diff --git a/docs/source/de/index.md b/docs/source/de/index.md index 5ddabb4e7382e1..864ec697199dc5 100644 --- a/docs/source/de/index.md +++ b/docs/source/de/index.md @@ -99,6 +99,7 @@ Die Bibliothek enthält derzeit JAX-, PyTorch- und TensorFlow-Implementierungen, 1. **[FLAVA](model_doc/flava)** (from Facebook AI) released with the paper [FLAVA: A Foundational Language And Vision Alignment Model](https://arxiv.org/abs/2112.04482) by Amanpreet Singh, Ronghang Hu, Vedanuj Goswami, Guillaume Couairon, Wojciech Galuba, Marcus Rohrbach, and Douwe Kiela. 1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon. 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. +1. **[GLM](model_doc/glm)** (von THU/ZhipuAI) veröffentlicht mit dem Paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) von Team GLM, Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou und Zihan Wang. 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. 1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. 1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. @@ -209,126 +210,127 @@ Flax), PyTorch, und/oder TensorFlow haben. | Model | Tokenizer slow | Tokenizer fast | PyTorch support | TensorFlow support | Flax Support | |:---------------------------:|:--------------:|:--------------:|:---------------:|:------------------:|:------------:| -| ALBERT | ✅ | ✅ | ✅ | ✅ | ✅ | -| BART | ✅ | ✅ | ✅ | ✅ | ✅ | -| BEiT | ❌ | ❌ | ✅ | ❌ | ✅ | -| BERT | ✅ | ✅ | ✅ | ✅ | ✅ | -| Bert Generation | ✅ | ❌ | ✅ | ❌ | ❌ | -| BigBird | ✅ | ✅ | ✅ | ❌ | ✅ | -| BigBird-Pegasus | ❌ | ❌ | ✅ | ❌ | ❌ | -| Blenderbot | ✅ | ✅ | ✅ | ✅ | ✅ | -| BlenderbotSmall | ✅ | ✅ | ✅ | ✅ | ✅ | -| BLOOM | ❌ | ✅ | ✅ | ❌ | ✅ | -| CamemBERT | ✅ | ✅ | ✅ | ✅ | ❌ | -| CANINE | ✅ | ❌ | ✅ | ❌ | ❌ | -| CLIP | ✅ | ✅ | ✅ | ✅ | ✅ | -| CodeGen | ✅ | ✅ | ✅ | ❌ | ❌ | -| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ | -| ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ | -| CTRL | ✅ | ❌ | ✅ | ✅ | ❌ | -| CvT | ❌ | ❌ | ✅ | ❌ | ❌ | -| Data2VecAudio | ❌ | ❌ | ✅ | ❌ | ❌ | -| Data2VecText | ❌ | ❌ | ✅ | ❌ | ❌ | -| Data2VecVision | ❌ | ❌ | ✅ | ✅ | ❌ | -| DeBERTa | ✅ | ✅ | ✅ | ✅ | ❌ | -| DeBERTa-v2 | ✅ | ✅ | ✅ | ✅ | ❌ | -| Decision Transformer | ❌ | ❌ | ✅ | ❌ | ❌ | -| DeiT | ❌ | ❌ | ✅ | ✅ | ❌ | -| DETR | ❌ | ❌ | ✅ | ❌ | ❌ | -| DistilBERT | ✅ | ✅ | ✅ | ✅ | ✅ | -| DPR | ✅ | ✅ | ✅ | ✅ | ❌ | -| DPT | ❌ | ❌ | ✅ | ❌ | ❌ | -| ELECTRA | ✅ | ✅ | ✅ | ✅ | ✅ | -| Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ | -| FairSeq Machine-Translation | ✅ | ❌ | ✅ | ❌ | ❌ | -| FlauBERT | ✅ | ❌ | ✅ | ✅ | ❌ | -| FLAVA | ❌ | ❌ | ✅ | ❌ | ❌ | -| FNet | ✅ | ✅ | ✅ | ❌ | ❌ | -| Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | -| GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | -| GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | -| GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | -| GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | -| GroupViT | ❌ | ❌ | ✅ | ❌ | ❌ | -| Hubert | ❌ | ❌ | ✅ | ✅ | ❌ | -| I-BERT | ❌ | ❌ | ✅ | ❌ | ❌ | -| ImageGPT | ❌ | ❌ | ✅ | ❌ | ❌ | -| LayoutLM | ✅ | ✅ | ✅ | ✅ | ❌ | -| LayoutLMv2 | ✅ | ✅ | ✅ | ❌ | ❌ | -| LayoutLMv3 | ✅ | ✅ | ✅ | ❌ | ❌ | -| LED | ✅ | ✅ | ✅ | ✅ | ❌ | -| LeViT | ❌ | ❌ | ✅ | ❌ | ❌ | -| Longformer | ✅ | ✅ | ✅ | ✅ | ❌ | -| LongT5 | ❌ | ❌ | ✅ | ❌ | ✅ | -| LUKE | ✅ | ❌ | ✅ | ❌ | ❌ | -| LXMERT | ✅ | ✅ | ✅ | ✅ | ❌ | -| M-CTC-T | ❌ | ❌ | ✅ | ❌ | ❌ | -| M2M100 | ✅ | ❌ | ✅ | ❌ | ❌ | -| Marian | ✅ | ❌ | ✅ | ✅ | ✅ | -| MaskFormer | ❌ | ❌ | ✅ | ❌ | ❌ | -| mBART | ✅ | ✅ | ✅ | ✅ | ✅ | -| Megatron-BERT | ❌ | ❌ | ✅ | ❌ | ❌ | -| MobileBERT | ✅ | ✅ | ✅ | ✅ | ❌ | -| MobileViT | ❌ | ❌ | ✅ | ❌ | ❌ | -| MPNet | ✅ | ✅ | ✅ | ✅ | ❌ | -| MT5 | ✅ | ✅ | ✅ | ✅ | ✅ | -| MVP | ✅ | ✅ | ✅ | ❌ | ❌ | -| Nezha | ❌ | ❌ | ✅ | ❌ | ❌ | -| Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ | -| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ | -| OpenAI GPT-2 | ✅ | ✅ | ✅ | ✅ | ✅ | -| OPT | ❌ | ❌ | ✅ | ✅ | ✅ | -| OWL-ViT | ❌ | ❌ | ✅ | ❌ | ❌ | -| Pegasus | ✅ | ✅ | ✅ | ✅ | ✅ | -| Perceiver | ✅ | ❌ | ✅ | ❌ | ❌ | -| PLBart | ✅ | ❌ | ✅ | ❌ | ❌ | -| PoolFormer | ❌ | ❌ | ✅ | ❌ | ❌ | -| ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ | -| QDQBert | ❌ | ❌ | ✅ | ❌ | ❌ | -| RAG | ✅ | ❌ | ✅ | ✅ | ❌ | -| REALM | ✅ | ✅ | ✅ | ❌ | ❌ | -| Reformer | ✅ | ✅ | ✅ | ❌ | ❌ | -| RegNet | ❌ | ❌ | ✅ | ✅ | ✅ | -| RemBERT | ✅ | ✅ | ✅ | ✅ | ❌ | -| ResNet | ❌ | ❌ | ✅ | ✅ | ✅ | -| RetriBERT | ✅ | ✅ | ✅ | ❌ | ❌ | -| RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ | -| RoFormer | ✅ | ✅ | ✅ | ✅ | ✅ | -| SegFormer | ❌ | ❌ | ✅ | ✅ | ❌ | -| SEW | ❌ | ❌ | ✅ | ❌ | ❌ | -| SEW-D | ❌ | ❌ | ✅ | ❌ | ❌ | -| Speech Encoder decoder | ❌ | ❌ | ✅ | ❌ | ✅ | -| Speech2Text | ✅ | ❌ | ✅ | ✅ | ❌ | -| Speech2Text2 | ✅ | ❌ | ❌ | ❌ | ❌ | -| Splinter | ✅ | ✅ | ✅ | ❌ | ❌ | -| SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ | -| Swin Transformer | ❌ | ❌ | ✅ | ✅ | ❌ | -| Swin Transformer V2 | ❌ | ❌ | ✅ | ❌ | ❌ | -| T5 | ✅ | ✅ | ✅ | ✅ | ✅ | -| TAPAS | ✅ | ❌ | ✅ | ✅ | ❌ | -| Trajectory Transformer | ❌ | ❌ | ✅ | ❌ | ❌ | -| Transformer-XL | ✅ | ❌ | ✅ | ✅ | ❌ | -| TrOCR | ❌ | ❌ | ✅ | ❌ | ❌ | -| UniSpeech | ❌ | ❌ | ✅ | ❌ | ❌ | -| UniSpeechSat | ❌ | ❌ | ✅ | ❌ | ❌ | -| VAN | ❌ | ❌ | ✅ | ❌ | ❌ | -| VideoMAE | ❌ | ❌ | ✅ | ❌ | ❌ | -| ViLT | ❌ | ❌ | ✅ | ❌ | ❌ | -| Vision Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ | -| VisionTextDualEncoder | ❌ | ❌ | ✅ | ❌ | ✅ | -| VisualBERT | ❌ | ❌ | ✅ | ❌ | ❌ | -| ViT | ❌ | ❌ | ✅ | ✅ | ✅ | -| ViTMAE | ❌ | ❌ | ✅ | ✅ | ❌ | -| Wav2Vec2 | ✅ | ❌ | ✅ | ✅ | ✅ | -| Wav2Vec2-Conformer | ❌ | ❌ | ✅ | ❌ | ❌ | -| WavLM | ❌ | ❌ | ✅ | ❌ | ❌ | -| XGLM | ✅ | ✅ | ✅ | ❌ | ✅ | -| XLM | ✅ | ❌ | ✅ | ✅ | ❌ | -| XLM-ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ | -| XLM-RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ | -| XLM-RoBERTa-XL | ❌ | ❌ | ✅ | ❌ | ❌ | -| XLNet | ✅ | ✅ | ✅ | ✅ | ❌ | -| YOLOS | ❌ | ❌ | ✅ | ❌ | ❌ | -| YOSO | ❌ | ❌ | ✅ | ❌ | ❌ | +| ALBERT | ✅ | ✅ | ✅ | ✅ | ✅ | +| BART | ✅ | ✅ | ✅ | ✅ | ✅ | +| BEiT | ❌ | ❌ | ✅ | ❌ | ✅ | +| BERT | ✅ | ✅ | ✅ | ✅ | ✅ | +| Bert Generation | ✅ | ❌ | ✅ | ❌ | ❌ | +| BigBird | ✅ | ✅ | ✅ | ❌ | ✅ | +| BigBird-Pegasus | ❌ | ❌ | ✅ | ❌ | ❌ | +| Blenderbot | ✅ | ✅ | ✅ | ✅ | ✅ | +| BlenderbotSmall | ✅ | ✅ | ✅ | ✅ | ✅ | +| BLOOM | ❌ | ✅ | ✅ | ❌ | ✅ | +| CamemBERT | ✅ | ✅ | ✅ | ✅ | ❌ | +| CANINE | ✅ | ❌ | ✅ | ❌ | ❌ | +| CLIP | ✅ | ✅ | ✅ | ✅ | ✅ | +| CodeGen | ✅ | ✅ | ✅ | ❌ | ❌ | +| ConvBERT | ✅ | ✅ | ✅ | ✅ | ❌ | +| ConvNeXT | ❌ | ❌ | ✅ | ✅ | ❌ | +| CTRL | ✅ | ❌ | ✅ | ✅ | ❌ | +| CvT | ❌ | ❌ | ✅ | ❌ | ❌ | +| Data2VecAudio | ❌ | ❌ | ✅ | ❌ | ❌ | +| Data2VecText | ❌ | ❌ | ✅ | ❌ | ❌ | +| Data2VecVision | ❌ | ❌ | ✅ | ✅ | ❌ | +| DeBERTa | ✅ | ✅ | ✅ | ✅ | ❌ | +| DeBERTa-v2 | ✅ | ✅ | ✅ | ✅ | ❌ | +| Decision Transformer | ❌ | ❌ | ✅ | ❌ | ❌ | +| DeiT | ❌ | ❌ | ✅ | ✅ | ❌ | +| DETR | ❌ | ❌ | ✅ | ❌ | ❌ | +| DistilBERT | ✅ | ✅ | ✅ | ✅ | ✅ | +| DPR | ✅ | ✅ | ✅ | ✅ | ❌ | +| DPT | ❌ | ❌ | ✅ | ❌ | ❌ | +| ELECTRA | ✅ | ✅ | ✅ | ✅ | ✅ | +| Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ | +| FairSeq Machine-Translation | ✅ | ❌ | ✅ | ❌ | ❌ | +| FlauBERT | ✅ | ❌ | ✅ | ✅ | ❌ | +| FLAVA | ❌ | ❌ | ✅ | ❌ | ❌ | +| FNet | ✅ | ✅ | ✅ | ❌ | ❌ | +| Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | +| GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | +| GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | +| GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | +| GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | +| GroupViT | ❌ | ❌ | ✅ | ❌ | ❌ | +| Hubert | ❌ | ❌ | ✅ | ✅ | ❌ | +| I-BERT | ❌ | ❌ | ✅ | ❌ | ❌ | +| ImageGPT | ❌ | ❌ | ✅ | ❌ | ❌ | +| LayoutLM | ✅ | ✅ | ✅ | ✅ | ❌ | +| LayoutLMv2 | ✅ | ✅ | ✅ | ❌ | ❌ | +| LayoutLMv3 | ✅ | ✅ | ✅ | ❌ | ❌ | +| LED | ✅ | ✅ | ✅ | ✅ | ❌ | +| LeViT | ❌ | ❌ | ✅ | ❌ | ❌ | +| Longformer | ✅ | ✅ | ✅ | ✅ | ❌ | +| LongT5 | ❌ | ❌ | ✅ | ❌ | ✅ | +| LUKE | ✅ | ❌ | ✅ | ❌ | ❌ | +| LXMERT | ✅ | ✅ | ✅ | ✅ | ❌ | +| M-CTC-T | ❌ | ❌ | ✅ | ❌ | ❌ | +| M2M100 | ✅ | ❌ | ✅ | ❌ | ❌ | +| Marian | ✅ | ❌ | ✅ | ✅ | ✅ | +| MaskFormer | ❌ | ❌ | ✅ | ❌ | ❌ | +| mBART | ✅ | ✅ | ✅ | ✅ | ✅ | +| Megatron-BERT | ❌ | ❌ | ✅ | ❌ | ❌ | +| MobileBERT | ✅ | ✅ | ✅ | ✅ | ❌ | +| MobileViT | ❌ | ❌ | ✅ | ❌ | ❌ | +| MPNet | ✅ | ✅ | ✅ | ✅ | ❌ | +| MT5 | ✅ | ✅ | ✅ | ✅ | ✅ | +| MVP | ✅ | ✅ | ✅ | ❌ | ❌ | +| Nezha | ❌ | ❌ | ✅ | ❌ | ❌ | +| Nyströmformer | ❌ | ❌ | ✅ | ❌ | ❌ | +| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ | +| OpenAI GPT-2 | ✅ | ✅ | ✅ | ✅ | ✅ | +| OPT | ❌ | ❌ | ✅ | ✅ | ✅ | +| OWL-ViT | ❌ | ❌ | ✅ | ❌ | ❌ | +| Pegasus | ✅ | ✅ | ✅ | ✅ | ✅ | +| Perceiver | ✅ | ❌ | ✅ | ❌ | ❌ | +| PLBart | ✅ | ❌ | ✅ | ❌ | ❌ | +| PoolFormer | ❌ | ❌ | ✅ | ❌ | ❌ | +| ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ | +| QDQBert | ❌ | ❌ | ✅ | ❌ | ❌ | +| RAG | ✅ | ❌ | ✅ | ✅ | ❌ | +| REALM | ✅ | ✅ | ✅ | ❌ | ❌ | +| Reformer | ✅ | ✅ | ✅ | ❌ | ❌ | +| RegNet | ❌ | ❌ | ✅ | ✅ | ✅ | +| RemBERT | ✅ | ✅ | ✅ | ✅ | ❌ | +| ResNet | ❌ | ❌ | ✅ | ✅ | ✅ | +| RetriBERT | ✅ | ✅ | ✅ | ❌ | ❌ | +| RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ | +| RoFormer | ✅ | ✅ | ✅ | ✅ | ✅ | +| SegFormer | ❌ | ❌ | ✅ | ✅ | ❌ | +| SEW | ❌ | ❌ | ✅ | ❌ | ❌ | +| SEW-D | ❌ | ❌ | ✅ | ❌ | ❌ | +| Speech Encoder decoder | ❌ | ❌ | ✅ | ❌ | ✅ | +| Speech2Text | ✅ | ❌ | ✅ | ✅ | ❌ | +| Speech2Text2 | ✅ | ❌ | ❌ | ❌ | ❌ | +| Splinter | ✅ | ✅ | ✅ | ❌ | ❌ | +| SqueezeBERT | ✅ | ✅ | ✅ | ❌ | ❌ | +| Swin Transformer | ❌ | ❌ | ✅ | ✅ | ❌ | +| Swin Transformer V2 | ❌ | ❌ | ✅ | ❌ | ❌ | +| T5 | ✅ | ✅ | ✅ | ✅ | ✅ | +| TAPAS | ✅ | ❌ | ✅ | ✅ | ❌ | +| Trajectory Transformer | ❌ | ❌ | ✅ | ❌ | ❌ | +| Transformer-XL | ✅ | ❌ | ✅ | ✅ | ❌ | +| TrOCR | ❌ | ❌ | ✅ | ❌ | ❌ | +| UniSpeech | ❌ | ❌ | ✅ | ❌ | ❌ | +| UniSpeechSat | ❌ | ❌ | ✅ | ❌ | ❌ | +| VAN | ❌ | ❌ | ✅ | ❌ | ❌ | +| VideoMAE | ❌ | ❌ | ✅ | ❌ | ❌ | +| ViLT | ❌ | ❌ | ✅ | ❌ | ❌ | +| Vision Encoder decoder | ❌ | ❌ | ✅ | ✅ | ✅ | +| VisionTextDualEncoder | ❌ | ❌ | ✅ | ❌ | ✅ | +| VisualBERT | ❌ | ❌ | ✅ | ❌ | ❌ | +| ViT | ❌ | ❌ | ✅ | ✅ | ✅ | +| ViTMAE | ❌ | ❌ | ✅ | ✅ | ❌ | +| Wav2Vec2 | ✅ | ❌ | ✅ | ✅ | ✅ | +| Wav2Vec2-Conformer | ❌ | ❌ | ✅ | ❌ | ❌ | +| WavLM | ❌ | ❌ | ✅ | ❌ | ❌ | +| XGLM | ✅ | ✅ | ✅ | ❌ | ✅ | +| XLM | ✅ | ❌ | ✅ | ✅ | ❌ | +| XLM-ProphetNet | ✅ | ❌ | ✅ | ❌ | ❌ | +| XLM-RoBERTa | ✅ | ✅ | ✅ | ✅ | ✅ | +| XLM-RoBERTa-XL | ❌ | ❌ | ✅ | ❌ | ❌ | +| XLNet | ✅ | ✅ | ✅ | ✅ | ❌ | +| YOLOS | ❌ | ❌ | ✅ | ❌ | ❌ | +| YOSO | ❌ | ❌ | ✅ | ❌ | ❌ | diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index dc88bbd45ab23e..6701f9978e7d0c 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -394,6 +394,8 @@ title: Gemma - local: model_doc/gemma2 title: Gemma2 + - local: model_doc/glm + title: GLM - local: model_doc/openai-gpt title: GPT - local: model_doc/gpt_neo diff --git a/docs/source/en/index.md b/docs/source/en/index.md index ac73d9ab70fc33..be9f9aacf6e440 100644 --- a/docs/source/en/index.md +++ b/docs/source/en/index.md @@ -150,6 +150,7 @@ Flax), PyTorch, and/or TensorFlow. | [Gemma](model_doc/gemma) | ✅ | ❌ | ✅ | | [Gemma2](model_doc/gemma2) | ✅ | ❌ | ❌ | | [GIT](model_doc/git) | ✅ | ❌ | ❌ | +| [GLM](model_doc/glm) | ✅ | ❌ | ❌ | | [GLPN](model_doc/glpn) | ✅ | ❌ | ❌ | | [GPT Neo](model_doc/gpt_neo) | ✅ | ❌ | ✅ | | [GPT NeoX](model_doc/gpt_neox) | ✅ | ❌ | ❌ | diff --git a/docs/source/en/model_doc/glm.md b/docs/source/en/model_doc/glm.md new file mode 100644 index 00000000000000..ac9bdc0e48ce60 --- /dev/null +++ b/docs/source/en/model_doc/glm.md @@ -0,0 +1,108 @@ + + +# GLM + +## Overview + +The GLM Model was proposed +in [ChatGLM: A Family of Large Language Models from GLM-130B to GLM-4 All Tools](https://arxiv.org/html/2406.12793v1) +by GLM Team, THUDM & ZhipuAI. + +The abstract from the paper is the following: + +*We introduce ChatGLM, an evolving family of large language models that we have been developing over time. This report +primarily focuses on the GLM-4 language series, which includes GLM-4, GLM-4-Air, and GLM-4-9B. They represent our most +capable models that are trained with all the insights and lessons gained from the preceding three generations of +ChatGLM. To date, the GLM-4 models are pre-trained on ten trillions of tokens mostly in Chinese and English, along with +a small set of corpus from 24 languages, and aligned primarily for Chinese and English usage. The high-quality alignment +is achieved via a multi-stage post-training process, which involves supervised fine-tuning and learning from human +feedback. Evaluations show that GLM-4 1) closely rivals or outperforms GPT-4 in terms of general metrics such as MMLU, +GSM8K, MATH, BBH, GPQA, and HumanEval, 2) gets close to GPT-4-Turbo in instruction following as measured by IFEval, 3) +matches GPT-4 Turbo (128K) and Claude 3 for long context tasks, and 4) outperforms GPT-4 in Chinese alignments as +measured by AlignBench. The GLM-4 All Tools model is further aligned to understand user intent and autonomously decide +when and which tool(s) to use—including web browser, Python interpreter, text-to-image model, and user-defined +functions—to effectively complete complex tasks. In practical applications, it matches and even surpasses GPT-4 All +Tools in tasks like accessing online information via web browsing and solving math problems using Python interpreter. +Over the course, we have open-sourced a series of models, including ChatGLM-6B (three generations), GLM-4-9B (128K, 1M), +GLM-4V-9B, WebGLM, and CodeGeeX, attracting over 10 million downloads on Hugging face in the year 2023 alone.* + +Tips: + +- This model was contributed by [THUDM](https://huggingface.co/THUDM). The most recent code can be + found [here](https://github.com/thudm/GLM-4). + + +## Usage tips + +`GLM-4` can be found on the [Huggingface Hub](https://huggingface.co/collections/THUDM/glm-4-665fcf188c414b03c2f7e3b7) + +In the following, we demonstrate how to use `glm-4-9b-chat` 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("THUDM/glm-4-9b-chat", device_map="auto") +>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-4-9b-chat") + +>>> 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] +``` + +## GLMConfig + +[[autodoc]] GLMConfig + +## GLMTokenizer + +[[autodoc]] GLMTokenizer + - save_vocabulary + +## GLMTokenizerFast + +[[autodoc]] GLMTokenizerFast + +## GLMModel + +[[autodoc]] GLMModel + - forward + +## GLMForCausalLM + +[[autodoc]] GLMForCausalLM + - forward + +## GLMForSequenceClassification + +[[autodoc]] GLMForSequenceClassification + - forward + +## GLMForTokenClassification + +[[autodoc]] GLMForTokenClassification + - forward diff --git a/docs/source/en/perf_infer_gpu_one.md b/docs/source/en/perf_infer_gpu_one.md index df1e64e36877aa..c1a8a5886c0377 100644 --- a/docs/source/en/perf_infer_gpu_one.md +++ b/docs/source/en/perf_infer_gpu_one.md @@ -46,6 +46,7 @@ FlashAttention-2 is currently supported for the following architectures: * [DistilBert](https://huggingface.co/docs/transformers/model_doc/distilbert#transformers.DistilBertModel) * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel) * [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model) +* [GLM](https://huggingface.co/docs/transformers/model_doc/glm#transformers.GLMModel) * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2) * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel) * [GPTNeo](https://huggingface.co/docs/transformers/model_doc/gpt_neo#transformers.GPTNeoModel) @@ -211,6 +212,7 @@ For now, Transformers supports SDPA inference and training for the following arc * [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel) * [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel) * [Gemma2](https://huggingface.co/docs/transformers/model_doc/gemma2#transformers.Gemma2Model) +* [GLM](https://huggingface.co/docs/transformers/model_doc/glm#transformers.GLMModel) * [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2) * [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel) * [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel) diff --git a/docs/source/es/index.md b/docs/source/es/index.md index fe7d65d94e356c..2c666d6ccba973 100644 --- a/docs/source/es/index.md +++ b/docs/source/es/index.md @@ -90,6 +90,7 @@ La biblioteca actualmente contiene implementaciones de JAX, PyTorch y TensorFlow 1. **[FNet](model_doc/fnet)** (de Google Research) publicado con el paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) por James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon. 1. **[Funnel Transformer](model_doc/funnel)** (de CMU/Google Brain) publicado con el paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) por Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GLPN](model_doc/glpn)** (de KAIST) publicado con el paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) por Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (de OpenAI) publicado con el paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) por Alec Radford, Karthik Narasimhan, Tim Salimans y Ilya Sutskever. 1. **[GPT-2](model_doc/gpt2)** (de OpenAI) publicado con el paper [Language Models are Unsupervised Multitask Learners](https://openai.com/research/better-language-models/) por Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei y Ilya Sutskever. 1. **[GPT-J](model_doc/gptj)** (de EleutherAI) publicado con el repositorio [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) por Ben Wang y Aran Komatsuzaki. @@ -208,6 +209,7 @@ Flax), PyTorch y/o TensorFlow. | FNet | ✅ | ✅ | ✅ | ❌ | ❌ | | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | | Hubert | ❌ | ❌ | ✅ | ✅ | ❌ | diff --git a/docs/source/fr/index.md b/docs/source/fr/index.md index 51d35b76e877db..4e4179e2b0dc65 100644 --- a/docs/source/fr/index.md +++ b/docs/source/fr/index.md @@ -116,6 +116,7 @@ La documentation est organisée en 5 parties: 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GIT](model_doc/git)** (from Microsoft Research) released with the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) by Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang. 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. 1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. 1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach @@ -298,6 +299,7 @@ Le tableau ci-dessous représente la prise en charge actuelle dans la bibliothè | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GIT | ❌ | ❌ | ✅ | ❌ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | | GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ | diff --git a/docs/source/it/index.md b/docs/source/it/index.md index 76cdc0ad246104..e31e02da48584c 100644 --- a/docs/source/it/index.md +++ b/docs/source/it/index.md @@ -97,6 +97,7 @@ La libreria attualmente contiene implementazioni in JAX, PyTorch e TensorFlow, p 1. **[FNet](model_doc/fnet)** (da Google Research) rilasciato con il paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) da James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon. 1. **[Funnel Transformer](model_doc/funnel)** (da CMU/Google Brain) rilasciato con il paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) da Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GLPN](model_doc/glpn)** (da KAIST) rilasciato con il paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) da Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (da OpenAI) rilasciato con il paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) da Alec Radford, Karthik Narasimhan, Tim Salimans e Ilya Sutskever. 1. **[GPT-2](model_doc/gpt2)** (da OpenAI) rilasciato con il paper [Language Models are Unsupervised Multitask Learners](https://openai.com/research/better-language-models/) da Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei e Ilya Sutskever. 1. **[GPT-J](model_doc/gptj)** (da EleutherAI) rilasciato nel repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) da Ben Wang e Aran Komatsuzaki. @@ -222,6 +223,7 @@ tokenizer (chiamato "slow"). Un tokenizer "fast" supportato dalla libreria 🤗 | FNet | ✅ | ✅ | ✅ | ❌ | ❌ | | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | | GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | diff --git a/docs/source/ja/index.md b/docs/source/ja/index.md index c3baa0888fc887..2ccbe78a1a251b 100644 --- a/docs/source/ja/index.md +++ b/docs/source/ja/index.md @@ -112,6 +112,7 @@ rendered properly in your Markdown viewer. 1. **[Funnel Transformer](https://huggingface.co/docs/transformers/model_doc/funnel)** (CMU/Google Brain から) Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le から公開された研究論文: [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) 1. **[GIT](https://huggingface.co/docs/transformers/main/model_doc/git)** (Microsoft Research から) Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang. から公開された研究論文 [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) 1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (KAIST から) Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim から公開された研究論文: [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) +1. **[GLM](model_doc/glm)** (THU/ZhipuAIより) は、チームGLM（Aohan Zeng、Bin Xu、Bowen Wang、Chenhui Zhang、Da Yin、Diego Rojas、Guanyu Feng、Hanlin Zhao、Hanyu Lai、Hao Yu、Hongning Wang、Jiadai Sun、Jiajie Zhang、Jiale Cheng、Jiayi Gui、Jie Tang、Jing Zhang、Juanzi Li、Lei Zhao、Lindong Wu、Lucen Zhong、Mingdao Liu、Minlie Huang、Peng Zhang、Qinkai Zheng、Rui Lu、Shuaiqi Duan、Shudan Zhang、Shulin Cao、Shuxun Yang、Weng Lam Tam、Wenyi Zhao、Xiao Liu、Xiao Xia、Xiaohan Zhang、Xiaotao Gu、Xin Lv、Xinghan Liu、Xinyi Liu、Xinyue Yang、Xixuan Song、Xunkai Zhang、Yifan An、Yifan Xu、Yilin Niu、Yuantao Yang、Yueyan Li、Yushi Bai、Yuxiao Dong、Zehan Qi、Zhaoyu Wang、Zhen Yang、Zhengxiao Du、Zhenyu Hou、Zihan Wang）が執筆した論文 [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) とともに発表されました。 1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (OpenAI から) Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever から公開された研究論文: [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) 1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (EleutherAI から) Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy から公開されたレポジトリー : [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) 1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (EleutherAI から) Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach から公開された研究論文: [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) @@ -288,6 +289,7 @@ rendered properly in your Markdown viewer. | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GIT | ❌ | ❌ | ✅ | ❌ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | | GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ | diff --git a/docs/source/ko/index.md b/docs/source/ko/index.md index 0726085c5b3ae7..3c3a611050af20 100644 --- a/docs/source/ko/index.md +++ b/docs/source/ko/index.md @@ -104,6 +104,7 @@ rendered properly in your Markdown viewer. 1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon. 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. 1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. 