Add qwen2 (#28436)
* add config, modeling, and tokenization * add auto and init * update readme * update readme * update team name * fixup * fixup * update config * update code style * update for fixup * update for fixup * update for fixup * update for testing * update for testing * fix bug for config and tokenization * fix bug for bos token * not doctest * debug tokenizer * not doctest * debug tokenization * debug init for tokenizer * fix style * update init * delete if in token auto * add tokenizer doc * add tokenizer in init * Update dummy_tokenizers_objects.py * update * update * debug * Update tokenization_qwen2.py * debug * Update convert_slow_tokenizer.py * add copies * add copied from and make style * update files map * update test * fix style * fix merge reading and update tests * fix tests * fix tests * fix style * debug a variable in readme * Update src/transformers/models/qwen2/configuration_qwen2.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * update test and copied from * fix style * update qwen2 tokenization and tests * Update tokenization_qwen2.py * delete the copied from after property * fix style * update tests * update tests * add copied from * fix bugs * update doc * add warning for sliding window attention * update qwen2 tokenization * fix style * Update src/transformers/models/qwen2/modeling_qwen2.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * fix tokenizer fast --------- Co-authored-by: Ren Xuancheng <jklj077@users.noreply.github.com> Co-authored-by: renxuancheng.rxc <renxuancheng.rxc@alibaba-inc.com> Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
Junyang Lin
@@ -458,6 +458,7 @@ Current number of checkpoints: ** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (from the Qwen team, Alibaba Group) released with the paper [Qwen Technical Report](https://arxiv.org/abs/2309.16609) by Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
|
||||
|
||||
@@ -433,6 +433,7 @@ Número actual de puntos de control: ** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (from the Qwen team, Alibaba Group) released with the paper [Qwen Technical Report](https://arxiv.org/abs/2309.16609) by Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
|
||||
|
||||
@@ -407,6 +407,7 @@ conda install conda-forge::transformers
|
||||
1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (माइक्रोसॉफ्ट रिसर्च से) साथ में पेपर [ProphetNet: प्रेडिक्टिंग फ्यूचर एन-ग्राम फॉर सीक्वेंस-टू-सीक्वेंस प्री-ट्रेनिंग ](https://arxiv.org/abs/2001.04063) यू यान, वीज़ेन क्यूई, येयुन गोंग, दयाहेंग लियू, नान डुआन, जिउशेंग चेन, रुओफ़ेई झांग और मिंग झोउ द्वारा पोस्ट किया गया।
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. से) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. द्वाराअनुसंधान पत्र [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) के साथ जारी किया गया
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA से) साथ वाला पेपर [डीप लर्निंग इंफ़ेक्शन के लिए इंटीजर क्वांटिज़ेशन: प्रिंसिपल्स एंड एम्पिरिकल इवैल्यूएशन](https:// arxiv.org/abs/2004.09602) हाओ वू, पैट्रिक जुड, जिआओजी झांग, मिखाइल इसेव और पॉलियस माइकेविसियस द्वारा।
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (the Qwen team, Alibaba Group से) Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu. द्वाराअनुसंधान पत्र [Qwen Technical Report](https://arxiv.org/abs/2309.16609) के साथ जारी किया गया
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (फेसबुक से) साथ में कागज [रिट्रीवल-ऑगमेंटेड जेनरेशन फॉर नॉलेज-इंटेंसिव एनएलपी टास्क](https://arxiv .org/abs/2005.11401) पैट्रिक लुईस, एथन पेरेज़, अलेक्जेंड्रा पिक्टस, फैबियो पेट्रोनी, व्लादिमीर कारपुखिन, नमन गोयल, हेनरिक कुटलर, माइक लुईस, वेन-ताउ यिह, टिम रॉकटाशेल, सेबस्टियन रिडेल, डौवे कीला द्वारा।
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google अनुसंधान से) केल्विन गु, केंटन ली, ज़ोरा तुंग, पानुपोंग पसुपत और मिंग-वेई चांग द्वारा साथ में दिया गया पेपर [REALM: रिट्रीवल-ऑगमेंटेड लैंग्वेज मॉडल प्री-ट्रेनिंग](https://arxiv.org/abs/2002.08909)।
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
|
||||
|
||||
@@ -467,6 +467,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
|
||||
1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research から) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou から公開された研究論文: [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063)
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. から) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. から公開された研究論文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA から) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius から公開された研究論文: [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602)
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (the Qwen team, Alibaba Group から) Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu. から公開された研究論文 [Qwen Technical Report](https://arxiv.org/abs/2309.16609)
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (Facebook から) Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela から公開された研究論文: [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401)
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google Research から) Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang から公開された研究論文: [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909)
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (Google Research から) Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya から公開された研究論文: [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451)
|
||||
|
||||
@@ -382,6 +382,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
|
||||
1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research 에서) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 의 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 논문과 함께 발표했습니다.
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. 에서 제공)은 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.의 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)논문과 함께 발표했습니다.
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA 에서) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 의 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 논문과 함께 발표했습니다.
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (the Qwen team, Alibaba Group 에서 제공)은 Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.의 [Qwen Technical Report](https://arxiv.org/abs/2309.16609)논문과 함께 발표했습니다.
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (Facebook 에서) Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela 의 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 논문과 함께 발표했습니다.
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google Research 에서) Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 의 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 논문과 함께 발표했습니다.
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (Google Research 에서) Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya 의 [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 논문과 함께 발표했습니다.
|
||||
|
||||
@@ -406,6 +406,7 @@ conda install conda-forge::transformers
|
||||
1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (来自 Nanjing University, The University of Hong Kong etc.) 伴随论文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) 由 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao 发布。
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (来自 NVIDIA) 伴随论文 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 由 Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 发布。
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (来自 the Qwen team, Alibaba Group) 伴随论文 [Qwen Technical Report](https://arxiv.org/abs/2309.16609) 由 Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu 发布。
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (来自 Facebook) 伴随论文 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 由 Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela 发布。
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (来自 Google Research) 伴随论文 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 由 Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 发布。
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (来自 Google Research) 伴随论文 [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 由 Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya 发布。
|
||||
|
||||
@@ -418,6 +418,7 @@ conda install conda-forge::transformers
|
||||
1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
|
||||
1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
|
||||
1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
|
||||
1. **[Qwen2](https://huggingface.co/docs/transformers/main/model_doc/qwen2)** (from the Qwen team, Alibaba Group) released with the paper [Qwen Technical Report](https://arxiv.org/abs/2309.16609) by Jinze Bai, Shuai Bai, Yunfei Chu, Zeyu Cui, Kai Dang, Xiaodong Deng, Yang Fan, Wenbin Ge, Yu Han, Fei Huang, Binyuan Hui, Luo Ji, Mei Li, Junyang Lin, Runji Lin, Dayiheng Liu, Gao Liu, Chengqiang Lu, Keming Lu, Jianxin Ma, Rui Men, Xingzhang Ren, Xuancheng Ren, Chuanqi Tan, Sinan Tan, Jianhong Tu, Peng Wang, Shijie Wang, Wei Wang, Shengguang Wu, Benfeng Xu, Jin Xu, An Yang, Hao Yang, Jian Yang, Shusheng Yang, Yang Yao, Bowen Yu, Hongyi Yuan, Zheng Yuan, Jianwei Zhang, Xingxuan Zhang, Yichang Zhang, Zhenru Zhang, Chang Zhou, Jingren Zhou, Xiaohuan Zhou and Tianhang Zhu.
|
||||
1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
|
||||
1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
|
||||
1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
|
||||
|
||||
@@ -446,6 +446,8 @@
|
||||
title: ProphetNet
|
||||
- local: model_doc/qdqbert
|
||||
title: QDQBert
|
||||
- local: model_doc/qwen2
|
||||
title: Qwen2
|
||||
- local: model_doc/rag
|
||||
title: RAG
|
||||
- local: model_doc/realm
|
||||
|
||||
@@ -232,6 +232,7 @@ Flax), PyTorch, and/or TensorFlow.
|
||||
| [ProphetNet](model_doc/prophetnet) | ✅ | ❌ | ❌ |
|
||||
| [PVT](model_doc/pvt) | ✅ | ❌ | ❌ |
|
||||
| [QDQBert](model_doc/qdqbert) | ✅ | ❌ | ❌ |
|
||||
| [Qwen2](model_doc/qwen2) | ✅ | ❌ | ❌ |
|
||||
| [RAG](model_doc/rag) | ✅ | ✅ | ❌ |
|
||||
| [REALM](model_doc/realm) | ✅ | ❌ | ❌ |
|
||||
| [Reformer](model_doc/reformer) | ✅ | ❌ | ❌ |
|
||||
|
||||
82
docs/source/en/model_doc/qwen2.md
Normal file
82
docs/source/en/model_doc/qwen2.md
Normal file
@@ -0,0 +1,82 @@
|
||||
<!--Copyright 2024 The Qwen Team and The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
|
||||
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
|
||||
rendered properly in your Markdown viewer.
|
||||
|
||||
-->
|
||||
|
||||
# Qwen2
|
||||
|
||||
## Overview
|
||||
|
||||
Qwen2 is the new model series of large language models from the Qwen team. Previously, we released the Qwen series, including Qwen-72B, Qwen-1.8B, Qwen-VL, Qwen-Audio, etc.
|
||||
|
||||
### Model Details
|
||||
|
||||
Qwen2 is a language model series including decoder language models of different model sizes. For each size, we release the base language model and the aligned chat model. It is based on the Transformer architecture with SwiGLU activation, attention QKV bias, group query attention, mixture of sliding window attention and full attention, etc. Additionally, we have an improved tokenizer adaptive to multiple natural languages and codes.
