266
 
267
  ### 长序列评测(Long-Context Evaluation)
268
 
269
+ 我们引入NTK插值,LogN注意力缩放,窗口注意力等技巧,将模型的上下文长度扩展到8K以上。在arXiv数据上使用PPL指标测试Qwen-7B在不同长度下的表现,结果如下:
270
+
271
+ **(若要启用NTK和LogN注意力缩放,请将config.json里的`use_dynamc_ntk`和`use_logn_attn`设置为true)**
272
 
273
  We introduce NTK-aware interpolation, LogN attention scaling, Window attention, etc. to extend the context length to over 8K tokens. We conduct language modeling experiments on the arXiv dataset with the PPL evaluation. Results are demonstrated below:
274
 
275
+ **(To use NTK interpolation and LogN scaling, please set `use_dynamic_ntk` and `use_long_attn` to true in config.json.)**
276
+
277
  <table>
278
  <tr>
279
  <th rowspan="2">Model</th><th colspan="5" align="center">序列长度 Sequence Length</th>
 
291
  <td>+ dynamic_ntk + logn</td><td align="center"><b>4.23</b></td><td align="center"><b>3.78</b></td><td align="center"><b>3.58</b></td><td align="center">3.56</td><td align="center">4.62</td>
292
  </tr>
293
  <tr>
294
+ <td>+ dynamic_ntk + logn + window_attn</td><td align="center"><b>4.23</b></td><td align="center"><b>3.78</b></td><td align="center"><b>3.58</b></td><td align="center"><b>3.49</b></td><td align="center"><b>4.32</b></td>
295
  </tr>
296
  </table>
297
 
 
325
 
326
  上述方法可以让我们将模型量化成`NF4`和`Int8`精度的模型进行读取,帮助我们节省显存开销。我们也提供了相关性能数据。我们发现尽管模型在效果上存在损失,但模型的显存开销大幅降低。
327
 
328
+ With this method, it is available to load Qwen-7B in `NF4` and `Int8`, which saves you memory usage. We provide related statistics of model performance below. We find that the quantization downgrades the effectiveness slightly but significantly increases inference efficiency and reduces memory costs.
329
 
330
  | Precision | MMLU | Memory |
331
  | :---------: | -------: | -----: |
 
333
  | Int8 | 52.8 | 10.1G |
334
  | NF4 | 48.9 | 7.4G |
335
 
 
 
336
  ## 评测复现(Reproduction)
337
 
338
  我们提供了评测脚本,方便大家复现模型效果,详见[链接](eval/EVALUATION.md)。提示:由于硬件和框架造成的舍入误差,复现结果如有小幅波动属于正常现象。
 
343
 
344
  我们的代码和模型权重对学术研究完全开放,并支持商用。请查看LICENSE了解具体的开源协议细节。
345
 
346
+ Our code and checkpoints are open to research purpose, and they are allowed for commercial purposes. Check [LICENSE](LICENSE) for more details about the license.
347
 
348
  ## 联系我们(Contact Us)
349
 
350
  如果你想给我们的研发团队和产品团队留言,请通过邮件(qianwen_opensource@alibabacloud.com)联系我们。
351
 
 
352
  If you are interested to leave a message to either our research team or product team, feel free to send an email to qianwen_opensource@alibabacloud.com.
353
 
