From d6ffe74dfa577b5e7d12e48aa1c686ad8d3ef557 Mon Sep 17 00:00:00 2001 From: Junyang Lin Date: Wed, 17 Jan 2024 23:02:22 +0800 Subject: [PATCH] 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 Co-authored-by: renxuancheng.rxc Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> --- README.md | 1 + README_es.md | 1 + README_hd.md | 1 + README_ja.md | 1 + README_ko.md | 1 + README_zh-hans.md | 1 + README_zh-hant.md | 1 + docs/source/en/_toctree.yml | 2 + docs/source/en/index.md | 1 + docs/source/en/model_doc/qwen2.md | 82 + docs/source/en/perf_infer_gpu_one.md | 2 + docs/source/en/tasks/language_modeling.md | 2 +- .../en/tasks/sequence_classification.md | 2 +- src/transformers/__init__.py | 22 + src/transformers/convert_slow_tokenizer.py | 43 + src/transformers/models/__init__.py | 1 + .../models/auto/configuration_auto.py | 3 + src/transformers/models/auto/modeling_auto.py | 3 + .../models/auto/tokenization_auto.py | 7 + src/transformers/models/qwen2/__init__.py | 80 + .../models/qwen2/configuration_qwen2.py | 144 ++ .../models/qwen2/modeling_qwen2.py | 1401 +++++++++++++++++ .../models/qwen2/tokenization_qwen2.py | 345 ++++ .../models/qwen2/tokenization_qwen2_fast.py | 143 ++ src/transformers/utils/dummy_pt_objects.py | 28 + .../utils/dummy_tokenizers_objects.py | 7 + tests/models/qwen2/__init__.py | 0 tests/models/qwen2/test_modeling_qwen2.py | 604 +++++++ tests/models/qwen2/test_tokenization_qwen2.py | 204 +++ utils/not_doctested.txt | 5 + 30 files changed, 3136 insertions(+), 2 deletions(-) create mode 100644 docs/source/en/model_doc/qwen2.md create mode 100644 src/transformers/models/qwen2/__init__.py create mode 100644 src/transformers/models/qwen2/configuration_qwen2.py create mode 100644 src/transformers/models/qwen2/modeling_qwen2.py create mode 100644 src/transformers/models/qwen2/tokenization_qwen2.py create mode 100644 src/transformers/models/qwen2/tokenization_qwen2_fast.py create mode 100644 tests/models/qwen2/__init__.py create mode 100644 tests/models/qwen2/test_modeling_qwen2.py create mode 100644 tests/models/qwen2/test_tokenization_qwen2.py diff --git a/README.md b/README.md index daab3d1f9d..806738d64e 100644 --- a/README.md +++ b/README.md @@ -458,6 +458,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h 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. diff --git a/README_es.md b/README_es.md index 9e1ac93b4a..c5aa57f226 100644 --- a/README_es.md +++ b/README_es.md @@ -433,6 +433,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt 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. diff --git a/README_hd.md b/README_hd.md index 92935efb58..5924c31627 100644 --- a/README_hd.md +++ b/README_hd.md @@ -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. diff --git a/README_ja.md b/README_ja.md index f43dda021c..e54ba07359 100644 --- a/README_ja.md +++ b/README_ja.md @@ -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) diff --git a/README_ko.md b/README_ko.md index c2e53a1b81..35d838346c 100644 --- a/README_ko.md +++ b/README_ko.md @@ -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. diff --git a/README_zh-hans.md b/README_zh-hans.md index 972f3a386f..3999da2bc1 100644 --- a/README_zh-hans.md +++ b/README_zh-hans.md @@ -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。 diff --git a/README_zh-hant.md b/README_zh-hant.md index b17c8946bc..83ff0b1532 100644 --- a/README_zh-hant.md +++ b/README_zh-hant.md @@ -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. diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 86cffb9a7e..02c13c2776 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -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 diff --git a/docs/source/en/index.md b/docs/source/en/index.md index 52b5df6e59..1f43ec7d54 100644 --- a/docs/source/en/index.md +++ b/docs/source/en/index.md @@ -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) | ✅ | ❌ | ❌ | diff --git a/docs/source/en/model_doc/qwen2.md b/docs/source/en/model_doc/qwen2.md new file mode 100644 index 0000000000..61e45fd9c2 --- /dev/null +++ b/docs/source/en/model_doc/qwen2.md @@ -0,0 +1,82 @@ + + +# 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 diff --git a/docs/source/en/perf_infer_gpu_one.md b/docs/source/en/perf_infer_gpu_one.md index 5cc9cd208d..899e5b52f0 100644 --- a/docs/source/en/perf_infer_gpu_one.md +++ b/docs/source/en/perf_infer_gpu_one.md @@ -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) diff --git a/docs/source/en/tasks/language_modeling.md b/docs/source/en/tasks/language_modeling.md index a50555dfcf..02b5f2ca73 100644 --- a/docs/source/en/tasks/language_modeling.md +++ b/docs/source/en/tasks/language_modeling.md @@ -37,7 +37,7 @@ You can finetune other architectures for causal language modeling following the Choose one of the following architectures: -[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) diff --git a/docs/source/en/tasks/sequence_classification.md b/docs/source/en/tasks/sequence_classification.md index 4a0e5b611c..0acbf7bfb1 100644 --- a/docs/source/en/tasks/sequence_classification.md +++ b/docs/source/en/tasks/sequence_classification.md @@ -33,7 +33,7 @@ The task illustrated in this tutorial is supported by the following