1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach @@ -264,6 +265,7 @@ rendered properly in your Markdown viewer. | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | | GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ | | GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | diff --git a/docs/source/ms/index.md b/docs/source/ms/index.md index f51c43c9bd01a6..407a14b57af82c 100644 --- a/docs/source/ms/index.md +++ b/docs/source/ms/index.md @@ -125,6 +125,7 @@ Dokumentasi disusun kepada lima bahagian: 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GIT](model_doc/git)** (from Microsoft Research) released with the paper [GIT: A Generative Image-to-text Transformer for Vision and Language](https://arxiv.org/abs/2205.14100) by Jianfeng Wang, Zhengyuan Yang, Xiaowei Hu, Linjie Li, Kevin Lin, Zhe Gan, Zicheng Liu, Ce Liu, Lijuan Wang. 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. 1. **[GPT Neo](model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy. 1. **[GPT NeoX](model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach @@ -335,6 +336,7 @@ Flax), PyTorch, dan/atau TensorFlow. | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GIT | ❌ | ❌ | ✅ | ❌ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT NeoX | ❌ | ✅ | ✅ | ❌ | ❌ | | GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ | diff --git a/docs/source/pt/index.md b/docs/source/pt/index.md index 18dbcbc06b8048..a18f815acace99 100644 --- a/docs/source/pt/index.md +++ b/docs/source/pt/index.md @@ -103,6 +103,7 @@ Atualmente a biblioteca contém implementações do PyTorch, TensorFlow e JAX, p 1. **[FNet](model_doc/fnet)** (from Google Research) released with the paper [FNet: Mixing Tokens with Fourier Transforms](https://arxiv.org/abs/2105.03824) by James Lee-Thorp, Joshua Ainslie, Ilya Eckstein, Santiago Ontanon. 1. **[Funnel Transformer](model_doc/funnel)** (from CMU/Google Brain) released with the paper [Funnel-Transformer: Filtering out Sequential Redundancy for Efficient Language Processing](https://arxiv.org/abs/2006.03236) by Zihang Dai, Guokun Lai, Yiming Yang, Quoc V. Le. 1. **[GLPN](model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim. +1. **[GLM](model_doc/glm)** (from THU/ZhipuAI) released with the paper [GLM: General Language Model Pretraining with Autoregressive Blank Infilling](https://arxiv.org/abs/2103.10360) by Team GLM, including Aohan Zeng, Bin Xu, Bowen Wang, Chenhui Zhang, Da Yin, Diego Rojas, Guanyu Feng, Hanlin Zhao, Hanyu Lai, Hao Yu, Hongning Wang, Jiadai Sun, Jiajie Zhang, Jiale Cheng, Jiayi Gui, Jie Tang, Jing Zhang, Juanzi Li, Lei Zhao, Lindong Wu, Lucen Zhong, Mingdao Liu, Minlie Huang, Peng Zhang, Qinkai Zheng, Rui Lu, Shuaiqi Duan, Shudan Zhang, Shulin Cao, Shuxun Yang, Weng Lam Tam, Wenyi Zhao, Xiao Liu, Xiao Xia, Xiaohan Zhang, Xiaotao Gu, Xin Lv, Xinghan Liu, Xinyi Liu, Xinyue Yang, Xixuan Song, Xunkai Zhang, Yifan An, Yifan Xu, Yilin Niu, Yuantao Yang, Yueyan Li, Yushi Bai, Yuxiao Dong, Zehan Qi, Zhaoyu Wang, Zhen Yang, Zhengxiao Du, Zhenyu Hou, and Zihan Wang. 1. **[GPT](model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://openai.com/research/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever. 1. **[GPT-2](model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://openai.com/research/better-language-models/) by Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei and Ilya Sutskever. 1. **[GPT-J](model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki. @@ -147,7 +148,7 @@ Atualmente a biblioteca contém implementações do PyTorch, TensorFlow e JAX, p 1. **[RoFormer](model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/abs/2104.09864) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu. 1. **[SegFormer](model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo. 1. **[SEW](model_doc/sew)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. -1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [Performance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. +1. **[SEW-D](model_doc/sew_d)** (from ASAPP) released with the paper [PFerformance-Efficiency Trade-offs in Unsupervised Pre-training for Speech Recognition](https://arxiv.org/abs/2109.06870) by Felix Wu, Kwangyoun Kim, Jing Pan, Kyu Han, Kilian Q. Weinberger, Yoav Artzi. 1. **[SpeechToTextTransformer](model_doc/speech_to_text)** (from Facebook), released together with the paper [fairseq S2T: Fast Speech-to-Text Modeling with fairseq](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Dmytro Okhonko, Juan Pino. 1. **[SpeechToTextTransformer2](model_doc/speech_to_text_2)** (from Facebook), released together with the paper [Large-Scale Self- and Semi-Supervised Learning for Speech Translation](https://arxiv.org/abs/2104.06678) by Changhan Wang, Anne Wu, Juan Pino, Alexei Baevski, Michael Auli, Alexis Conneau. 1. **[Splinter](model_doc/splinter)** (from Tel Aviv University), released together with the paper [Few-Shot Question Answering by Pretraining Span Selection](https://arxiv.org/abs/2101.00438) by Ori Ram, Yuval Kirstain, Jonathan Berant, Amir Globerson, Omer Levy. @@ -223,6 +224,7 @@ disso, são diferenciados pelo suporte em diferentes frameworks: JAX (por meio d | FNet | ✅ | ✅ | ✅ | ❌ | ❌ | | Funnel Transformer | ✅ | ✅ | ✅ | ✅ | ❌ | | GLPN | ❌ | ❌ | ✅ | ❌ | ❌ | +| GLM | ✅ | ✅ | ✅ | ❌ | ❌ | | GPT Neo | ❌ | ❌ | ✅ | ❌ | ✅ | | GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ | | Hubert | ❌ | ❌ | ✅ | ✅ | ❌ | diff --git a/docs/source/te/index.md b/docs/source/te/index.md index 3e23f8f5eb1392..28b76ee2387515 100644 --- a/docs/source/te/index.md +++ b/docs/source/te/index.md @@ -139,6 +139,7 @@ rendered properly in your Markdown viewer. | [Funnel Transformer](model_doc/funnel) | ✅ | ✅ | ❌ | | [GIT](model_doc/git) | ✅ | ❌ | ❌ | | [GLPN](model_doc/glpn) | ✅ | ❌ | ❌ | +| [GLM](../en/model_doc/glm) | ✅ | ❌ | ❌ | | [GPT Neo](model_doc/gpt_neo) | ✅ | ❌ | ✅ | | [GPT NeoX](model_doc/gpt_neox) | ✅ | ❌ | ❌ | | [GPT NeoX Japanese](model_doc/gpt_neox_japanese) | ✅ | ❌ | ❌ | diff --git a/docs/source/tr/index.md b/docs/source/tr/index.md index 1b2c665e169d80..c6254badf668d7 100644 --- a/docs/source/tr/index.md +++ b/docs/source/tr/index.md @@ -134,6 +134,7 @@ Aşağıdaki tablo, her bir model için kütüphanede yer alan mevcut desteği t | [Fuyu](model_doc/fuyu) | ✅ | ❌ | ❌ | | [GIT](model_doc/git) | ✅ | ❌ | ❌ | | [GLPN](model_doc/glpn) | ✅ | ❌ | ❌ | +| [GLM](../en/model_doc/glm) | ✅ | ❌ | ❌ | | [GPT Neo](model_doc/gpt_neo) | ✅ | ❌ | ✅ | | [GPT NeoX](model_doc/gpt_neox) | ✅ | ❌ | ❌ | | [GPT NeoX Japanese](model_doc/gpt_neox_japanese) | ✅ | ❌ | ❌ | diff --git a/docs/source/zh/index.md b/docs/source/zh/index.md index 3750e506b0ea04..230cb1ffaf2b0e 100644 --- a/docs/source/zh/index.md +++ b/docs/source/zh/index.md @@ -143,6 +143,7 @@ rendered properly in your Markdown viewer. | [Gemma](../en/model_doc/gemma) | ✅ | ❌ | ✅ | | [GIT](../en/model_doc/git) | ✅ | ❌ | ❌ | | [GLPN](../en/model_doc/glpn) | ✅ | ❌ | ❌ | +| [GLM](../en/model_doc/glm) | ✅ | ❌ | ❌ | | [GPT Neo](../en/model_doc/gpt_neo) | ✅ | ❌ | ✅ | | [GPT NeoX](../en/model_doc/gpt_neox) | ✅ | ❌ | ❌ | | [GPT NeoX Japanese](../en/model_doc/gpt_neox_japanese) | ✅ | ❌ | ❌ | diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index 1d36e7f8c74637..24068f0be79218 100755 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -450,6 +450,7 @@ "GitProcessor", "GitVisionConfig", ], + "models.glm": ["GLMConfig", "GLMTokenizer"], "models.glpn": ["GLPNConfig"], "models.gpt2": [ "GPT2Config", @@ -1045,6 +1046,7 @@ _import_structure["models.fnet"].append("FNetTokenizerFast") _import_structure["models.funnel"].append("FunnelTokenizerFast") _import_structure["models.gemma"].append("GemmaTokenizerFast") + _import_structure["models.glm"].append("GLMTokenizerFast") _import_structure["models.gpt2"].append("GPT2TokenizerFast") _import_structure["models.gpt_neox"].append("GPTNeoXTokenizerFast") _import_structure["models.gpt_neox_japanese"].append("GPTNeoXJapaneseTokenizer") @@ -2251,6 +2253,15 @@ "GitVisionModel", ] ) + _import_structure["models.glm"].extend( + [ + "GLMForCausalLM", + "GLMForSequenceClassification", + "GLMForTokenClassification", + "GLMModel", + "GLMPreTrainedModel", + ] + ) _import_structure["models.glpn"].extend( [ "GLPNForDepthEstimation", @@ -5188,6 +5199,10 @@ GitProcessor, GitVisionConfig, ) + from .models.glm import ( + GLMConfig, + GLMTokenizer, + ) from .models.glpn import GLPNConfig from .models.gpt2 import ( GPT2Config, @@ -5828,6 +5843,7 @@ from .models.fnet import FNetTokenizerFast from .models.funnel import FunnelTokenizerFast from .models.gemma import GemmaTokenizerFast + from .models.glm import GLMTokenizerFast from .models.gpt2 import GPT2TokenizerFast from .models.gpt_neox import GPTNeoXTokenizerFast from .models.gpt_neox_japanese import GPTNeoXJapaneseTokenizer @@ -6864,6 +6880,13 @@ GitPreTrainedModel, GitVisionModel, ) + from .models.glm import ( + GLMForCausalLM, + GLMForSequenceClassification, + GLMForTokenClassification, + GLMModel, + GLMPreTrainedModel, + ) from .models.glpn import ( GLPNForDepthEstimation, GLPNModel, diff --git a/src/transformers/convert_slow_tokenizer.py b/src/transformers/convert_slow_tokenizer.py index 2d0302d3f6bf02..667d82b10baaf2 100644 --- a/src/transformers/convert_slow_tokenizer.py +++ b/src/transformers/convert_slow_tokenizer.py @@ -1226,6 +1226,49 @@ def converted(self) -> Tokenizer: return tokenizer +class GLMConverter(Converter): + def converted(self, vocab: Dict[str, int] = None, merges: List[Tuple[str, str]] = None) -> Tokenizer: + if not vocab: + vocab = self.original_tokenizer.encoder + if not merges: + 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}{1,3}| ?[^\\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 BlenderbotConverter(Converter): def converted(self) -> Tokenizer: ot = self.original_tokenizer @@ -1529,6 +1572,7 @@ def converted(self) -> Tokenizer: "ElectraTokenizer": BertConverter, "FNetTokenizer": AlbertConverter, "FunnelTokenizer": FunnelConverter, + "GLMTokenizer": GLMConverter, "GPT2Tokenizer": GPT2Converter, "HerbertTokenizer": HerbertConverter, "LayoutLMTokenizer": BertConverter, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index f60f72a2361451..0bf0e7cebd7751 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -97,6 +97,7 @@ gemma, gemma2, git, + glm, glpn, gpt2, gpt_bigcode, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index ecd0a6674041bc..251e2538ceebaa 100644 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -113,6 +113,7 @@ ("gemma", "GemmaConfig"), ("gemma2", "Gemma2Config"), ("git", "GitConfig"), + ("glm", "GLMConfig"), ("glpn", "GLPNConfig"), ("gpt-sw3", "GPT2Config"), ("gpt2", "GPT2Config"), @@ -401,6 +402,7 @@ ("gemma", "Gemma"), ("gemma2", "Gemma2"), ("git", "GIT"), + ("glm", "GLM"), ("glpn", "GLPN"), ("gpt-sw3", "GPT-Sw3"), ("gpt2", "OpenAI GPT-2"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index 38086aa0f2e962..73bbe632e0247c 100644 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -110,6 +110,7 @@ ("gemma", "GemmaModel"), ("gemma2", "Gemma2Model"), ("git", "GitModel"), + ("glm", "GLMModel"), ("glpn", "GLPNModel"), ("gpt-sw3", "GPT2Model"), ("gpt2", "GPT2Model"), @@ -276,7 +277,6 @@ ("yoso", "YosoModel"), ] ) - MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict( [ # Model for pre-training mapping @@ -471,6 +471,7 @@ ("gemma", "GemmaForCausalLM"), ("gemma2", "Gemma2ForCausalLM"), ("git", "GitForCausalLM"), + ("glm", "GLMForCausalLM"), ("gpt-sw3", "GPT2LMHeadModel"), ("gpt2", "GPT2LMHeadModel"), ("gpt_bigcode", "GPTBigCodeForCausalLM"), @@ -887,6 +888,7 @@ ("funnel", "FunnelForSequenceClassification"), ("gemma", "GemmaForSequenceClassification"), ("gemma2", "Gemma2ForSequenceClassification"), + ("glm", "GLMForSequenceClassification"), ("gpt-sw3", "GPT2ForSequenceClassification"), ("gpt2", "GPT2ForSequenceClassification"), ("gpt_bigcode", "GPTBigCodeForSequenceClassification"), @@ -1071,6 +1073,7 @@ ("funnel", "FunnelForTokenClassification"), ("gemma", "GemmaForTokenClassification"), ("gemma2", "Gemma2ForTokenClassification"), + ("glm", "GLMForTokenClassification"), ("gpt-sw3", "GPT2ForTokenClassification"), ("gpt2", "GPT2ForTokenClassification"), ("gpt_bigcode", "GPTBigCodeForTokenClassification"), diff --git a/src/transformers/models/auto/tokenization_auto.py b/src/transformers/models/auto/tokenization_auto.py index b094f50b5e97ad..2c1be877c22cea 100644 --- a/src/transformers/models/auto/tokenization_auto.py +++ b/src/transformers/models/auto/tokenization_auto.py @@ -204,6 +204,7 @@ ), ), ("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)), + ("glm", ("GLMTokenizer", "GLMTokenizerFast" if is_tokenizers_available() else None)), ("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)), ("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)), ("gpt_bigcode", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)), diff --git a/src/transformers/models/glm/__init__.py b/src/transformers/models/glm/__init__.py new file mode 100644 index 00000000000000..46525053954a36 --- /dev/null +++ b/src/transformers/models/glm/__init__.py @@ -0,0 +1,83 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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_glm": ["GLMConfig"], + "tokenization_glm": ["GLMTokenizer"], +} + +try: + if not is_tokenizers_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["tokenization_glm_fast"] = ["GLMTokenizerFast"] + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["modeling_glm"] = [ + "GLMPreTrainedModel", + "GLMModel", + "GLMForCausalLM", + "GLMForSequenceClassification", + "GLMForTokenClassification", + ] + +if TYPE_CHECKING: + from .configuration_glm import GLMConfig + from .tokenization_glm import GLMTokenizer + + try: + if not is_tokenizers_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .tokenization_glm_fast import GLMTokenizerFast + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .modeling_glm import ( + GLMForCausalLM, + GLMForSequenceClassification, + GLMForTokenClassification, + GLMModel, + GLMPreTrainedModel, + ) + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) diff --git a/src/transformers/models/glm/configuration_glm.py b/src/transformers/models/glm/configuration_glm.py new file mode 100644 index 00000000000000..16d3b85d079246 --- /dev/null +++ b/src/transformers/models/glm/configuration_glm.py @@ -0,0 +1,138 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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. + +"""GLM model configuration""" + +from ...configuration_utils import PretrainedConfig +from ...utils import logging + + +logger = logging.get_logger(__name__) + + +class GLMConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`GLMModel`]. It is used to instantiate a GLM + model according to the specified arguments, defining the model architecture. Instantiating a configuration with the + defaults will yield a similar configuration to that of the + [THUDM/glm-4-9b-chat](https://huggingface.co/THUDM/glm-4-9b-chat). + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + Args: + num_hidden_layers (`int`, *optional*, defaults to 40): + Number of hidden layers in the Transformer decoder. + vocab_size (`int`, *optional*, defaults to 151552): + Vocabulary size of the Phi-3 model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`GLMModel`]. + hidden_size (`int`, *optional*, defaults to 4096): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 13696): + Dimension of the MLP representations. + kv_channels (`int`, *optional*, defaults to 128): + Defines the number of channels for the key and value tensors. + num_attention_heads (`int`, *optional*, defaults to 32): + Number of attention heads for each attention layer in the Transformer decoder. + max_position_embeddings (`int`, *optional*, defaults to 131072): + The maximum sequence length that this model might ever be used with. + hidden_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the hidden layer. + classifier_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for classifier. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio after computing the attention scores. + 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-05): + The epsilon value used for the RMSNorm. + add_qkv_bias (`bool`, *optional*, defaults to `True`): + Whether to add bias to the query, key, value tensors. + Whether to use multi query attention or not. + multi_query_attention (`bool`, *optional*, defaults to `False`): + Whether to use multi query attention or not. + multi_query_group_num (`int`, *optional*, defaults to 2): + The number of groups in the multi query attention + rope_theta (`float`, *optional*, defaults to 1.0): + The base period of the RoPE embeddings. + apply_query_key_layer_scaling (`bool`, *optional*, defaults to `True`): + Whether to apply layer scaling to query and key. + attention_softmax_in_fp32 (`bool`, *optional*, defaults to `True`): + Whether to use fp32 for softmax in attention. + Whether to use fp32 for residual connection. + fp32_residual_connection (`bool`, *optional*, defaults to `False`): + Whether to use fp32 for residual connection. + 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`. Whether to tie weight embeddings or not. + Example: + + ```python + >>> from transformers import GLMModel, GLMConfig + >>> configuration = GLMConfig.from_pretrained("THUDM/glm-4-9b-chat") + >>> model = GLMModel(configuration) + >>> configuration = model.config + ```""" + + model_type = "glm" + keys_to_ignore_at_inference = ["past_key_values"] + + def __init__( + self, + vocab_size=151552, + hidden_size=4096, + intermediate_size=13696, + kv_channels=128, + num_hidden_layers=40, + num_attention_heads=32, + max_position_embeddings=131072, + hidden_dropout=0.0, + classifier_dropout=None, + attention_dropout=0.0, + initializer_range=0.02, + rms_norm_eps=1e-5, + use_cache=True, + add_qkv_bias=True, + multi_query_attention=False, + multi_query_group_num=2, + rope_theta=1.0, + apply_query_key_layer_scaling=True, + attention_softmax_in_fp32=True, + fp32_residual_connection=False, + **kwargs, + ): + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.initializer_range = initializer_range + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.kv_channels = kv_channels + + self.add_qkv_bias = add_qkv_bias + self.hidden_dropout = hidden_dropout + self.classifier_dropout = classifier_dropout + self.attention_dropout = attention_dropout + self.rms_norm_eps = rms_norm_eps + self.multi_query_attention = multi_query_attention + self.multi_query_group_num = multi_query_group_num + self.rope_theta = rope_theta + self.apply_query_key_layer_scaling = apply_query_key_layer_scaling + self.attention_softmax_in_fp32 = attention_softmax_in_fp32 + self.fp32_residual_connection = fp32_residual_connection + self.use_cache = use_cache + + super().__init__(**kwargs) diff --git a/src/transformers/models/glm/modeling_glm.py b/src/transformers/models/glm/modeling_glm.py new file mode 100644 index 00000000000000..fcd58be392ed26 --- /dev/null +++ b/src/transformers/models/glm/modeling_glm.py @@ -0,0 +1,1532 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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. +"""PyTorch GLM model.""" + +import math +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 ...cache_utils import Cache, DynamicCache, StaticCache +from ...modeling_attn_mask_utils import AttentionMaskConverter +from ...modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +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_glm import GLMConfig + + +if is_flash_attn_2_available(): + from ...modeling_flash_attention_utils import _flash_attention_forward + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "THUDM/glm-4-9b-chat" +_CONFIG_FOR_DOC = "GLMConfig" + + +# Copied from transformers.models.llama.modeling_llama.LlamaRMSNorm with Llama->GLM +class GLMRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + GLMRMSNorm 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) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +class GLMRotaryEmbedding(nn.Module): + def __init__(self, dim, rope_theta=1, original_impl=False, device=None, dtype=None): + super().__init__() + inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2, device=device).to(dtype=dtype) / dim)) + self.register_buffer("inv_freq", inv_freq) + self.dim = dim + self.original_impl = original_impl + self.rope_theta = rope_theta + + def forward_impl( + self, + seq_len: int, + n_elem: int, + dtype: torch.dtype, + device: torch.device, + base: int = 10000, + ): + """Enhanced Transformer with Rotary Position Embedding. + Derived from: https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/labml_nn/ + transformers/rope/__init__.py. MIT License: + https://github.com/labmlai/annotated_deep_learning_paper_implementations/blob/master/license. + """ + # $\Theta = {\theta_i = 10000^{\frac{2(i-1)}{d}}, i \in [1, 2, ..., \frac{d}{2}]}$ + base = base * self.rope_theta + theta = 1.0 / (base ** (torch.arange(0, n_elem, 2, dtype=torch.float, device=device) / n_elem)) + + # Create position indexes `[0, 1, ..., seq_len - 1]` + seq_idx = torch.arange(seq_len, dtype=torch.float, device=device) + + # Calculate the product of position index and $\theta_i$ + idx_theta = torch.outer(seq_idx, theta).float() + + cache = torch.stack([torch.cos(idx_theta), torch.sin(idx_theta)], dim=-1).to(dtype=dtype) + return cache + + def forward(self, seq_len): + return self.forward_impl( + seq_len, + self.dim, + dtype=self.inv_freq.dtype, + device=self.inv_freq.device, + ) + + +def split_tensor_along_last_dim( + tensor: torch.Tensor, + num_partitions: int, + contiguous_split_chunks: bool = False, +) -> List[torch.Tensor]: + """Split a tensor along its last dimension. + + Arguments: + tensor: input tensor. + num_partitions: number of partitions to split the tensor + contiguous_split_chunks: If True, make each chunk contiguous + in memory. + + Returns: + A list of Tensors + """ + # Get the size and dimension. + last_dim = tensor.dim() - 1 + last_dim_size = tensor.size()[last_dim] // num_partitions + # Split. + tensor_list = torch.split(tensor, last_dim_size, dim=last_dim) + # Note: torch.split does not create contiguous tensors by default. + if contiguous_split_chunks: + return tuple(chunk.contiguous() for chunk in tensor_list) + + return tensor_list + + +class GLMSelfAttention(torch.nn.Module): + """Parallel self-attention layer abstract class. + + Self-attention layer takes input with size [s, b, h] + and returns output of the same size. + """ + + def __init__(self, config: GLMConfig, layer_number, device=None): + super(GLMSelfAttention, self).