assets/logo.jpg ADDED
assets/qwen_tokenizer.png ADDED
config.json ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "activation": "swiglu",
3
+ "apply_residual_connection_post_layernorm": false,
4
+ "architectures": [
5
+ "QWenLMHeadModel"
6
+ ],
7
+ "auto_map": {
8
+ "AutoConfig": "configuration_qwen.QWenConfig",
9
+ "AutoModelForCausalLM": "modeling_qwen.QWenLMHeadModel"
10
+ },
11
+ "attn_pdrop": 0.0,
12
+ "bf16": true,
13
+ "bias_dropout_fusion": true,
14
+ "bos_token_id": 151643,
15
+ "embd_pdrop": 0.1,
16
+ "eos_token_id": 151643,
17
+ "ffn_hidden_size": 22016,
18
+ "fp16": false,
19
+ "initializer_range": 0.02,
20
+ "kv_channels": 128,
21
+ "layer_norm_epsilon": 1e-05,
22
+ "model_type": "qwen",
23
+ "n_embd": 4096,
24
+ "n_head": 32,
25
+ "n_layer": 32,
26
+ "n_positions": 6144,
27
+ "no_bias": true,
28
+ "onnx_safe": null,
29
+ "padded_vocab_size": 151936,
30
+ "params_dtype": "torch.bfloat16",
31
+ "pos_emb": "rotary",
32
+ "resid_pdrop": 0.1,
33
+ "rotary_emb_base": 10000,
34
+ "rotary_pct": 1.0,
35
+ "scale_attn_weights": true,
36
+ "seq_length": 2048,
37
+ "tie_word_embeddings": false,
38
+ "tokenizer_type": "QWenTokenizer",
39
+ "transformers_version": "4.31.0",
40
+ "use_cache": true,
41
+ "use_flash_attn": false,
42
+ "vocab_size": 151936,
43
+ "use_dynamic_ntk": false,
44
+ "use_logn_attn": false
45
+ }
configuration_qwen.py ADDED
@@ -0,0 +1,74 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ from transformers import PretrainedConfig
7
+
8
+
9
+ class QWenConfig(PretrainedConfig):
10
+ model_type = "qwen"
11
+ keys_to_ignore_at_inference = ["past_key_values"]
12
+ attribute_map = {
13
+ "hidden_size": "n_embd",
14
+ "num_attention_heads": "n_head",
15
+ "max_position_embeddings": "n_positions",
16
+ "num_hidden_layers": "n_layer",
17
+ }
18
+
19
+ def __init__(
20
+ self,
21
+ vocab_size=151851,
22
+ n_embd=4096,
23
+ n_layer=32,
24
+ n_head=32,
25
+ n_inner=None,
26
+ embd_pdrop=0.0,
27
+ attn_pdrop=0.0,
28
+ layer_norm_epsilon=1e-5,
29
+ initializer_range=0.02,
30
+ scale_attn_weights=True,
31
+ use_cache=True,
32
+ eos_token_id=151643,
33
+ apply_residual_connection_post_layernorm=False,
34
+ bf16=True,
35
+ kv_channels=128,
36
+ rotary_pct=1.0,
37
+ rotary_emb_base=10000,
38
+ use_dynamic_ntk=False,
39
+ use_logn_attn=False,
40
+ use_flash_attn=True,
41
+ ffn_hidden_size=22016,
42
+ no_bias=True,
43
+ tie_word_embeddings=False,
44
+ **kwargs,
45
+ ):
46
+ self.eos_token_id = eos_token_id
47
+ super().__init__(
48
+ eos_token_id=eos_token_id, tie_word_embeddings=tie_word_embeddings, **kwargs
49
+ )
50
+
51
+ self.vocab_size = vocab_size
52
+ self.n_embd = n_embd
53
+ self.n_layer = n_layer
54
+ self.n_head = n_head
55
+ self.n_inner = n_inner
56
+ self.embd_pdrop = embd_pdrop
57
+ self.attn_pdrop = attn_pdrop
58
+ self.layer_norm_epsilon = layer_norm_epsilon
59
+ self.initializer_range = initializer_range
60
+ self.scale_attn_weights = scale_attn_weights
61
+ self.use_cache = use_cache
62
+ self.apply_residual_connection_post_layernorm = (
63
+ apply_residual_connection_post_layernorm
64
+ )
65
+ self.bf16 = bf16
66
+ self.kv_channels = kv_channels
67
+ self.rotary_pct = rotary_pct
68
+ self.rotary_emb_base = rotary_emb_base
69
+ self.use_dynamic_ntk = use_dynamic_ntk
70
+ self.use_logn_attn = use_logn_attn
71
+ self.use_flash_attn = use_flash_attn
72
+ self.ffn_hidden_size = ffn_hidden_size
73
+ self.no_bias = no_bias
74
+ self.tie_word_embeddings = tie_word_embeddings
generation_config.json ADDED
@@ -0,0 +1,16 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "chat_format": "raw",
3
+ "decay_bound": 0.0,
4
+ "decay_factor": 1.0,
5
+ "eos_token_id": 151643,
6
+ "factual_nucleus_sampling": false,
7
+ "max_context_size": 1024,
8
+ "max_generate_size": 512,
9
+ "max_new_tokens": 512,
10
+ "pad_token_id": 151643,
11
+ "stop_words_ids": [[151643]],
12
+ "do_sample": true,
13
+ "top_k": 0,
14
+ "top_p": 0.8,
15
+ "transformers_version": "4.31.0"
16
+ }
modeling_qwen.py ADDED
@@ -0,0 +1,1027 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ import importlib
7
+ import math
8
+ from typing import TYPE_CHECKING, Optional, Tuple, Union, Callable, List
9
+
10
+ import torch
11
+ import torch.nn.functional as F
12
+ import torch.utils.checkpoint
13
+ from torch.cuda.amp import autocast
14
+
15
+ from torch.nn import CrossEntropyLoss
16
+ from transformers import PreTrainedTokenizer, GenerationConfig, StoppingCriteriaList
17
+ from transformers.generation.logits_process import LogitsProcessorList
18
+ if TYPE_CHECKING:
19
+ from transformers.generation.streamers import BaseStreamer
20
+ from transformers.generation.utils import GenerateOutput
21
+ from transformers.modeling_outputs import (
22
+ BaseModelOutputWithPast,
23
+ CausalLMOutputWithPast,
24
+ )
25
+ from transformers.modeling_utils import PreTrainedModel
26
+ from transformers.utils import logging
27
+
28
+ try:
29
+ from einops import rearrange
30
+ except ImportError:
31
+ rearrange = None
32
+ from torch import nn
33
+
34
+ try:
35
+ from flash_attn.layers.rotary import apply_rotary_emb_func
36
+ from einops import rearrange
37
+
38
+ use_flash_rotary = True
39
+ print("use flash_attn rotary")
40
+ except ImportError:
41
+ use_flash_rotary = False
42
+ print("import flash_attn rotary fail")
43
+
44
+ try:
45
+ from flash_attn.ops.rms_norm import rms_norm
46
+
47
+ print("use flash_attn rms_norm")
48
+ except ImportError:
49
+ rms_norm = None
50
+ print("import flash_attn rms_norm fail")
51
+
52
+ from .configuration_qwen import QWenConfig
53
+ from .qwen_generation_utils import (
54
+ HistoryType,
55
+ make_context,
56
+ decode_tokens,
57
+ get_stop_words_ids,
58
+ StopWordsLogitsProcessor,
59
+ )
60
+
61
+
62
+ logger = logging.get_logger(__name__)
63
+
64
+ _CHECKPOINT_FOR_DOC = "qwen"
65
+ _CONFIG_FOR_DOC = "QWenConfig"
66
+
67
+ QWen_PRETRAINED_MODEL_ARCHIVE_LIST = ["qwen-7b"]
68
+
69
+ try:
70
+ from flash_attn.flash_attn_interface import flash_attn_unpadded_func
71
+ except ImportError:
72
+ flash_attn_unpadded_func = None
73
+
74
+
75
+ class FlashSelfAttention(torch.nn.Module):
76
+ def __init__(
77
+ self,
78
+ causal=False,
79
+ softmax_scale=None,
80
+ attention_dropout=0.0,
81
+ ):
82
+ super().__init__()
83
+ assert flash_attn_unpadded_func is not None, (
84
+ "Please install FlashAttention first, " "e.g., with pip install flash-attn"
85
+ )
86
+ assert (
87
+ rearrange is not None
88
+ ), "Please install einops first, e.g., with pip install einops"
89
+ self.causal = causal
90
+ self.softmax_scale = softmax_scale
91
+ self.dropout_p = attention_dropout
92
+
93
+ def forward(self, q, k, v):
94
+ assert all((i.dtype in [torch.float16, torch.bfloat16] for i in (q, k, v)))
95
+ assert all((i.is_cuda for i in (q, k, v)))
96
+ batch_size, seqlen_q = q.shape[0], q.shape[1]
97
+ seqlen_k = k.shape[1]
98
+ q, k, v = [rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v]]
99
+ cu_seqlens_q = torch.arange(
100
+ 0,
101
+ (batch_size + 1) * seqlen_q,
102
+ step=seqlen_q,
103
+ dtype=torch.int32,
104
+ device=q.device,
105
+ )
106
+
107
+ if self.training:
108
+ assert seqlen_k == seqlen_q
109
+
110
+ is_causal = self.causal
111
+ cu_seqlens_k = cu_seqlens_q
112
+ else:
113
+ is_causal = seqlen_q == seqlen_k
114
+ cu_seqlens_k = torch.arange(
115
+ 0,
116
+ (batch_size + 1) * seqlen_k,
117
+ step=seqlen_k,
118
+ dtype=torch.int32,
119
+ device=q.device,
120
+ )
121
+ self.dropout_p = 0
122
+ output = flash_attn_unpadded_func(
123
+ q,
124
+ k,
125
+ v,
126
+ cu_seqlens_q,
127
+ cu_seqlens_k,
128
+ seqlen_q,
129
+ seqlen_k,
130
+ self.dropout_p,
131
+ softmax_scale=self.softmax_scale,
132
+ causal=is_causal,
133
+ )