|
||||
|
||||
|
||||
## Usage tips
|
||||
|
||||
`Qwen2-7B-beta` and `Qwen2-7B-Chat-beta` can be found on the [Huggingface Hub](https://huggingface.co/Qwen)
|
||||
|
||||
In the following, we demonstrate how to use `Qwen2-7B-Chat-beta` for the inference. Note that we have used the ChatML format for dialog, in this demo we show how to leverage `apply_chat_template` for this purpose.
|
||||
|
||||
```python
|
||||
>>> from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
>>> device = "cuda" # the device to load the model onto
|
||||
|
||||
>>> model = AutoModelForCausalLM.from_pretrained("Qwen2/Qwen2-7B-Chat-beta", device_map="auto")
|
||||
>>> tokenizer = AutoTokenizer.from_pretrained("Qwen2/Qwen2-7B-Chat-beta")
|
||||
|
||||
>>> prompt = "Give me a short introduction to large language model."
|
||||
|
||||
>>> messages = [{"role": "user", "content": prompt}]
|
||||
|
||||
>>> text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
||||
|
||||
>>> model_inputs = tokenizer([text], return_tensors="pt").to(device)
|
||||
|
||||
>>> generated_ids = model.generate(model_inputs.input_ids, max_new_tokens=512, do_sample=True)
|
||||
|
||||
>>> generated_ids = [output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)]
|
||||
|
||||
>>> response = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
|
||||
```
|
||||
|
||||
## Qwen2Config
|
||||
|
||||
[[autodoc]] Qwen2Config
|
||||
|
||||
## Qwen2Tokenizer
|
||||
|
||||
[[autodoc]] Qwen2Tokenizer
|
||||
- save_vocabulary
|
||||
|
||||
## Qwen2TokenizerFast
|
||||
|
||||
[[autodoc]] Qwen2TokenizerFast
|
||||
|
||||
## Qwen2Model
|
||||
|
||||
[[autodoc]] Qwen2Model
|
||||
- forward
|
||||
|
||||
## Qwen2ForCausalLM
|
||||
|
||||
[[autodoc]] Qwen2ForCausalLM
|
||||
- forward
|
||||
|
||||
## Qwen2ForSequenceClassification
|
||||
|
||||
[[autodoc]] Qwen2ForSequenceClassification
|
||||
- forward
|
||||
@@ -52,6 +52,7 @@ FlashAttention-2 is currently supported for the following architectures:
|
||||
* [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
|
||||
* [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel)
|
||||
* [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
|
||||
* [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
|
||||
* [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
|
||||
|
||||
You can request to add FlashAttention-2 support for another model by opening a GitHub Issue or Pull Request.
|
||||
@@ -174,6 +175,7 @@ For now, Transformers supports SDPA inference and training for the following arc
|
||||
* [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
|
||||
* [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
|
||||
* [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
|
||||
* [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
|
||||
|
||||
<Tip>
|
||||
|
||||
|
||||
@@ -37,7 +37,7 @@ You can finetune other architectures for causal language modeling following the
|
||||
Choose one of the following architectures:
|
||||
|
||||
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
|
||||
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [Fuyu](../model_doc/fuyu), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [Whisper](../model_doc/whisper), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
|
||||
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [Fuyu](../model_doc/fuyu), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [Whisper](../model_doc/whisper), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -33,7 +33,7 @@ The task illustrated in this tutorial is supported by the following model archit
|
||||
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
|
||||
|
||||
|
||||
[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
|
||||
|
||||
|
||||
|
||||
@@ -711,6 +711,11 @@ _import_structure = {
|
||||
],
|
||||
"models.pvt": ["PVT_PRETRAINED_CONFIG_ARCHIVE_MAP", "PvtConfig"],
|
||||
"models.qdqbert": ["QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "QDQBertConfig"],
|
||||
"models.qwen2": [
|
||||
"QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP",
|
||||
"Qwen2Config",
|
||||
"Qwen2Tokenizer",
|
||||
],
|
||||
"models.rag": ["RagConfig", "RagRetriever", "RagTokenizer"],
|
||||
"models.realm": [
|
||||
"REALM_PRETRAINED_CONFIG_ARCHIVE_MAP",
|
||||
@@ -1185,6 +1190,7 @@ else:
|
||||
_import_structure["models.nougat"].append("NougatTokenizerFast")
|
||||
_import_structure["models.openai"].append("OpenAIGPTTokenizerFast")
|
||||
_import_structure["models.pegasus"].append("PegasusTokenizerFast")
|
||||
_import_structure["models.qwen2"].append("Qwen2TokenizerFast")
|
||||
_import_structure["models.realm"].append("RealmTokenizerFast")
|
||||
_import_structure["models.reformer"].append("ReformerTokenizerFast")
|
||||
_import_structure["models.rembert"].append("RemBertTokenizerFast")
|
||||
@@ -2971,6 +2977,14 @@ else:
|
||||
"load_tf_weights_in_qdqbert",
|
||||
]
|
||||
)
|
||||
_import_structure["models.qwen2"].extend(
|
||||
[
|
||||
"Qwen2ForCausalLM",
|
||||
"Qwen2ForSequenceClassification",
|
||||
"Qwen2Model",
|
||||
"Qwen2PreTrainedModel",
|
||||
]
|
||||
)
|
||||
_import_structure["models.rag"].extend(
|
||||
[
|
||||
"RagModel",
|
||||
@@ -5404,6 +5418,7 @@ if TYPE_CHECKING:
|
||||
)
|
||||
from .models.pvt import PVT_PRETRAINED_CONFIG_ARCHIVE_MAP, PvtConfig
|
||||
from .models.qdqbert import QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, QDQBertConfig
|
||||
from .models.qwen2 import QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP, Qwen2Config, Qwen2Tokenizer
|
||||
from .models.rag import RagConfig, RagRetriever, RagTokenizer
|
||||
from .models.realm import (
|
||||
REALM_PRETRAINED_CONFIG_ARCHIVE_MAP,
|
||||
@@ -5871,6 +5886,7 @@ if TYPE_CHECKING:
|
||||
from .models.nougat import NougatTokenizerFast
|
||||
from .models.openai import OpenAIGPTTokenizerFast
|
||||
from .models.pegasus import PegasusTokenizerFast
|
||||
from .models.qwen2 import Qwen2TokenizerFast
|
||||
from .models.realm import RealmTokenizerFast
|
||||
from .models.reformer import ReformerTokenizerFast
|
||||
from .models.rembert import RemBertTokenizerFast
|
||||
@@ -7373,6 +7389,12 @@ if TYPE_CHECKING:
|
||||
QDQBertPreTrainedModel,
|
||||
load_tf_weights_in_qdqbert,
|
||||
)
|
||||
from .models.qwen2 import (
|
||||
Qwen2ForCausalLM,
|
||||
Qwen2ForSequenceClassification,
|
||||
Qwen2Model,
|
||||
Qwen2PreTrainedModel,
|
||||
)
|
||||
from .models.rag import (
|
||||
RagModel,
|
||||
RagPreTrainedModel,
|
||||
|
||||
@@ -355,6 +355,48 @@ class HerbertConverter(Converter):
|
||||
return tokenizer
|
||||
|
||||
|
||||
class Qwen2Converter(Converter):
|
||||
def converted(self) -> Tokenizer:
|
||||
vocab = self.original_tokenizer.encoder
|
||||
merges = list(self.original_tokenizer.bpe_ranks.keys())
|
||||
|
||||
tokenizer = Tokenizer(
|
||||
BPE(
|
||||
vocab=vocab,
|
||||
merges=merges,
|
||||
dropout=None,
|
||||
unk_token=None,
|
||||
continuing_subword_prefix="",
|
||||
end_of_word_suffix="",
|
||||
fuse_unk=False,
|
||||
byte_fallback=False,
|
||||
)
|
||||
)
|
||||
|
||||
tokenizer.normalizer = normalizers.NFC()
|
||||
|
||||
tokenizer.pre_tokenizer = pre_tokenizers.Sequence(
|
||||
[
|
||||
pre_tokenizers.Split(
|
||||
Regex(
|
||||
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+"""
|
||||
),
|
||||
behavior="isolated",
|
||||
invert=False,
|
||||
),
|
||||
pre_tokenizers.ByteLevel(
|
||||
add_prefix_space=getattr(self.original_tokenizer, "add_prefix_space", False),
|
||||
use_regex=False,
|
||||
),
|
||||
]
|
||||
)
|
||||
|
||||
tokenizer.decoder = decoders.ByteLevel()
|
||||
tokenizer.post_processor = processors.ByteLevel(trim_offsets=False)
|
||||
|
||||
return tokenizer
|
||||
|
||||
|
||||
class RobertaConverter(Converter):
|
||||
def converted(self) -> Tokenizer:
|
||||
ot = self.original_tokenizer
|
||||
@@ -1289,6 +1331,7 @@ SLOW_TO_FAST_CONVERTERS = {
|
||||
"NllbTokenizer": NllbConverter,
|
||||
"OpenAIGPTTokenizer": OpenAIGPTConverter,
|
||||
"PegasusTokenizer": PegasusConverter,
|
||||
"Qwen2Tokenizer": Qwen2Converter,
|
||||
"RealmTokenizer": BertConverter,
|
||||
"ReformerTokenizer": ReformerConverter,
|
||||
"RemBertTokenizer": RemBertConverter,
|
||||
|
||||
@@ -176,6 +176,7 @@ from . import (
|
||||
prophetnet,
|
||||
pvt,
|
||||
qdqbert,
|
||||
qwen2,
|
||||
rag,
|
||||
realm,
|
||||
reformer,
|
||||
|
||||
@@ -182,6 +182,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
|
||||
("prophetnet", "ProphetNetConfig"),
|
||||
("pvt", "PvtConfig"),
|
||||
("qdqbert", "QDQBertConfig"),
|
||||
("qwen2", "Qwen2Config"),
|
||||
("rag", "RagConfig"),
|
||||
("realm", "RealmConfig"),
|
||||
("reformer", "ReformerConfig"),
|
||||
@@ -405,6 +406,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
|
||||
("prophetnet", "PROPHETNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("pvt", "PVT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("qdqbert", "QDQBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("qwen2", "QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("realm", "REALM_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("regnet", "REGNET_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("rembert", "REMBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