model archit -[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) diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index 4941d72445..78cef80fea 100644 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -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, diff --git a/src/transformers/convert_slow_tokenizer.py b/src/transformers/convert_slow_tokenizer.py index 76ac66ceb9..46f2b2dc23 100644 --- a/src/transformers/convert_slow_tokenizer.py +++ b/src/transformers/convert_slow_tokenizer.py @@ -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, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index 2c20873c2e..3fea0edb9a 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -176,6 +176,7 @@ from . import ( prophetnet, pvt, qdqbert, + qwen2, rag, realm, reformer, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index 9eb3f1985c..d439e5bb61 100755 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -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"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index 7bf50a4518..25e5699201 100755 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -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"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index ac09eecd1e..27998e6c0f 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -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)), ( diff --git a/src/transformers/models/qwen2/__init__.py b/src/transformers/models/qwen2/__init__.py new file mode 100644 index 0000000000..9fd51aaffe --- /dev/null +++ b/src/transformers/models/qwen2/__init__.py @@ -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__) diff --git a/src/transformers/models/qwen2/configuration_qwen2.py b/src/transformers/models/qwen2/configuration_qwen2.py new file mode 100644 index 0000000000..0bbfd1cf16 --- /dev/null +++ b/src/transformers/models/qwen2/configuration_qwen2.py @@ -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, + ) diff --git a/src/transformers/models/qwen2/modeling_qwen2.py b/src/transformers/models/qwen2/modeling_qwen2.py new file mode 100644 index 0000000000..f8290928a5 --- /dev/null +++ b/src/transformers/models/qwen2/modeling_qwen2.py @@ -0,0 +1,1401 @@ +# coding=utf-8 +# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved. +# +# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX +# and OPT implementations in this library. It has been modified from its +# original forms to accommodate minor architectural differences compared +# to GPT-NeoX and OPT used by the Meta AI team that trained the model. +# +# 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. +""" PyTorch Qwen2 model.""" +import inspect +import math +import warnings +from typing import List, Optional, Tuple, Union + +import torch +import torch.nn.functional as F +import torch.utils.checkpoint +from torch import nn +from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss + +from ...activations import ACT2FN +from ...cache_utils import Cache, DynamicCache +from ...modeling_attn_mask_utils import _prepare_4d_causal_attention_mask, _prepare_4d_causal_attention_mask_for_sdpa +from ...modeling_outputs import BaseModelOutputWithPast, CausalLMOutputWithPast, SequenceClassifierOutputWithPast +from ...modeling_utils import PreTrainedModel +from ...utils import ( + add_start_docstrings, + add_start_docstrings_to_model_forward, + is_flash_attn_2_available, + is_flash_attn_greater_or_equal_2_10, + logging, + replace_return_docstrings, +) +from .configuration_qwen2 import Qwen2Config + + +if is_flash_attn_2_available(): + from flash_attn import flash_attn_func, flash_attn_varlen_func + from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input # noqa + + _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters) + + +logger = logging.get_logger(__name__) + + +_CHECKPOINT_FOR_DOC = "Qwen/Qwen2-7B-beta" +_CONFIG_FOR_DOC = "Qwen2Config" + +QWEN2_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "Qwen/Qwen2-7B-beta", + # See all Qwen2 models at https://huggingface.co/models?filter=qwen2 +] + + +# Copied from transformers.models.llama.modeling_llama._get_unpad_data +def _get_unpad_data(attention_mask): + seqlens_in_batch = attention_mask.sum(dim=-1, dtype=torch.int32) + indices = torch.nonzero(attention_mask.flatten(), as_tuple=False).flatten() + max_seqlen_in_batch = seqlens_in_batch.max().item() + cu_seqlens = F.pad(torch.cumsum(seqlens_in_batch, dim=0, dtype=torch.torch.int32), (1, 0)) + return ( + indices, + cu_seqlens, + max_seqlen_in_batch, + ) + + +# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->Qwen2 +class Qwen2RMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + Qwen2RMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + +# Copied from transformers.models.llama.modeling_llama.LlamaRotaryEmbedding with Llama->Qwen2 +class Qwen2RotaryEmbedding(nn.Module): + def __init__(self, dim, max_position_embeddings=2048, base=10000, device=None): + super().