134
+
135
+ output = rearrange(output, "(b s) ... -> b s ...", b=batch_size)
136
+ return output
137
+
138
+
139
+ class QWenAttention(nn.Module):
140
+ def __init__(self, config, layer_number=None):
141
+ super().__init__()
142
+
143
+ max_positions = config.max_position_embeddings
144
+ self.register_buffer(
145
+ "bias",
146
+ torch.tril(
147
+ torch.ones((max_positions, max_positions), dtype=torch.bool)
148
+ ).view(1, 1, max_positions, max_positions),
149
+ persistent=False,
150
+ )
151
+ self.register_buffer("masked_bias", torch.tensor(-1e4), persistent=False)
152
+ self.layer_number = max(1, layer_number)
153
+ self.params_dtype = config.params_dtype
154
+ self.seq_length = config.seq_length
155
+
156
+ self.hidden_size = config.hidden_size
157
+ self.split_size = config.hidden_size
158
+ self.num_heads = config.num_attention_heads
159
+ self.head_dim = self.hidden_size // self.num_heads
160
+
161
+ self.use_flash_attn = config.use_flash_attn
162
+ self.scale_attn_weights = True
163
+
164
+ self.layer_idx = None
165
+
166
+ self.projection_size = config.kv_channels * config.num_attention_heads
167
+
168
+ assert self.projection_size % config.num_attention_heads == 0
169
+ self.hidden_size_per_attention_head = (
170
+ self.projection_size // config.num_attention_heads
171
+ )
172
+
173
+ self.c_attn = nn.Linear(config.hidden_size, 3 * self.projection_size)
174
+
175
+ self.c_proj = nn.Linear(
176
+ config.hidden_size, self.projection_size, bias=not config.no_bias
177
+ )
178
+
179
+ if self.use_flash_attn:
180
+ self.core_attention_flash = FlashSelfAttention(
181
+ causal=True, attention_dropout=config.attn_pdrop
182
+ )
183
+
184
+ self.bf16 = config.bf16
185
+
186
+ if config.rotary_pct == 1.0:
187
+ self.rotary_ndims = None
188
+ else:
189
+ assert config.rotary_pct < 1
190
+ self.rotary_ndims = int(
191
+ self.hidden_size_per_attention_head * config.rotary_pct
192
+ )
193
+ dim = (
194
+ self.rotary_ndims
195
+ if self.rotary_ndims is not None
196
+ else self.hidden_size_per_attention_head
197
+ )
198
+ self.rotary_emb = RotaryEmbedding(
199
+ dim, base=config.rotary_emb_base
200
+ )
201
+
202
+ self.use_dynamic_ntk = config.use_dynamic_ntk
203
+ self.use_logn_attn = config.use_logn_attn
204
+
205
+ logn_list = [math.log(i, self.seq_length) if i > self.seq_length else 1 for i in range(1, 32768)]
206
+ self.logn_tensor = torch.Tensor(logn_list)[None, :, None, None]
207
+ self._ntk_cached = 1.0
208
+
209
+ self.attn_dropout = nn.Dropout(config.attn_pdrop)
210
+
211
+ def _attn(self, query, key, value, attention_mask=None, head_mask=None):
212
+ attn_weights = torch.matmul(query, key.transpose(-1, -2))
213
+
214
+ if self.scale_attn_weights:
215
+ attn_weights = attn_weights / torch.full(
216
+ [],
217
+ value.size(-1) ** 0.5,
218
+ dtype=attn_weights.dtype,
219
+ device=attn_weights.device,
220
+ )
221
+
222
+ query_length, key_length = query.size(-2), key.size(-2)
223
+ causal_mask = self.bias[
224
+ :, :, key_length - query_length : key_length, :key_length
225
+ ]
226
+ mask_value = torch.finfo(attn_weights.dtype).min
227
+ mask_value = torch.full([], mask_value, dtype=attn_weights.dtype).to(
228
+ attn_weights.device
229
+ )
230
+ attn_weights = torch.where(
231
+ causal_mask, attn_weights.to(attn_weights.dtype), mask_value
232
+ )
233
+
234
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
235
+
236
+ attn_weights = attn_weights.type(value.dtype)
237
+ attn_weights = self.attn_dropout(attn_weights)
238
+
239
+ if head_mask is not None:
240
+ attn_weights = attn_weights * head_mask
241
+
242
+ attn_output = torch.matmul(attn_weights, value)
243
+ attn_output = attn_output.transpose(1, 2)
244
+
245
+ return attn_output, attn_weights
246
+
247
+ def _upcast_and_reordered_attn(
248
+ self, query, key, value, attention_mask=None, head_mask=None
249
+ ):
250
+ bsz, num_heads, q_seq_len, dk = query.size()
251
+ _, _, k_seq_len, _ = key.size()
252
+
253
+ attn_weights = torch.empty(
254
+ bsz * num_heads,
255
+ q_seq_len,
256
+ k_seq_len,
257
+ dtype=torch.float32,
258
+ device=query.device,
259
+ )
260
+
261
+ scale_factor = 1.0
262
+ if self.scale_attn_weights:
263
+ scale_factor /= float(value.size(-1)) ** 0.5
264
+
265
+ with autocast(enabled=False):
266
+ q, k = query.reshape(-1, q_seq_len, dk), key.transpose(-1, -2).reshape(
267
+ -1, dk, k_seq_len
268
+ )
269
+ attn_weights = torch.baddbmm(
270
+ attn_weights, q.float(), k.float(), beta=0, alpha=scale_factor
271
+ )
272
+ attn_weights = attn_weights.reshape(bsz, num_heads, q_seq_len, k_seq_len)
273
+
274
+ query_length, key_length = query.size(-2), key.size(-2)
275
+ causal_mask = self.bias[
276
+ :, :, key_length - query_length : key_length, :key_length
277
+ ]
278
+ mask_value = torch.finfo(attn_weights.dtype).min
279
+ mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(
280
+ attn_weights.device
281
+ )
282
+ attn_weights = torch.where(causal_mask, attn_weights, mask_value)
283
+
284
+ if attention_mask is not None:
285
+ attn_weights = attn_weights + attention_mask
286
+
287
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1)
288
+
289
+ if attn_weights.dtype != torch.float32:
290
+ raise RuntimeError(
291
+ "Error with upcasting, attn_weights does not have dtype torch.float32"
292
+ )
293
+ attn_weights = attn_weights.type(value.dtype)
294
+ attn_weights = self.attn_dropout(attn_weights)
295
+
296
+ if head_mask is not None:
297
+ attn_weights = attn_weights * head_mask
298
+
299
+ attn_output = torch.matmul(attn_weights, value)
300
+
301
+ return attn_output, attn_weights
302
+
303
+ def _split_heads(self, tensor, num_heads, attn_head_size):
304
+ new_shape = tensor.size()[:-1] + (num_heads, attn_head_size)
305
+ tensor = tensor.view(new_shape)
306
+ return tensor
307
+
308
+ def _merge_heads(self, tensor, num_heads, attn_head_size):
309
+ tensor = tensor.contiguous()
310
+ new_shape = tensor.size()[:-2] + (num_heads * attn_head_size,)
311
+ return tensor.view(new_shape)
312
+
313
+ def forward(
314
+ self,
315
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
316
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
317
+ attention_mask: Optional[torch.FloatTensor] = None,
318
+ head_mask: Optional[torch.FloatTensor] = None,
319
+ encoder_hidden_states: Optional[torch.Tensor] = None,
320
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
321
+ output_attentions: Optional[bool] = False,
322
+ use_cache: Optional[bool] = False,
323
+ ):
324
+
325
+ mixed_x_layer = self.c_attn(hidden_states)
326
+ query, key, value = mixed_x_layer.split(self.split_size, dim=2)
327
+
328
+ query = self._split_heads(query, self.num_heads, self.head_dim)
329
+ key = self._split_heads(key, self.num_heads, self.head_dim)
330
+ value = self._split_heads(value, self.num_heads, self.head_dim)
331
+
332
+ kv_seq_len = hidden_states.size()[1]
333
+ if layer_past:
334
+ # layer past[0] shape: bs * seq_len * head_num * dim
335
+ kv_seq_len += layer_past[0].shape[1]
336
+ if self.use_dynamic_ntk and kv_seq_len == hidden_states.size()[1]:
337
+ context_value = math.log(kv_seq_len / self.seq_length, 2) + 1
338
+ ntk_alpha = 2 ** math.ceil(context_value) - 1
339
+ ntk_alpha = max(ntk_alpha, 1)
340
+ self._ntk_cached = ntk_alpha
341
+ else:
342
+ ntk_alpha = self._ntk_cached
343
+ rotary_pos_emb = self.rotary_emb(kv_seq_len, ntk_alpha=ntk_alpha).to(hidden_states.device)
344
+
345
+ if rotary_pos_emb is not None:
346
+ if isinstance(rotary_pos_emb, tuple):
347
+ rotary_pos_emb = rotary_pos_emb
348
+ else:
349
+ rotary_pos_emb = (rotary_pos_emb,) * 2
350
+
351
+ if rotary_pos_emb is not None:
352
+ q_pos_emb, k_pos_emb = rotary_pos_emb
353