@@ -649,6 +651,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
|
||||
("prophetnet", "ProphetNet"),
|
||||
("pvt", "PVT"),
|
||||
("qdqbert", "QDQBert"),
|
||||
("qwen2", "Qwen2"),
|
||||
("rag", "RAG"),
|
||||
("realm", "REALM"),
|
||||
("reformer", "Reformer"),
|
||||
|
||||
@@ -177,6 +177,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
|
||||
("prophetnet", "ProphetNetModel"),
|
||||
("pvt", "PvtModel"),
|
||||
("qdqbert", "QDQBertModel"),
|
||||
("qwen2", "Qwen2Model"),
|
||||
("reformer", "ReformerModel"),
|
||||
("regnet", "RegNetModel"),
|
||||
("rembert", "RemBertModel"),
|
||||
@@ -449,6 +450,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
|
||||
("plbart", "PLBartForCausalLM"),
|
||||
("prophetnet", "ProphetNetForCausalLM"),
|
||||
("qdqbert", "QDQBertLMHeadModel"),
|
||||
("qwen2", "Qwen2ForCausalLM"),
|
||||
("reformer", "ReformerModelWithLMHead"),
|
||||
("rembert", "RemBertForCausalLM"),
|
||||
("roberta", "RobertaForCausalLM"),
|
||||
@@ -792,6 +794,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
|
||||
("phi", "PhiForSequenceClassification"),
|
||||
("plbart", "PLBartForSequenceClassification"),
|
||||
("qdqbert", "QDQBertForSequenceClassification"),
|
||||
("qwen2", "Qwen2ForSequenceClassification"),
|
||||
("reformer", "ReformerForSequenceClassification"),
|
||||
("rembert", "RemBertForSequenceClassification"),
|
||||
("roberta", "RobertaForSequenceClassification"),
|
||||
|
||||
@@ -333,6 +333,13 @@ else:
|
||||
("plbart", ("PLBartTokenizer" if is_sentencepiece_available() else None, None)),
|
||||
("prophetnet", ("ProphetNetTokenizer", None)),
|
||||
("qdqbert", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
|
||||
(
|
||||
"qwen2",
|
||||
(
|
||||
"Qwen2Tokenizer",
|
||||
"Qwen2TokenizerFast" if is_tokenizers_available() else None,
|
||||
),
|
||||
),
|
||||
("rag", ("RagTokenizer", None)),
|
||||
("realm", ("RealmTokenizer", "RealmTokenizerFast" if is_tokenizers_available() else None)),
|
||||
(
|
||||
|
||||
80
src/transformers/models/qwen2/__init__.py
Normal file
80
src/transformers/models/qwen2/__init__.py
Normal file
@@ -0,0 +1,80 @@
|
||||
# Copyright 2024 The Qwen Team and The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from ...utils import (
|
||||
OptionalDependencyNotAvailable,
|
||||
_LazyModule,
|
||||
is_tokenizers_available,
|
||||
is_torch_available,
|
||||
)
|
||||
|
||||
|
||||
_import_structure = {
|
||||
"configuration_qwen2": ["QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP", "Qwen2Config"],
|
||||
"tokenization_qwen2": ["Qwen2Tokenizer"],
|
||||
}
|
||||
|
||||
try:
|
||||
if not is_tokenizers_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
_import_structure["tokenization_qwen2_fast"] = ["Qwen2TokenizerFast"]
|
||||
|
||||
try:
|
||||
if not is_torch_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
_import_structure["modeling_qwen2"] = [
|
||||
"Qwen2ForCausalLM",
|
||||
"Qwen2Model",
|
||||
"Qwen2PreTrainedModel",
|
||||
"Qwen2ForSequenceClassification",
|
||||
]
|
||||
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .configuration_qwen2 import QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP, Qwen2Config
|
||||
from .tokenization_qwen2 import Qwen2Tokenizer
|
||||
|
||||
try:
|
||||
if not is_tokenizers_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
from .tokenization_qwen2_fast import Qwen2TokenizerFast
|
||||
|
||||
try:
|
||||
if not is_torch_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
from .modeling_qwen2 import (
|
||||
Qwen2ForCausalLM,
|
||||
Qwen2ForSequenceClassification,
|
||||
Qwen2Model,
|
||||
Qwen2PreTrainedModel,
|
||||
)
|
||||
|
||||
|
||||
else:
|
||||
import sys
|
||||
|
||||
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
||||
144
src/transformers/models/qwen2/configuration_qwen2.py
Normal file
144
src/transformers/models/qwen2/configuration_qwen2.py
Normal file
@@ -0,0 +1,144 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Qwen2 model configuration"""
|
||||
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
QWEN2_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
"Qwen/Qwen2-7B-beta": "https://huggingface.co/Qwen/Qwen2-7B-beta/resolve/main/config.json",
|
||||
}
|
||||
|
||||
|
||||
class Qwen2Config(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`Qwen2Model`]. It is used to instantiate a
|
||||
Qwen2 model according to the specified arguments, defining the model architecture. Instantiating a configuration
|
||||
with the defaults will yield a similar configuration to that of
|
||||
Qwen2-7B-beta [Qwen/Qwen2-7B-beta](https://huggingface.co/Qwen/Qwen2-7B-beta).
|
||||
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 151936):
|
||||
Vocabulary size of the Qwen2 model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`Qwen2Model`]
|
||||
hidden_size (`int`, *optional*, defaults to 4096):
|
||||
Dimension of the hidden representations.
|
||||
intermediate_size (`int`, *optional*, defaults to 22016):
|
||||
Dimension of the MLP representations.
|
||||
num_hidden_layers (`int`, *optional*, defaults to 32):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
num_attention_heads (`int`, *optional*, defaults to 32):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
num_key_value_heads (`int`, *optional*, defaults to 32):
|
||||
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
|
||||
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
|
||||
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
|
||||
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
|
||||
by meanpooling all the original heads within that group. For more details checkout [this
|
||||
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
|
||||
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
|
||||
The non-linear activation function (function or string) in the decoder.
|
||||
max_position_embeddings (`int`, *optional*, defaults to 32768):
|
||||
The maximum sequence length that this model might ever be used with.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
|
||||
The epsilon used by the rms normalization layers.
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models). Only
|
||||
relevant if `config.is_decoder=True`.
|
||||
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
|
||||
Whether the model's input and output word embeddings should be tied.
|
||||
rope_theta (`float`, *optional*, defaults to 10000.0):
|
||||
The base period of the RoPE embeddings.
|
||||
use_sliding_window (`bool`, *optional*, defaults to `False`):
|
||||
Whether to use sliding window attention.
|
||||
sliding_window (`int`, *optional*, defaults to 4096):
|
||||
Sliding window attention (SWA) window size. If not specified, will default to `4096`.
|
||||
max_window_layers (`int`, *optional*, defaults to 28):
|
||||
The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
|
||||
attention_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for the attention probabilities.
|
||||
|
||||
```python
|
||||
>>> from transformers import Qwen2Model, Qwen2Config
|
||||
|
||||
>>> # Initializing a Qwen2 style configuration
|
||||
>>> configuration = Qwen2Config()
|
||||
|
||||
>>> # Initializing a model from the Qwen2-7B style configuration
|
||||
>>> model = Qwen2Model(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "qwen2"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=151936,
|
||||
hidden_size=4096,
|
||||
intermediate_size=22016,
|
||||
num_hidden_layers=32,
|
||||
num_attention_heads=32,
|
||||
num_key_value_heads=32,
|
||||
hidden_act="silu",
|
||||
max_position_embeddings=32768,
|
||||
initializer_range=0.02,
|
||||
rms_norm_eps=1e-6,
|
||||
use_cache=True,
|
||||
tie_word_embeddings=False,
|
||||
rope_theta=10000.0,
|
||||
use_sliding_window=False,
|
||||
sliding_window=4096,
|
||||
max_window_layers=28,
|
||||
attention_dropout=0.0,
|
||||
**kwargs,
|
||||
):
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.hidden_size = hidden_size
|
||||
self.intermediate_size = intermediate_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.use_sliding_window = use_sliding_window
|
||||
self.sliding_window = sliding_window
|
||||
self.max_window_layers = max_window_layers
|
||||
|
||||
# for backward compatibility
|
||||
if num_key_value_heads is None:
|
||||
num_key_value_heads = num_attention_heads
|
||||
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.hidden_act = hidden_act
|
||||
self.initializer_range = initializer_range
|
||||
self.rms_norm_eps = rms_norm_eps
|
||||
self.use_cache = use_cache
|
||||
self.rope_theta = rope_theta
|
||||
self.attention_dropout = attention_dropout
|
||||
|
||||
super().__init__(
|
||||
tie_word_embeddings=tie_word_embeddings,
|
||||
**kwargs,
|
||||
)
|
||||
1401
src/transformers/models/qwen2/modeling_qwen2.py
Normal file
1401
src/transformers/models/qwen2/modeling_qwen2.py
Normal file
File diff suppressed because it is too large
Load Diff
345
src/transformers/models/qwen2/tokenization_qwen2.py
Normal file
345
src/transformers/models/qwen2/tokenization_qwen2.py
Normal file
@@ -0,0 +1,345 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Tokenization classes for Qwen2."""
