From b3ab3fac1df7c534be17f97c3584744a97599bfd Mon Sep 17 00:00:00 2001 From: Matt Date: Tue, 11 Jul 2023 13:36:31 +0100 Subject: [PATCH] Falcon port (#24523) * Initial commit * Update src/transformers/models/falcon/configuration_falcon.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/falcon/configuration_falcon.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Cleanup config docstring * Update src/transformers/models/falcon/configuration_falcon.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Convert to relative imports * Remove torch < 1.8 warning * Restructure cos_sin header * qkv -> query, key, value * Refactor attention calculation * Add a couple of config variables to account for the different checkpoints * Successful merging of the code paths! * Fix misplaced line in the non-parallel attention path * Update config and tests * Add a pad_token_id when testing * Support output_attentions when alibi is None * make fixup * Skip KV cache shape test * No more _keys_to_ignore_on_load_missing * Simplify self attention a bit * Simplify self attention a bit * make fixup * stash commit * Some more attention mask updates * Should pass all tests except assisted generation! * Add big model generation test * make fixup * Add temporary workaround for test * Test overrides for assisted generation * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update tests/models/falcon/test_modeling_falcon.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Test overrides for assisted generation * Add generation demo * Update copyright * Make the docstring model actually small * Add module-level docstring * Remove all assertions * Add copied from bloom * Reformat the QKV layer * Add copied from bloom * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Remove unused line and reformat * No single letter variables * Cleanup return names * Add copied from line * Remove the deprecated arguments blocks * Change the embeddings test to an alibi on/off test * Remove position_ids from FalconForQA * Remove old check for token type IDs * Fix the alibi path when multi_query is False * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update src/transformers/models/falcon/modeling_falcon.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update tests/models/falcon/test_modeling_falcon.py Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> * Update config naming * Fix typo for new_decoder_architecture * Add some comments * Fix docstring * Fix docstring * Create range in the right dtype from the start * Review comment cleanup * n_head_kv -> num_kv_heads * self.alibi -> self.use_alibi * self.num_kv -> self.num_kv_heads * Reorder config args * Made alibi arguments Optional * Add all model docstrings * Add extra checkpoints * Add author info for Falcon * Stop removing token_type_ids because our checkpoints shouldn't return it anymore * Add one hopeful comment for the future * Fix typo * Update tests, fix cache issue for generation * Use -1e9 instead of -inf to avoid float overflow * Recompute the rotary embeddings much less often * Re-enable disabled tests * One final fix to attention mask calculation, and update tests * Cleanup targeting falcon-40b equivalency * Post-rebase docs update * Update docstrings, especially in the config * More descriptive variable names, and comments where we can't rename them --------- Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com> --- README.md | 1 + README_es.md | 1 + README_hd.md | 1 + README_ja.md | 1 + README_ko.md | 1 + README_zh-hans.md | 1 + README_zh-hant.md | 1 + docs/source/en/_toctree.yml | 2 + docs/source/en/index.md | 2 + docs/source/en/model_doc/falcon.md | 66 + docs/source/en/tasks/language_modeling.md | 2 +- docs/source/en/tasks/question_answering.md | 2 +- .../en/tasks/sequence_classification.md | 2 +- docs/source/en/tasks/token_classification.md | 2 +- src/transformers/__init__.py | 22 + src/transformers/models/__init__.py | 1 + .../models/auto/configuration_auto.py | 3 + src/transformers/models/auto/modeling_auto.py | 5 + .../models/bloom/modeling_bloom.py | 2 +- src/transformers/models/falcon/__init__.py | 68 + .../models/falcon/configuration_falcon.py | 147 ++ .../models/falcon/modeling_falcon.py | 1262 +++++++++++++++++ src/transformers/utils/dummy_pt_objects.py | 45 + tests/models/falcon/__init__.py | 0 tests/models/falcon/test_modeling_falcon.py | 469 ++++++ 25 files changed, 2104 insertions(+), 5 deletions(-) create mode 100644 docs/source/en/model_doc/falcon.md create mode 100644 src/transformers/models/falcon/__init__.py create mode 100644 src/transformers/models/falcon/configuration_falcon.py create mode 100644 src/transformers/models/falcon/modeling_falcon.py create mode 100644 tests/models/falcon/__init__.py create mode 100644 tests/models/falcon/test_modeling_falcon.py diff --git a/README.md b/README.md index 7b2b1cb3d5..d0fddfd1b6 100644 --- a/README.md +++ b/README.md @@ -351,6 +351,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu. 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2 and ESMFold** were released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab. diff --git a/README_es.md b/README_es.md index 84ae55f0cf..62ba9546f4 100644 --- a/README_es.md +++ b/README_es.md @@ -326,6 +326,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu. 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2** was released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab. diff --git a/README_hd.md b/README_hd.md index 44850339f6..0c1dcb3bfd 100644 --- a/README_hd.md +++ b/README_hd.md @@ -298,6 +298,7 @@ conda install -c huggingface transformers 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)**(Baidu से) साथ देने वाला पेपर [ERNIE: एन्हांस्ड रिप्रेजेंटेशन थ्रू नॉलेज इंटीग्रेशन](https://arxiv.org/abs/1904.09223) यू सन, शुओहुआन वांग, युकुन ली, शिकुन फेंग, ज़ुई चेन, हान झांग, शिन तियान, डैनक्सियांग झू, हाओ तियान, हुआ वू द्वारा पोस्ट किया गया। 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu से) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. द्वाराअनुसंधान पत्र [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) के साथ जारी किया गया 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (मेटा AI से) ट्रांसफॉर्मर प्रोटीन भाषा मॉडल हैं। **ESM-1b** पेपर के साथ जारी किया गया था [ अलेक्जेंडर राइव्स, जोशुआ मेयर, टॉम सर्कु, सिद्धार्थ गोयल, ज़ेमिंग लिन द्वारा जैविक संरचना और कार्य असुरक्षित सीखने को 250 मिलियन प्रोटीन अनुक्रमों तक स्केल करने से उभरता है] (https://www.pnas.org/content/118/15/e2016239118) जेसन लियू, डेमी गुओ, मायल ओट, सी. लॉरेंस ज़िटनिक, जेरी मा और रॉब फर्गस। **ESM-1v** को पेपर के साथ जारी किया गया था [भाषा मॉडल प्रोटीन फ़ंक्शन पर उत्परिवर्तन के प्रभावों की शून्य-शॉट भविष्यवाणी को सक्षम करते हैं] (https://doi.org/10.1101/2021.07.09.450648) जोशुआ मेयर, रोशन राव, रॉबर्ट वेरकुइल, जेसन लियू, टॉम सर्कु और अलेक्जेंडर राइव्स द्वारा। **ESM-2** को पेपर के साथ जारी किया गया था [भाषा मॉडल विकास के पैमाने पर प्रोटीन अनुक्रम सटीक संरचना भविष्यवाणी को सक्षम करते हैं](https://doi.org/10.1101/2022.07.20.500902) ज़ेमिंग लिन, हलील अकिन, रोशन राव, ब्रायन ही, झोंगकाई झू, वेंटिंग लू, ए द्वारा लान डॉस सैंटोस कोस्टा, मरियम फ़ज़ल-ज़रंडी, टॉम सर्कू, साल कैंडिडो, अलेक्जेंडर राइव्स। +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (CNRS से) साथ वाला पेपर [FlauBERT: Unsupervised Language Model Pre-training for फ़्रेंच](https://arxiv .org/abs/1912.05372) Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, बेंजामिन लेकोउटेक्स, अलेक्जेंड्रे अल्लाउज़ेन, बेनोइट क्रैबे, लॉरेंट बेसेसियर, डिडिएर श्वाब द्वारा। diff --git a/README_ja.md b/README_ja.md index 2306e2a1af..8c5b67adab 100644 --- a/README_ja.md +++ b/README_ja.md @@ -360,6 +360,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu から) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu から公開された研究論文: [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu から) Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. から公開された研究論文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (Meta AI から) はトランスフォーマープロテイン言語モデルです. **ESM-1b** は Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus から公開された研究論文: [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118). **ESM-1v** は Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives　から公開された研究論文: [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648). **ESM-2** と　**ESMFold** は Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives から公開された研究論文: [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (Google AI から) Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V から公開されたレポジトリー [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (CNRS から) Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab から公開された研究論文: [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) diff --git a/README_ko.md b/README_ko.md index 9081b83c70..53b619d118 100644 --- a/README_ko.md +++ b/README_ko.md @@ -275,6 +275,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (Baidu 에서) Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 의 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) 논문과 함께 발표했습니다. 