+ # Slice the pos emb for current inference
354
+ cur_len = query.shape[1]
355
+ q_pos_emb = q_pos_emb[:, -cur_len:, :, :]
356
+ k_pos_emb = k_pos_emb[:, -cur_len:, :, :]
357
+ query = apply_rotary_pos_emb(query, q_pos_emb)
358
+ key = apply_rotary_pos_emb(key, k_pos_emb)
359
+
360
+ if layer_past is not None:
361
+ past_key, past_value = layer_past[0], layer_past[1]
362
+ key = torch.cat((past_key, key), dim=1)
363
+ value = torch.cat((past_value, value), dim=1)
364
+
365
+ if use_cache:
366
+ present = (key, value)
367
+ else:
368
+ present = None
369
+
370
+ if self.use_logn_attn:
371
+ if self.logn_tensor.device != query.device:
372
+ self.logn_tensor = self.logn_tensor.to(query.device).type_as(query)
373
+ seq_start = key.size(0) - query.size(0)
374
+ seq_end = key.size(0)
375
+ logn_tensor = self.logn_tensor[:, seq_start:seq_end, :, :]
376
+ query = query * logn_tensor.expand_as(query)
377
+
378
+ if self.use_flash_attn:
379
+ q, k, v = query, key, value
380
+ context_layer = self.core_attention_flash(q, k, v)
381
+
382
+ context_layer = rearrange(
383
+ context_layer, "b s h d -> b s (h d)"
384
+ ).contiguous()
385
+ else:
386
+ query = query.permute(0, 2, 1, 3)
387
+ key = key.permute(0, 2, 1, 3)
388
+ value = value.permute(0, 2, 1, 3)
389
+ attn_output, attn_weight = self._attn(
390
+ query, key, value, attention_mask, head_mask
391
+ )
392
+ context_layer = self._merge_heads(
393
+ attn_output, self.num_heads, self.head_dim
394
+ )
395
+
396
+ attn_output = self.c_proj(context_layer)
397
+ outputs = (attn_output, present)
398
+ if output_attentions:
399
+ if self.use_flash_attn:
400
+ raise ValueError("Cannot output attentions while using flash-attn")
401
+ else:
402
+ outputs += (attn_weight,)
403
+
404
+ return outputs
405
+
406
+
407
+ class QWenMLP(nn.Module):
408
+ def __init__(self, config):
409
+ super().__init__()
410
+ self.w1 = nn.Linear(
411
+ config.hidden_size, config.ffn_hidden_size // 2, bias=not config.no_bias
412
+ )
413
+ self.w2 = nn.Linear(
414
+ config.hidden_size, config.ffn_hidden_size // 2, bias=not config.no_bias
415
+ )
416
+ ff_dim_in = config.ffn_hidden_size // 2
417
+ self.c_proj = nn.Linear(ff_dim_in, config.hidden_size, bias=not config.no_bias)
418
+
419
+ def forward(self, hidden_states):
420
+ a1 = self.w1(hidden_states)
421
+ a2 = self.w2(hidden_states)
422
+ intermediate_parallel = a1 * F.silu(a2)
423
+ output = self.c_proj(intermediate_parallel)
424
+ return output
425
+
426
+
427
+ class QWenBlock(nn.Module):
428
+ def __init__(self, config, layer_idx=None, num_expert=1):
429
+ super().__init__()
430
+ self.num_expert = num_expert
431
+ self.layer_number = layer_idx
432
+ self.apply_residual_connection_post_layernorm = (
433
+ config.apply_residual_connection_post_layernorm
434
+ )
435
+ hidden_size = config.hidden_size
436
+ self.apply_residual_connection_post_layernorm = (
437
+ config.apply_residual_connection_post_layernorm
438
+ )
439
+ self.bf16 = config.bf16
440
+
441
+ self.ln_1 = RMSNorm(
442
+ hidden_size,
443
+ eps=config.layer_norm_epsilon,
444
+ )
445
+ self.attn = QWenAttention(config, layer_number=layer_idx)
446
+ self.ln_2 = RMSNorm(
447
+ hidden_size,
448
+ eps=config.layer_norm_epsilon,
449
+ )
450
+
451
+ self.mlp = QWenMLP(config)
452
+
453
+ def forward(
454
+ self,
455
+ hidden_states: Optional[Tuple[torch.FloatTensor]],
456
+ layer_past: Optional[Tuple[torch.Tensor]] = None,
457
+ attention_mask: Optional[torch.FloatTensor] = None,
458
+ head_mask: Optional[torch.FloatTensor] = None,
459
+ encoder_hidden_states: Optional[torch.Tensor] = None,
460
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
461
+ use_cache: Optional[bool] = False,
462
+ output_attentions: Optional[bool] = False,
463
+ ):
464
+ layernorm_output = self.ln_1(hidden_states)
465
+
466
+ attn_outputs = self.attn(
467
+ layernorm_output,
468
+ layer_past=layer_past,
469
+ attention_mask=attention_mask,
470
+ head_mask=head_mask,
471
+ use_cache=use_cache,
472
+ output_attentions=output_attentions,
473
+ )
474
+ attn_output = attn_outputs[0]
475
+
476
+ outputs = attn_outputs[1:]
477
+
478
+ if self.apply_residual_connection_post_layernorm:
479
+ residual = layernorm_output
480
+ else:
481
+ residual = hidden_states
482
+ layernorm_input = attn_output + residual
483
+
484
+ layernorm_output = self.ln_2(layernorm_input)
485
+
486
+ if self.apply_residual_connection_post_layernorm:
487
+ residual = layernorm_output
488
+ else:
489
+ residual = layernorm_input
490
+
491
+ mlp_output = self.mlp(layernorm_output)
492
+ hidden_states = residual + mlp_output
493
+
494
+ if use_cache:
495
+ outputs = (hidden_states,) + outputs
496
+ else:
497
+ outputs = (hidden_states,) + outputs[1:]
498
+
499
+ return outputs
500
+
501
+
502
+ class QWenPreTrainedModel(PreTrainedModel):
503
+ config_class = QWenConfig
504
+ base_model_prefix = "transformer"
505
+ is_parallelizable = False
506
+ supports_gradient_checkpointing = True
507
+ _no_split_modules = ["QWenBlock"]
508
+
509
+ def __init__(self, *inputs, **kwargs):
510
+ super().__init__(*inputs, **kwargs)
511
+
512
+ def _init_weights(self, module):
513
+ """Initialize the weights."""
514
+ if isinstance(module, nn.Linear):
515
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
516
+ if module.bias is not None:
517
+ module.bias.data.zero_()
518
+ elif isinstance(module, nn.Embedding):
519
+ module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
520
+ if module.padding_idx is not None:
521
+ module.weight.data[module.padding_idx].zero_()
522
+ elif isinstance(module, RMSNorm):
523
+ module.weight.data.fill_(1.0)
524
+
525
+ for name, p in module.named_parameters():
526
+ if name == "c_proj.weight":
527
+ p.data.normal_(
528
+ mean=0.0,
529
+ std=(
530
+ self.config.initializer_range
531
+ / math.sqrt(2 * self.config.n_layer)
532
+ ),
533
+ )
534
+
535
+ def _set_gradient_checkpointing(self, module, value=False):
536
+ if isinstance(module, QWenModel):
537
+ module.gradient_checkpointing = value
538
+
539
+
540
+ class QWenModel(QWenPreTrainedModel):
541
+ _keys_to_ignore_on_load_missing = ["attn.masked_bias"]
542
+
543
+ def __init__(self, config):
544
+ super().__init__(config)
545
+ self.vocab_size = config.padded_vocab_size
546
+ self.num_hidden_layers = config.num_hidden_layers
547
+ self.embed_dim = config.hidden_size
548
+
549
+ max_sequence_length = config.max_position_embeddings
550
+ self.position_embedding_type = config.pos_emb
551
+ self.gradient_checkpointing = False
552
+
553
+ if self.position_embedding_type == "learned":
554
+ self.wpe = nn.Embedding(max_sequence_length, self.embed_dim)
555
+ self.init_method(self.position_embeddings.weight)
556
+ self._position_embeddings_key = "position_embeddings"
557
+ self.init_method(self.position_embeddings.weight)
558
+ else:
559
+ self.wpe = None
560
+ self._position_embeddings_key = ""
561
+
562
+ self.wte = nn.Embedding(self.vocab_size, self.embed_dim)
563
+
564
+ self.drop = nn.Dropout(config.embd_pdrop)
565
+ self.h = nn.ModuleList(
566
+ [
567
+ QWenBlock(
568
+ config,
569
+ layer_idx=i,
570
+ )
571
+ for i in range(config.num_hidden_layers)
572
+ ]
573
+ )
574
+ self.ln_f = RMSNorm(
575
+ self.embed_dim,
576
+ eps=config.layer_norm_epsilon,
577
+ )
578
+
579
+ self.post_init()
580
+
581
+ def get_input_embeddings(self):
582
+ return self.wte
583
+
584
+ def set_input_embeddings(self, new_embeddings):
585
+ self.wte = new_embeddings
586
+
587
+ def forward(
588
+ self,
589
+ input_ids: Optional[torch.LongTensor] = None,
590
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
591
+ attention_mask: Optional[torch.FloatTensor] = None,
592
+ token_type_ids: Optional[torch.LongTensor] = None,