|
||||
|
||||
import json
|
||||
import os
|
||||
import unicodedata
|
||||
from functools import lru_cache
|
||||
from typing import Optional, Tuple
|
||||
|
||||
import regex as re
|
||||
|
||||
from ...tokenization_utils import AddedToken, PreTrainedTokenizer
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
VOCAB_FILES_NAMES = {
|
||||
"vocab_file": "vocab.json",
|
||||
"merges_file": "merges.txt",
|
||||
}
|
||||
|
||||
PRETRAINED_VOCAB_FILES_MAP = {
|
||||
"vocab_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/vocab.json"},
|
||||
"merges_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/merges.txt"},
|
||||
}
|
||||
|
||||
MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
|
||||
|
||||
PRETOKENIZE_REGEX = 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+"""
|
||||
|
||||
|
||||
@lru_cache()
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.bytes_to_unicode
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a mapping to unicode strings. We specifically avoids mapping to whitespace/control
|
||||
characters the bpe code barfs on.
|
||||
|
||||
The reversible bpe codes work on unicode strings. This means you need a large # of unicode characters in your vocab
|
||||
if you want to avoid UNKs. When you're at something like a 10B token dataset you end up needing around 5K for
|
||||
decent coverage. This is a significant percentage of your normal, say, 32K bpe vocab. To avoid that, we want lookup
|
||||
tables between utf-8 bytes and unicode strings.
|
||||
"""
|
||||
bs = (
|
||||
list(range(ord("!"), ord("~") + 1)) + list(range(ord("¡"), ord("¬") + 1)) + list(range(ord("®"), ord("ÿ") + 1))
|
||||
)
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8 + n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.get_pairs
|
||||
def get_pairs(word):
|
||||
"""
|
||||
Return set of symbol pairs in a word.
|
||||
|
||||
Word is represented as tuple of symbols (symbols being variable-length strings).
|
||||
"""
|
||||
pairs = set()
|
||||
prev_char = word[0]
|
||||
for char in word[1:]:
|
||||
pairs.add((prev_char, char))
|
||||
prev_char = char
|
||||
return pairs
|
||||
|
||||
|
||||
class Qwen2Tokenizer(PreTrainedTokenizer):
|
||||
"""
|
||||
Construct a Qwen2 tokenizer. Based on byte-level Byte-Pair-Encoding.
|
||||
|
||||
Same with GPT2Tokenzier, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
|
||||
be encoded differently whether it is at the beginning of the sentence (without space) or not:
|
||||
|
||||
```python
|
||||
>>> from transformers import Qwen2Tokenizer
|
||||
|
||||
>>> tokenizer = Qwen2Tokenizer.from_pretrained("Qwen/Qwen-tokenizer")
|
||||
>>> tokenizer("Hello world")["input_ids"]
|
||||
[9707, 1879]
|
||||
|
||||
>>> tokenizer(" Hello world")["input_ids"]
|
||||
[21927, 1879]
|
||||
```
|
||||
This is expected.
|
||||
|
||||
You should not use GPT2Tokenizer instead, because of the different pretokenization rules.
|
||||
|
||||
This tokenizer inherits from [`PreTrainedTokenizer`] which contains most of the main methods. Users should refer to
|
||||
this superclass for more information regarding those methods.
|
||||
|
||||
Args:
|
||||
vocab_file (`str`):
|
||||
Path to the vocabulary file.
|
||||
merges_file (`str`):
|
||||
Path to the merges file.
|
||||
errors (`str`, *optional*, defaults to `"replace"`):
|
||||
Paradigm to follow when decoding bytes to UTF-8. See
|
||||
[bytes.decode](https://docs.python.org/3/library/stdtypes.html#bytes.decode) for more information.
|
||||
unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
|
||||
token instead.
|
||||
bos_token (`str`, *optional*):
|
||||
The beginning of sequence token. Not applicable for this tokenizer.
|
||||
eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The end of sequence token.
|
||||
pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The token used for padding, for example when batching sequences of different lengths.
|
||||
clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
|
||||
Whether or not the model should cleanup the spaces that were added when splitting the input text during the
|
||||
tokenization process. Not applicable to this tokenizer, since tokenization does not add spaces.
|
||||
split_special_tokens (`bool`, *optional*, defaults to `False`):
|
||||
Whether or not the special tokens should be split during the tokenization process. The default behavior is
|
||||
to not split special tokens. This means that if `<|endoftext|>` is the `eos_token`, then `tokenizer.tokenize("<|endoftext|>") =
|
||||
['<|endoftext|>`]. Otherwise, if `split_special_tokens=True`, then `tokenizer.tokenize("<|endoftext|>")` will be give `['<',
|
||||
'|', 'endo', 'ft', 'ext', '|', '>']`. This argument is only supported for `slow` tokenizers for the moment.
|
||||
"""
|
||||
|
||||
vocab_files_names = VOCAB_FILES_NAMES
|
||||
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
||||
max_model_input_sizes = MAX_MODEL_INPUT_SIZES
|
||||
model_input_names = ["input_ids", "attention_mask"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_file,
|
||||
merges_file,
|
||||
errors="replace",
|
||||
unk_token="<|endoftext|>",
|
||||
bos_token=None,
|
||||
eos_token="<|endoftext|>",
|
||||
pad_token="<|endoftext|>",
|
||||
clean_up_tokenization_spaces=False,
|
||||
split_special_tokens=False,
|
||||
**kwargs,
|
||||
):
|
||||
# Qwen vocab does not contain control tokens; added tokens need to be special
|
||||
bos_token = (
|
||||
AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(bos_token, str)
|
||||
else bos_token
|
||||
)
|
||||
eos_token = (
|
||||
AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(eos_token, str)
|
||||
else eos_token
|
||||
)
|
||||
unk_token = (
|
||||
AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(unk_token, str)
|
||||
else unk_token
|
||||
)
|
||||
pad_token = (
|
||||
AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(pad_token, str)
|
||||
else pad_token
|
||||
)
|
||||
|
||||
with open(vocab_file, encoding="utf-8") as vocab_handle:
|
||||
self.encoder = json.load(vocab_handle)
|
||||
self.decoder = {v: k for k, v in self.encoder.items()}
|
||||
self.errors = errors # how to handle errors in decoding
|
||||
self.byte_encoder = bytes_to_unicode()
|
||||
self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
|
||||
bpe_merges = []
|
||||
with open(merges_file, encoding="utf-8") as merges_handle:
|
||||
for line in merges_handle:
|
||||
line = line.strip()
|
||||
if not line or line.startswith("#"):
|
||||
continue
|
||||
bpe_merges.append(tuple(line.split()))
|
||||
self.bpe_ranks = dict(zip(bpe_merges, range(len(bpe_merges))))
|
||||
# NOTE: the cache can grow without bound and will get really large for long running processes
|
||||
# (esp. for texts of language that do not use space between word, e.g. Chinese); technically
|
||||
# not a memory leak but appears as one.
|
||||
# GPT2Tokenizer has the same problem, so let's be consistent.
|
||||
self.cache = {}
|
||||
|
||||
self.pat = re.compile(PRETOKENIZE_REGEX)
|
||||
|
||||
if kwargs.get("add_prefix_space", False):
|
||||
logger.warning_once(
|
||||
f"{self.__class__.__name} does not support `add_prefix_space`, setting it to True has no effect."
|
||||
)
|
||||
|
||||
super().__init__(
|
||||
errors=errors,
|
||||
bos_token=bos_token,
|
||||
eos_token=eos_token,
|
||||
pad_token=pad_token,
|
||||
unk_token=unk_token,
|
||||
clean_up_tokenization_spaces=clean_up_tokenization_spaces,
|
||||
split_special_tokens=split_special_tokens,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
@property
|
||||
def vocab_size(self) -> int:
|
||||
return len(self.encoder)
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.get_vocab
|
||||
def get_vocab(self):
|
||||
return dict(self.encoder, **self.added_tokens_encoder)
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.bpe
|
||||
def bpe(self, token):
|
||||
if token in self.cache:
|
||||
return self.cache[token]
|
||||
word = tuple(token)
|
||||
pairs = get_pairs(word)
|
||||
|
||||
if not pairs:
|
||||
return token
|
||||
|
||||
while True:
|
||||
bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
|
||||
if bigram not in self.bpe_ranks:
|
||||
break
|
||||
first, second = bigram
|
||||
new_word = []
|
||||
i = 0
|
||||
while i < len(word):
|
||||
try:
|
||||
j = word.index(first, i)
|
||||
except ValueError:
|
||||
new_word.extend(word[i:])
|
||||
break
|
||||
else:
|
||||
new_word.extend(word[i:j])
|
||||
i = j
|
||||
|
||||
if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
|
||||
new_word.append(first + second)
|
||||
i += 2
|
||||
else:
|
||||
new_word.append(word[i])
|
||||
i += 1
|
||||
new_word = tuple(new_word)
|
||||
word = new_word
|
||||
if len(word) == 1:
|
||||
break
|
||||
else:
|
||||
pairs = get_pairs(word)
|
||||
word = " ".join(word)
|
||||
self.cache[token] = word
|
||||
return word
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._tokenize
|
||||
def _tokenize(self, text):
|
||||
"""Tokenize a string."""
|
||||
bpe_tokens = []
|
||||
for token in re.findall(self.pat, text):
|
||||
token = "".join(
|
||||
self.byte_encoder[b] for b in token.encode("utf-8")
|
||||
) # Maps all our bytes to unicode strings, avoiding control tokens of the BPE (spaces in our case)
|
||||
bpe_tokens.extend(bpe_token for bpe_token in self.bpe(token).split(" "))
|
||||
return bpe_tokens
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_token_to_id
|
||||
def _convert_token_to_id(self, token):
|
||||
"""Converts a token (str) in an id using the vocab."""