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (Baidu 에서 제공)은 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang.의 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674)논문과 함께 발표했습니다. 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2** was released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab. diff --git a/README_zh-hans.md b/README_zh-hans.md index e325896353..d8b23fa833 100644 --- a/README_zh-hans.md +++ b/README_zh-hans.md @@ -299,6 +299,7 @@ conda install -c huggingface transformers 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (来自 Baidu) 伴随论文 [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu 发布。 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (来自 Baidu) 伴随论文 [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) 由 Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang 发布。 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2** was released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (来自 CNRS) 伴随论文 [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) 由 Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab 发布。 diff --git a/README_zh-hant.md b/README_zh-hant.md index d91e1c06d8..da52c58311 100644 --- a/README_zh-hant.md +++ b/README_zh-hant.md @@ -311,6 +311,7 @@ conda install -c huggingface transformers 1. **[ERNIE](https://huggingface.co/docs/transformers/model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu. 1. **[ErnieM](https://huggingface.co/docs/transformers/model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. 1. **[ESM](https://huggingface.co/docs/transformers/model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2** was released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](https://huggingface.co/docs/transformers/main/model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](https://huggingface.co/docs/transformers/model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](https://huggingface.co/docs/transformers/model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](https://huggingface.co/docs/transformers/model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab. diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml index 9e0d612377..0a98f79313 100644 --- a/docs/source/en/_toctree.yml +++ b/docs/source/en/_toctree.yml @@ -295,6 +295,8 @@ title: ErnieM - local: model_doc/esm title: ESM + - local: model_doc/falcon + title: Falcon - local: model_doc/flan-t5 title: FLAN-T5 - local: model_doc/flan-ul2 diff --git a/docs/source/en/index.md b/docs/source/en/index.md index cbde131668..d3632aea5c 100644 --- a/docs/source/en/index.md +++ b/docs/source/en/index.md @@ -115,6 +115,7 @@ The documentation is organized into five sections: 1. **[ERNIE](model_doc/ernie)** (from Baidu) released with the paper [ERNIE: Enhanced Representation through Knowledge Integration](https://arxiv.org/abs/1904.09223) by Yu Sun, Shuohuan Wang, Yukun Li, Shikun Feng, Xuyi Chen, Han Zhang, Xin Tian, Danxiang Zhu, Hao Tian, Hua Wu. 1. **[ErnieM](model_doc/ernie_m)** (from Baidu) released with the paper [ERNIE-M: Enhanced Multilingual Representation by Aligning Cross-lingual Semantics with Monolingual Corpora](https://arxiv.org/abs/2012.15674) by Xuan Ouyang, Shuohuan Wang, Chao Pang, Yu Sun, Hao Tian, Hua Wu, Haifeng Wang. 1. **[ESM](model_doc/esm)** (from Meta AI) are transformer protein language models. **ESM-1b** was released with the paper [Biological structure and function emerge from scaling unsupervised learning to 250 million protein sequences](https://www.pnas.org/content/118/15/e2016239118) by Alexander Rives, Joshua Meier, Tom Sercu, Siddharth Goyal, Zeming Lin, Jason Liu, Demi Guo, Myle Ott, C. Lawrence Zitnick, Jerry Ma, and Rob Fergus. **ESM-1v** was released with the paper [Language models enable zero-shot prediction of the effects of mutations on protein function](https://doi.org/10.1101/2021.07.09.450648) by Joshua Meier, Roshan Rao, Robert Verkuil, Jason Liu, Tom Sercu and Alexander Rives. **ESM-2 and ESMFold** were released with the paper [Language models of protein sequences at the scale of evolution enable accurate structure prediction](https://doi.org/10.1101/2022.07.20.500902) by Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Sal Candido, Alexander Rives. +1. **[Falcon](model_doc/falcon)** (from Technology Innovation Institute) by Almazrouei, Ebtesam and Alobeidli, Hamza and Alshamsi, Abdulaziz and Cappelli, Alessandro and Cojocaru, Ruxandra and Debbah, Merouane and Goffinet, Etienne and Heslow, Daniel and Launay, Julien and Malartic, Quentin and Noune, Badreddine and Pannier, Baptiste and Penedo, Guilherme. 1. **[FLAN-T5](model_doc/flan-t5)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-t5-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FLAN-UL2](model_doc/flan-ul2)** (from Google AI) released in the repository [google-research/t5x](https://github.com/google-research/t5x/blob/main/docs/models.md#flan-ul2-checkpoints) by Hyung Won Chung, Le Hou, Shayne Longpre, Barret Zoph, Yi Tay, William Fedus, Eric Li, Xuezhi Wang, Mostafa Dehghani, Siddhartha Brahma, Albert Webson, Shixiang Shane Gu, Zhuyun Dai, Mirac Suzgun, Xinyun Chen, Aakanksha Chowdhery, Sharan Narang, Gaurav Mishra, Adams Yu, Vincent Zhao, Yanping Huang, Andrew Dai, Hongkun Yu, Slav Petrov, Ed H. Chi, Jeff Dean, Jacob Devlin, Adam Roberts, Denny Zhou, Quoc V. Le, and Jason Wei 1. **[FlauBERT](model_doc/flaubert)** (from CNRS) released with the paper [FlauBERT: Unsupervised Language Model Pre-training for French](https://arxiv.org/abs/1912.05372) by Hang Le, Loïc Vial, Jibril Frej, Vincent Segonne, Maximin Coavoux, Benjamin Lecouteux, Alexandre Allauzen, Benoît Crabbé, Laurent Besacier, Didier Schwab. @@ -332,6 +333,7 @@ Flax), PyTorch, and/or TensorFlow. | ErnieM | ✅ | ❌ | ✅ | ❌ | ❌ | | ESM | ✅ | ❌ | ✅ | ✅ | ❌ | | FairSeq Machine-Translation | ✅ | ❌ | ✅ | ❌ | ❌ | +| Falcon | ❌ | ❌ | ✅ | ❌ | ❌ | | FlauBERT | ✅ | ❌ | ✅ | ✅ | ❌ | | FLAVA | ❌ | ❌ | ✅ | ❌ | ❌ | | FNet | ✅ | ✅ | ✅ | ❌ | ❌ | diff --git a/docs/source/en/model_doc/falcon.md b/docs/source/en/model_doc/falcon.md new file mode 100644 index 0000000000..8b6c655a4e --- /dev/null +++ b/docs/source/en/model_doc/falcon.md @@ -0,0 +1,66 @@ + + +# Falcon + +## Overview + +Falcon is a state-of-the-art language model trained on the [RefinedWeb dataset](https://arxiv.org/abs/2306.01116). At the time of writing, it is the leading model on the [OpenLLM leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard). + +There is no paper associated with Falcon yet, but for citation information please see [the repository for Falcon-40B](https://huggingface.co/tiiuae/falcon-40b#citation), the highest-performance Falcon model. + +- The model and tokenizer can be loaded via: + +```python +from transformers import AutoModelForCausalLM, AutoTokenizer + +tokenizer = AutoTokenizer.from_pretrained("tiiuae/falcon-40b-instruct") +model = AutoModelForCausalLM.from_pretrained("tiiuae/falcon-40b-instruct") + +inputs = tokenizer("What's the best way to divide a pizza between three people?", return_tensors="pt") +outputs = model.generate(**inputs, max_length=50) +``` + +- The Falcon tokenizer is a BPE model. + +## FalconConfig + +[[autodoc]] FalconConfig + +## FalconModel + +[[autodoc]] FalconModel + - forward + +## FalconForCausalLM + +[[autodoc]] FalconForCausalLM + - forward + +## FalconForSequenceClassification + +[[autodoc]] FalconForSequenceClassification + - forward + +## FalconForTokenClassification + +[[autodoc]] FalconForTokenClassification + - forward + +## FalconForQuestionAnswering + +[[autodoc]] FalconForQuestionAnswering + - forward \ No newline at end of file diff --git a/docs/source/en/tasks/language_modeling.md b/docs/source/en/tasks/language_modeling.md index 4986d90b28..172ac5786e 100644 --- a/docs/source/en/tasks/language_modeling.md +++ b/docs/source/en/tasks/language_modeling.md @@ -37,7 +37,7 @@ You can finetune other architectures for causal language modeling following the Choose one of the following architectures: -[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod) +[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod) diff --git a/docs/source/en/tasks/question_answering.md b/docs/source/en/tasks/question_answering.md index 3a3eda904e..6009cef493 100644 --- a/docs/source/en/tasks/question_answering.md +++ b/docs/source/en/tasks/question_answering.md @@ -35,7 +35,7 @@ The task illustrated in this tutorial is supported by the following model archit -[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [OpenAI GPT-2](../model_doc/gpt2), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [LXMERT](../model_doc/lxmert), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OPT](../model_doc/opt), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [Splinter](../model_doc/splinter), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) +[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [OpenAI GPT-2](../model_doc/gpt2), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [LXMERT](../model_doc/lxmert), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OPT](../model_doc/opt), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [Splinter](../model_doc/splinter), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) diff --git a/docs/source/en/tasks/sequence_classification.md b/docs/source/en/tasks/sequence_classification.md index bf07878a63..25649b03de 100644 --- a/docs/source/en/tasks/sequence_classification.md +++ b/docs/source/en/tasks/sequence_classification.md @@ -32,7 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit -[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) +[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) diff --git a/docs/source/en/tasks/token_classification.md b/docs/source/en/tasks/token_classification.md index 66d06b9f38..399b7c515d 100644 --- a/docs/source/en/tasks/token_classification.md +++ b/docs/source/en/tasks/token_classification.md @@ -32,7 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit -[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) +[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MRA](../model_doc/mra), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso) diff --git a/src/transformers/__init__.py b/src/transformers/__init__.py index ac9ab14f5d..82f461e0fd 100644 --- a/src/transformers/__init__.py +++ b/src/transformers/__init__.py @@ -294,6 +294,7 @@ _import_structure = { ], "models.ernie_m": ["ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP", "ErnieMConfig"], "models.esm": ["ESM_PRETRAINED_CONFIG_ARCHIVE_MAP", "EsmConfig", "EsmTokenizer"], + "models.falcon": ["FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP", "FalconConfig"], "models.flaubert": ["FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP", "FlaubertConfig", "FlaubertTokenizer"], "models.flava": [ "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP", @@ -1638,6 +1639,17 @@ else: "EsmPreTrainedModel", ] ) + _import_structure["models.falcon"].extend( + [ + "FALCON_PRETRAINED_MODEL_ARCHIVE_LIST", + "FalconForCausalLM", + "FalconForQuestionAnswering", + "FalconForSequenceClassification", + "FalconForTokenClassification", + "FalconModel", + "FalconPreTrainedModel", + ] + ) _import_structure["models.flaubert"].extend( [ "FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST", @@ -4188,6 +4200,7 @@ if TYPE_CHECKING: from .models.ernie import ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieConfig from .models.ernie_m import ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP, ErnieMConfig from .models.esm import ESM_PRETRAINED_CONFIG_ARCHIVE_MAP, EsmConfig, EsmTokenizer + from .models.falcon import FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP, FalconConfig from .models.flaubert import FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP, FlaubertConfig, FlaubertTokenizer from .models.flava import ( FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP, @@ -5349,6 +5362,15 @@ if TYPE_CHECKING: EsmModel, EsmPreTrainedModel, ) + from .models.falcon import ( + FALCON_PRETRAINED_MODEL_ARCHIVE_LIST, + FalconForCausalLM, + FalconForQuestionAnswering, + FalconForSequenceClassification, + FalconForTokenClassification, + FalconModel, + FalconPreTrainedModel, + ) from .models.flaubert import ( FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST, FlaubertForMultipleChoice, diff --git a/src/transformers/models/__init__.py b/src/transformers/models/__init__.py index c13b202975..7acaee7685 100644 --- a/src/transformers/models/__init__.py +++ b/src/transformers/models/__init__.py @@ -77,6 +77,7 @@ from . import ( ernie, ernie_m, esm, + falcon, flaubert, flava, fnet, diff --git a/src/transformers/models/auto/configuration_auto.py b/src/transformers/models/auto/configuration_auto.py index 70ad350565..3f5c88be52 100755 --- a/src/transformers/models/auto/configuration_auto.py +++ b/src/transformers/models/auto/configuration_auto.py @@ -85,6 +85,7 @@ CONFIG_MAPPING_NAMES = OrderedDict( ("ernie", "ErnieConfig"), ("ernie_m", "ErnieMConfig"), ("esm", "EsmConfig"), + ("falcon", "FalconConfig"), ("flaubert", "FlaubertConfig"), ("flava", "FlavaConfig"), ("fnet", "FNetConfig"), @@ -282,6 +283,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict( ("ernie", "ERNIE_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("ernie_m", "ERNIE_M_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("esm", "ESM_PRETRAINED_CONFIG_ARCHIVE_MAP"), + ("falcon", "FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("flaubert", "FLAUBERT_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("flava", "FLAVA_PRETRAINED_CONFIG_ARCHIVE_MAP"), ("fnet", "FNET_PRETRAINED_CONFIG_ARCHIVE_MAP"), @@ -475,6 +477,7 @@ MODEL_NAMES_MAPPING = OrderedDict( ("ernie", "ERNIE"), ("ernie_m", "ErnieM"), ("esm", "ESM"), + ("falcon", "Falcon"), ("flan-t5", "FLAN-T5"), ("flan-ul2", "FLAN-UL2"), ("flaubert", "FlauBERT"), diff --git a/src/transformers/models/auto/modeling_auto.py b/src/transformers/models/auto/modeling_auto.py index b72d263f9b..415ee5e20c 100755 --- a/src/transformers/models/auto/modeling_auto.py +++ b/src/transformers/models/auto/modeling_auto.py @@ -83,6 +83,7 @@ MODEL_MAPPING_NAMES = OrderedDict( ("ernie", "ErnieModel"), ("ernie_m", "ErnieMModel"), ("esm", "EsmModel"), + ("falcon", "FalconModel"), ("flaubert", "FlaubertModel"), ("flava", "FlavaModel"), ("fnet", "FNetModel"), @@ -380,6 +381,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict( ("data2vec-text", "Data2VecTextForCausalLM"), ("electra", "ElectraForCausalLM"), ("ernie", "ErnieForCausalLM"), + ("falcon", "FalconForCausalLM"), ("git", "GitForCausalLM"), ("gpt-sw3", "GPT2LMHeadModel"), ("gpt2", "GPT2LMHeadModel"), @@ -684,6 +686,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("ernie", "ErnieForSequenceClassification"), ("ernie_m", "ErnieMForSequenceClassification"), ("esm", "EsmForSequenceClassification"), + ("falcon", "FalconForSequenceClassification"), ("flaubert", "FlaubertForSequenceClassification"), ("fnet", "FNetForSequenceClassification"), ("funnel", "FunnelForSequenceClassification"), @@ -755,6 +758,7 @@ MODEL_FOR_QUESTION_ANSWERING_MAPPING_NAMES = OrderedDict( ("electra", "ElectraForQuestionAnswering"), ("ernie", "ErnieForQuestionAnswering"), ("ernie_m", "ErnieMForQuestionAnswering"), + ("falcon", "FalconForQuestionAnswering"), ("flaubert", "FlaubertForQuestionAnsweringSimple"), ("fnet", "FNetForQuestionAnswering"), ("funnel", "FunnelForQuestionAnswering"), @@ -842,6 +846,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict( ("ernie", "ErnieForTokenClassification"), ("ernie_m", "ErnieMForTokenClassification"), ("esm", "EsmForTokenClassification"), + ("falcon", "FalconForTokenClassification"), ("flaubert", "FlaubertForTokenClassification"), ("fnet", "FNetForTokenClassification"), ("funnel", "FunnelForTokenClassification"), diff --git a/src/transformers/models/bloom/modeling_bloom.py b/src/transformers/models/bloom/modeling_bloom.py index d37972a429..48a8db9186 100644 --- a/src/transformers/models/bloom/modeling_bloom.py +++ b/src/transformers/models/bloom/modeling_bloom.py @@ -135,7 +135,7 @@ def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: x (`torch.tensor`, *required*): input tensor residual (`torch.tensor`, *required*): - esidual tensor + residual tensor prob (`float`, *required*): dropout probability training (`bool`, *required*): diff --git a/src/transformers/models/falcon/__init__.py b/src/transformers/models/falcon/__init__.py new file mode 100644 index 0000000000..070e0cc033 --- /dev/null +++ b/src/transformers/models/falcon/__init__.py @@ -0,0 +1,68 @@ +# coding=utf-8 +# Copyright 2023 the Falcon authors and HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +from typing import TYPE_CHECKING + +from ...utils import ( + OptionalDependencyNotAvailable, + _LazyModule, + is_torch_available, +) + + +_import_structure = { + "configuration_falcon": ["FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP", "FalconConfig"], +} + +try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() +except OptionalDependencyNotAvailable: + pass +else: + _import_structure["modeling_falcon"] = [ + "FALCON_PRETRAINED_MODEL_ARCHIVE_LIST", + "FalconForCausalLM", + "FalconModel", + "FalconPreTrainedModel", + "FalconForSequenceClassification", + "FalconForTokenClassification", + "FalconForQuestionAnswering", + ] + + +if TYPE_CHECKING: + from .configuration_falcon import FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP, FalconConfig + + try: + if not is_torch_available(): + raise OptionalDependencyNotAvailable() + except OptionalDependencyNotAvailable: + pass + else: + from .modeling_falcon import ( + FALCON_PRETRAINED_MODEL_ARCHIVE_LIST, + FalconForCausalLM, + FalconForQuestionAnswering, + FalconForSequenceClassification, + FalconForTokenClassification, + FalconModel, + FalconPreTrainedModel, + ) + + +else: + import sys + + sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__) diff --git a/src/transformers/models/falcon/configuration_falcon.py b/src/transformers/models/falcon/configuration_falcon.py new file mode 100644 index 0000000000..b239f88799 --- /dev/null +++ b/src/transformers/models/falcon/configuration_falcon.py @@ -0,0 +1,147 @@ +# coding=utf-8 +# Copyright 2023 the Falcon authors and HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Falcon configuration""" +from ...configuration_utils import PretrainedConfig +from ...utils import logging + + +logger = logging.get_logger(__name__) + +FALCON_PRETRAINED_CONFIG_ARCHIVE_MAP = { + "tiiuae/falcon-40b": "https://huggingface.co/tiiuae/falcon-40b/resolve/main/config.json", + "tiiuae/falcon-7b": "https://huggingface.co/tiiuae/falcon-7b/resolve/main/config.json", +} + + +class FalconConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`FalconModel`]. It is used to instantiate a Falcon + 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 + [tiiuae/falcon-7b](https://huggingface.co/tiiuae/falcon-7b) architecture. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + + Args: + vocab_size (`int`, *optional*, defaults to 65024): + Vocabulary size of the Falcon model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`FalconModel`] + hidden_size (`int`, *optional*, defaults to 4544): + Dimension of the hidden representations. + num_hidden_layers (`int`, *optional*, defaults to 32): + Number of hidden layers in the Transformer decoder. + num_attention_heads (`int`, *optional*, defaults to 71): + Number of attention heads for each attention layer in the Transformer encoder. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + use_cache (`bool`, *optional*, defaults to `True`): + Whether the model should return the last key/values attentions (not used by all models). Only relevant if + `config.is_decoder=True`. + layer_norm_epsilon (`float`, *optional*, defaults to 1e-5): + The epsilon used by the layer normalization layers. + hidden_dropout (`float`, *optional*, defaults to 0.0): + The dropout probability for MLP layers. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout probability for attention layers. + num_kv_heads (`int`, *optional*): + Number of key-value heads to use per attention layer. If unset, defaults to the same value as + `num_attention_heads`. + alibi (`bool`, *optional*, defaults to `False`): + Whether to use ALiBi positional biases during self-attention. + new_decoder_architecture (`bool`, *optional*, defaults to `False`): + Whether to use the new (Falcon-40B) decoder architecture. If `True`, the `multi_query` and `parallel_attn` + arguments are ignored, as the new decoder always uses parallel attention. + multi_query (`bool`, *optional*, defaults to `True`): + Whether to use multi-query attention in the decoder. Ignored when `new_decoder_architecture` is `True`. + parallel_attn (`bool`, *optional*, defaults to `True`): + Whether to compute attention in parallel with the feedforward layer. If False, they are consecutive + instead, as in the original Transformer architecture. Ignored when `new_decoder_architecture` is `True`. + bias (`bool`, *optional*, defaults to `False`): + Whether to use bias on Linear layers. + bos_token_id (`int`, *optional*, defaults to 11): + The id of the "beginning-of-sequence" token. + eos_token_id (`int`, *optional*, defaults to 11): + The id of the "end-of-sequence" token. + + Example: + + ```pytho + >>> from transformers import FalconModel, FalconConfig + + >>> # Initializing a small (2-layer) Falcon configuration + >>> configuration = FalconConfig(num_hidden_layers=2) + + >>> # Initializing a model from the small configuration + >>> model = FalconModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + model_type = "falcon" + keys_to_ignore_at_inference = ["past_key_values"] + + def __init__( + self, + vocab_size=65024, + hidden_size=4544, + num_hidden_layers=32, + num_attention_heads=71, + layer_norm_epsilon=1e-5, + initializer_range=0.02, + use_cache=True, + hidden_dropout=0.0, + attention_dropout=0.0, + num_kv_heads=None, + alibi=False, + new_decoder_architecture=False, + multi_query=True, + parallel_attn=True, + bias=False, + bos_token_id=11, + eos_token_id=11, + **kwargs, + ): + self.vocab_size = vocab_size + # Backward compatibility with n_embed kwarg + n_embed = kwargs.pop("n_embed", None) + self.hidden_size = hidden_size if n_embed is None else n_embed + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + self.layer_norm_epsilon = layer_norm_epsilon + self.initializer_range = initializer_range + self.use_cache = use_cache + self.hidden_dropout = hidden_dropout + self.attention_dropout = attention_dropout + + self.bos_token_id = bos_token_id + self.eos_token_id = eos_token_id + self.num_kv_heads = num_attention_heads if num_kv_heads is None else num_kv_heads + self.alibi = alibi + self.new_decoder_architecture = new_decoder_architecture + self.multi_query = multi_query # Ignored when new_decoder_architecture is True + self.parallel_attn = parallel_attn + self.bias = bias + + super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs) + + @property + def head_dim(self): + return self.hidden_size // self.num_attention_heads + + @property + def rotary(self): + return not self.alibi diff --git a/src/transformers/models/falcon/modeling_falcon.py b/src/transformers/models/falcon/modeling_falcon.py new file mode 100644 index 0000000000..9ab9030a54 --- /dev/null +++ b/src/transformers/models/falcon/modeling_falcon.py @@ -0,0 +1,1262 @@ +# coding=utf-8 +# Copyright 2023 the Falcon authors and HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""PyTorch