593
+ position_ids: Optional[torch.LongTensor] = None,
594
+ head_mask: Optional[torch.FloatTensor] = None,
595
+ inputs_embeds: Optional[torch.FloatTensor] = None,
596
+ encoder_hidden_states: Optional[torch.Tensor] = None,
597
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
598
+ use_cache: Optional[bool] = None,
599
+ output_attentions: Optional[bool] = None,
600
+ output_hidden_states: Optional[bool] = None,
601
+ return_dict: Optional[bool] = None,
602
+ ):
603
+ output_attentions = (
604
+ output_attentions
605
+ if output_attentions is not None
606
+ else self.config.output_attentions
607
+ )
608
+ output_hidden_states = (
609
+ output_hidden_states
610
+ if output_hidden_states is not None
611
+ else self.config.output_hidden_states
612
+ )
613
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
614
+ return_dict = (
615
+ return_dict if return_dict is not None else self.config.use_return_dict
616
+ )
617
+
618
+ if input_ids is not None and inputs_embeds is not None:
619
+ raise ValueError(
620
+ "You cannot specify both input_ids and inputs_embeds at the same time"
621
+ )
622
+ elif input_ids is not None:
623
+ input_shape = input_ids.size()
624
+ input_ids = input_ids.view(-1, input_shape[-1])
625
+ batch_size = input_ids.shape[0]
626
+ elif inputs_embeds is not None:
627
+ input_shape = inputs_embeds.size()[:-1]
628
+ batch_size = inputs_embeds.shape[0]
629
+ else:
630
+ raise ValueError("You have to specify either input_ids or inputs_embeds")
631
+
632
+ device = input_ids.device if input_ids is not None else inputs_embeds.device
633
+
634
+ if token_type_ids is not None:
635
+ token_type_ids = token_type_ids.view(-1, input_shape[-1])
636
+ if position_ids is not None:
637
+ position_ids = position_ids.view(-1, input_shape[-1])
638
+
639
+ if past_key_values is None:
640
+ past_length = 0
641
+ past_key_values = tuple([None] * len(self.h))
642
+ else:
643
+ past_length = past_key_values[0][0].size(-2)
644
+
645
+ if position_ids is None:
646
+ position_ids = torch.arange(
647
+ past_length,
648
+ input_shape[-1] + past_length,
649
+ dtype=torch.long,
650
+ device=device,
651
+ )
652
+ position_ids = position_ids.unsqueeze(0).view(-1, input_shape[-1])
653
+
654
+ if attention_mask is not None:
655
+ if batch_size <= 0:
656
+ raise ValueError("batch_size has to be defined and > 0")
657
+ attention_mask = attention_mask.view(batch_size, -1)
658
+ attention_mask = attention_mask[:, None, None, :]
659
+ attention_mask = attention_mask.to(dtype=self.dtype)
660
+ attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
661
+
662
+ encoder_attention_mask = None
663
+ head_mask = self.get_head_mask(head_mask, self.config.n_layer)
664
+
665
+ if inputs_embeds is None:
666
+ inputs_embeds = self.wte(input_ids)
667
+ hidden_states = inputs_embeds
668
+ if self.wpe is not None:
669
+ position_embeds = self.wpe(position_ids)
670
+ hidden_states = hidden_states + position_embeds
671
+
672
+ hidden_states = self.drop(hidden_states)
673
+ output_shape = input_shape + (hidden_states.size(-1),)
674
+
675
+ if self.gradient_checkpointing and self.training:
676
+ if use_cache:
677
+ logger.warning_once(
678
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
679
+ )
680
+ use_cache = False
681
+
682
+ presents = () if use_cache else None
683
+ all_self_attentions = () if output_attentions else None
684
+ all_hidden_states = () if output_hidden_states else None
685
+ for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
686
+
687
+ if output_hidden_states:
688
+ all_hidden_states = all_hidden_states + (hidden_states,)
689
+
690
+ if self.gradient_checkpointing and self.training:
691
+
692
+ def create_custom_forward(module):
693
+ def custom_forward(*inputs):
694
+ # None for past_key_value
695
+ return module(*inputs, use_cache, output_attentions)
696
+
697
+ return custom_forward
698
+
699
+ outputs = torch.utils.checkpoint.checkpoint(
700
+ create_custom_forward(block),
701
+ hidden_states,
702
+ None,
703
+ attention_mask,
704
+ head_mask[i],
705
+ encoder_hidden_states,
706
+ encoder_attention_mask,
707
+ )
708
+ else:
709
+ outputs = block(
710
+ hidden_states,
711
+ layer_past=layer_past,
712
+ attention_mask=attention_mask,
713
+ head_mask=head_mask[i],
714
+ encoder_hidden_states=encoder_hidden_states,
715
+ encoder_attention_mask=encoder_attention_mask,
716
+ use_cache=use_cache,
717
+ output_attentions=output_attentions,
718
+ )
719
+
720
+ hidden_states = outputs[0]
721
+ if use_cache is True:
722
+ presents = presents + (outputs[2 if output_attentions else 1],)
723
+
724
+ if output_attentions:
725
+ all_self_attentions = all_self_attentions + (outputs[1],)
726
+
727
+ hidden_states = self.ln_f(hidden_states)
728
+ hidden_states = hidden_states.view(output_shape)
729
+
730
+ if not return_dict:
731
+ return tuple(
732
+ v for v in [hidden_states, presents, all_hidden_states] if v is not None
733
+ )
734
+
735
+ return BaseModelOutputWithPast(
736
+ last_hidden_state=hidden_states,
737
+ past_key_values=presents,
738
+ hidden_states=all_hidden_states,
739
+ attentions=all_self_attentions,
740
+ )
741
+
742
+
743
+ class QWenLMHeadModel(QWenPreTrainedModel):
744
+ _keys_to_ignore_on_load_missing = [r"h\.\d+\.attn\.rotary_emb\.inv_freq"]
745
+ _keys_to_ignore_on_load_unexpected = [r"h\.\d+\.attn\.masked_bias"]
746
+
747
+ def __init__(self, config):
748
+ super().__init__(config)
749
+ self.transformer = QWenModel(config)
750
+ self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
751
+ self.post_init()
752
+
753
+ def get_output_embeddings(self):
754
+ return self.lm_head
755
+
756
+ def set_output_embeddings(self, new_embeddings):
757
+ self.lm_head = new_embeddings
758
+
759
+ def prepare_inputs_for_generation(
760
+ self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs
761
+ ):
762
+ token_type_ids = kwargs.get("token_type_ids", None)
763
+ if past_key_values:
764
+ input_ids = input_ids[:, -1].unsqueeze(-1)
765
+ if token_type_ids is not None:
766
+ token_type_ids = token_type_ids[:, -1].unsqueeze(-1)
767
+
768
+ attention_mask = kwargs.get("attention_mask", None)
769
+ position_ids = kwargs.get("position_ids", None)
770
+
771
+ if attention_mask is not None and position_ids is None:
772
+ position_ids = attention_mask.long().cumsum(-1) - 1
773
+ position_ids.masked_fill_(attention_mask == 0, 1)
774
+ if past_key_values:
775
+ position_ids = position_ids[:, -1].unsqueeze(-1)
776
+ else:
777
+ position_ids = None
778
+
779
+ if inputs_embeds is not None and past_key_values is None:
780
+ model_inputs = {"inputs_embeds": inputs_embeds}
781
+ else:
782
+ model_inputs = {"input_ids": input_ids}
783
+
784
+ model_inputs.update(
785
+ {
786
+ "past_key_values": past_key_values,
787
+ "use_cache": kwargs.get("use_cache"),
788
+ "position_ids": position_ids,
789
+ "attention_mask": attention_mask,
790
+ "token_type_ids": token_type_ids,
791
+ }
792
+ )
793
+ return model_inputs
794
+
795
+ def forward(
796
+ self,
797
+ input_ids: Optional[torch.LongTensor] = None,
798
+ past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
799
+ attention_mask: Optional[torch.FloatTensor] = None,
800
+ token_type_ids: Optional[torch.LongTensor] = None,
801
+ position_ids: Optional[torch.LongTensor] = None,
802
+ head_mask: Optional[torch.FloatTensor] = None,
803
+ inputs_embeds: Optional[torch.FloatTensor] = None,
804
+ encoder_hidden_states: Optional[torch.Tensor] = None,
805
+ encoder_attention_mask: Optional[torch.FloatTensor] = None,
806
+ labels: Optional[torch.LongTensor] = None,
807
+ use_cache: Optional[bool] = None,
808
+ output_attentions: Optional[bool] = None,
809