__init__() + + self.dim = dim + self.max_position_embeddings = max_position_embeddings + self.base = base + inv_freq = 1.0 / (self.base ** (torch.arange(0, self.dim, 2).float().to(device) / self.dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + + # Build here to make `torch.jit.trace` work. + self._set_cos_sin_cache( + seq_len=max_position_embeddings, device=self.inv_freq.device, dtype=torch.get_default_dtype() + ) + + def _set_cos_sin_cache(self, seq_len, device, dtype): + self.max_seq_len_cached = seq_len + t = torch.arange(self.max_seq_len_cached, device=device, dtype=self.inv_freq.dtype) + + freqs = torch.outer(t, self.inv_freq) + # Different from paper, but it uses a different permutation in order to obtain the same calculation + emb = torch.cat((freqs, freqs), dim=-1) + self.register_buffer("cos_cached", emb.cos().to(dtype), persistent=False) + self.register_buffer("sin_cached", emb.sin().to(dtype), persistent=False) + + def forward(self, x, seq_len=None): + # x: [bs, num_attention_heads, seq_len, head_size] + if seq_len > self.max_seq_len_cached: + self._set_cos_sin_cache(seq_len=seq_len, device=x.device, dtype=x.dtype) + + return ( + self.cos_cached[:seq_len].to(dtype=x.dtype), + self.sin_cached[:seq_len].to(dtype=x.dtype), + ) + + +# Copied from transformers.models.llama.modeling_llama.rotate_half +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +# Copied from transformers.models.llama.modeling_llama.apply_rotary_pos_emb +def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`): + The position indices of the tokens corresponding to the query and key tensors. For example, this can be + used to pass offsetted position ids when working with a KV-cache. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos[position_ids].unsqueeze(unsqueeze_dim) + sin = sin[position_ids].unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + + +# Copied from transformers.models.mistral.modeling_mistral.MistralMLP with Mistral->Qwen2 +class Qwen2MLP(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = config.intermediate_size + self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False) + self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) + self.act_fn = ACT2FN[config.hidden_act] + + def forward(self, x): + return self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + + +# Copied from transformers.models.llama.modeling_llama.repeat_kv +def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: + """ + This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, + num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) + """ + batch, num_key_value_heads, slen, head_dim = hidden_states.shape + if n_rep == 1: + return hidden_states + hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) + return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) + + +class Qwen2Attention(nn.Module): + """ + Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer + and "Generating Long Sequences with Sparse Transformers". + """ + + def __init__(self, config: Qwen2Config, layer_idx: Optional[int] = None): + super().__init__() + self.config = config + self.layer_idx = layer_idx + if layer_idx is None: + logger.warning_once( + f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will " + "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` " + "when creating this class." + ) + + self.hidden_size = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = self.hidden_size // self.num_heads + self.num_key_value_heads = config.num_key_value_heads + self.num_key_value_groups = self.num_heads // self.num_key_value_heads + self.max_position_embeddings = config.max_position_embeddings + self.rope_theta = config.rope_theta + self.is_causal = True + self.attention_dropout = config.attention_dropout + + if (self.head_dim * self.num_heads) != self.hidden_size: + raise ValueError( + f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}" + f" and `num_heads`: {self.num_heads})." + ) + self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True) + self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) + self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True) + self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False) + + self.rotary_emb = Qwen2RotaryEmbedding( + self.head_dim, + max_position_embeddings=self.max_position_embeddings, + base=self.rope_theta, + ) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + **kwargs, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if "padding_mask" in kwargs: + warnings.warn( + "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" + ) + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + if self.layer_idx is None: + raise ValueError( + f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " + "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " + "with a layer index." + ) + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # repeat k/v heads if n_kv_heads < n_heads + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) + + if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len): + raise ValueError( + f"Attention weights should be of size {(bsz, self.num_heads, q_len, kv_seq_len)}, but is" + f" {attn_weights.size()}" + ) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" + ) + + attn_weights = attn_weights + attention_mask + + # upcast attention to fp32 + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) + attn_output = torch.matmul(attn_weights, value_states) + + if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) + + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + +class Qwen2FlashAttention2(Qwen2Attention): + """ + Qwen2 flash attention module, following Qwen2 attention module. This module inherits from `Qwen2Attention` + as the weights of the module stays untouched. The only required change would be on the forward pass + where it needs to correctly call the public API of flash attention and deal with padding tokens + in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom + config.max_window_layers layers. + """ + + # Copied from transformers.models.llama.modeling_llama.LlamaFlashAttention2.