Falcon model.""" + +import math +from typing import Optional, Tuple, Union + +import torch +import torch.utils.checkpoint +from torch import nn +from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, LayerNorm, MSELoss +from torch.nn import functional as F + +from ...modeling_outputs import ( + BaseModelOutputWithPastAndCrossAttentions, + CausalLMOutputWithCrossAttentions, + QuestionAnsweringModelOutput, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from ...modeling_utils import PreTrainedModel +from ...utils import add_code_sample_docstrings, add_start_docstrings, add_start_docstrings_to_model_forward, logging +from .configuration_falcon import FalconConfig + + +logger = logging.get_logger(__name__) + +FALCON_PRETRAINED_MODEL_ARCHIVE_LIST = [ + "tiiuae/falcon-40b", + "tiiuae/falcon-40b-instruct", + "tiiuae/falcon-7b", + "tiiuae/falcon-7b-instruct", + "tiiuae/falcon-rw-7b", + "tiiuae/falcon-rw-1b", +] +_CHECKPOINT_FOR_DOC = "Rocketknight1/falcon-rw-1b" +_CONFIG_FOR_DOC = "FalconConfig" + + +# NOTE(Hesslow): Unfortunately we did not fuse matmul and bias during training, this means that there's one additional quantization to bfloat16 between the operations. +# In order not to degrade the quality of our HF-port, we keep these characteristics in the final model. +class FalconLinear(nn.Linear): + def forward(self, input: torch.Tensor) -> torch.Tensor: + hidden_states = input @ self.weight.T + if self.bias is None: + return hidden_states + return hidden_states + self.bias + + +# rotary pos emb helpers (torch.jit.script does not seem to support staticmethod...) +def rotate_half(x): + x1, x2 = x[..., : x.shape[-1] // 2], x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +class FalconRotaryEmbedding(nn.Module): + """Implementation of RotaryEmbedding from GPT-NeoX. + This implementation is designed to operate on queries and keys that are compatible with `[batch_size, + n_heads_per_partition, seq_len, head_dim]` (e.g. MinGPTAttention format). + """ + + def __init__(self, head_dim: int, base=10000): + super().__init__() + inv_freq = 1.0 / (base ** (torch.arange(0, head_dim, 2).float() / head_dim)) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.head_dim = head_dim + self.seq_len_cached = -1 + self.cos_cached: torch.Tensor | None = None + self.sin_cached: torch.Tensor | None = None + + def cos_sin(self, seq_len: int, past_key_values_length: int, device="cpu", dtype=torch.bfloat16) -> torch.Tensor: + total_length = seq_len + past_key_values_length + if total_length > self.seq_len_cached: + self.seq_len_cached = total_length + t = torch.arange(total_length, device=device, dtype=self.inv_freq.dtype) + freqs = torch.einsum("i,j->ij", t, self.inv_freq) + emb = torch.cat((freqs, freqs), dim=-1).to(device) + + if dtype in [torch.float16, torch.bfloat16]: + emb = emb.float() + + self.cos_cached = emb.cos()[None, :, :] + self.sin_cached = emb.sin()[None, :, :] + + self.cos_cached = self.cos_cached.type(dtype) + self.sin_cached = self.sin_cached.type(dtype) + + return ( + self.cos_cached[:, past_key_values_length : seq_len + past_key_values_length], + self.sin_cached[:, past_key_values_length : seq_len + past_key_values_length], + ) + + def forward(self, query, key, past_key_values_length=0): + batch, seq_len, head_dim = query.shape + cos, sin = self.cos_sin(seq_len, past_key_values_length, query.device, query.dtype) + return (query * cos) + (rotate_half(query) * sin), (key * cos) + (rotate_half(key) * sin) + + +def _make_causal_mask( + input_ids_shape: torch.Size, device: torch.device, past_key_values_length: int +) -> torch.BoolTensor: + """ + Make causal mask used for self-attention. This mask does not take the existing attention mask into account - it + just blocks tokens from attending forwards in the sequence. The output shape will be `[batch_size, 1, + target_length, target_length+past_key_values_length]`. + """ + batch_size, target_length = input_ids_shape + + mask = torch.triu(torch.ones((target_length, target_length), dtype=torch.bool, device=device), diagonal=1) + # If past_key_values_length is 0 this is an empty tensor and the concatenation is a no-op. + # This code style is an unfortunate consequence of getting your TF engineer to port models; doing it this + # way avoids a data-dependent conditional, which will help me when I have to port this to XLA later. + past_mask = torch.zeros((target_length, past_key_values_length), dtype=torch.bool, device=device) + mask = torch.cat([past_mask, mask], dim=-1) + expanded_mask = mask[None, None, :, :].expand(batch_size, 1, target_length, target_length + past_key_values_length) + return expanded_mask + + +def _expand_mask(mask: torch.Tensor, past_key_values_length: int) -> torch.BoolTensor: + """ + Expands attention_mask from `[batch_size, seq_length]` to `[batch_size, 1, seq_length, seq_length + past_length]`. + """ + batch_size, total_length = mask.shape + seq_length = total_length - past_key_values_length if past_key_values_length is not None else total_length + + expanded_mask = ~(mask[:, None, None, :].to(torch.bool)) + return expanded_mask.expand(batch_size, 1, seq_length, total_length) + + +def build_alibi_tensor(attention_mask: torch.Tensor, num_heads: int, dtype: torch.dtype) -> torch.Tensor: + batch_size, seq_length = attention_mask.shape + closest_power_of_2 = 2 ** math.floor(math.log2(num_heads)) + base = torch.tensor( + 2 ** (-(2 ** -(math.log2(closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32 + ) + powers = torch.arange(1, 1 + closest_power_of_2, device=attention_mask.device, dtype=torch.int32) + slopes = torch.pow(base, powers) + + if closest_power_of_2 != num_heads: + extra_base = torch.tensor( + 2 ** (-(2 ** -(math.log2(2 * closest_power_of_2) - 3))), device=attention_mask.device, dtype=torch.float32 + ) + num_remaining_heads = min(closest_power_of_2, num_heads - closest_power_of_2) + extra_powers = torch.arange(1, 1 + 2 * num_remaining_heads, 2, device=attention_mask.device, dtype=torch.int32) + slopes = torch.cat([slopes, torch.pow(extra_base, extra_powers)], dim=0) + + # Note: alibi will added to the attention bias that will be applied to the query, key product of attention + # => therefore alibi will have to be of shape (batch_size, num_heads, query_length, key_length) + # => here we set (batch_size=1, num_heads=num_heads, query_length=1, key_length=max_length) + # => the query_length dimension will then be broadcasted correctly + # This is more or less identical to T5's relative position bias: + # https://github.com/huggingface/transformers/blob/f681437203baa7671de3174b0fa583c349d9d5e1/src/transformers/models/t5/modeling_t5.py#L527 + arange_tensor = ((attention_mask.cumsum(dim=-1) - 1) * attention_mask)[:, None, :] + alibi = slopes[..., None].bfloat16() * arange_tensor + return alibi.reshape(batch_size * num_heads, 1, seq_length).to(dtype) + + +# Copied from transformers.models.bloom.modeling_bloom.dropout_add +def dropout_add(x: torch.Tensor, residual: torch.Tensor, prob: float, training: bool) -> torch.Tensor: + """ + Dropout add function + + Args: + x (`torch.tensor`, *required*): + input tensor + residual (`torch.tensor`, *required*): + residual tensor + prob (`float`, *required*): + dropout probability + training (`bool`, *required*): + training mode + """ + out = F.dropout(x, p=prob, training=training) + out = residual + out + return out + + +class FalconAttention(nn.Module): + def __init__(self, config: FalconConfig): + super().