+ output_hidden_states: Optional[bool] = None,
810
+ return_dict: Optional[bool] = None,
811
+ ) -> Union[Tuple, CausalLMOutputWithPast]:
812
+
813
+ return_dict = (
814
+ return_dict if return_dict is not None else self.config.use_return_dict
815
+ )
816
+
817
+ transformer_outputs = self.transformer(
818
+ input_ids,
819
+ past_key_values=past_key_values,
820
+ attention_mask=attention_mask,
821
+ token_type_ids=token_type_ids,
822
+ position_ids=position_ids,
823
+ head_mask=head_mask,
824
+ inputs_embeds=inputs_embeds,
825
+ encoder_hidden_states=encoder_hidden_states,
826
+ encoder_attention_mask=encoder_attention_mask,
827
+ use_cache=use_cache,
828
+ output_attentions=output_attentions,
829
+ output_hidden_states=output_hidden_states,
830
+ return_dict=return_dict,
831
+ )
832
+ hidden_states = transformer_outputs[0]
833
+
834
+ lm_logits = self.lm_head(hidden_states)
835
+
836
+ loss = None
837
+ if labels is not None:
838
+ labels = labels.to(lm_logits.device)
839
+ shift_logits = lm_logits[..., :-1, :].contiguous()
840
+ shift_labels = labels[..., 1:].contiguous()
841
+ loss_fct = CrossEntropyLoss()
842
+ loss = loss_fct(
843
+ shift_logits.view(-1, shift_logits.size(-1)), shift_labels.view(-1)
844
+ )
845
+
846
+ if not return_dict:
847
+ output = (lm_logits,) + transformer_outputs[1:]
848
+ return ((loss,) + output) if loss is not None else output
849
+
850
+ return CausalLMOutputWithPast(
851
+ loss=loss,
852
+ logits=lm_logits,
853
+ past_key_values=transformer_outputs.past_key_values,
854
+ hidden_states=transformer_outputs.hidden_states,
855
+ attentions=transformer_outputs.attentions,
856
+ )
857
+
858
+ @staticmethod
859
+ def _reorder_cache(
860
+ past_key_values: Tuple[Tuple[torch.Tensor]], beam_idx: torch.Tensor
861
+ ) -> Tuple[Tuple[torch.Tensor]]:
862
+
863
+ return tuple(
864
+ tuple(
865
+ past_state.index_select(0, beam_idx.to(past_state.device))
866
+ for past_state in layer_past
867
+ )
868
+ for layer_past in past_key_values
869
+ )
870
+
871
+ def chat(
872
+ self,
873
+ tokenizer: PreTrainedTokenizer,
874
+ query: str,
875
+ history: Optional[HistoryType],
876
+ system: str = "You are a helpful assistant.",
877
+ append_history: bool = True,
878
+ ) -> Tuple[str, HistoryType]:
879
+
880
+ if history is None:
881
+ history = []
882
+
883
+ raw_text, context_tokens = make_context(
884
+ tokenizer,
885
+ query,
886
+ history=history,
887
+ system=system,
888
+ max_window_size=6144,
889
+ chat_format=self.generation_config.chat_format,
890
+ )
891
+
892
+ stop_words_ids = get_stop_words_ids(
893
+ self.generation_config.chat_format, tokenizer
894
+ )
895
+ input_ids = torch.tensor([context_tokens]).to(self.device)
896
+
897
+ outputs = self.generate(
898
+ input_ids,
899
+ stop_words_ids=stop_words_ids,
900
+ return_dict_in_generate=False,
901
+ )
902
+
903
+ response = decode_tokens(
904
+ outputs[0],
905
+ tokenizer,
906
+ raw_text_len=len(raw_text),
907
+ context_length=len(context_tokens),
908
+ chat_format=self.generation_config.chat_format,
909
+ verbose=False,
910
+ )
911
+
912
+ if append_history:
913
+ history.append((query, response))
914
+
915
+ return response, history
916
+
917
+ def generate(
918
+ self,
919
+ inputs: Optional[torch.Tensor] = None,
920
+ generation_config: Optional[GenerationConfig] = None,
921
+ logits_processor: Optional[LogitsProcessorList] = None,
922
+ stopping_criteria: Optional[StoppingCriteriaList] = None,
923
+ prefix_allowed_tokens_fn: Optional[Callable[[int, torch.Tensor], List[int]]] = None,
924
+ synced_gpus: Optional[bool] = None,
925
+ streamer: Optional["BaseStreamer"] = None,
926
+ **kwargs,
927
+ ) -> Union[GenerateOutput, torch.LongTensor]:
928
+ # Process stop_words_ids.
929
+ stop_words_ids = kwargs.pop('stop_words_ids', None)
930
+ if stop_words_ids is None and generation_config is not None:
931
+ stop_words_ids = getattr(generation_config, 'stop_words_ids', None)
932
+ if stop_words_ids is None:
933
+ stop_words_ids = getattr(self.generation_config, 'stop_words_ids', None)
934
+
935
+ if stop_words_ids is not None:
936
+ stop_words_logits_processor = StopWordsLogitsProcessor(
937
+ stop_words_ids=stop_words_ids, eos_token_id=self.generation_config.eos_token_id)
938
+ if logits_processor is None:
939
+ logits_processor = LogitsProcessorList([stop_words_logits_processor])
940
+ else:
941
+ logits_processor.append(stop_words_logits_processor)
942
+
943
+ return super().generate(
944
+ inputs,
945
+ generation_config,
946
+ logits_processor,
947
+ stopping_criteria,
948
+ prefix_allowed_tokens_fn,
949
+ synced_gpus,
950
+ streamer,
951
+ **kwargs,
952
+ )
953
+
954
+
955
+ class RotaryEmbedding(torch.nn.Module):
956
+ def __init__(self, dim, base=10000):
957
+ super().__init__()
958
+ self.dim = dim
959
+ self.base = base
960
+ inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float() / dim))
961
+ self.register_buffer("inv_freq", inv_freq)
962
+ if importlib.util.find_spec("einops") is None:
963
+ raise RuntimeError("einops is required for Rotary Embedding")
964
+
965
+ self._rotary_pos_emb_cache = None
966
+ self._seq_len_cached = 0
967
+ self._ntk_alpha_cached = 1.0
968
+
969
+ def update_rotary_pos_emb_cache(self, max_seq_len, offset=0, ntk_alpha=1.0):
970
+ seqlen = max_seq_len + offset
971
+ if seqlen > self._seq_len_cached or ntk_alpha != self._ntk_alpha_cached:
972
+ base = self.base * ntk_alpha ** (self.dim / (self.dim - 2))
973
+ self.inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, device=self.inv_freq.device).float() / self.dim))
974
+ self._seq_len_cached = seqlen
975
+ self._ntk_alpha_cached = ntk_alpha
976
+ seq = torch.arange(seqlen, device=self.inv_freq.device)
977
+ freqs = torch.outer(seq.type_as(self.inv_freq), self.inv_freq)
978
+ emb = torch.cat((freqs, freqs), dim=-1)
979
+ from einops import rearrange
980
+
981
+ self._rotary_pos_emb_cache = rearrange(emb, "n d -> 1 n 1 d")
982
+
983
+ def forward(self, max_seq_len, offset=0, ntk_alpha=1.0):
984
+ self.update_rotary_pos_emb_cache(max_seq_len, offset, ntk_alpha)
985
+ return self._rotary_pos_emb_cache[:, offset : offset + max_seq_len]
986
+
987
+
988
+ def _rotate_half(x):
989
+ from einops import rearrange
990
+
991
+ x = rearrange(x, "... (j d) -> ... j d", j=2)
992
+ x1, x2 = x.unbind(dim=-2)
993
+ return torch.cat((-x2, x1), dim=-1)
994
+
995
+
996
+ def apply_rotary_pos_emb(t, freqs, use_flash_rotary=False):
997
+ if use_flash_rotary:
998
+ t_ = t.float()
999
+ freqs = freqs.squeeze(0).squeeze(1)
1000
+ cos = freqs[:, : freqs.shape[-1] // 2].cos()
1001
+ sin = freqs[:, : freqs.shape[-1] // 2].sin()
1002
+ output = apply_rotary_emb_func(t_, cos, sin).type_as(t)
1003
+ return output
1004
+ else:
1005
+ rot_dim = freqs.shape[-1]
1006
+ t_, t_pass_ = t[..., :rot_dim], t[..., rot_dim:]
1007
+ t_ = t_.float()
1008
+ t_pass_ = t_pass_.float()
1009
+ t_ = (t_ * freqs.cos()) + (_rotate_half(t_) * freqs.sin())
1010
+ return torch.cat((t_, t_pass_), dim=-1).type_as(t)
1011
+
1012
+
1013
+ class RMSNorm(torch.nn.Module):
1014
+ def __init__(self, dim: int, eps: float = 1e-6):
1015
+ super().__init__()
1016
+ self.eps = eps
1017
+ self.weight = nn.Parameter(torch.ones(dim))
1018
+
1019
+ def _norm(self, x):
1020
+ return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
1021
+
1022
+ def forward(self, x):
1023
+ if rms_norm is not None:
1024
+ return rms_norm(x, self.weight, self.eps)
1025
+ else:
1026
+ output = self._norm(x.float()).type_as(x)
1027
+ return output * self.weight
pytorch_model.bin CHANGED
@@ -1,3 +1,3 @@
1
  version https://git-lfs.github.com/spec/v1
2
- oid sha256:b59e725947bb45fd0574f6b2f653dc89d283e024ec47bef6e7caba0711669b63
3
  size 15442733145
 