__init__() + self.layer_number = max(1, layer_number) + self.num_heads = config.num_attention_heads + self.projection_size = config.kv_channels * self.num_heads + self.hidden_size_per_attention_head = self.projection_size // self.num_heads + self.multi_query_group_num = config.multi_query_group_num + self.multi_query_attention = config.multi_query_attention + self.hidden_size = config.hidden_size + self.qkv_hidden_size = 3 * self.projection_size + self.add_qkv_bias = config.add_qkv_bias + + if self.multi_query_attention: + self.num_multi_query_groups_per_partition = self.multi_query_group_num + self.qkv_hidden_size = ( + self.projection_size + 2 * self.hidden_size_per_attention_head * self.multi_query_group_num + ) + self.query_key_value = nn.Linear( + self.hidden_size, + self.qkv_hidden_size, + bias=self.add_qkv_bias, + device=device, + ) + + self.core_attention = GLM_ATTENTION_CLASSES[config._attn_implementation](config, self.layer_number) + + # Output. + self.dense = nn.Linear( + self.projection_size, + self.hidden_size, + bias=False, + device=device, + ) + + def _allocate_memory(self, inference_max_sequence_len, batch_size, device=None, dtype=None): + if self.multi_query_attention: + num_attention_heads = self.num_multi_query_groups_per_partition + else: + num_attention_heads = self.num_heads + return torch.empty( + inference_max_sequence_len, + batch_size, + num_attention_heads, + self.hidden_size_per_attention_head, + dtype=dtype, + device=device, + ) + + def forward( + self, + hidden_states, + attention_mask, + rotary_pos_emb, + past_key_value=None, + use_cache=True, + ): + # hidden_states: [b, sq, h] + + # ================================================= + # Pre-allocate memory for key-values for inference. + # ================================================= + # ===================== + # Query, Key, and Value + # ===================== + + # Attention heads [b, sq, h] --> [b, sq, (np * 3 * hn)] + mixed_x_layer = self.query_key_value(hidden_states) + + if self.multi_query_attention: + (query_layer, key_layer, value_layer) = mixed_x_layer.split( + [ + self.num_heads * self.hidden_size_per_attention_head, + self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head, + self.num_multi_query_groups_per_partition * self.hidden_size_per_attention_head, + ], + dim=-1, + ) + query_layer = query_layer.view( + query_layer.size()[:-1] + + ( + self.num_heads, + self.hidden_size_per_attention_head, + ) + ) + key_layer = key_layer.view( + key_layer.size()[:-1] + + ( + self.num_multi_query_groups_per_partition, + self.hidden_size_per_attention_head, + ) + ) + value_layer = value_layer.view( + value_layer.size()[:-1] + + ( + self.num_multi_query_groups_per_partition, + self.hidden_size_per_attention_head, + ) + ) + else: + new_tensor_shape = mixed_x_layer.size()[:-1] + ( + self.num_heads, + 3 * self.hidden_size_per_attention_head, + ) + mixed_x_layer = mixed_x_layer.view(*new_tensor_shape) + + # [b, sq, np, 3 * hn] --> 3 [b, sq, np, hn] + (query_layer, key_layer, value_layer) = split_tensor_along_last_dim(mixed_x_layer, 3) + # [b, sq, np, hn] -> [b, np, sq, hn] + query_layer, key_layer, value_layer = [k.transpose(1, 2) for k in [query_layer, key_layer, value_layer]] + + # apply relative positional encoding (rotary embedding) + if rotary_pos_emb is not None: + query_layer = apply_rotary_pos_emb(query_layer, rotary_pos_emb) + key_layer = apply_rotary_pos_emb(key_layer, rotary_pos_emb) + + # adjust key and value for inference + if past_key_value is not None: + key_layer, value_layer = past_key_value.update(key_layer, value_layer, self.layer_number - 1) + if self.multi_query_attention: + key_layer = key_layer.unsqueeze(2) + key_layer = key_layer.expand( + -1, + -1, + self.num_heads // self.num_multi_query_groups_per_partition, + -1, + -1, + ) + key_layer = key_layer.contiguous().view(key_layer.size()[:1] + (self.num_heads,) + key_layer.size()[3:]) + value_layer = value_layer.unsqueeze(2) + value_layer = value_layer.expand( + -1, + -1, + self.num_heads // self.num_multi_query_groups_per_partition, + -1, + -1, + ) + value_layer = value_layer.contiguous().view( + value_layer.size()[:1] + (self.num_heads,) + value_layer.size()[3:] + ) + # ================================== + # core attention computation + # ================================== + + context_layer = self.core_attention(query_layer, key_layer, value_layer, attention_mask) + + # ================= + # Output. [sq, b, h] + # ================= + + output = self.dense(context_layer) + + return output, past_key_value + + +class GLMMLP(nn.Module): + def __init__(self, config: GLMConfig): + super().__init__() + + self.dense_h_to_4h = nn.Linear(config.hidden_size, config.intermediate_size * 2, bias=False) + self.dense_4h_to_h = nn.Linear(config.intermediate_size, config.hidden_size, bias=False) + + def swiglu(x): + x = torch.chunk(x, 2, dim=-1) + return F.silu(x[0]) * x[1] + + self.act = swiglu + + def forward(self, hidden_states): + hidden_states = self.dense_h_to_4h(hidden_states) + hidden_states = self.act(hidden_states) + hidden_states = self.dense_4h_to_h(hidden_states) + return hidden_states + + +# 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 GLMAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper, modified to include features from CoreAttention.""" + + def __init__(self, config: GLMConfig, layer_number): + super(GLMAttention, self).__init__() + self.config = config + self.apply_query_key_layer_scaling = config.apply_query_key_layer_scaling + self.attention_softmax_in_fp32 = config.attention_softmax_in_fp32 + self.num_heads = config.num_attention_heads + self.kv_channels = config.kv_channels + self.attention_dropout = config.attention_dropout + if self.apply_query_key_layer_scaling: + self.attention_softmax_in_fp32 = True + self.layer_number = max(1, layer_number) + self.is_causal = True + + projection_size = self.kv_channels * self.num_heads + + # Per attention head and per partition values. + self.hidden_size_per_partition = projection_size + self.hidden_size_per_attention_head = projection_size // self.num_heads + + self.layer_scaling_coefficient = None + self.norm_factor = math.sqrt(self.hidden_size_per_attention_head) + if self.apply_query_key_layer_scaling: + # Scale the norm factor by the layer number to adjust attention dynamics across layers + self.layer_scaling_coefficient = self.layer_number + self.norm_factor *= self.layer_scaling_coefficient + + self.attention_dropout = torch.nn.Dropout(self.attention_dropout) + + def forward(self, query_layer, key_layer, value_layer, attention_mask): + # [b, np, sq, sk] + output_size = ( + query_layer.size(0), + query_layer.size(1), + query_layer.size(2), + key_layer.size(2), + ) + + # [b, np, sq, hn] -> [b * np, sq, hn] + query_layer = query_layer.reshape(output_size[0] * output_size[1], output_size[2], -1) + # [b, np, sk, hn] -> [b * np, sk, hn] + key_layer = key_layer.reshape(output_size[0] * output_size[1], output_size[3], -1) + + # preallocating input tensor: [b * np, sq, sk] + matmul_input_buffer = torch.empty( + output_size[0] * output_size[1], + output_size[2], + output_size[3], + dtype=query_layer.dtype, + device=query_layer.device, + ) + + # Raw attention scores. [b * np, sq, sk] + matmul_result = torch.baddbmm( + matmul_input_buffer, + query_layer, # [b * np, sq, hn] + key_layer.transpose(1, 2), # [b * np, hn, sk] + beta=0.0, + alpha=(1.0 / self.norm_factor), + ) + + # change view to [b, np, sq, sk] + attention_scores = matmul_result.reshape(*output_size) + + # =========================== + # Attention probs and dropout + # =========================== + + # attention scores and attention mask [b, np, sq, sk] + if self.attention_softmax_in_fp32: + attention_scores = attention_scores.float() + if self.layer_scaling_coefficient is not None: + attention_scores = attention_scores * self.layer_scaling_coefficient + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_layer.shape[-2]] + attention_scores = attention_scores + causal_mask + attention_probs = F.softmax(attention_scores, dim=-1) + attention_probs = attention_probs.type_as(value_layer) + + # This is actually dropping out entire tokens to attend to, which might + # seem a bit unusual, but is taken from the original Transformer paper. + attention_probs = self.attention_dropout(attention_probs) + + # query layer shape: [b * np, sq, hn] + # value layer shape: [b, np, sk, hn] + # attention shape: [b, np, sq, sk] + # context layer shape: [b, np, sq, hn] + output_size = ( + value_layer.size(0), + value_layer.size(1), + query_layer.size(1), + value_layer.size(3), + ) + # change view [b * np, sk, hn] + value_layer = value_layer.reshape(output_size[0] * output_size[1], value_layer.size(2), -1) + # change view [b * np, sq, sk] + attention_probs = attention_probs.reshape(output_size[0] * output_size[1], output_size[2], -1) + # matmul: [b * np, sq, hn] + context_layer = torch.bmm(attention_probs, value_layer) + # change view [b, np, sq, hn] + context_layer = context_layer.reshape(*output_size) + # [b, np, sq, hn] --> [b, sq, np, hn] + context_layer = context_layer.transpose(1, 2).contiguous() + # [b, sq, np, hn] --> [b, sq, hp] + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.reshape(*new_context_layer_shape) + return context_layer + + +def apply_rotary_pos_emb(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor: + # x: [b, np, sq, hn] and hn is not used in here. + b, np, sq, _ = x.size(0), x.size(1), x.size(2), x.size(3) + rot_dim = rope_cache.shape[-2] * 2 + x, x_pass = x[..., :rot_dim], x[..., rot_dim:] + # truncate to support variable sizes + rope_cache = rope_cache[:, :sq] + xshaped = x.reshape(b, np, sq, rot_dim // 2, 2) + rope_cache = rope_cache.view(-1, 1, sq, xshaped.size(3), 2) + x_out2 = torch.stack( + [ + xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1], + xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1], + ], + -1, + ) + x_out2 = x_out2.flatten(3) + return torch.cat((x_out2, x_pass), dim=-1) + + +class GLMFlashAttention2(GLMAttention): + """ + GLM flash attention module. This module inherits from `GLMAttention` as the weights of the module stay + 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. + """ + + def __init__(self, config: GLMConfig, layer_number): + super().__init__(config, layer_number) + 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, + cache_position: Optional[torch.LongTensor] = None, + ): + 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.hidden_size_per_attention_head).transpose( + 1, 2 + ) + key_states = key_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_heads, self.hidden_size_per_attention_head).