__init__() + + self.hidden_size = config.hidden_size + self.num_heads = config.num_attention_heads + self.head_dim = self.hidden_size // self.num_heads + self.split_size = self.hidden_size + self.hidden_dropout = config.hidden_dropout + + if self.head_dim * self.num_heads != self.hidden_size: + raise ValueError( + f"`hidden_size` must be divisible by num_heads (got `hidden_size`: {self.hidden_size} and `num_heads`:" + f" {self.num_heads})." + ) + + self.maybe_rotary = FalconRotaryEmbedding(config.head_dim) if config.rotary else lambda q, k, t: (q, k) + + # Layer-wise attention scaling + self.inv_norm_factor = 1.0 / math.sqrt(self.head_dim) + self.beta = self.inv_norm_factor + if config.new_decoder_architecture: + qkv_out_dim = (config.num_kv_heads * 2 + config.num_attention_heads) * self.head_dim + elif config.multi_query: + qkv_out_dim = self.hidden_size + 2 * self.head_dim + else: + qkv_out_dim = 3 * self.hidden_size + self.query_key_value = FalconLinear(self.hidden_size, qkv_out_dim, bias=config.bias) + self.new_decoder_architecture = config.new_decoder_architecture + self.multi_query = config.multi_query + self.dense = FalconLinear(self.hidden_size, self.hidden_size, bias=config.bias) + self.attention_dropout = nn.Dropout(config.attention_dropout) + self.num_kv_heads = config.num_kv_heads if (self.new_decoder_architecture or not self.multi_query) else 1 + + def _split_heads(self, fused_qkv: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]: + """ + Split the last dimension into (num_heads, head_dim), results share same memory storage as `fused_qkv` + + Args: + fused_qkv (`torch.tensor`, *required*): [batch_size, seq_length, num_heads * 3 * head_dim] + + Returns: + query: [batch_size, seq_length, num_heads, head_dim] key: [batch_size, seq_length, num_heads, head_dim] + value: [batch_size, seq_length, num_heads, head_dim] + """ + if self.new_decoder_architecture: + batch, seq_len, _ = fused_qkv.shape + qkv = fused_qkv.view(batch, seq_len, -1, self.num_heads // self.num_kv_heads + 2, self.head_dim) + query = qkv[:, :, :, :-2] + key = qkv[:, :, :, [-2]] + value = qkv[:, :, :, [-1]] + key = torch.broadcast_to(key, query.shape) + value = torch.broadcast_to(value, query.shape) + + query, key, value = [x.flatten(2, 3) for x in (query, key, value)] + return query, key, value + elif not self.multi_query: + batch_size, seq_length, three_times_hidden_size = fused_qkv.shape + fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads, 3, self.head_dim) + return fused_qkv[..., 0, :], fused_qkv[..., 1, :], fused_qkv[..., 2, :] + else: + batch_size, seq_length, three_times_hidden_size = fused_qkv.shape + fused_qkv = fused_qkv.view(batch_size, seq_length, self.num_heads + 2, self.head_dim) + return fused_qkv[..., :-2, :], fused_qkv[..., [-2], :], fused_qkv[..., [-1], :] + + # Copied from transformers.models.bloom.modeling_bloom.BloomAttention._merge_heads + def _merge_heads(self, x: torch.Tensor) -> torch.Tensor: + """ + Merge heads together over the last dimenstion + + Args: + x (`torch.tensor`, *required*): [batch_size * num_heads, seq_length, head_dim] + + Returns: + torch.tensor: [batch_size, seq_length, num_heads * head_dim] + """ + # What we want to achieve is: + # batch_size * num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads * head_dim + batch_size_and_num_heads, seq_length, _ = x.shape + batch_size = batch_size_and_num_heads // self.num_heads + + # First view to decompose the batch size + # batch_size * num_heads, seq_length, head_dim -> batch_size, num_heads, seq_length, head_dim + x = x.view(batch_size, self.num_heads, seq_length, self.head_dim) + + # batch_size, num_heads, seq_length, head_dim -> batch_size, seq_length, num_heads, head_dim + x = x.permute(0, 2, 1, 3) + + # batch_size, seq_length, num_heads, head_dim -> batch_size, seq_length, num_heads * head_dim + return x.reshape(batch_size, seq_length, self.num_heads * self.head_dim) + + def forward( + self, + hidden_states: torch.Tensor, + alibi: Optional[torch.Tensor], + attention_mask: torch.Tensor, + layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + head_mask: Optional[torch.Tensor] = None, + use_cache: bool = False, + output_attentions: bool = False, + ): + fused_qkv = self.query_key_value(hidden_states) # [batch_size, seq_length, 3 x hidden_size] + num_kv_heads = self.num_heads if self.new_decoder_architecture else self.num_kv_heads + # 3 x [batch_size, seq_length, num_heads, head_dim] + (query_layer, key_layer, value_layer) = self._split_heads(fused_qkv) + + batch_size, query_length, _, _ = query_layer.shape + + query_layer = query_layer.transpose(1, 2).reshape(batch_size * self.num_heads, query_length, self.head_dim) + key_layer = key_layer.transpose(1, 2).reshape( + batch_size * num_kv_heads, + query_length, + self.head_dim, + ) + value_layer = value_layer.transpose(1, 2).reshape(batch_size * num_kv_heads, query_length, self.head_dim) + + past_kv_length = 0 if layer_past is None else layer_past[0].shape[1] + query_layer, key_layer = self.maybe_rotary(query_layer, key_layer, past_kv_length) + + if layer_past is not None: + past_key, past_value = layer_past + # concatenate along seq_length dimension: + # - key: [batch_size * self.num_heads, kv_length, head_dim] + # - value: [batch_size * self.num_heads, kv_length, head_dim] + key_layer = torch.cat((past_key, key_layer), dim=1) + value_layer = torch.cat((past_value, value_layer), dim=1) + + _, kv_length, _ = key_layer.shape + if use_cache: + present = (key_layer, value_layer) + else: + present = None + + attention_mask_float = (attention_mask * 1.0).masked_fill(attention_mask, float("-1e9")).to(query_layer.dtype) + + query_layer_ = query_layer.reshape(batch_size, self.num_heads, -1, self.head_dim) + key_layer_ = key_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim) + value_layer_ = value_layer.reshape(batch_size, num_kv_heads, -1, self.head_dim) + + if alibi is None: + if output_attentions: + # F.scaled_dot_product_attention doesn't return the attention weights, so we have + # to do it by hand if we want them + attention_scores = query_layer_ @ key_layer_.transpose(-1, -2) + attention_scores /= math.sqrt(self.head_dim) + + attention_scores = F.softmax( + attention_scores + attention_mask_float, dim=-1, dtype=hidden_states.dtype + ) + attn_output = attention_scores @ value_layer_ + else: + attn_output = F.scaled_dot_product_attention( + query_layer_, key_layer_, value_layer_, attention_mask_float, 0.0, is_causal=False + ) + attention_scores = None + + attn_output = attn_output.view(batch_size, self.num_heads, query_length, self.head_dim) + attn_output = attn_output.permute(0, 2, 1, 3) + attn_output = attn_output.reshape(batch_size, query_length, self.num_heads * self.head_dim) + + output_tensor = self.dense(attn_output) + + if output_attentions: + return output_tensor, present, attention_scores + else: + return output_tensor, present + + else: + matmul_result = query_layer_ @ key_layer_.transpose(-1, -2) + + # change view to [batch_size, num_heads, q_length, kv_length] + attention_scores = matmul_result.view(batch_size, self.num_heads, query_length, kv_length) + + # cast attention scores to fp32, compute scaled softmax and cast back to initial dtype - [batch_size, num_heads, q_length, kv_length] + input_dtype = attention_scores.dtype + # `float16` has a minimum value of -65504.0, whereas `bfloat16` and `float32` have a minimum value of `-3.4e+38` + if input_dtype == torch.float16 or input_dtype == torch.bfloat16: + attention_scores = attention_scores.to(torch.float32) + # Matt (HF) note: We could possibly use F.scaled_dot_product_attention here too, by + # adding (alibi * self.inv_norm_factor) to attention_mask_float. I think this would be mathematically + # equivalent and more performant, but there might be a numerical difference. If you're reading this + # and you'd like to experiment and maybe file a PR, feel free! + attention_logits = attention_scores + alibi.view(batch_size, self.num_heads, 1, -1) + attention_logits *= self.inv_norm_factor + attention_probs = F.softmax(attention_logits + attention_mask_float, dim=-1, dtype=hidden_states.dtype) + # [batch_size, num_heads, q_length, kv_length] + attention_probs = self.attention_dropout(attention_probs) + + if head_mask is not None: + attention_probs = attention_probs * head_mask + + # change view [batch_size, num_heads, q_length, kv_length] + attention_probs_reshaped = attention_probs.view(batch_size, self.num_heads, query_length, kv_length) + + # matmul: [batch_size * num_heads, q_length, head_dim] + context_layer = (attention_probs_reshaped @ value_layer_).flatten(0, 1) + + # change view [batch_size, num_heads, q_length, head_dim] + context_layer = self._merge_heads(context_layer) + + output_tensor = self.dense(context_layer) + + if output_attentions: + return output_tensor, present, attention_probs + else: + return output_tensor, present + + +class FalconMLP(nn.Module): + def __init__(self, config: FalconConfig): + super().