1
  version https://git-lfs.github.com/spec/v1
2
+ oid sha256:7361b298a82b284129276f586dec570b5d41259130d190960321cc0db92d958f
3
  size 15442733145
qwen.tiktoken ADDED
The diff for this file is too large to render. See raw diff
 
qwen_generation_utils.py ADDED
@@ -0,0 +1,411 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ """Generation support."""
7
+
8
+ from typing import Tuple, List, Union, Iterable
9
+
10
+ import numpy as np
11
+ import torch
12
+ import torch.nn.functional as F
13
+ from transformers import PreTrainedTokenizer
14
+ from transformers import logging
15
+ from transformers.generation import LogitsProcessor
16
+
17
+ logger = logging.get_logger(__name__)
18
+
19
+ # Types.
20
+ HistoryType = List[Tuple[str, str]]
21
+ TokensType = List[int]
22
+ BatchTokensType = List[List[int]]
23
+
24
+
25
+ def pad_batch(batch: BatchTokensType, pad_id: int, seq_length: int) -> BatchTokensType:
26
+ for tokens in batch:
27
+ context_length = len(tokens)
28
+ if context_length < seq_length:
29
+ tokens.extend([pad_id] * (seq_length - context_length))
30
+ return batch
31
+
32
+
33
+ def get_ltor_masks_and_position_ids(
34
+ data,
35
+ eod_token,
36
+ reset_position_ids,
37
+ reset_attention_mask,
38
+ eod_mask_loss,
39
+ ):
40
+ """Build masks and position id for left to right model."""
41
+
42
+ # Extract batch size and sequence length.
43
+ micro_batch_size, seq_length = data.size()
44
+
45
+ # Attention mask (lower triangular).
46
+ if reset_attention_mask:
47
+ att_mask_batch = micro_batch_size
48
+ else:
49
+ att_mask_batch = 1
50
+ attention_mask = torch.tril(
51
+ torch.ones((att_mask_batch, seq_length, seq_length), device=data.device)
52
+ ).view(att_mask_batch, 1, seq_length, seq_length)
53
+
54
+ # Loss mask.
55
+ loss_mask = torch.ones(data.size(), dtype=torch.float, device=data.device)
56
+ if eod_mask_loss:
57
+ loss_mask[data == eod_token] = 0.0
58
+
59
+ # Position ids.
60
+ position_ids = torch.arange(seq_length, dtype=torch.long, device=data.device)
61
+ position_ids = position_ids.unsqueeze(0).expand_as(data)
62
+ # We need to clone as the ids will be modifed based on batch index.
63
+ if reset_position_ids:
64
+ position_ids = position_ids.clone()
65
+
66
+ if reset_position_ids or reset_attention_mask:
67
+ # Loop through the batches:
68
+ for b in range(micro_batch_size):
69
+
70
+ # Find indecies where EOD token is.
71
+ eod_index = position_ids[b, data[b] == eod_token]
72
+ # Detach indecies from positions if going to modify positions.
73
+ if reset_position_ids:
74
+ eod_index = eod_index.clone()
75
+
76
+ # Loop through EOD indecies:
77
+ prev_index = 0
78
+ for j in range(eod_index.size()[0]):
79
+ i = eod_index[j]
80
+ # Mask attention loss.
81
+ if reset_attention_mask:
82
+ attention_mask[b, 0, (i + 1) :, : (i + 1)] = 0
83
+ # Reset positions.
84
+ if reset_position_ids:
85
+ position_ids[b, (i + 1) :] -= i + 1 - prev_index
86
+ prev_index = i + 1
87
+
88
+ # Convert attention mask to binary:
89
+ attention_mask = attention_mask < 0.5
90
+
91
+ return attention_mask, loss_mask, position_ids
92
+
93
+
94
+ def get_batch(context_tokens: torch.LongTensor, eod_id: int):
95
+ """Generate batch from context tokens."""
96
+ # Move to GPU.
97
+ tokens = context_tokens.contiguous().to(context_tokens.device)
98
+ # Get the attention mask and postition ids.
99
+ attention_mask, _, position_ids = get_ltor_masks_and_position_ids(
100
+ tokens,
101
+ eod_id,
102
+ reset_position_ids=False,
103
+ reset_attention_mask=False,
104
+ eod_mask_loss=False,
105
+ )
106
+ return tokens, attention_mask, position_ids
107
+
108
+
109
+ def get_stop_words_ids(chat_format, tokenizer):
110
+ if chat_format == "raw":
111
+ stop_words_ids = [tokenizer.encode("Human:"), [tokenizer.eod_id]]
112
+ elif chat_format == "chatml":
113
+ stop_words_ids = [[tokenizer.im_end_id], [tokenizer.im_start_id]]
114
+ else:
115
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
116
+ return stop_words_ids
117
+
118
+
119
+ def make_context(
120
+ tokenizer: PreTrainedTokenizer,
121
+ query: str,
122
+ history: List[Tuple[str, str]] = None,
123
+ system: str = "",
124
+ max_window_size: int = 6144,
125
+ chat_format: str = "chatml",
126
+ ):
127
+ if history is None:
128
+ history = []
129
+
130
+ if chat_format == "chatml":
131
+ im_start, im_end = "<|im_start|>", "<|im_end|>"
132
+ im_start_tokens = [tokenizer.im_start_id]
133
+ im_end_tokens = [tokenizer.im_end_id]
134
+ nl_tokens = tokenizer.encode("\n")
135
+
136
+ def _tokenize_str(role, content):
137
+ return f"{role}\n{content}", tokenizer.encode(
138
+ role
139
+ ) + nl_tokens + tokenizer.encode(content)
140
+
141
+ system_text, system_tokens_part = _tokenize_str("system", system)
142
+ system_tokens = im_start_tokens + system_tokens_part + im_end_tokens
143
+
144
+ raw_text = ""
145
+ context_tokens = []
146
+
147
+ for turn_query, turn_response in reversed(history):
148
+ query_text, query_tokens_part = _tokenize_str("user", turn_query)
149
+ query_tokens = im_start_tokens + query_tokens_part + im_end_tokens
150
+ response_text, response_tokens_part = _tokenize_str(
151
+ "assistant", turn_response
152
+ )
153
+ response_tokens = im_start_tokens + response_tokens_part + im_end_tokens
154
+
155
+ next_context_tokens = nl_tokens + query_tokens + nl_tokens + response_tokens
156
+ prev_chat = (
157
+ f"\n{im_start}{query_text}{im_end}\n{im_start}{response_text}{im_end}"
158
+ )
159
+
160
+ current_context_size = (
161
+ len(system_tokens) + len(next_context_tokens) + len(context_tokens)
162
+ )
163
+ if current_context_size < max_window_size:
164
+ context_tokens = next_context_tokens + context_tokens
165
+ raw_text = prev_chat + raw_text
166
+ else:
167
+ break
168
+
169
+ context_tokens = system_tokens + context_tokens
170
+ raw_text = f"{im_start}{system_text}{im_end}" + raw_text
171
+ context_tokens += (
172
+ nl_tokens
173
+ + im_start_tokens
174
+ + _tokenize_str("user", query)[1]
175
+ + im_end_tokens
176
+ + nl_tokens
177
+ + im_start_tokens
178
+ + tokenizer.encode("assistant")
179
+ + nl_tokens
180
+ )
181
+ raw_text += f"\n{im_start}user\n{query}{im_end}\n{im_start}assistant\n"
182
+
183
+ elif chat_format == "raw":
184
+ raw_text = query
185
+ context_tokens = tokenizer.encode(raw_text)
186
+ else:
187
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
188
+
189
+ return raw_text, context_tokens
190
+
191
+
192
+ def _decode_default(
193
+ tokens: List[int],
194
+ *,
195
+ stop_words: List[str],
196
+ eod_words: List[str],
197
+ tokenizer: PreTrainedTokenizer,
198
+ raw_text_len: int,
199
+ verbose: bool = False,
200
+ return_end_reason: bool = False,
201
+ ):
202
+ trim_decode_tokens = tokenizer.decode(tokens)[raw_text_len:]
203
+ if verbose:
204
+ print("\nRaw Generate: ", trim_decode_tokens)
205
+
206
+ end_reason = f"Gen length {len(tokens)}"
207
+ for stop_word in stop_words:
208
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
209
+ for eod_word in eod_words:
210
+ if eod_word in trim_decode_tokens:
211
+ end_reason = f"Gen {eod_word!r}"
212
+ trim_decode_tokens = trim_decode_tokens.split(eod_word)[0]
213
+ trim_decode_tokens = trim_decode_tokens.strip()
214
+ if verbose:
215
+ print("\nEnd Reason:", end_reason)
216
+ print("\nGenerate: ", trim_decode_tokens)
217
+
218
+ if return_end_reason:
219
+ return trim_decode_tokens, end_reason
220
+ else:
221
+ return trim_decode_tokens
222
+
223
+
224
+ def _decode_chatml(
225
+ tokens: List[int],
226
+ *,
227
+ stop_words: List[str],
228
+ eod_token_ids: List[int],
229
+ tokenizer: PreTrainedTokenizer,
230
+ raw_text_len: int,
231
+ context_length: int,
232
+ verbose: bool = False,
233
+ return_end_reason: bool = False,
234
+ ):
235
+ end_reason = f"Gen length {len(tokens)}"
236
+ eod_token_idx = context_length
237
+ for eod_token_idx in range(context_length, len(tokens)):
238
+ if tokens[eod_token_idx] in eod_token_ids:
239
+ end_reason = f"Gen {tokenizer.decode([tokens[eod_token_idx]])!r}"
240
+ break
241
+
242
+ trim_decode_tokens = tokenizer.decode(tokens[:eod_token_idx])[raw_text_len:]
243
+ if verbose:
244
+ print("\nRaw Generate w/o EOD:", tokenizer.decode(tokens)[raw_text_len:])
245
+ print("\nRaw Generate:", trim_decode_tokens)
246
+ print("\nEnd Reason:", end_reason)
247
+ for stop_word in stop_words:
248
+ trim_decode_tokens = trim_decode_tokens.replace(stop_word, "").strip()
249
+ trim_decode_tokens = trim_decode_tokens.strip()
250
+ if verbose:
251
+ print("\nGenerate:", trim_decode_tokens)
252
+
253
+ if return_end_reason:
254
+ return trim_decode_tokens, end_reason
255
+ else:
256
+ return trim_decode_tokens
257
+
258
+
259
+ def decode_tokens(
260
+ tokens: Union[torch.LongTensor, TokensType],
261
+ tokenizer: PreTrainedTokenizer,
262
+ raw_text_len: int,
263
+ context_length: int,
264
+ chat_format: str,
265
+ verbose: bool = False,
266
+ return_end_reason: bool = False,
267
+ ) -> str:
268
+ if torch.is_tensor(tokens):
269
+ tokens = tokens.cpu().numpy().tolist()
270
+
271
+ if chat_format == "chatml":
272
+ return _decode_chatml(
273
+ tokens,
274
+ stop_words=[],
275
+ eod_token_ids=[tokenizer.im_start_id, tokenizer.im_end_id],
276
+ tokenizer=tokenizer,
277
+ raw_text_len=raw_text_len,
278
+ context_length=context_length,
279
+ verbose=verbose,
280
+ return_end_reason=return_end_reason,
281
+ )
282
+ elif chat_format == "raw":
283
+ return _decode_default(
284
+ tokens,
285
+ stop_words=["<|endoftext|>"],
286
+ eod_words=["<|endoftext|>"],
287
+ tokenizer=tokenizer,
288
+ raw_text_len=raw_text_len,
289
+ verbose=verbose,
290
+ return_end_reason=return_end_reason,
291
+ )
292
+ else:
293
+ raise NotImplementedError(f"Unknown chat format {chat_format!r}")
294
+
295
+
296
+ class StopWordsLogitsProcessor(LogitsProcessor):
297
+ """
298
+ :class:`transformers.LogitsProcessor` that enforces that when specified sequences appear, stop geration.
299
+
300
+ Args:
301
+ stop_words_ids (:obj:`List[List[int]]`):
302
+ List of list of token ids of stop ids. In order to get the tokens of the words
303
+ that should not appear in the generated text, use :obj:`tokenizer(bad_word,
304
+ add_prefix_space=True).input_ids`.
305
+ eos_token_id (:obj:`int`):
306
+ The id of the `end-of-sequence` token.
307
+ """
308
+
309
+ def __init__(self, stop_words_ids: Iterable[Iterable[int]], eos_token_id: int):
310
+
311
+ if not isinstance(stop_words_ids, List) or len(stop_words_ids) == 0:
312
+ raise ValueError(
313
+ f"`stop_words_ids` has to be a non-emtpy list, but is {stop_words_ids}."
314
+ )
315
+ if any(not isinstance(bad_word_ids, list) for bad_word_ids in stop_words_ids):
316
+ raise ValueError(
317
+ f"`stop_words_ids` has to be a list of lists, but is {stop_words_ids}."
318
+ )
319
+ if any(
320
+ any(
321
+ (not isinstance(token_id, (int, np.integer)) or token_id < 0)
322
+ for token_id in stop_word_ids
323
+ )
324
+ for stop_word_ids in stop_words_ids
325
+ ):
326
+ raise ValueError(
327
+ f"Each list in `stop_words_ids` has to be a list of positive integers, but is {stop_words_ids}."
328
+ )
329
+
330
+ self.stop_words_ids = list(