|
||||
return self.encoder.get(token, self.encoder.get(self.unk_token))
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer._convert_id_to_token
|
||||
def _convert_id_to_token(self, index):
|
||||
"""Converts an index (integer) in a token (str) using the vocab."""
|
||||
return self.decoder.get(index)
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.convert_tokens_to_string
|
||||
def convert_tokens_to_string(self, tokens):
|
||||
"""Converts a sequence of tokens (string) in a single string."""
|
||||
text = "".join(tokens)
|
||||
text = bytearray([self.byte_decoder[c] for c in text]).decode("utf-8", errors=self.errors)
|
||||
return text
|
||||
|
||||
def decode(
|
||||
self,
|
||||
token_ids,
|
||||
skip_special_tokens: bool = False,
|
||||
clean_up_tokenization_spaces: Optional[bool] = False,
|
||||
spaces_between_special_tokens: bool = False,
|
||||
**kwargs,
|
||||
) -> str:
|
||||
# `spaces_between_special_tokens` defaults to True for _decode in slow tokenizers
|
||||
# and cannot be configured elsewhere, but it should default to False for Qwen2Tokenizer
|
||||
return super().decode(
|
||||
token_ids,
|
||||
skip_special_tokens=skip_special_tokens,
|
||||
clean_up_tokenization_spaces=clean_up_tokenization_spaces,
|
||||
spaces_between_special_tokens=spaces_between_special_tokens,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2.GPT2Tokenizer.save_vocabulary
|
||||
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
|
||||
if not os.path.isdir(save_directory):
|
||||
logger.error(f"Vocabulary path ({save_directory}) should be a directory")
|
||||
return
|
||||
vocab_file = os.path.join(
|
||||
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["vocab_file"]
|
||||
)
|
||||
merge_file = os.path.join(
|
||||
save_directory, (filename_prefix + "-" if filename_prefix else "") + VOCAB_FILES_NAMES["merges_file"]
|
||||
)
|
||||
|
||||
with open(vocab_file, "w", encoding="utf-8") as f:
|
||||
f.write(json.dumps(self.encoder, indent=2, sort_keys=True, ensure_ascii=False) + "\n")
|
||||
|
||||
index = 0
|
||||
with open(merge_file, "w", encoding="utf-8") as writer:
|
||||
writer.write("#version: 0.2\n")
|
||||
for bpe_tokens, token_index in sorted(self.bpe_ranks.items(), key=lambda kv: kv[1]):
|
||||
if index != token_index:
|
||||
logger.warning(
|
||||
f"Saving vocabulary to {merge_file}: BPE merge indices are not consecutive."
|
||||
" Please check that the tokenizer is not corrupted!"
|
||||
)
|
||||
index = token_index
|
||||
writer.write(" ".join(bpe_tokens) + "\n")
|
||||
index += 1
|
||||
|
||||
return vocab_file, merge_file
|
||||
|
||||
def prepare_for_tokenization(self, text, **kwargs):
|
||||
text = unicodedata.normalize("NFC", text)
|
||||
return (text, kwargs)
|
||||
143
src/transformers/models/qwen2/tokenization_qwen2_fast.py
Normal file
143
src/transformers/models/qwen2/tokenization_qwen2_fast.py
Normal file
@@ -0,0 +1,143 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
"""Tokenization classes for Qwen2."""
|
||||
|
||||
from typing import Optional, Tuple
|
||||
|
||||
from ...tokenization_utils import AddedToken
|
||||
from ...tokenization_utils_fast import PreTrainedTokenizerFast
|
||||
from ...utils import logging
|
||||
from .tokenization_qwen2 import Qwen2Tokenizer
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
VOCAB_FILES_NAMES = {
|
||||
"vocab_file": "vocab.json",
|
||||
"merges_file": "merges.txt",
|
||||
"tokenizer_file": "tokenizer.json",
|
||||
}
|
||||
|
||||
PRETRAINED_VOCAB_FILES_MAP = {
|
||||
"vocab_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/vocab.json"},
|
||||
"merges_file": {"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/merges.txt"},
|
||||
"tokenizer_file": {
|
||||
"qwen/qwen-tokenizer": "https://huggingface.co/qwen/qwen-tokenizer/resolve/main/tokenizer.json"
|
||||
},
|
||||
}
|
||||
|
||||
MAX_MODEL_INPUT_SIZES = {"qwen/qwen-tokenizer": 32768}
|
||||
|
||||
|
||||
class Qwen2TokenizerFast(PreTrainedTokenizerFast):
|
||||
"""
|
||||
Construct a "fast" Qwen2 tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level
|
||||
Byte-Pair-Encoding.
|
||||
|
||||
Same with GPT2Tokenzier, this tokenizer has been trained to treat spaces like parts of the tokens so a word will
|
||||
be encoded differently whether it is at the beginning of the sentence (without space) or not:
|
||||
|
||||
```python
|
||||
>>> from transformers import Qwen2TokenizerFast
|
||||
|
||||
>>> tokenizer = Qwen2TokenizerFast.from_pretrained("Qwen/Qwen-tokenizer")
|
||||
>>> tokenizer("Hello world")["input_ids"]
|
||||
[9707, 1879]
|
||||
|
||||
>>> tokenizer(" Hello world")["input_ids"]
|
||||
[21927, 1879]
|
||||
```
|
||||
This is expected.
|
||||
|
||||
This tokenizer inherits from [`PreTrainedTokenizerFast`] which contains most of the main methods. Users should
|
||||
refer to this superclass for more information regarding those methods.
|
||||
|
||||
Args:
|
||||
vocab_file (`str`, *optional*):
|
||||
Path to the vocabulary file.
|
||||
merges_file (`str`, *optional*):
|
||||
Path to the merges file.
|
||||
tokenizer_file (`str`, *optional*):
|
||||
Path to [tokenizers](https://github.com/huggingface/tokenizers) file (generally has a .json extension) that
|
||||
contains everything needed to load the tokenizer.
|
||||
unk_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The unknown token. A token that is not in the vocabulary cannot be converted to an ID and is set to be this
|
||||
token instead. Not applicable to this tokenizer.
|
||||
bos_token (`str`, *optional*):
|
||||
The beginning of sequence token. Not applicable for this tokenizer.
|
||||
eos_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The end of sequence token.
|
||||
pad_token (`str`, *optional*, defaults to `"<|endoftext|>"`):
|
||||
The token used for padding, for example when batching sequences of different lengths.
|
||||
"""
|
||||
|
||||
vocab_files_names = VOCAB_FILES_NAMES
|
||||
pretrained_vocab_files_map = PRETRAINED_VOCAB_FILES_MAP
|
||||
max_model_input_sizes = MAX_MODEL_INPUT_SIZES
|
||||
model_input_names = ["input_ids", "attention_mask"]
|
||||
slow_tokenizer_class = Qwen2Tokenizer
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_file=None,
|
||||
merges_file=None,
|
||||
tokenizer_file=None,
|
||||
unk_token="<|endoftext|>",
|
||||
bos_token=None,
|
||||
eos_token="<|endoftext|>",
|
||||
pad_token="<|endoftext|>",
|
||||
**kwargs,
|
||||
):