__init__() + hidden_size = config.hidden_size + + self.dense_h_to_4h = FalconLinear(hidden_size, 4 * hidden_size, bias=config.bias) + self.act = nn.GELU() + self.dense_4h_to_h = FalconLinear(4 * hidden_size, hidden_size, bias=config.bias) + self.hidden_dropout = config.hidden_dropout + + def forward(self, x: torch.Tensor) -> torch.Tensor: + x = self.act(self.dense_h_to_4h(x)) + x = self.dense_4h_to_h(x) + return x + + +class FalconDecoderLayer(nn.Module): + def __init__(self, config: FalconConfig): + super().__init__() + hidden_size = config.hidden_size + self.num_heads = config.num_attention_heads + self.self_attention = FalconAttention(config) + self.mlp = FalconMLP(config) + self.hidden_dropout = config.hidden_dropout + self.config = config + + if config.new_decoder_architecture: + # The layer norm before self-attention + self.ln_attn = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) + # The layer norm before the MLP + self.ln_mlp = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) + else: + self.input_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) + if not config.parallel_attn: + self.post_attention_layernorm = LayerNorm(hidden_size, eps=config.layer_norm_epsilon) + + def forward( + self, + hidden_states: torch.Tensor, + alibi: Optional[torch.Tensor], + attention_mask: torch.Tensor, + layer_past: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, + head_mask: Optional[torch.Tensor] = None, + use_cache: bool = False, + output_attentions: bool = False, + ): + residual = hidden_states + + if self.config.new_decoder_architecture: + attention_layernorm_out = self.ln_attn(hidden_states) + mlp_layernorm_out = self.ln_mlp(hidden_states) + else: + attention_layernorm_out = self.input_layernorm(hidden_states) + + # Self attention. + attn_outputs = self.self_attention( + attention_layernorm_out, + layer_past=layer_past, + attention_mask=attention_mask, + alibi=alibi, + head_mask=head_mask, + use_cache=use_cache, + output_attentions=output_attentions, + ) + + attention_output = attn_outputs[0] + + if not self.config.new_decoder_architecture: + if self.config.parallel_attn: + mlp_layernorm_out = attention_layernorm_out + else: + residual = dropout_add( + attention_output, residual, self.config.attention_dropout, training=self.training + ) + mlp_layernorm_out = self.post_attention_layernorm(residual) + + outputs = attn_outputs[1:] + + # MLP. + mlp_output = self.mlp(mlp_layernorm_out) + + if self.config.new_decoder_architecture or self.config.parallel_attn: + mlp_output += attention_output + + output = dropout_add(mlp_output, residual, self.config.hidden_dropout, training=self.training) + + if use_cache: + outputs = (output,) + outputs + else: + outputs = (output,) + outputs[1:] + + return outputs # hidden_states, present, attentions + + +FALCON_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 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 ([`FalconConfig`]): 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. +""" + +FALCON_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, input_ids_length)`): + `input_ids_length` = `sequence_length` if `past_key_values` is `None` else `past_key_values[0][0].shape[2]` + (`sequence_length` of input past key value states). Indices of input sequence tokens in the vocabulary. + + If `past_key_values` is used, only `input_ids` that do not have their past calculated should be passed as + `input_ids`. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + past_key_values (`Tuple[Tuple[torch.Tensor]]` of length `config.num_hidden_layers`): + Contains precomputed hidden-states (key and values in the attention blocks) as computed by the model (see + `past_key_values` output below). Can be used to speed up sequential decoding. The `input_ids` which have + their past given to this model should not be passed as `input_ids` as they have already been computed. + + Each element of `past_key_values` is a tuple (past_key, past_value): + - past_key: [batch_size * num_heads, head_dim, kv_length] + - past_value: [batch_size * num_heads, kv_length, head_dim] + attention_mask (`torch.FloatTensor` 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) + head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): + Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + + 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. + + If `past_key_values` is used, optionally only the last `inputs_embeds` have to be input (see + `past_key_values`). + 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 [`~file_utils.ModelOutput`] instead of a plain tuple. +""" + + +class FalconPreTrainedModel(PreTrainedModel): + """ + An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained + models. + """ + + config_class = FalconConfig + base_model_prefix = "transformer" + supports_gradient_checkpointing = True + _no_split_modules = ["FalconDecoderLayer"] + + def __init__(self, *inputs, **kwargs): + super().__init__(*inputs, **kwargs) + + def _init_weights(self, module: nn.Module): + """Initialize the weights.""" + if isinstance(module, nn.Linear) or isinstance(module, FalconLinear): + # Slightly different from the TF version which uses truncated_normal for initialization + # cf https://github.com/pytorch/pytorch/pull/5617 + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=self.config.initializer_range) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + elif isinstance(module, LayerNorm): + module.bias.data.zero_() + module.weight.data.fill_(1.0) + + # Copied from transformers.models.bloom.modeling_bloom.BloomPreTrainedModel._set_gradient_checkpointing with BloomModel->FalconModel + def _set_gradient_checkpointing(self, module: nn.Module, value: bool = False): + if isinstance(module, FalconModel): + module.gradient_checkpointing = value + + @staticmethod + def _convert_cache_to_standard_format( + past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]], batch_size: int + ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]: + """ + Standardizes the format of the cache so as to match most implementations, i.e. to tuple(tuple([batch_size, + num_heads, ...])) + """ + batch_size_times_num_heads, kv_length, head_dim = past_key_value[0][0].shape + # [batch_size * self.num_heads, kv_length, head_dim] -> [batch_size, num_heads, kv_length, head_dim] + # Note that don't want to use self.num_attention_heads because the number of heads may vary depending + # on whether we use multi_query attention. + num_heads = batch_size_times_num_heads // batch_size + return tuple( + ( + layer_past[0].view(batch_size, num_heads, kv_length, head_dim), + layer_past[1].view(batch_size, num_heads, kv_length, head_dim), + ) + for layer_past in past_key_value + ) + + @staticmethod + def _convert_to_rw_cache( + past_key_value: Tuple[Tuple[torch.Tensor, torch.Tensor]] + ) -> Tuple[Tuple[torch.Tensor, torch.Tensor]]: + batch_size, num_heads, kv_length, head_dim = past_key_value[0][0].shape + batch_size_times_num_heads = batch_size * num_heads + # [batch_size, num_heads, kv_length, head_dim] -> [batch_size * num_heads, kv_length, head_dim] + return tuple( + ( + layer_past[0].view(batch_size_times_num_heads, kv_length, head_dim), + layer_past[1].view(batch_size_times_num_heads, kv_length, head_dim), + ) + for layer_past in past_key_value + ) + + +@add_start_docstrings( + "The bare Falcon Model transformer outputting raw hidden-states without any specific head on top.", + FALCON_START_DOCSTRING, +) +class FalconModel(FalconPreTrainedModel): + def __init__(self, config: FalconConfig): + super().