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_number) + + rotary_seq_len = max(kv_seq_len, position_ids[:, -1].max().item()) + 1 + cos, sin = self.rotary_pos_emb(value_states, seq_len=rotary_seq_len) + + query_states, key_states = apply_rotary_pos_emb(query_states, key_states) + + if past_key_value is not None: + cache_has_contents = past_key_value.get_seq_length(self.layer_number) > 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_number][0] + past_value = past_key_value[self.layer_number][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, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_number, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_heads) + value_states = repeat_kv(value_states, self.num_heads) + dropout_rate = 0.0 if not self.training else self.attention_dropout + + input_dtype = query_states.dtype + if input_dtype == torch.float32: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + elif hasattr(self.config, "_pre_quantization_dtype"): + target_dtype = self.config._pre_quantization_dtype + else: + target_dtype = self.q_proj.weight.dtype + + query_states = query_states.to(target_dtype) + key_states = key_states.to(target_dtype) + value_states = value_states.to(target_dtype) + + query_states = query_states.transpose(1, 2) + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + sliding_window = None + if ( + self.config.use_sliding_window + and getattr(self.config, "sliding_window", None) is not None + and self.layer_number >= self.config.max_window_layers + ): + sliding_window = self.config.sliding_window + + attn_output = _flash_attention_forward( + query_states, + key_states, + value_states, + attention_mask, + q_len, + position_ids=position_ids, + dropout=dropout_rate, + sliding_window=sliding_window, + is_causal=self.is_causal, + use_top_left_mask=self._flash_attn_uses_top_left_mask, + ) + + attn_output = attn_output.reshape(bsz, q_len, self.hidden_size_per_partition).contiguous() + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + +class GLMSdpaAttention(GLMAttention): + """ + GLM attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from + `GLMAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to + SDPA API. + """ + + def forward(self, query_layer, key_layer, value_layer, attention_mask): + if attention_mask is None and query_layer.shape[2] == key_layer.shape[2]: + context_layer = torch.nn.functional.scaled_dot_product_attention( + query_layer, + key_layer, + value_layer, + is_causal=True, + dropout_p=self.config.attention_dropout if self.training else 0.0, + ) + else: + context_layer = torch.nn.functional.scaled_dot_product_attention( + query_layer, + key_layer, + value_layer, + attention_mask, + dropout_p=self.config.attention_dropout if self.training else 0.0, + ) + context_layer = context_layer.transpose(1, 2).contiguous() + new_context_layer_shape = context_layer.size()[:-2] + (self.hidden_size_per_partition,) + context_layer = context_layer.reshape(*new_context_layer_shape) + return context_layer + + +GLM_ATTENTION_CLASSES = { + "eager": GLMAttention, + "flash_attention_2": GLMFlashAttention2, + "sdpa": GLMSdpaAttention, +} + +GLM_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 ([`GLMConfig`]): + 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 GLM Model outputting raw hidden-states without any specific head on top.", + GLM_START_DOCSTRING, +) +class GLMPreTrainedModel(PreTrainedModel): + config_class = GLMConfig + base_model_prefix = "transformer" + supports_gradient_checkpointing = True + _no_split_modules = ["GLMDecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_cache_class = True + + _version = "0.0.5" + + 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_() + + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + # TODO: As of torch==2.2.0, the `attention_mask` passed to the model in `generate` is 2D and of dynamic length even when the static + # KV cache is used. This is an issue for torch.compile which then recaptures cudagraphs at each decode steps due to the dynamic shapes. + # (`recording cudagraph tree for symint key 13`, etc.), which is VERY slow. A workaround is `@torch.compiler.disable`, but this prevents using + # `fullgraph=True`. See more context in https://github.com/huggingface/transformers/pull/29114 + + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method + # calls the eager implementation's forward + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + dtype, device = input_tensor.dtype, input_tensor.device + min_dtype = torch.finfo(dtype).min + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_length() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + if attention_mask is not None and attention_mask.dim() == 4: + # in this case we assume that the mask comes already in inverted + # form and requires no inversion or slicing + if attention_mask.max() != 0: + raise ValueError("Custom 4D attention mask should be passed in inverted form with max==0`") + causal_mask = attention_mask + else: + causal_mask = torch.full( + (sequence_length, target_length), + fill_value=min_dtype, + dtype=dtype, + device=device, + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + causal_mask *= torch.arange(target_length, device=device) > cache_position.reshape(-1, 1) + causal_mask = causal_mask[None, None, :, :].expand(input_tensor.shape[0], 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + + +class Embedding(torch.nn.Module): + """Language model embeddings.""" + + def __init__(self, config: GLMConfig, device=None): + super(Embedding, self).__init__() + self.vocab_size = config.vocab_size + self.hidden_size = config.hidden_size + # Word embeddings (parallel). + self.word_embeddings = nn.Embedding(self.vocab_size, self.hidden_size, device=device) + self.fp32_residual_connection = config.fp32_residual_connection + + def forward(self, input_ids): + words_embeddings = self.word_embeddings(input_ids) + embeddings = words_embeddings + # If the input flag for fp32 residual connection is set, convert for + # float. + if self.fp32_residual_connection: + embeddings = embeddings.float() + return embeddings + + +class GLMDecoderLayer(torch.nn.Module): + """A single transformer layer. + + Transformer layer takes input with size [s, b, h] and returns an + output of the same size. + """ + + def __init__(self, config: GLMConfig, layer_number, device=None): + super(GLMDecoderLayer, self).__init__() + self.layer_number = layer_number + self.fp32_residual_connection = config.fp32_residual_connection + self.self_attention = GLMSelfAttention(config, layer_number, device=device) + self.hidden_dropout = config.hidden_dropout + + self.mlp = GLMMLP(config) + self.input_layernorm = GLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = GLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + + def forward( + self, + hidden_states, + attention_mask, + rotary_pos_emb, + past_key_value=None, + use_cache=True, + ): + # hidden_states: [s, b, h] + + # Layer norm at the beginning of the transformer layer. + layernorm_output = self.input_layernorm(hidden_states) + # Self attention. + attention_output, past_key_value = self.self_attention( + layernorm_output, + attention_mask, + rotary_pos_emb, + past_key_value=past_key_value, + use_cache=use_cache, + ) + + # Residual connection. + residual = hidden_states + + layernorm_input = torch.nn.functional.dropout(attention_output, p=self.hidden_dropout, training=self.training) + layernorm_input = residual + layernorm_input + + # Layer norm post the self attention. + layernorm_output = self.post_attention_layernorm(layernorm_input) + + # MLP. + mlp_output = self.mlp(layernorm_output) + + # Second residual connection. + residual = layernorm_input + + output = torch.nn.functional.dropout(mlp_output, p=self.hidden_dropout, training=self.training) + output = residual + output + + return output, past_key_value + + +class GLMTransformer(torch.nn.Module): + """Transformer class.""" + + def __init__(self, config: GLMConfig, device=None): + super(GLMTransformer, self).__init__() + + self.fp32_residual_connection = config.fp32_residual_connection + + # Number of layers. + self.num_hidden_layers = config.num_hidden_layers + + # Transformer layers. + def build_layer(layer_number): + return GLMDecoderLayer(config, layer_number, device=device) + + self.layers = torch.nn.ModuleList([build_layer(i + 1) for i in range(self.num_hidden_layers)]) + + self.final_layernorm = GLMRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.gradient_checkpointing = False + + def _get_layer(self, layer_number): + return self.layers[layer_number] + + def forward( + self, + hidden_states, + attention_mask, + rotary_pos_emb, + past_key_values, + output_attentions: bool = False, + use_cache: Optional[bool] = True, + output_hidden_states: Optional[bool] = False, + ): + if self.gradient_checkpointing and self.training and use_cache: + logger.warning( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = None + + for index in range(self.num_hidden_layers): + if output_hidden_states: + all_hidden_states += (hidden_states,) + + layer = self._get_layer(index) + if self.gradient_checkpointing and self.training: + layer_ret = self._gradient_checkpointing_func( + layer.__call__, + hidden_states, + attention_mask, + rotary_pos_emb, + past_key_values, + use_cache, + use_reentrant=False, + ) + else: + layer_ret = layer( + hidden_states, + attention_mask=attention_mask, + rotary_pos_emb=rotary_pos_emb, + past_key_value=past_key_values, + use_cache=use_cache, + ) + + hidden_states, next_decoder_cache = layer_ret + + if output_attentions: + all_self_attns += (hidden_states,) + + hidden_states = self.final_layernorm(hidden_states) + + if output_hidden_states: + all_hidden_states += (hidden_states,) + + return hidden_states, next_decoder_cache, all_hidden_states, all_self_attns + + +GLM_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. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + + +@add_start_docstrings( + "The bare GLM Model outputting raw hidden-states without any specific head on top.", + GLM_START_DOCSTRING, + """ + device ([`str`], *optional*): + The device on which this model will be run. + add_lm_head ([`bool`], *optional*, defaults to `False`): + Whether or not to add a language modeling head on top of the model. The language modeling head is composed + of two dense layers. +""", +) +class GLMModel(GLMPreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`GLMBlock`] + + Args: + config: GLMConfig + device: The device on which the model should be run. + add_lm_head: Whether to add a language modeling head on top of the model. + """ + + def __init__(self, config: GLMConfig, device: str = None, add_lm_head: bool = False): + super().__init__(config) + + def default_init(cls, *args, **kwargs): + return cls(*args, **kwargs) + + init_method = default_init + init_kwargs = {} + if device is not None: + init_kwargs["device"] = device + self.embedding = init_method(Embedding, config, **init_kwargs) + self.num_hidden_layers = config.num_hidden_layers + self.multi_query_group_num = config.multi_query_group_num + self.kv_channels = config.kv_channels + self.num_heads = config.num_attention_heads + self.hidden_size = config.hidden_size + self.vocab_size = config.vocab_size + # Rotary positional embeddings + self.max_position_embeddings = config.max_position_embeddings + rotary_dim = config.hidden_size // self.num_heads if self.kv_channels is None else self.kv_channels + + self.rotary_pos_emb = GLMRotaryEmbedding( + rotary_dim // 2, + rope_theta=config.rope_theta, + original_impl=True, + device=device, + ) + self.encoder = init_method(GLMTransformer, config, **init_kwargs) + if add_lm_head: + self.output_layer = init_method( + nn.Linear, + self.hidden_size, + self.vocab_size, + bias=False, + **init_kwargs, + ) + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embedding.word_embeddings + + def set_input_embeddings(self, value): + self.embedding.word_embeddings = value + + @add_start_docstrings_to_model_forward(GLM_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, + cache_position: Optional[torch.LongTensor] = 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 + return_legacy_cache = False + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError( + "You cannot specify both input_ids and inputs_embeds at the same time, and must specify either one" + ) + if inputs_embeds is None: + inputs_embeds = self.embedding(input_ids) + + batch_size, seq_length = inputs_embeds.shape[:2] + + if use_cache and not isinstance(past_key_values, Cache): # kept for BC (non `Cache` `past_key_values` inputs) + return_legacy_cache = True + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. " + "Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)" + ) + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, + past_seen_tokens + inputs_embeds.shape[1], + device=inputs_embeds.device, + ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, + inputs_embeds, + cache_position, + past_key_values, + output_attentions, + ) + hidden_states = inputs_embeds + + # create position embeddings to be shared across the decoder layers + rotary_pos_emb = self.rotary_pos_emb(self.max_position_embeddings) + + if position_ids is not None: + rotary_pos_emb = rotary_pos_emb[position_ids] + else: + rotary_pos_emb = rotary_pos_emb[None, :seq_length] + + # Run encoder. + hidden_states, next_cache, all_hidden_states, all_self_attns = self.encoder( + hidden_states=hidden_states, + attention_mask=causal_mask, + rotary_pos_emb=rotary_pos_emb, + past_key_values=past_key_values, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + ) + + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if return_legacy_cache: + next_cache = next_cache.to_legacy_cache() + + if not use_cache: + next_cache = None + + 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 GLMForCausalLM(GLMPreTrainedModel): + _tied_weights_keys = ["transformer.output_layer.weight"] + + def __init__(self, config: GLMConfig, device=None): + super().