|
||||
# We need to at least pass vocab_file and merges_file to base class
|
||||
# in case a slow tokenizer needs to be initialized; other can be
|
||||
# configured through files.
|
||||
# following GPT2TokenizerFast, also adding unk_token, bos_token, and eos_token
|
||||
|
||||
bos_token = (
|
||||
AddedToken(bos_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(bos_token, str)
|
||||
else bos_token
|
||||
)
|
||||
eos_token = (
|
||||
AddedToken(eos_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(eos_token, str)
|
||||
else eos_token
|
||||
)
|
||||
unk_token = (
|
||||
AddedToken(unk_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(unk_token, str)
|
||||
else unk_token
|
||||
)
|
||||
pad_token = (
|
||||
AddedToken(pad_token, lstrip=False, rstrip=False, special=True, normalized=False)
|
||||
if isinstance(pad_token, str)
|
||||
else pad_token
|
||||
)
|
||||
|
||||
super().__init__(
|
||||
vocab_file,
|
||||
merges_file,
|
||||
tokenizer_file=tokenizer_file,
|
||||
unk_token=unk_token,
|
||||
bos_token=bos_token,
|
||||
eos_token=eos_token,
|
||||
pad_token=pad_token,
|
||||
**kwargs,
|
||||
)
|
||||
|
||||
# Copied from transformers.models.gpt2.tokenization_gpt2_fast.GPT2TokenizerFast.save_vocabulary
|
||||
def save_vocabulary(self, save_directory: str, filename_prefix: Optional[str] = None) -> Tuple[str]:
|
||||
files = self._tokenizer.model.save(save_directory, name=filename_prefix)
|
||||
return tuple(files)
|
||||
@@ -6715,6 +6715,34 @@ def load_tf_weights_in_qdqbert(*args, **kwargs):
|
||||
requires_backends(load_tf_weights_in_qdqbert, ["torch"])
|
||||
|
||||
|
||||
class Qwen2ForCausalLM(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class Qwen2ForSequenceClassification(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class Qwen2Model(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class Qwen2PreTrainedModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class RagModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
|
||||
@@ -331,6 +331,13 @@ class PegasusTokenizerFast(metaclass=DummyObject):
|
||||
requires_backends(self, ["tokenizers"])
|
||||
|
||||
|
||||
class Qwen2TokenizerFast(metaclass=DummyObject):
|
||||
_backends = ["tokenizers"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["tokenizers"])
|
||||
|
||||
|
||||
class RealmTokenizerFast(metaclass=DummyObject):
|
||||
_backends = ["tokenizers"]
|
||||
|
||||
|
||||
0
tests/models/qwen2/__init__.py
Normal file
0
tests/models/qwen2/__init__.py
Normal file
604
tests/models/qwen2/test_modeling_qwen2.py
Normal file
604
tests/models/qwen2/test_modeling_qwen2.py
Normal file
@@ -0,0 +1,604 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" Testing suite for the PyTorch Qwen2 model. """
|
||||
|
||||
|
||||
import gc
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
import pytest
|
||||
|
||||
from transformers import AutoTokenizer, Qwen2Config, is_torch_available, set_seed
|
||||
from transformers.testing_utils import (
|
||||
backend_empty_cache,
|
||||
require_bitsandbytes,
|
||||
require_flash_attn,
|
||||
require_torch,
|
||||
require_torch_gpu,
|
||||
require_torch_sdpa,
|
||||
slow,
|
||||
torch_device,
|
||||
)
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
||||
from ...test_pipeline_mixin import PipelineTesterMixin
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import (
|
||||
Qwen2ForCausalLM,
|
||||
Qwen2ForSequenceClassification,
|
||||
Qwen2Model,
|
||||
)
|
||||
|
||||
|
||||
class Qwen2ModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_input_mask=True,
|
||||
use_token_type_ids=True,
|
||||
use_labels=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
max_window_layers=3,
|
||||
use_sliding_window=True,
|
||||
sliding_window=2,
|
||||
num_attention_heads=4,
|
||||
num_key_value_heads=2,
|
||||
intermediate_size=37,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=512,
|
||||
type_vocab_size=16,
|
||||
type_sequence_label_size=2,
|
||||
initializer_range=0.02,
|
||||
num_labels=3,
|
||||
num_choices=4,
|
||||
pad_token_id=0,
|
||||
bos_token_id=1,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_token_type_ids = use_token_type_ids
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.max_window_layers = max_window_layers
|
||||
self.use_sliding_window = use_sliding_window
|
||||
self.sliding_window = sliding_window
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.num_key_value_heads = num_key_value_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
self.scope = scope
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = torch.tril(torch.ones(self.batch_size, self.seq_length)).to(torch_device)
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
sequence_labels = None
|
||||
token_labels = None
|
||||
choice_labels = None
|
||||
if self.use_labels:
|
||||
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
|
||||
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
|
||||
choice_labels = ids_tensor([self.batch_size], self.num_choices)
|
||||
|
||||
config = self.get_config()
|
||||
|
||||
return config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
|
||||
def get_config(self):
|
||||
return Qwen2Config(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
max_window_layers=self.max_window_layers,
|
||||
use_sliding_window=self.use_sliding_window,
|
||||
sliding_window=self.sliding_window,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
num_key_value_heads=self.num_key_value_heads,
|
||||
intermediate_size=self.intermediate_size,
|
||||
hidden_act=self.hidden_act,
|
||||
hidden_dropout_prob=self.hidden_dropout_prob,
|
||||
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
is_decoder=False,
|
||||
initializer_range=self.initializer_range,
|
||||
pad_token_id=self.pad_token_id,
|
||||
bos_token_id=self.bos_token_id,
|
||||
)
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->Qwen2
|
||||
def create_and_check_model(
|
||||
self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels
|
||||
):
|
||||
model = Qwen2Model(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask)
|
||||
result = model(input_ids)
|
||||
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model_as_decoder with Llama->Qwen2
|
||||
def create_and_check_model_as_decoder(
|
||||
self,
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask,
|
||||
):
|
||||
config.add_cross_attention = True
|
||||
model = Qwen2Model(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(
|
||||
input_ids,
|
||||
attention_mask=input_mask,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_attention_mask=encoder_attention_mask,
|
||||
)
|
||||
result = model(
|
||||
input_ids,
|
||||
attention_mask=input_mask,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
)
|
||||
result = model(input_ids, attention_mask=input_mask)
|
||||
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_for_causal_lm with Llama->Qwen2
|
||||
def create_and_check_for_causal_lm(
|
||||
self,
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask,
|
||||
):
|
||||
model = Qwen2ForCausalLM(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
|
||||
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_decoder_model_past_large_inputs with Llama->Qwen2
|
||||
def create_and_check_decoder_model_past_large_inputs(
|
||||
self,
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask,
|
||||
):
|
||||
config.is_decoder = True
|
||||
config.add_cross_attention = True
|
||||
model = Qwen2ForCausalLM(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
# first forward pass
|
||||
outputs = model(
|
||||
input_ids,
|
||||
attention_mask=input_mask,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_attention_mask=encoder_attention_mask,
|
||||
use_cache=True,
|
||||
)
|
||||
past_key_values = outputs.past_key_values
|
||||
|
||||
# create hypothetical multiple next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
|
||||
|
||||
output_from_no_past = model(
|
||||
next_input_ids,
|
||||
attention_mask=next_attention_mask,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_attention_mask=encoder_attention_mask,
|
||||
output_hidden_states=True,
|
||||
)["hidden_states"][0]
|
||||
output_from_past = model(
|
||||
next_tokens,
|
||||
attention_mask=next_attention_mask,
|
||||
encoder_hidden_states=encoder_hidden_states,
|
||||
encoder_attention_mask=encoder_attention_mask,
|
||||
past_key_values=past_key_values,
|
||||
output_hidden_states=True,
|
||||
)["hidden_states"][0]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
|
||||
|
||||
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
|
||||
|
||||
# test that outputs are equal for slice
|
||||
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
|
||||
|
||||
# Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.prepare_config_and_inputs_for_common
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
token_type_ids,
|
||||
input_mask,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = config_and_inputs
|
||||
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
# Copied from tests.models.mistral.test_modeling_mistral.MistralModelTest with Mistral->Qwen2
|
||||
class Qwen2ModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (Qwen2Model, Qwen2ForCausalLM, Qwen2ForSequenceClassification) if is_torch_available() else ()
|
||||
all_generative_model_classes = (Qwen2ForCausalLM,) if is_torch_available() else ()
|
||||
pipeline_model_mapping = (
|
||||
{
|
||||
"feature-extraction": Qwen2Model,
|
||||
"text-classification": Qwen2ForSequenceClassification,
|
||||
"text-generation": Qwen2ForCausalLM,
|
||||
"zero-shot": Qwen2ForSequenceClassification,
|
||||
}
|
||||
if is_torch_available()
|
||||
else {}
|
||||
)
|
||||
test_headmasking = False
|
||||
test_pruning = False
|
||||
|
||||
# TODO (ydshieh): Check this. See https://app.circleci.com/pipelines/github/huggingface/transformers/79245/workflows/9490ef58-79c2-410d-8f51-e3495156cf9c/jobs/1012146
|
||||
def is_pipeline_test_to_skip(
|
||||
self, pipeline_test_casse_name, config_class, model_architecture, tokenizer_name, processor_name
|
||||
):
|
||||
return True
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = Qwen2ModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=Qwen2Config, hidden_size=37)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
def test_model_various_embeddings(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
for type in ["absolute", "relative_key", "relative_key_query"]:
|
||||
config_and_inputs[0].position_embedding_type = type
|
||||
self.model_tester.create_and_check_model(*config_and_inputs)
|
||||
|
||||
def test_Qwen2_sequence_classification_model(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
print(config)
|
||||
config.num_labels = 3
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = Qwen2ForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
def test_Qwen2_sequence_classification_model_for_single_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "single_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size)
|
||||
model = Qwen2ForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
def test_Qwen2_sequence_classification_model_for_multi_label(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
config.num_labels = 3
|
||||
config.problem_type = "multi_label_classification"
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
sequence_labels = ids_tensor(
|
||||
[self.model_tester.batch_size, config.num_labels], self.model_tester.type_sequence_label_size
|
||||
).to(torch.float)
|
||||
model = Qwen2ForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=attention_mask, labels=sequence_labels)
|
||||
self.assertEqual(result.logits.shape, (self.model_tester.batch_size, self.model_tester.num_labels))
|
||||
|
||||
@unittest.skip("Qwen2 buffers include complex numbers, which breaks this test")
|
||||
def test_save_load_fast_init_from_base(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Qwen2 uses GQA on all models so the KV cache is a non standard format")
|
||||
def test_past_key_values_format(self):
|
||||
pass
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_generate_padding_right(self):
|
||||
import torch
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
|
||||
torch_device
|
||||
)
|
||||
|
||||
dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
|
||||
dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)
|
||||
|
||||
model.generate(dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False)
|
||||
|
||||
model = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
attn_implementation="flash_attention_2",
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
with self.assertRaises(ValueError):
|
||||
_ = model.generate(
|
||||
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
|
||||
)
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_generate_use_cache(self):
|
||||
import torch
|
||||
|
||||
max_new_tokens = 30
|
||||
|
||||
for model_class in self.all_generative_model_classes:
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
dummy_input = inputs_dict[model_class.main_input_name]
|
||||
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
|
||||
dummy_input = dummy_input.to(torch.float16)
|
||||
|
||||
# make sure that all models have enough positions for generation
|
||||
if hasattr(config, "max_position_embeddings"):