__init__ + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. + # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. + # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). + self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + **kwargs, + ): + if "padding_mask" in kwargs: + warnings.warn( + "Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`" + ) + + # overwrite attention_mask with padding_mask + attention_mask = kwargs.pop("padding_mask") + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + if self.layer_idx is None: + raise ValueError( + f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} " + "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class " + "with a layer index." + ) + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + + # Because the input can be padded, the absolute sequence length depends on the max position id. + rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1 + cos, sin = self.rotary_emb(value_states, seq_len=rotary_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + use_sliding_windows = ( + _flash_supports_window_size + and getattr(self.config, "sliding_window", None) is not None + and kv_seq_len > self.config.sliding_window + and self.config.use_sliding_window + ) + + if not _flash_supports_window_size: + logger.warning_once( + "The current flash attention version does not support sliding window attention, for a more memory efficient implementation" + " make sure to upgrade flash-attn library." + ) + + if past_key_value is not None: + # Activate slicing cache only if the config has a value `sliding_windows` attribute + cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0 + if ( + getattr(self.config, "sliding_window", None) is not None + and kv_seq_len > self.config.sliding_window + and cache_has_contents + ): + slicing_tokens = 1 - self.config.sliding_window + + past_key = past_key_value[self.layer_idx][0] + past_value = past_key_value[self.layer_idx][1] + + past_key = past_key[:, :, slicing_tokens:, :].contiguous() + past_value = past_value[:, :, slicing_tokens:, :].contiguous() + + if past_key.shape[-2] != self.config.sliding_window - 1: + raise ValueError( + f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got" + f" {past_key.shape}" + ) + + if attention_mask is not None: + attention_mask = attention_mask[:, slicing_tokens:] + attention_mask = torch.cat([attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1) + + cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # repeat k/v heads if n_kv_heads < n_heads + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + dropout_rate = 0.0 if not self.training else self.attention_dropout + + # In PEFT, usually we cast the layer norms in float32 for training stability reasons + # therefore the input hidden states gets silently casted in float32. Hence, we need + # cast them back in float16 just to be sure everything works as expected. + input_dtype = query_states.dtype + if input_dtype == torch.float32: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + # Handle the case where the model is quantized + elif hasattr(self.config, "_pre_quantization_dtype"): + target_dtype = self.config._pre_quantization_dtype + else: + target_dtype = self.q_proj.weight.dtype + + logger.warning_once( + f"The input hidden states seems to be silently casted in float32, this might be related to" + f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" + f" {target_dtype}." + ) + + query_states = query_states.to(target_dtype) + key_states = key_states.to(target_dtype) + value_states = value_states.to(target_dtype) + + # Reashape to the expected shape for Flash Attention + query_states = query_states.transpose(1, 2) + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + attn_output = self._flash_attention_forward( + query_states, + key_states, + value_states, + attention_mask, + q_len, + dropout=dropout_rate, + use_sliding_windows=use_sliding_windows, + ) + + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous() + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + def _flash_attention_forward( + self, + query_states, + key_states, + value_states, + attention_mask, + query_length, + dropout=0.0, + softmax_scale=None, + use_sliding_windows=False, + ): + """ + Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token + first unpad the input, then computes the attention scores and pad the final attention scores. + + Args: + query_states (`torch.Tensor`): + Input query states to be passed to Flash Attention API + key_states (`torch.Tensor`): + Input key states to be passed to Flash Attention API + value_states (`torch.Tensor`): + Input value states to be passed to Flash Attention API + attention_mask (`torch.Tensor`): + The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the + position of padding tokens and 1 for the position of non-padding tokens. + dropout (`int`, *optional*): + Attention dropout + softmax_scale (`float`, *optional*): + The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim) + use_sliding_windows (`bool`, *optional*): + Whether to activate sliding window attention. + """ + if not self._flash_attn_uses_top_left_mask: + causal = self.is_causal + else: + # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in LlamaFlashAttention2 __init__. + causal = self.is_causal and query_length != 1 + + # Decide whether to use SWA or not by layer index. + if use_sliding_windows and self.layer_idx >= self.config.max_window_layers: + use_sliding_windows = False + + # Contains at least one padding token in the sequence + if attention_mask is not None: + batch_size = query_states.shape[0] + query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input( + query_states, key_states, value_states, attention_mask, query_length + ) + + cu_seqlens_q, cu_seqlens_k = cu_seq_lens + max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens + + if not use_sliding_windows: + attn_output_unpad = flash_attn_varlen_func( + query_states, + key_states, + value_states, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_k=cu_seqlens_k, + max_seqlen_q=max_seqlen_in_batch_q, + max_seqlen_k=max_seqlen_in_batch_k, + dropout_p=dropout, + softmax_scale=softmax_scale, + causal=causal, + ) + else: + attn_output_unpad = flash_attn_varlen_func( + query_states, + key_states, + value_states, + cu_seqlens_q=cu_seqlens_q, + cu_seqlens_k=cu_seqlens_k, + max_seqlen_q=max_seqlen_in_batch_q, + max_seqlen_k=max_seqlen_in_batch_k, + dropout_p=dropout, + softmax_scale=softmax_scale, + causal=causal, + window_size=(self.config.sliding_window, self.config.sliding_window), + ) + + attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length) + else: + if not use_sliding_windows: + attn_output = flash_attn_func( + query_states, + key_states, + value_states, + dropout, + softmax_scale=softmax_scale, + causal=causal, + ) + else: + attn_output = flash_attn_func( + query_states, + key_states, + value_states, + dropout, + softmax_scale=softmax_scale, + causal=causal, + window_size=(self.config.sliding_window, self.config.sliding_window), + ) + + return attn_output + + # Copied from transformers.models.mistral.modeling_mistral.MistralFlashAttention2._upad_input + def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length): + batch_size, kv_seq_len, num_heads, head_dim = key_layer.shape + + # On the first iteration we need to properly re-create the padding mask + # by slicing it on the proper place + if kv_seq_len != attention_mask.shape[-1]: + attention_mask_num_tokens = attention_mask.shape[-1] + attention_mask = attention_mask[:, attention_mask_num_tokens - kv_seq_len :] + + indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask) + + key_layer = index_first_axis(key_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) + value_layer = index_first_axis(value_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k) + + if query_length == kv_seq_len: + query_layer = index_first_axis( + query_layer.reshape(batch_size * kv_seq_len, num_heads, head_dim), indices_k + ) + cu_seqlens_q = cu_seqlens_k + max_seqlen_in_batch_q = max_seqlen_in_batch_k + indices_q = indices_k + elif query_length == 1: + max_seqlen_in_batch_q = 1 + cu_seqlens_q = torch.arange( + batch_size + 1, dtype=torch.int32, device=query_layer.device + ) # There is a memcpy here, that is very bad. + indices_q = cu_seqlens_q[:-1] + query_layer = query_layer.squeeze(1) + else: + # The -q_len: slice assumes left padding. + attention_mask = attention_mask[:, -query_length:] + query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask) + + return ( + query_layer, + key_layer, + value_layer, + indices_q, + (cu_seqlens_q, cu_seqlens_k), + (max_seqlen_in_batch_q, max_seqlen_in_batch_k), + ) + + +# Copied from transformers.models.llama.modeling_llama.LlamaSdpaAttention with Llama->Qwen2 +class Qwen2SdpaAttention(Qwen2Attention): + """ + Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from + `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to + SDPA API. + """ + + # Adapted from Qwen2Attention.forward + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if output_attentions: + # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. + logger.warning_once( + "Qwen2Model is using Qwen2SdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " + 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + return super().forward( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + kv_seq_len = key_states.shape[-2] + if past_key_value is not None: + kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx) + cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids) + + if past_key_value is not None: + cache_kwargs = {"sin": sin, "cos": cos} # Specific to RoPE models + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + if attention_mask is not None: + if attention_mask.size() != (bsz, 