__init__(config) + + self.max_sequence_length = config.max_length + self.transformer = GLMModel(config, add_lm_head=True, device=device) + self.config = config + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.transformer.embedding.word_embeddings + + def set_input_embeddings(self, value): + self.transformer.embedding.word_embeddings = value + + def get_output_embeddings(self): + return self.transformer.output_layer + + def set_output_embeddings(self, new_embeddings): + self.transformer.output_layer = new_embeddings + + def set_decoder(self, decoder): + self.transformer = decoder + + def get_decoder(self): + return self.transformer + + @add_start_docstrings_to_model_forward(GLM_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: bool = False, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = 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 GLMTokenizer, GLMForCausalLM + + >>> model = GLMForCausalLM.from_pretrained(PATH_TO_CONVERTED_WEIGHTS) + >>> tokenizer = GLMTokenizer.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.transformer( + 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, + cache_position=cache_position, + ) + + hidden_states = outputs[0] + logits = self.transformer.output_layer(hidden_states) + + 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, + ) + + # Copied from transformers.models.llama.modeling_llama.LlamaForCausalLM.prepare_inputs_for_generation + def prepare_inputs_for_generation( + self, + input_ids, + past_key_values=None, + attention_mask=None, + inputs_embeds=None, + cache_position=None, + position_ids=None, + use_cache=True, + **kwargs, + ): + # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens + # Exception 1: when passing input_embeds, input_ids may be missing entries + # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here + if past_key_values is not None: + if inputs_embeds is not None: # Exception 1 + input_ids = input_ids[:, -cache_position.shape[0] :] + elif input_ids.shape[1] != cache_position.shape[0]: # Default case (the "else", a no op, is Exception 2) + input_ids = input_ids[:, cache_position] + + 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 cache_position[0] == 0: + model_inputs = {"inputs_embeds": inputs_embeds} + else: + model_inputs = {"input_ids": input_ids.contiguous()} # `contiguous()` needed for compilation use cases + + model_inputs.update( + { + "position_ids": position_ids, + "cache_position": cache_position, + "past_key_values": past_key_values, + "use_cache": use_cache, + "attention_mask": attention_mask, + } + ) + return model_inputs + + +@add_start_docstrings( + """ + The GLM Model transformer with a sequence classification head on top (linear layer). + + [`GLMForSequenceClassification`] 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). + """, + GLM_START_DOCSTRING, +) +class GLMForSequenceClassification(GLMPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.transformer = GLMModel(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.transformer.embedding.word_embeddings + + def set_input_embeddings(self, value): + self.transformer.embedding.word_embeddings = value + + @add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, 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.transformer( + 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, + ) + + +@add_start_docstrings( + """ + The GLM Model transformer with a token classification head on top (a linear layer on top of the hidden-states + output) e.g. for Named-Entity-Recognition (NER) tasks. + """, + GLM_START_DOCSTRING, +) +class GLMForTokenClassification(GLMPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.transformer = GLMModel(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.transformer.embedding.word_embeddings + + def set_input_embeddings(self, value): + self.transformer.embedding.word_embeddings = value + + @add_start_docstrings_to_model_forward(GLM_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[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, TokenClassifierOutput]: + 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 + + outputs = self.transformer( + 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, + ) + sequence_output = outputs[0] + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss_fct = CrossEntropyLoss() + loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1)) + + if not return_dict: + output = (logits,) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) diff --git a/src/transformers/models/glm/tokenization_glm.py b/src/transformers/models/glm/tokenization_glm.py new file mode 100644 index 00000000000000..080c23cfc41234 --- /dev/null +++ b/src/transformers/models/glm/tokenization_glm.py @@ -0,0 +1,344 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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 GLM.""" + +import json +import os +from functools import lru_cache +from typing import Optional, Tuple, Type + +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", +} + +PRETOKENIZE_REGEX = "(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\\r\\n\\p{L}\\p{N}]?\\p{L}+|\\p{N}{1,3}| ?[^\\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 GLMTokenizer(PreTrainedTokenizer): + """ + Construct a GLM tokenizer. Based on byte-level Byte-Pair-Encoding. + + Same with GPT2Tokenizer, 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 GLMTokenizer + + >>> tokenizer = GLMTokenizer.from_pretrained("THUDM/GLM-tokenizer") + >>> tokenizer("Hello world")["input_ids"] + [9703, 1879] + + >>> tokenizer(" Hello world")["input_ids"] + [21873, 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. + use_default_system_prompt (`bool`, *optional*, defaults to `False`): + Whether or not the default system prompt for Cohere tokenizer should be used. + 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. + spaces_between_special_tokens (`bool`, *optional*, defaults to `False`): + Whether or not to add spaces between special tokens. + add_prefix_space (`bool`, *optional*, defaults to `True`): + Whether or not to add a space to the beginning of the text. This allows to treat the leading word just as any other word. + """ + + vocab_files_names = VOCAB_FILES_NAMES + 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, + use_default_system_prompt=False, + split_special_tokens=False, + spaces_between_special_tokens=False, + add_prefix_space=True, + **kwargs, + ): + 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 + self.byte_encoder = bytes_to_unicode() + self.byte_decoder = {v: k for k, v in self.byte_encoder.items()} + self.add_prefix_space = add_prefix_space + self.use_default_system_prompt = use_default_system_prompt + self.spaces_between_special_tokens = spaces_between_special_tokens + + bpe_merges = [] + with open(merges_file, encoding="utf-8") as merges_handle: + for i, line in enumerate(merges_handle): + line = line.strip() + if (i == 0 and line.startswith("#version:")) or not line: + 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) + 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, + use_default_system_prompt=use_default_system_prompt, + split_special_tokens=split_special_tokens, + add_prefix_space=add_prefix_space, + **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 save_vocabulary( + self, save_directory: str, filename_prefix: Optional[str] = None + ) -> Type[tuple] | tuple[str, str]: + if not os.path.isdir(save_directory): + logger.error(f"Vocabulary path ({save_directory}) should be a directory") + return Tuple[None] + + 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 + + @property + def default_chat_template(self): + """ + GLM uses [gMASK] and to indicate user messages. The system message is included as part of the first user + message. The assistant messages do not have special tokens, as they can be identified by their order. + + We add a system prompt to make GLM-4 can be used in Function Calling and GLM All Tools capability. + + Here is an example of output： + + [gMASK]<|system|>\nSystemPrompt<|user|>\nPrompt<|assistant|>n\\Answer<|user|>\nPrompt<|assistant|>\nAnswer<|user|> + + """ + template = "[gMASK]{% for item in messages %}{% if item['tools'] is defined %}<|system|>\n你是一个名为 GLM-4 的人工智能助手。你是基于智谱AI训练的语言模型 GLM-4 模型开发的，你的任务是针对用户的问题和要求提供适当的答复和支持。\n\n# 可用工具{% set tools = item['tools'] %}{% for tool in tools %}{% if tool['type'] == 'function' %}\n\n## {{ tool['function']['name'] }}\n\n{{ tool['function'] | tojson(indent=4) }}\n在调用上述函数时，请使用 Json 格式表示调用的参数。{% elif tool['type'] == 'python' %}\n\n## python\n\n当你向 `python` 发送包含 Python 代码的消息时，该代码将会在一个有状态的 Jupyter notebook 环境中执行。\n`python` 返回代码执行的输出，或在执行 60 秒后返回超时。\n`/mnt/data` 将会持久化存储你的文件。在此会话中，`python` 无法访问互联网。不要使用 `python` 进行任何网络请求或者在线 API 调用，这些在线内容的访问将不会成功。{% elif tool['type'] == 'simple_browser' %}\n\n## simple_browser\n\n你可以使用 `simple_browser` 工具。该工具支持以下函数：\n`search(query: str, recency_days: int)`：使用搜索引擎进行查询并显示结果，可以使用 `recency_days` 参数控制搜索内容的时效性。\n`mclick(ids: list[int])`：获取一系列指定 id 的页面内容。每次调用时，须选择3-10个页面。选择多个角度的页面，同时尽可能选择可信任的信息来源。考虑到部分页面是无法加载的，你也可以多打开一些可能有用的页面而不用担心内容过多。\n`open_url(url: str)`：打开指定的 URL。\n\n使用 `【{引用 id}†{引用文本}】` 来引用内容。\n\n操作步骤：1. 使用 `search` 来获得信息列表; 2. 使用 `mclick` 来获取指定 ID 页面的内容; 3. 根据获得的内容进行回复。在回复中应当引用信息来源。\n 如果用户提供了 URL，也可以用 `open_url` 直接打开页面。\n如果初次搜索结果没有找到合适的信息，也可以再次使用 `search` 进行搜索。{% elif tool['type'] == 'cogview' %}\n\n## cogview\n\n如果用户的请求中包含了对图像的描述，你可以使用 `cogview` 来生成图像并展示给用户。你需要向 `cogview` 发送图像描述，规则：\n- 发送给 `cogview` 的消息必须使用英语。用户的中文描述必须完全翻译为英语。\n- 应当尽可能详细地描述图像生成的需求，需求描述约 100 英文单词。\n- 保持用户原始描述的意图。不要虚构内容或者没见过的人物。\n- 如无特殊说明，所在地为中国，持有中国立场并遵循中国社会主义价值观。