331
+ filter(
332
+ lambda bad_token_seq: bad_token_seq != [eos_token_id], stop_words_ids
333
+ )
334
+ )
335
+ self.eos_token_id = eos_token_id
336
+ for stop_token_seq in self.stop_words_ids:
337
+ assert (
338
+ len(stop_token_seq) > 0
339
+ ), "Stop words token sequences {} cannot have an empty list".format(
340
+ stop_words_ids
341
+ )
342
+
343
+ def __call__(
344
+ self, input_ids: torch.LongTensor, scores: torch.FloatTensor
345
+ ) -> torch.FloatTensor:
346
+ stopped_samples = self._calc_stopped_samples(input_ids)
347
+ for i, should_stop in enumerate(stopped_samples):
348
+ if should_stop:
349
+ scores[i, self.eos_token_id] = float(2**30)
350
+ return scores
351
+
352
+ def _tokens_match(self, prev_tokens: torch.LongTensor, tokens: List[int]) -> bool:
353
+ if len(tokens) == 0:
354
+ # if bad word tokens is just one token always ban it
355
+ return True
356
+ elif len(tokens) > len(prev_tokens):
357
+ # if bad word tokens are longer then prev input_ids they can't be equal
358
+ return False
359
+ elif prev_tokens[-len(tokens) :].tolist() == tokens:
360
+ # if tokens match
361
+ return True
362
+ else:
363
+ return False
364
+
365
+ def _calc_stopped_samples(self, prev_input_ids: Iterable[int]) -> Iterable[int]:
366
+ stopped_samples = []
367
+ for prev_input_ids_slice in prev_input_ids:
368
+ match = False
369
+ for stop_token_seq in self.stop_words_ids:
370
+ if self._tokens_match(prev_input_ids_slice, stop_token_seq):
371
+ # if tokens do not match continue
372
+ match = True
373
+ break
374
+ stopped_samples.append(match)
375
+
376
+ return stopped_samples
377
+
378
+
379
+ def top_k_logits(logits, top_k=0, top_p=0.0, filter_value=-float("Inf")):
380
+ """This function has been mostly taken from huggingface conversational
381
+ ai code at
382
+ https://medium.com/huggingface/how-to-build-a-state-of-the-art-
383
+ conversational-ai-with-transfer-learning-2d818ac26313"""
384
+
385
+ if top_k > 0:
386
+ # Remove all tokens with a probability less than the
387
+ # last token of the top-k
388
+ indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
389
+ logits[indices_to_remove] = filter_value
390
+
391
+ if top_p > 0.0:
392
+ # Cconvert to 1D
393
+ sorted_logits, sorted_indices = torch.sort(logits, descending=True, dim=-1)
394
+ cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
395
+
396
+ # Remove tokens with cumulative probability above the threshold
397
+ sorted_indices_to_remove = cumulative_probs > top_p
398
+ # Shift the indices to the right to keep also the first token
399
+ # above the threshold
400
+ sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
401
+ sorted_indices_to_remove[..., 0] = 0
402
+ for i in range(sorted_indices.size(0)):
403
+ indices_to_remove = sorted_indices[i][sorted_indices_to_remove[i]]
404
+ logits[i][indices_to_remove] = filter_value
405
+
406
+ return logits
407
+
408
+
409
+ def switch(val1, val2, boolean):
410
+ boolean = boolean.type_as(val1)
411
+ return (1 - boolean) * val1 + boolean * val2
tokenization_qwen.py ADDED
@@ -0,0 +1,243 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Alibaba Cloud.
2
+ #
3
+ # This source code is licensed under the license found in the
4
+ # LICENSE file in the root directory of this source tree.
5
+
6
+ """Tokenization classes for QWen."""
7
+
8
+ from __future__ import absolute_import, division, print_function, unicode_literals
9
+
10
+ import json
11
+ import logging
12
+ import os
13
+ import unicodedata
14
+ from io import open
15
+ import base64
16
+ import tiktoken
17
+ from typing import List, Optional, Tuple, Union
18
+
19
+ from transformers import PreTrainedTokenizer, AddedToken
20
+
21
+ logger = logging.getLogger(__name__)
22
+
23
+ TIKTOKEN_NAME = "qwen.tiktoken"
24
+
25
+
26
+ class QWenTokenizer(PreTrainedTokenizer):
27
+ """QWen tokenizer."""
28
+
29
+ """NOTE: This tokenizer will not handle special tokens to avoid injection attacks"""
30
+
31
+ @classmethod
32
+ def from_pretrained(
33
+ cls, pretrained_model_name_or_path, cache_dir=None, *inputs, **kwargs
34
+ ):
35
+ merges_file = os.path.join(pretrained_model_name_or_path, TIKTOKEN_NAME)
36
+ tokenizer = cls(merges_file, *inputs, **kwargs)
37
+ return tokenizer
38
+
39
+ def __init__(
40
+ self,
41
+ merges_file,
42
+ errors="replace",
43
+ max_len=None,
44
+ unk_token="<|endoftext|>",
45
+ bos_token="<|endoftext|>",
46
+ eos_token="<|endoftext|>",
47
+ pad_token=None,
48
+ add_prefix_space=False,
49
+ add_bos_token=False,
50
+ add_more_sp_tokens=True,
51
+ **kwargs,
52
+ ):
53
+ bos_token = (
54
+ AddedToken(bos_token, lstrip=False, rstrip=False)
55
+ if isinstance(bos_token, str)
56
+ else bos_token
57
+ )
58
+ eos_token = (
59
+ AddedToken(eos_token, lstrip=False, rstrip=False)
60
+ if isinstance(eos_token, str)
61
+ else eos_token
62
+ )
63
+ unk_token = (
64
+ AddedToken(unk_token, lstrip=False, rstrip=False)
65
+ if isinstance(unk_token, str)
66
+ else unk_token
67
+ )
68
+ pad_token = (
69
+ AddedToken(pad_token, lstrip=False, rstrip=False)
70
+ if isinstance(pad_token, str)
71
+ else pad_token
72
+ )
73
+ super().__init__(
74
+ errors=errors,
75
+ unk_token=unk_token,
76
+ bos_token=bos_token,
77
+ eos_token=eos_token,
78
+ pad_token=pad_token,
79
+ add_prefix_space=add_prefix_space,
80
+ add_bos_token=add_bos_token,
81
+ )
82
+ self.add_bos_token = add_bos_token
83
+ self.max_len = max_len if max_len is not None else int(1e12)
84
+
85
+ self.errors = errors # how to handle errors in decoding
86
+
87
+ name = "QWen"
88
+ ENDOFTEXT = "<|endoftext|>"
89
+ IMSTART = "<|im_start|>"
90
+ IMEND = "<|im_end|>"
91
+ if add_more_sp_tokens:
92
+ special_tokens = (
93
+ ENDOFTEXT,
94
+ IMSTART,
95
+ IMEND,
96
+ "<R>",
97
+ "<S>",
98
+ "<X>",
99
+ "<mask>",
100
+ "<sep>",
101
+ ) + tuple([f"<extra_{i}>" for i in range(200)])
102
+ else:
103
+ special_tokens = (ENDOFTEXT, IMSTART, IMEND)
104
+
105
+ PAT_STR = r"""(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+"""
106
+
107
+ def load_tiktoken_bpe(tiktoken_bpe_file: str) -> "dict[bytes, int]":
108
+ contents = open(tiktoken_bpe_file, "rb").read()
109
+ return {
110
+ base64.b64decode(token): int(rank)
111
+ for token, rank in (
112
+ line.split() for line in contents.splitlines() if line
113
+ )
114
+ }
115
+
116
+ mergeable_ranks = load_tiktoken_bpe(merges_file)
117
+ special_tokens = {
118
+ token: index
119
+ for index, token in enumerate(special_tokens, start=len(mergeable_ranks))
120
+ }
121
+ self.special_tokens = special_tokens
122
+ enc = tiktoken.Encoding(
123
+ name,
124
+ pat_str=PAT_STR,
125
+ mergeable_ranks=mergeable_ranks,
126
+ special_tokens=special_tokens,
127
+ )
128
+ assert (
129
+ len(mergeable_ranks) + len(special_tokens) == enc.n_vocab
130
+ ), f"{len(mergeable_ranks) + len(special_tokens)} != {enc.n_vocab} in encoding"
131
+
132
+ self.mergeable_ranks = mergeable_ranks
133
+ self.encoder = self.mergeable_ranks
134
+ self.decoder = {v: k for k, v in self.encoder.items()}
135
+ self.tokenizer = enc # type: tiktoken.Encoding
136
+ self.eod_id = self.tokenizer.eot_token
137
+ self.im_start_id = special_tokens[IMSTART]
138
+ self.im_end_id = special_tokens[IMEND]
139
+
140
+ def __len__(self):
141
+ return self.tokenizer.n_vocab
142
+
143
+ def get_vocab(self):
144
+ return self.mergeable_ranks
145
+
146
+ def convert_tokens_to_ids(self, tokens):
147
+ ids = []
148
+ # Remove support for py2
149
+ if isinstance(tokens, str):
150
+ if tokens in self.special_tokens:
151
+ return self.special_tokens[tokens]
152
+ else:
153
+ return self.encoder.get(tokens)
154
+ for token in tokens:
155
+ if token in self.special_tokens:
156
+ ids.append(self.special_tokens[token])
157
+ else:
158
+ ids.append(self.encoder.get(token))
159
+ if len(ids) > self.max_len:
160
+ logger.warning(
161
+ "Token indices sequence length is longer than the specified maximum "
162
+ " sequence length for this OpenAI GPT model ({} > {}). Running this"
163
+ " sequence through the model will result in indexing errors".format(
164
+ len(ids), self.max_len
165
+ )
166
+ )
167
+ return ids
168
+
169
+ def save_vocabulary(self, save_directory: str) -> Tuple[str]:
170
+ """
171
+ Save only the vocabulary of the tokenizer (vocabulary + added tokens).
172
+
173
+ Returns:
174
+ `Tuple(str)`: Paths to the files saved.
175
+ """
176
+ file_path = os.path.join(save_directory, "qwen.tiktoken")
177
+ with open(file_path, "w", encoding="utf8") as w:
178
+ for k, v in self.mergeable_ranks.items():
179
+ line = base64.b64encode(k).decode("utf8") + " " + str(v) + "\n"
180
+ w.write(line)
181
+ return (file_path,)
182
+
183
+ def tokenize(self, text: str, **kwargs) -> List[str]:
184
+ """
185
+ Converts a string in a sequence of tokens, replacing unknown tokens with the `unk_token`.
186
+
187
+ Args:
188
+ text (`str`):
189
+ The sequence to be encoded.
190
+ pair (`str`, *optional*):
191
+ A second sequence to be encoded with the first.
192
+ add_special_tokens (`bool`, *optional*, defaults to `False`):
193
+ Whether or not to add the special tokens associated with the corresponding model.
194
+ kwargs (additional keyword arguments, *optional*):
195
+ Will be passed to the underlying model specific encode method. See details in
196
+ [`~PreTrainedTokenizerBase.__call__`]
197
+
198
+ Returns:
199
+ `List[str]`: The list of tokens.
200
+ """
201
+ tokens = []
202
+ text = unicodedata.normalize("NFC", text)
203
+ for t in self.tokenizer.encode_ordinary(text):
204
+ tokens.append(self.decoder[t])
205
+ return tokens
206
+
207
+ def convert_tokens_to_string(self, tokens: List[str]) -> str:
208
+ """
209
+ Converts a sequence of tokens in a single string. The most simple way to do it is `" ".join(tokens)` but we
210
+ often want to remove sub-word tokenization artifacts at the same time.
211
+ """
212
+ text = "".join(tokens)
213
+ text = bytearray([self.byte_decoder[c] for c in text]).decode(
214
+ "utf-8", errors=self.errors
215
+ )
216
+ return text
217
+
218
+ @property
219
+ def vocab_size(self):
220
+ return self.tokenizer.n_vocab
221
+
222
+ def _convert_id_to_token(self, index: int) -> str:
223
+ raise NotImplementedError
224
+
225
+ def _tokenize(self, text, **kwargs):
226
+ """
227
+ Converts a string in a sequence of tokens (string), using the tokenizer. Split in words for word-based
228
+ vocabulary or sub-words for sub-word-based vocabularies (BPE/SentencePieces/WordPieces).
229
+
230
+ Do NOT take care of added tokens.
231
+ """
232
+ raise NotImplementedError
233
+
234
+ def _decode(
235
+ self,
236
+ token_ids: Union[int, List[int]],
237
+ skip_special_tokens: bool = False,
238
+ clean_up_tokenization_spaces: bool = None,
239
+ **kwargs,
240
+ ) -> str:
241
+ if isinstance(token_ids, int):
242
+ token_ids = [token_ids]
243
+ return self.tokenizer.decode(token_ids)
tokenizer_config.json ADDED
@@ -0,0 +1,11 @@
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "remove_space": false,
3
+ "do_lower_case": false,
4
+ "tokenizer_class": "QWenTokenizer",
5
+ "auto_map": {
6
+ "AutoTokenizer": [
7
+ "tokenization_qwen.QWenTokenizer",
8
+ null
9
+ ]
10
+ }
11
+ }