1, q_len, kv_seq_len): + raise ValueError( + f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}" + ) + + # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, + # Reference: https://github.com/pytorch/pytorch/issues/112577. + if query_states.device.type == "cuda" and attention_mask is not None: + query_states = query_states.contiguous() + key_states = key_states.contiguous() + value_states = value_states.contiguous() + + attn_output = torch.nn.functional.scaled_dot_product_attention( + query_states, + key_states, + value_states, + attn_mask=attention_mask, + dropout_p=self.attention_dropout if self.training else 0.0, + # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1. + is_causal=self.is_causal and attention_mask is None and q_len > 1, + ) + + attn_output = attn_output.transpose(1, 2).contiguous() + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size) + + attn_output = self.o_proj(attn_output) + + return attn_output, None, past_key_value + + +QWEN2_ATTENTION_CLASSES = { + "eager": Qwen2Attention, + "flash_attention_2": Qwen2FlashAttention2, + "sdpa": Qwen2SdpaAttention, +} + + +class Qwen2DecoderLayer(nn.Module): + def __init__(self, config: Qwen2Config, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + + if config.use_sliding_window and config._attn_implementation != "flash_attention_2": + logger.warning_once( + f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; " + "unexpected results may be encountered." + ) + self.self_attn = QWEN2_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx) + + self.mlp = Qwen2MLP(config) + self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Tuple[torch.Tensor]] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + **kwargs, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + if "padding_mask" in kwargs: + warnings.warn( + "Passing `padding_mask` is deprecated and will be removed in v4.37. " + "Please make sure use `attention_mask` instead.`" + ) + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`, *optional*): attention mask of size + `(batch, sequence_length)` where padding elements are indicated by 0. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + """ + + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if use_cache: + outputs += (present_key_value,) + + return outputs + + +QWEN2_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`Qwen2Config`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +@add_start_docstrings( + "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.", + QWEN2_START_DOCSTRING, +) +class Qwen2PreTrainedModel(PreTrainedModel): + config_class = Qwen2Config + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["Qwen2DecoderLayer"] + _skip_keys_device_placement = "past_key_values" + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_cache_class = True + + def _init_weights(self, module): + std = self.config.initializer_range + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + +QWEN2_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. + + [What are position IDs?](../glossary#position-ids) + past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): + Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` + returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. + + Two formats are allowed: + - a [`~cache_utils.Cache`] instance; + - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy + cache format. + + The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the + legacy cache format will be returned. + + If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't + have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` + of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert `input_ids` indices into associated vectors than the + model's internal embedding lookup matrix. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. +""" + + +@add_start_docstrings( + "The bare Qwen2 Model outputting raw hidden-states without any specific head on top.", + QWEN2_START_DOCSTRING, +) +class Qwen2Model(Qwen2PreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`Qwen2DecoderLayer`] + + Args: + config: Qwen2Config + """ + + def __init__(self, config: Qwen2Config): + super().__init__(config) + self.padding_idx = config.pad_token_id + self.vocab_size = config.vocab_size + + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) + self.layers = nn.ModuleList( + [Qwen2DecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self._attn_implementation = config._attn_implementation + self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + self.gradient_checkpointing = False + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # retrieve input_ids and inputs_embeds + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time") + elif input_ids is not None: + batch_size, seq_length = input_ids.shape + elif inputs_embeds is not None: + batch_size, seq_length, _ = inputs_embeds.shape + else: + raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds") + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + past_key_values_length = 0 + + if use_cache: + use_legacy_cache = not isinstance(past_key_values, Cache) + if use_legacy_cache: + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + past_key_values_length = past_key_values.get_usable_length(seq_length) + + if position_ids is None: + device = input_ids.device if input_ids is not None else inputs_embeds.device + position_ids = torch.arange( + past_key_values_length, seq_length + past_key_values_length, dtype=torch.long, device=device + ) + position_ids = position_ids.unsqueeze(0).view(-1, seq_length) + else: + position_ids = position_ids.view(-1, seq_length).long() + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + if attention_mask is not None and self._attn_implementation == "flash_attention_2" and use_cache: + is_padding_right = attention_mask[:, -1].sum().item() != batch_size + if is_padding_right: + raise ValueError( + "You are attempting to perform batched generation with padding_side='right'" + " this may lead to unexpected behaviour for Flash Attention version of Qwen2. Make sure to " + " call `tokenizer.padding_side = 'left'` before tokenizing the input. " + ) + + if self._attn_implementation == "flash_attention_2": + # 2d mask is passed through the layers + attention_mask = attention_mask if (attention_mask is not None and 0 in attention_mask) else None + elif self._attn_implementation == "sdpa" and not output_attentions: + # output_attentions=True can not be supported when using SDPA, and we fall back on + # the manual implementation that requires a 4D causal mask in all cases. + attention_mask = _prepare_4d_causal_attention_mask_for_sdpa( + attention_mask, + (batch_size, seq_length), + inputs_embeds, + past_key_values_length, + ) + else: + # 4d mask is passed through the layers + attention_mask = _prepare_4d_causal_attention_mask( + attention_mask, + (batch_size, seq_length), + inputs_embeds, + past_key_values_length, + sliding_window=self.config.sliding_window, + ) + + hidden_states = inputs_embeds + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = None + + for decoder_layer in self.layers: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + decoder_layer.__call__, + hidden_states, + attention_mask, + position_ids, + past_key_values, + output_attentions, + use_cache, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + use_cache=use_cache, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache = layer_outputs[2 if output_attentions else 1] + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = None + if use_cache: + next_cache = next_decoder_cache.to_legacy_cache() if use_legacy_cache else next_decoder_cache + + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + +class Qwen2ForCausalLM(Qwen2PreTrainedModel): + _tied_weights_keys = ["lm_head.weight"] + + def __init__(self, config): + super().__init__(config) + self.model = Qwen2Model(config) + self.vocab_size = config.vocab_size + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + Args: + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, Qwen2ForCausalLM + + >>> model = Qwen2ForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS) + >>> tokenizer = AutoTokenizer.from_pretrained(PATH_TO_CONVERTED_TOKENIZER) + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + outputs = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + hidden_states = outputs[0] + logits = self.lm_head(hidden_states) + logits = logits.float() + + loss = None + if labels is not None: + # Shift so that tokens < n predict n + shift_logits = logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + # Flatten the tokens + loss_fct = CrossEntropyLoss() + shift_logits = shift_logits.view(-1, self.config.vocab_size) + shift_labels = shift_labels.view(-1) + # Enable model parallelism + shift_labels = shift_labels.to(shift_logits.device) + loss = loss_fct(shift_logits, shift_labels) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + def prepare_inputs_for_generation( + self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs + ): + # Omit tokens covered by past_key_values + if past_key_values is not None: + if isinstance(past_key_values, Cache): + cache_length = past_key_values.get_seq_length() + past_length = past_key_values.seen_tokens + max_cache_length = past_key_values.get_max_length() + else: + cache_length = past_length = past_key_values[0][0].shape[2] + max_cache_length = None + + # Keep only the unprocessed tokens: + # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where + # some of the inputs are exclusively passed as part of the cache (e.g. when passing input_embeds as + # input) + if attention_mask is not None and attention_mask.shape[1] > input_ids.shape[1]: + input_ids = input_ids[:, -(attention_mask.shape[1] - past_length) :] + # 2 - If the past_length is smaller than input_ids', then input_ids holds all input tokens. We can discard + # input_ids based on the past_length. + elif past_length < input_ids.shape[1]: + input_ids = input_ids[:, past_length:] + # 3 - Otherwise (past_length >= input_ids.shape[1]), let's assume input_ids only has unprocessed tokens. + + # If we are about to go beyond the maximum cache length, we need to crop the input attention mask. + if ( + max_cache_length is not None + and attention_mask is not None + and cache_length + input_ids.shape[1] > max_cache_length + ): + attention_mask = attention_mask[:, -max_cache_length:] + + position_ids = kwargs.get("position_ids", None) + if attention_mask is not None and position_ids is None: + # create position_ids on the fly for batch generation + position_ids = attention_mask.long().cumsum(-1) - 1 + position_ids.masked_fill_(attention_mask == 0, 1) + if past_key_values: + position_ids = position_ids[:, -input_ids.shape[1] :] + + # if `inputs_embeds` are passed, we only want to use them in the 1st generation step + if inputs_embeds is not None and past_key_values is None: + model_inputs = {"inputs_embeds": inputs_embeds} + else: + model_inputs = {"input_ids": input_ids} + + model_inputs.update( + { + "position_ids": position_ids, + "past_key_values": past_key_values, + "use_cache": kwargs.get("use_cache"), + "attention_mask": attention_mask, + } + ) + return model_inputs + + @staticmethod + def _reorder_cache(past_key_values, beam_idx): + reordered_past = () + for layer_past in past_key_values: + reordered_past += ( + tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past), + ) + return reordered_past + + +@add_start_docstrings( + """ + The Qwen2 Model transformer with a sequence classification head on top (linear layer). + + [`Qwen2ForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-2) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """, + QWEN2_START_DOCSTRING, +) +class Qwen2ForSequenceClassification(Qwen2PreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = Qwen2Model(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @add_start_docstrings_to_model_forward(QWEN2_INPUTS_DOCSTRING) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, SequenceClassifierOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + transformer_outputs = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = transformer_outputs[0] + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") + if self.config.pad_token_id is None: + sequence_lengths = -1 + else: + if input_ids is not None: + # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility + sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 + sequence_lengths = sequence_lengths % input_ids.shape[-1] + sequence_lengths = sequence_lengths.to(logits.device) + else: + sequence_lengths = -1 + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] + + loss = None + if labels is not None: + labels = labels.to(logits.device) + if self.config.problem_type is None: + if self.num_labels == 1: + self.config.problem_type = "regression" + elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int): + self.config.problem_type = "single_label_classification" + else: + self.config.problem_type = "multi_label_classification" + + if self.config.problem_type == "regression": + loss_fct = MSELoss() + if self.num_labels == 1: + loss = loss_fct(pooled_logits.squeeze(), labels.squeeze()) + else: + loss = loss_fct(pooled_logits, labels) + elif self.config.problem_type == "single_label_classification": + loss_fct = CrossEntropyLoss() + loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1)) + elif self.config.problem_type == "multi_label_classification": + loss_fct = BCEWithLogitsLoss() + loss = loss_fct(pooled_logits, labels) + if not return_dict: + output = (pooled_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) diff --git a/src/transformers/models/qwen2/tokenization_qwen2.py b/src/transformers/models/qwen2/tokenization_qwen2.py new file mode 100644 index 0000000000..fe8e5ded83 --- /dev/null +++ b/src/transformers/models/qwen2/tokenization_qwen2.py @@ -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) diff --git a/src/transformers/models/qwen2/tokenization_qwen2_fast.py b/src/transformers/models/qwen2/tokenization_qwen2_fast.py new file mode 100644 index 0000000000..178af4e62f --- /dev/null +++ b/src/transformers/models/qwen2/tokenization_qwen2_fast.py @@ -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) diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index 4d89b2942f..da544a0beb 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -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"] diff --git a/src/transformers/utils/dummy_tokenizers_objects.py b/src/transformers/utils/dummy_tokenizers_objects.py index b8cc21303a..863cb3ad03 100644 --- a/src/transformers/utils/dummy_tokenizers_objects.py +++ b/src/transformers/utils/dummy_tokenizers_objects.py @@ -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"] diff --git a/tests/models/qwen2/__init__.py b/tests/models/qwen2/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/models/qwen2/test_modeling_qwen2.py b/tests/models/qwen2/test_modeling_qwen2.py new file mode 100644 index 0000000000..587312bfa2 --- /dev/null +++ b/tests/models/qwen2/test_modeling_qwen2.py @@ -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() diff --git a/tests/models/qwen2/test_tokenization_qwen2.py b/tests/models/qwen2/test_tokenization_qwen2.py new file mode 100644 index 0000000000..565520367f --- /dev/null +++ b/tests/models/qwen2/test_tokenization_qwen2.py @@ -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, + ) diff --git a/utils/not_doctested.txt b/utils/not_doctested.txt index 611c515b82..de64dfa2ce 100644 --- a/utils/not_doctested.txt +++ b/utils/not_doctested.txt @@ -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