{% endif %}{% endfor %}{% endif %}{% if item['content'] %}<|{{ item['role'] }}|>{{ item['metadata'] }}\n{{ item['content'] }}{% endif %}{% endfor %}{% if add_generation_prompt %}<|assistant|>{% endif %}" + template = template.replace("USE_DEFAULT_PROMPT", "true" if self.use_default_system_prompt else "false") + return template diff --git a/src/transformers/models/glm/tokenization_glm_fast.py b/src/transformers/models/glm/tokenization_glm_fast.py new file mode 100644 index 00000000000000..17c0d9779563d2 --- /dev/null +++ b/src/transformers/models/glm/tokenization_glm_fast.py @@ -0,0 +1,132 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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 GLM.""" + +from typing import Optional, Tuple + +from ...tokenization_utils import AddedToken +from ...tokenization_utils_fast import PreTrainedTokenizerFast +from ...utils import logging +from .tokenization_glm import GLMTokenizer + + +logger = logging.get_logger(__name__) + +VOCAB_FILES_NAMES = { + "vocab_file": "vocab.json", + "merges_file": "merges.txt", + "tokenizer_file": "tokenizer.json", +} + + +class GLMTokenizerFast(PreTrainedTokenizerFast): + """ + Construct a "fast" GLM tokenizer (backed by HuggingFace's *tokenizers* library). Based on byte-level + Byte-Pair-Encoding. + + Same with GPT2Tokenizer, 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 GLMTokenizerFast + + >>> tokenizer = GLMTokenizer.from_pretrained("THUDM/GLM-tokenizer") + >>> tokenizer("Hello world")["input_ids"] + [9703, 1879] + + >>> tokenizer(" Hello world")["input_ids"] + [21873, 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 + model_input_names = ["input_ids", "attention_mask"] + slow_tokenizer_class = GLMTokenizer + + 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=vocab_file, + merges_file=merges_file, + tokenizer_file=tokenizer_file, + unk_token=unk_token, + bos_token=bos_token, + eos_token=eos_token, + pad_token=pad_token, + **kwargs, + ) + + 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 4732ecea8611f0..fce5aea63ee427 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -4368,6 +4368,41 @@ def __init__(self, *args, **kwargs): requires_backends(self, ["torch"]) +class GLMForCausalLM(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class GLMForSequenceClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class GLMForTokenClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class GLMModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class GLMPreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + class GLPNForDepthEstimation(metaclass=DummyObject): _backends = ["torch"] diff --git a/src/transformers/utils/dummy_tokenizers_objects.py b/src/transformers/utils/dummy_tokenizers_objects.py index df83e6fa6478e6..317cad511a3cea 100644 --- a/src/transformers/utils/dummy_tokenizers_objects.py +++ b/src/transformers/utils/dummy_tokenizers_objects.py @@ -191,6 +191,13 @@ def __init__(self, *args, **kwargs): requires_backends(self, ["tokenizers"]) +class GLMTokenizerFast(metaclass=DummyObject): + _backends = ["tokenizers"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["tokenizers"]) + + class GPT2TokenizerFast(metaclass=DummyObject): _backends = ["tokenizers"] diff --git a/src/transformers/utils/fx.py b/src/transformers/utils/fx.py index c78b4c34c331f0..5df00b872c4119 100755 --- a/src/transformers/utils/fx.py +++ b/src/transformers/utils/fx.py @@ -135,6 +135,7 @@ def _generate_supported_model_class_names( "distilbert", "donut-swin", "electra", + "glm", "gpt2", "gpt_neo", "gptj", diff --git a/tests/models/glm/__init__.py b/tests/models/glm/__init__.py new file mode 100644 index 00000000000000..e69de29bb2d1d6 diff --git a/tests/models/glm/test_modeling_glm.py b/tests/models/glm/test_modeling_glm.py new file mode 100644 index 00000000000000..714ef5f0a967a8 --- /dev/null +++ b/tests/models/glm/test_modeling_glm.py @@ -0,0 +1,661 @@ +# coding=utf-8 +# Copyright 2024 GLM & ZhipuAI 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 GLM model.""" + +import gc +import tempfile +import unittest + +import pytest + +from transformers import AutoTokenizer, GLMConfig, is_torch_available +from transformers.testing_utils import ( + backend_empty_cache, + is_flaky, + 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 ( + GLMForCausalLM, + GLMForSequenceClassification, + GLMForTokenClassification, + GLMModel, + ) + + +class GLMModelTester: + 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=8, + num_hidden_layers=2, + num_attention_heads=4, + 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.num_attention_heads = num_attention_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 GLMConfig( + vocab_size=self.vocab_size, + hidden_size=self.hidden_size, + num_hidden_layers=self.num_hidden_layers, + num_attention_heads=self.num_attention_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, + initializer_range=self.initializer_range, + pad_token_id=self.pad_token_id, + bos_token_id=self.bos_token_id, + output_attentions=False, + ) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTester.create_and_check_model with Llama->GLM + def create_and_check_model( + self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels + ): + model = GLMModel(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->GLM + 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 = GLMModel(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->GLM + 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 = GLMForCausalLM(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->GLM + 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 = GLMForCausalLM(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 +class GLMModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase): + all_model_classes = ( + (GLMModel, GLMForCausalLM, GLMForSequenceClassification, GLMForTokenClassification) + if is_torch_available() + else () + ) + all_generative_model_classes = (GLMForCausalLM,) if is_torch_available() else () + pipeline_model_mapping = ( + { + "feature-extraction": GLMModel, + "text-classification": GLMForSequenceClassification, + "token-classification": GLMForTokenClassification, + "text-generation": GLMForCausalLM, + "zero-shot": GLMForSequenceClassification, + } + if is_torch_available() + else {} + ) + test_headmasking = False + test_pruning = False + test_hidden_states_output = False + fx_compatible = False + test_attention_outputs = 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 + + # Ignore copy + # TODO: @Fxmarty + @is_flaky(max_attempts=3, description="flaky on some models.") + @require_torch_sdpa + @slow + def test_eager_matches_sdpa_generate(self): + super().test_eager_matches_sdpa_generate() + + def setUp(self): + self.model_tester = GLMModelTester(self) + self.config_tester = ConfigTester(self, config_class=GLMConfig, 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_GLM_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 = GLMForSequenceClassification(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_GLM_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 = GLMForSequenceClassification(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_GLM_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 = GLMForSequenceClassification(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)) + + # Copied from tests.models.llama.test_modeling_llama.LlamaModelTest.test_llama_token_classification_model with Llama->GLM,llama->GLM + def test_GLM_token_classification_model(self): + config, input_dict = self.model_tester.prepare_config_and_inputs_for_common() + config.num_labels = 3 + input_ids = input_dict["input_ids"] + attention_mask = input_ids.ne(1).to(torch_device) + token_labels = ids_tensor([self.model_tester.batch_size, self.model_tester.seq_length], config.num_labels) + model = GLMForTokenClassification(config=config) + model.to(torch_device) + model.eval() + result = model(input_ids, attention_mask=attention_mask, labels=token_labels) + self.assertEqual( + result.logits.shape, + (self.model_tester.batch_size, self.model_tester.seq_length, self.model_tester.num_labels), + ) + + @unittest.skip(reason="GLM buffers include complex numbers, which breaks this test") + def test_save_load_fast_init_from_base(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: GLM 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_equivalence_right_padding(self): + self.skipTest(reason="GLM flash attention does not support right padding") + + @unittest.skip("GLM KV cache is a non standard format") + def test_past_key_values_format(self): + pass + + @slow + @require_torch + class GLMIntegrationTest(unittest.TestCase): + def test_glm_instruct_logits(self): + input_ids = [ + 151331, + 151333, + 151336, + 198, + 102162, + 220, + 16, + 10, + 16, + 100694, + 99312, + 3837, + 99558, + 104559, + 100295, + 151337, + ] + model = GLMForCausalLM.from_pretrained("THUDM/glm-4-9b-chat").to(torch_device) + 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( + [ + [ + -2.6504, + -0.0175, + -1.7773, + -1.9961, + -2.2734, + -2.8457, + -2.4512, + -2.6133, + -2.4199, + -2.3535, + -2.8203, + -2.5664, + -1.9512, + -3.4766, + -3.4395, + -3.0156, + ] + ] + ) + 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( + [ + 3.9199, + 6.3906, + 4.7812, + 4.1914, + -1.0078, + -1.2148, + 4.2109, + 5.5625, + 2.4121, + 2.2910, + 4.3438, + 5.7969, + 7.0859, + 4.5273, + 0.9565, + -1.8076, + 3.1582, + 3.7305, + 4.5977, + 5.7500, + 4.1211, + 4.2461, + 4.4883, + 2.9395, + 4.0703, + 7.1953, + 3.5430, + 2.4707, + 0.0379, + 2.0449, + ] + ) + + 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() + + def test_glm_instruct_generation(self): + model = GLMForCausalLM.from_pretrained("THUDM/glm-4-9b-chat") + tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-4-9b-chat") + messages = [ + { + "role": "system", + "content": "You are a helpful digital assistant. Please provide safe, ethical and accurate information to the user.", + }, + {"role": "user", "content": "Tell me the answer of 1 plus 1?"}, + ] + inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt") + outputs = model.generate(inputs, max_new_tokens=32) + output_text = tokenizer.batch_decode(outputs) + EXPECTED_OUTPUT = [ + "[gMASK] <|system|> \nYou are a helpful digital assistant. Please provide safe, ethical and accurate information to the user. <|user|> \nTell me the answer of 1 plus 1? <|assistant|> \nThe answer to 1 plus 1 is 2. <|user|>" + ] + self.assertListEqual(output_text, EXPECTED_OUTPUT) + + def _check_attentions_for_generate( + self, + batch_size, + attentions, + min_length, + max_length, + config, + use_cache=False, + num_beam_groups=1, + ): + self.assertIsInstance(attentions, tuple) + self.assertListEqual( + [isinstance(iter_attentions, tuple) for iter_attentions in attentions], + [True] * len(attentions), + ) + self.assertEqual(len(attentions), (max_length - min_length) * num_beam_groups) + + for idx, iter_attentions in enumerate(attentions): + tgt_len = min_length + idx if not use_cache else 1 + + expected_shape = ( + batch_size, + tgt_len, + config.hidden_size, + ) + + # check attn size + self.assertListEqual( + [layer_attention.shape for layer_attention in iter_attentions], + [expected_shape] * len(iter_attentions), + ) + + def _check_past_key_values_for_generate(self, batch_size, past_key_values, seq_length, config, num_beam_groups=1): + self.assertIsInstance(past_key_values, tuple) + self.assertListEqual( + [isinstance(iter_past_key_values, tuple) for iter_past_key_values in past_key_values], + [True] * len(past_key_values), + ) + + # (batch, head, seq_length, kv_channels) + expected_shape = ( + batch_size * num_beam_groups, + config.num_attention_heads, + seq_length, + config.kv_channels, + ) + # check shape key, value + self.assertListEqual( + [layer_past_key_values[0].shape for layer_past_key_values in past_key_values], + [expected_shape] * len(past_key_values), + ) + self.assertListEqual( + [layer_past_key_values[1].shape for layer_past_key_values in past_key_values], + [expected_shape] * len(past_key_values), + ) diff --git a/utils/not_doctested.txt b/utils/not_doctested.txt index cd87d09ec8ec6d..046a4b1801cf0a 100644 --- a/utils/not_doctested.txt +++ b/utils/not_doctested.txt @@ -589,6 +589,9 @@ src/transformers/models/gemma/modeling_flax_gemma.py src/transformers/models/gemma/modeling_gemma.py src/transformers/models/git/configuration_git.py src/transformers/models/git/convert_git_to_pytorch.py +src/transformers/models/glm/configuration_glm.py +src/transformers/models/glm/modeling_glm.py +src/transformers/models/glm/tokenization_glm.py src/transformers/models/glpn/configuration_glpn.py src/transformers/models/glpn/convert_glpn_to_pytorch.py src/transformers/models/gpt2/CONVERSION.md