|
||||
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
|
||||
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
|
||||
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
|
||||
# NOTE: Qwen2 apparently does not support right padding + use_cache with FA2.
|
||||
dummy_attention_mask[:, -1] = 1
|
||||
|
||||
model = model_class.from_pretrained(
|
||||
tmpdirname,
|
||||
torch_dtype=torch.float16,
|
||||
attn_implementation="flash_attention_2",
|
||||
low_cpu_mem_usage=True,
|
||||
).to(torch_device)
|
||||
|
||||
# Just test that a large cache works as expected
|
||||
_ = model.generate(
|
||||
dummy_input,
|
||||
attention_mask=dummy_attention_mask,
|
||||
max_new_tokens=max_new_tokens,
|
||||
do_sample=False,
|
||||
use_cache=True,
|
||||
)
|
||||
|
||||
@require_flash_attn
|
||||
@require_torch_gpu
|
||||
@pytest.mark.flash_attn_test
|
||||
@slow
|
||||
def test_flash_attn_2_inference_padding_right(self):
|
||||
self.skipTest("Qwen2 flash attention does not support right padding")
|
||||
|
||||
|
||||
@require_torch
|
||||
class Qwen2IntegrationTest(unittest.TestCase):
|
||||
@slow
|
||||
def test_model_450m_logits(self):
|
||||
input_ids = [1, 306, 4658, 278, 6593, 310, 2834, 338]
|
||||
model = Qwen2ForCausalLM.from_pretrained("Qwen/Qwen2-450m-beta", device_map="auto")
|
||||
input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
|
||||
with torch.no_grad():
|
||||
out = model(input_ids).logits.cpu()
|
||||
# Expected mean on dim = -1
|
||||
EXPECTED_MEAN = torch.tensor([[-2.5548, -2.5737, -3.0600, -2.5906, -2.8478, -2.8118, -2.9325, -2.7694]])
|
||||
torch.testing.assert_close(out.mean(-1), EXPECTED_MEAN, atol=1e-2, rtol=1e-2)
|
||||
# slicing logits[0, 0, 0:30]
|
||||
EXPECTED_SLICE = torch.tensor([-5.8781, -5.8616, -0.1052, -4.7200, -5.8781, -5.8774, -5.8773, -5.8777, -5.8781, -5.8780, -5.8781, -5.8779, -1.0787, 1.7583, -5.8779, -5.8780, -5.8783, -5.8778, -5.8776, -5.8781, -5.8784, -5.8778, -5.8778, -5.8777, -5.8779, -5.8778, -5.8776, -5.8780, -5.8779, -5.8781]) # fmt: skip
|
||||
print(out[0, 0, :30])
|
||||
torch.testing.assert_close(out[0, 0, :30], EXPECTED_SLICE, atol=1e-4, rtol=1e-4)
|
||||
|
||||
del model
|
||||
backend_empty_cache(torch_device)
|
||||
gc.collect()
|
||||
|
||||
@slow
|
||||
def test_model_450m_generation(self):
|
||||
EXPECTED_TEXT_COMPLETION = """My favourite condiment is 100% ketchup. I love it on everything. I’m not a big"""
|
||||
prompt = "My favourite condiment is "
|
||||
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-450m-beta", use_fast=False)
|
||||
model = Qwen2ForCausalLM.from_pretrained("Qwen/Qwen2-450m-beta", device_map="auto")
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)
|
||||
|
||||
# greedy generation outputs
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=20, temperature=0)
|
||||
text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
||||
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
|
||||
|
||||
del model
|
||||
backend_empty_cache(torch_device)
|
||||
gc.collect()
|
||||
|
||||
@require_bitsandbytes
|
||||
@slow
|
||||
@require_flash_attn
|
||||
def test_model_450m_long_prompt(self):
|
||||
EXPECTED_OUTPUT_TOKEN_IDS = [306, 338]
|
||||
# An input with 4097 tokens that is above the size of the sliding window
|
||||
input_ids = [1] + [306, 338] * 2048
|
||||
model = Qwen2ForCausalLM.from_pretrained(
|
||||
"Qwen/Qwen2-450m-beta",
|
||||
device_map="auto",
|
||||
load_in_4bit=True,
|
||||
attn_implementation="flash_attention_2",
|
||||
)
|
||||
input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
|
||||
self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())
|
||||
|
||||
# Assisted generation
|
||||
assistant_model = model
|
||||
assistant_model.generation_config.num_assistant_tokens = 2
|
||||
assistant_model.generation_config.num_assistant_tokens_schedule = "constant"
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
|
||||
self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())
|
||||
|
||||
del assistant_model
|
||||
del model
|
||||
backend_empty_cache(torch_device)
|
||||
gc.collect()
|
||||
|
||||
@slow
|
||||
@require_torch_sdpa
|
||||
def test_model_450m_long_prompt_sdpa(self):
|
||||
EXPECTED_OUTPUT_TOKEN_IDS = [306, 338]
|
||||
# An input with 4097 tokens that is above the size of the sliding window
|
||||
input_ids = [1] + [306, 338] * 2048
|
||||
model = Qwen2ForCausalLM.from_pretrained(
|
||||
"Qwen/Qwen2-450m-beta",
|
||||
device_map="auto",
|
||||
attn_implementation="sdpa",
|
||||
)
|
||||
input_ids = torch.tensor([input_ids]).to(model.model.embed_tokens.weight.device)
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=4, temperature=0)
|
||||
self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())
|
||||
|
||||
# Assisted generation
|
||||
assistant_model = model
|
||||
assistant_model.generation_config.num_assistant_tokens = 2
|
||||
assistant_model.generation_config.num_assistant_tokens_schedule = "constant"
|
||||
generated_ids = assistant_model.generate(input_ids, max_new_tokens=4, temperature=0)
|
||||
self.assertEqual(EXPECTED_OUTPUT_TOKEN_IDS, generated_ids[0][-2:].tolist())
|
||||
|
||||
del assistant_model
|
||||
|
||||
backend_empty_cache(torch_device)
|
||||
gc.collect()
|
||||
|
||||
EXPECTED_TEXT_COMPLETION = """My favourite condiment is 100% ketchup. I love it on everything. I’m not a big"""
|
||||
prompt = "My favourite condiment is "
|
||||
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-450m-beta", use_fast=False)
|
||||
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)
|
||||
|
||||
# greedy generation outputs
|
||||
generated_ids = model.generate(input_ids, max_new_tokens=20, temperature=0)
|
||||
text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
||||
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
|
||||
|
||||
@slow
|
||||
def test_speculative_generation(self):
|
||||
EXPECTED_TEXT_COMPLETION = (
|
||||
"My favourite condiment is 100% Sriracha. I love the heat, the tang and the fact costs"
|
||||
)
|
||||
prompt = "My favourite condiment is "
|
||||
tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-7B-beta", use_fast=False)
|
||||
model = Qwen2ForCausalLM.from_pretrained("Qwen/Qwen2-450m-beta", device_map="auto", torch_dtype=torch.float16)
|
||||
assistant_model = Qwen2ForCausalLM.from_pretrained(
|
||||
"Qwen/Qwen2-450m-beta", device_map="auto", torch_dtype=torch.float16
|
||||
)
|
||||
input_ids = tokenizer.encode(prompt, return_tensors="pt").to(model.model.embed_tokens.weight.device)
|
||||
|
||||
# greedy generation outputs
|
||||
set_seed(0)
|
||||
generated_ids = model.generate(
|
||||
input_ids, max_new_tokens=20, do_sample=True, temperature=0.3, assistant_model=assistant_model
|
||||
)
|
||||
text = tokenizer.decode(generated_ids[0], skip_special_tokens=True)
|
||||
self.assertEqual(EXPECTED_TEXT_COMPLETION, text)
|
||||
|
||||
del model
|
||||
backend_empty_cache(torch_device)
|
||||
gc.collect()
|
||||
204
tests/models/qwen2/test_tokenization_qwen2.py
Normal file
204
tests/models/qwen2/test_tokenization_qwen2.py
Normal file
@@ -0,0 +1,204 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
|
||||
import json
|
||||
import os
|
||||
import unittest
|
||||
|
||||
from transformers import AddedToken, Qwen2Tokenizer, Qwen2TokenizerFast
|
||||
from transformers.models.qwen2.tokenization_qwen2 import VOCAB_FILES_NAMES, bytes_to_unicode
|
||||
from transformers.testing_utils import require_tokenizers, slow
|
||||
|
||||
from ...test_tokenization_common import TokenizerTesterMixin
|
||||
|
||||
|
||||
@require_tokenizers
|
||||
class Qwen2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
tokenizer_class = Qwen2Tokenizer
|
||||
rust_tokenizer_class = Qwen2TokenizerFast
|
||||
test_slow_tokenizer = True
|
||||
test_rust_tokenizer = True
|
||||
space_between_special_tokens = False
|
||||
from_pretrained_kwargs = None
|
||||
test_seq2seq = False
|
||||
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
|
||||
# this make sure the vocabuary is complete at the byte level.
|
||||
vocab = list(bytes_to_unicode().values())
|
||||
# the vocabulary, note:
|
||||
# - `"\u0120n"`, `"\u0120lowest"`, `"\u0120newer"`, and `"\u0120wider"` are ineffective, because there are
|
||||
# not in the merges.
|
||||
# - `"01"` is ineffective, because the merge is ineffective due to pretokenization.
|
||||
vocab.extend(
|
||||
[
|
||||
"\u0120l",
|
||||
"\u0120n",
|
||||
"\u0120lo",
|
||||
"\u0120low",
|
||||
"er",
|
||||
"\u0120lowest",
|
||||
"\u0120newer",
|
||||
"\u0120wider",
|
||||
"01",
|
||||
";}",
|
||||
";}\u010a",
|
||||
"\u00cf\u0135",
|
||||
]
|
||||
)
|
||||
|
||||
vocab_tokens = dict(zip(vocab, range(len(vocab))))
|
||||
|
||||
# note: `"0 1"` is in the merges, but the pretokenization rules render it ineffective
|
||||
merges = [
|
||||
"#version: 0.2",
|
||||
"\u0120 l",
|
||||
"\u0120l o",
|
||||
"\u0120lo w",
|
||||
"e r",
|
||||
"0 1",
|
||||
"; }",
|
||||
";} \u010a",
|
||||
"\u00cf \u0135",
|
||||
]
|
||||
|
||||
self.special_tokens_map = {"eos_token": "<|endoftext|>"}
|
||||
|
||||
self.vocab_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["vocab_file"])
|
||||
self.merges_file = os.path.join(self.tmpdirname, VOCAB_FILES_NAMES["merges_file"])
|
||||
with open(self.vocab_file, "w", encoding="utf-8") as fp:
|
||||
fp.write(json.dumps(vocab_tokens) + "\n")
|
||||
with open(self.merges_file, "w", encoding="utf-8") as fp:
|
||||
fp.write("\n".join(merges))
|
||||
|
||||
def get_tokenizer(self, **kwargs):
|
||||
kwargs.update(self.special_tokens_map)
|
||||
return Qwen2Tokenizer.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_rust_tokenizer(self, **kwargs):
|
||||
kwargs.update(self.special_tokens_map)
|
||||
return Qwen2TokenizerFast.from_pretrained(self.tmpdirname, **kwargs)
|
||||
|
||||
def get_input_output_texts(self, tokenizer):
|
||||
# this case should cover
|
||||
# - NFC normalization (code point U+03D3 has different normalization forms under NFC, NFD, NFKC, and NFKD)
|
||||
# - the pretokenization rules (spliting digits and merging symbols with \n\r)
|
||||
input_text = "lower lower newer 010;}\n<|endoftext|>\u03d2\u0301"
|
||||
output_text = "lower lower newer 010;}\n<|endoftext|>\u03d3"
|
||||
return input_text, output_text
|
||||
|
||||
def test_python_full_tokenizer(self):
|
||||
tokenizer = self.get_tokenizer()
|
||||
sequence, _ = self.get_input_output_texts(tokenizer)
|
||||
bpe_tokens = [
|
||||
"l",
|
||||
"o",
|
||||
"w",
|
||||
"er",
|
||||
"\u0120low",
|
||||
"er",
|
||||
"\u0120",
|
||||
"n",
|
||||
"e",
|
||||
"w",
|
||||
"er",
|
||||
"\u0120",
|
||||
"0",
|
||||
"1",
|
||||
"0",
|
||||
";}\u010a",
|
||||
"<|endoftext|>",
|
||||
"\u00cf\u0135",
|
||||
]
|
||||
tokens = tokenizer.tokenize(sequence)
|
||||
self.assertListEqual(tokens, bpe_tokens)
|
||||
|
||||
input_tokens = tokens
|
||||
input_bpe_tokens = [75, 78, 86, 260, 259, 260, 220, 77, 68, 86, 260, 220, 15, 16, 15, 266, 268, 267]
|
||||
self.assertListEqual(tokenizer.convert_tokens_to_ids(input_tokens), input_bpe_tokens)
|
||||
|
||||
@unittest.skip("We disable the test of pretokenization as it is not reversible.")
|
||||
def test_pretokenized_inputs(self):
|
||||
# the test case in parent class uses str.split to "pretokenize",
|
||||
# which eats the whitespaces, which, in turn, is not reversible.
|
||||
# the results, by nature, should be different.
|
||||
pass
|
||||
|
||||
def test_nfc_normalization(self):
|
||||
# per https://unicode.org/faq/normalization.html, there are three characters whose normalization forms
|
||||
# under NFC, NFD, NFKC, and NFKD are all different
|
||||
# using these, we can make sure only NFC is applied
|
||||
input_string = "\u03d2\u0301\u03d2\u0308\u017f\u0307" # the NFD form
|
||||
output_string = "\u03d3\u03d4\u1e9b" # the NFC form
|
||||
|
||||
if self.test_slow_tokenizer:
|
||||
tokenizer = self.get_tokenizer()
|
||||
tokenizer_output_string, _ = tokenizer.prepare_for_tokenization(input_string)
|
||||
self.assertEqual(tokenizer_output_string, output_string)
|
||||
|
||||
if self.test_rust_tokenizer:
|
||||
tokenizer = self.get_rust_tokenizer()
|
||||
# we can check the class of the normalizer, but it would be okay if Sequence([NFD, NFC]) is used
|
||||
# let's check the output instead
|
||||
tokenizer_output_string = tokenizer.backend_tokenizer.normalizer.normalize_str(input_string)
|
||||
self.assertEqual(tokenizer_output_string, output_string)
|
||||
|
||||
def test_slow_tokenizer_decode_spaces_between_special_tokens_default(self):
|
||||
# Qwen2Tokenzier changes the default `spaces_between_special_tokens` in `decode` to False
|
||||
if not self.test_slow_tokenizer:
|
||||
return
|
||||
|
||||
# tokenizer has a special token: `"<|endfotext|>"` as eos, but it is not `legacy_added_tokens`
|
||||
# special tokens in `spaces_between_special_tokens` means spaces between `legacy_added_tokens`
|
||||
# that would be `"<|im_start|>"` and `"<|im_end|>"` in Qwen/Qwen2 Models
|
||||
token_ids = [259, 260, 268, 269, 26]
|
||||
sequence = " lower<|endoftext|><|im_start|>;"
|
||||
sequence_with_space = " lower<|endoftext|> <|im_start|> ;"
|
||||
|
||||
tokenizer = self.get_tokenizer()
|
||||
# let's add a legacy_added_tokens
|
||||
im_start = AddedToken(
|
||||
"<|im_start|>", single_word=False, lstrip=False, rstrip=False, special=True, normalized=False
|
||||
)
|
||||
tokenizer.add_tokens([im_start])
|
||||
|
||||
# `spaces_between_special_tokens` defaults to False
|
||||
self.assertEqual(tokenizer.decode(token_ids), sequence)
|
||||
|
||||
# but it can be set to True
|
||||
self.assertEqual(tokenizer.decode(token_ids, spaces_between_special_tokens=True), sequence_with_space)
|
||||
|
||||
@slow
|
||||
def test_tokenizer_integration(self):
|
||||
sequences = [
|
||||
"Transformers (formerly known as pytorch-transformers and pytorch-pretrained-bert) provides "
|
||||
"general-purpose architectures (BERT, GPT-2, RoBERTa, XLM, DistilBert, XLNet...) for Natural "
|
||||
"Language Understanding (NLU) and Natural Language Generation (NLG) with over 32+ pretrained "
|
||||
"models in 100+ languages and deep interoperability between Jax, PyTorch and TensorFlow.",
|
||||
"🤗 Transformers 提供了可以轻松地下载并且训练先进的预训练模型的 API 和工具。使用预训练模型可以减少计算消耗和碳排放，并且节省从头训练所需要的时间和资源。",
|
||||
"""```python\ntokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen-tokenizer")\n"""
|
||||
"""tokenizer("世界，你好！")```""",
|
||||
]
|
||||
|
||||
expected_encoding = {'input_ids': [[8963, 388, 320, 69514, 3881, 438, 4510, 27414, 32852, 388, 323, 4510, 27414, 21334, 35722, 1455, 529, 8, 5707, 4586, 58238, 77235, 320, 61437, 11, 479, 2828, 12, 17, 11, 11830, 61437, 64, 11, 1599, 10994, 11, 27604, 321, 33, 529, 11, 29881, 6954, 32574, 369, 18448, 11434, 45451, 320, 45, 23236, 8, 323, 18448, 11434, 23470, 320, 30042, 38, 8, 448, 916, 220, 18, 17, 10, 80669, 4119, 304, 220, 16, 15, 15, 10, 15459, 323, 5538, 94130, 2897, 1948, 619, 706, 11, 5355, 51, 21584, 323, 94986, 13], [144834, 80532, 93685, 83744, 34187, 73670, 104261, 29490, 62189, 103937, 104034, 102830, 98841, 104034, 104949, 9370, 5333, 58143, 102011, 1773, 37029, 98841, 104034, 104949, 73670, 101940, 100768, 104997, 33108, 100912, 105054, 90395, 100136, 106831, 45181, 64355, 104034, 113521, 101975, 33108, 85329, 1773, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643], [73594, 12669, 198, 85593, 284, 8979, 37434, 6387, 10442, 35722, 445, 48, 16948, 45274, 16948, 34841, 3135, 1138, 85593, 445, 99489, 3837, 108386, 6313, 899, 73594, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643]], 'attention_mask': [[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0], [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]} # fmt: off
|
||||
|
||||
self.tokenizer_integration_test_util(
|
||||
expected_encoding=expected_encoding,
|
||||
model_name="Qwen/Qwen-tokenizer",
|
||||
revision="5909c8222473b2c73b0b73fb054552cd4ef6a8eb",
|
||||
sequences=sequences,
|
||||
)
|
||||
@@ -198,6 +198,7 @@ docs/source/en/model_doc/pop2piano.md
|
||||
docs/source/en/model_doc/prophetnet.md
|
||||
docs/source/en/model_doc/pvt.md
|
||||
docs/source/en/model_doc/qdqbert.md
|
||||
docs/source/en/model_doc/qwen2.md
|
||||
docs/source/en/model_doc/rag.md
|
||||
docs/source/en/model_doc/realm.md
|
||||
docs/source/en/model_doc/reformer.md
|
||||
@@ -745,6 +746,10 @@ src/transformers/models/pvt/image_processing_pvt.py
|
||||
src/transformers/models/pvt/modeling_pvt.py
|
||||
src/transformers/models/qdqbert/configuration_qdqbert.py
|
||||
src/transformers/models/qdqbert/modeling_qdqbert.py
|
||||
src/transformers/models/qwen2/configuration_qwen2.py
|
||||
src/transformers/models/qwen2/modeling_qwen2.py
|
||||
src/transformers/models/qwen2/tokenization_qwen2.py
|
||||
src/transformers/models/qwen2/tokenization_qwen2_fast.py
|
||||
src/transformers/models/rag/configuration_rag.py
|
||||
src/transformers/models/rag/modeling_rag.py
|
||||
src/transformers/models/rag/modeling_tf_rag.py
|
||||
|
||||
Reference in New Issue
Block a user