__init__(config) + + self.embed_dim = config.hidden_size + self.num_heads = config.num_attention_heads + self.use_alibi = config.alibi + + # Embedding + LN Embedding + self.word_embeddings = nn.Embedding(config.vocab_size, self.embed_dim) + + # Transformer blocks + self.h = nn.ModuleList([FalconDecoderLayer(config) for _ in range(config.num_hidden_layers)]) + + # Final Layer Norm + self.ln_f = LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon) + + self.gradient_checkpointing = False + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.word_embeddings + + @staticmethod + def _prepare_attn_mask( + attention_mask: torch.Tensor, input_shape: Tuple[int, int], past_key_values_length: int + ) -> torch.BoolTensor: + # Create a causal mask + # The attention mask we receive as input should cover the whole extended sequence, including any past + # cache, so its shape should be [batch_size, seq_length + past_key_values_length] + # The output shape will be [batch_size, 1, seq_length, seq_length + past_key_values_length] + if input_shape[1] + past_key_values_length != attention_mask.shape[1]: + raise ValueError( + "Attention mask shape should be (batch_size, seq_length + past_key_values_length)" + f" but is {attention_mask.shape} with input_ids shape {input_shape} and past length" + f" {past_key_values_length}." + ) + combined_attention_mask = None + device = attention_mask.device + _, seq_length = input_shape + + if seq_length > 1: + combined_attention_mask = _make_causal_mask( + input_shape, device=device, past_key_values_length=past_key_values_length + ) + + # [batch_size, seq_length + past_key_values_length] -> [batch_size, 1, seq_length, seq_length + past_key_values_length] + expanded_attn_mask = _expand_mask(attention_mask, past_key_values_length=past_key_values_length) + combined_attention_mask = ( + expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask | combined_attention_mask + ) + + return combined_attention_mask + + def set_input_embeddings(self, new_embeddings: torch.Tensor): + self.word_embeddings = new_embeddings + + @add_start_docstrings_to_model_forward(FALCON_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=BaseModelOutputWithPastAndCrossAttentions, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.LongTensor] = None, + inputs_embeds: 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[torch.Tensor, ...], BaseModelOutputWithPastAndCrossAttentions]: + 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 + + if input_ids is not None and inputs_embeds is not None: + raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time") + elif input_ids is not None: + batch_size, seq_length = input_ids.shape + elif inputs_embeds is not None: + batch_size, seq_length, _ = inputs_embeds.shape + else: + raise ValueError("You have to specify either input_ids or inputs_embeds") + + if past_key_values is None: + past_key_values = tuple([None] * len(self.h)) + else: + past_key_values = self._convert_to_rw_cache(past_key_values) + + # Prepare head mask if needed + # 1.0 in head_mask indicate we keep the head + # attention_probs has shape batch_size x num_heads x N x N + # head_mask has shape n_layer x batch x num_heads x N x N + head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers) + + if inputs_embeds is None: + inputs_embeds = self.word_embeddings(input_ids) + + hidden_states = inputs_embeds + + presents = () if use_cache else None + all_self_attentions = () if output_attentions else None + all_hidden_states = () if output_hidden_states else None + + # Compute alibi tensor: check build_alibi_tensor documentation + past_key_values_length = 0 + if past_key_values[0] is not None: + past_key_values_length = past_key_values[0][0].shape[1] # 1 because RW-cache, not standard format + if attention_mask is None: + attention_mask = torch.ones((batch_size, seq_length + past_key_values_length), device=hidden_states.device) + else: + attention_mask = attention_mask.to(hidden_states.device) + + if self.use_alibi: + alibi = build_alibi_tensor(attention_mask, self.num_heads, dtype=hidden_states.dtype) + else: + alibi = None + + causal_mask = self._prepare_attn_mask( + attention_mask, + input_shape=(batch_size, seq_length), + past_key_values_length=past_key_values_length, + ) + + for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)): + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if self.gradient_checkpointing and self.training: + if use_cache: + logger.warning( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..." + ) + use_cache = False + + def create_custom_forward(module): + def custom_forward(*inputs): + # None for past_key_value + return module(*inputs, use_cache=use_cache, output_attentions=output_attentions) + + return custom_forward + + outputs = torch.utils.checkpoint.checkpoint( + create_custom_forward(block), + hidden_states, + alibi, + causal_mask, + head_mask[i], + ) + else: + outputs = block( + hidden_states, + layer_past=layer_past, + attention_mask=causal_mask, + head_mask=head_mask[i], + use_cache=use_cache, + output_attentions=output_attentions, + alibi=alibi, + ) + + hidden_states = outputs[0] + if use_cache is True: + presents = presents + (outputs[1],) + + if output_attentions: + all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],) + + # Add last hidden state + hidden_states = self.ln_f(hidden_states) + + if output_hidden_states: + all_hidden_states = all_hidden_states + (hidden_states,) + + if presents is not None: + presents = self._convert_cache_to_standard_format(presents, batch_size) + + if not return_dict: + return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions] if v is not None) + + return BaseModelOutputWithPastAndCrossAttentions( + last_hidden_state=hidden_states, + past_key_values=presents, + hidden_states=all_hidden_states, + attentions=all_self_attentions, + ) + + +@add_start_docstrings( + "The Falcon Model transformer with a language modeling head on top (linear layer with weights tied to the input embeddings).", + FALCON_START_DOCSTRING, +) +class FalconForCausalLM(FalconPreTrainedModel): + _tied_weights_keys = ["lm_head.weight"] + + def __init__(self, config: FalconConfig): + super().__init__(config) + self.transformer = FalconModel(config) + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings: torch.Tensor): + self.lm_head = new_embeddings + + def prepare_inputs_for_generation( + self, + input_ids: torch.LongTensor, + past_key_values: Optional[torch.Tensor] = None, + attention_mask: Optional[torch.Tensor] = None, + **kwargs, + ) -> dict: + if past_key_values is not None: + input_ids = input_ids[:, -1:] + + return { + "input_ids": input_ids, + "past_key_values": past_key_values, + "use_cache": kwargs.get("use_cache"), + "attention_mask": attention_mask, + } + + @add_start_docstrings_to_model_forward(FALCON_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=CausalLMOutputWithCrossAttentions, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple[torch.Tensor], CausalLMOutputWithCrossAttentions]: + r""" + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for language modeling. Note that the labels **are shifted** inside the model, i.e. you can set + `labels = input_ids` Indices are selected in `[-100, 0, ..., config.vocab_size]` All labels set to `-100` + are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` + """ + + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + transformer_outputs = self.transformer( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + head_mask=head_mask, + 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] + + lm_logits = self.lm_head(hidden_states) + + loss = None + if labels is not None: + # Shift so that tokens < n predict n + shift_logits = lm_logits[..., :-1, :].contiguous() + shift_labels = labels[..., 1:].contiguous() + batch_size, seq_length, vocab_size = shift_logits.shape + # Flatten the tokens + loss_fct = CrossEntropyLoss() + loss = loss_fct( + shift_logits.view(batch_size * seq_length, vocab_size), shift_labels.view(batch_size * seq_length) + ) + + if not return_dict: + output = (lm_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return CausalLMOutputWithCrossAttentions( + loss=loss, + logits=lm_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + def _reorder_cache( + self, past: Tuple[Tuple[torch.Tensor, torch.Tensor], ...], beam_idx: torch.LongTensor + ) -> Tuple[Tuple[torch.Tensor, torch.Tensor], ...]: + """ + This function is used to re-order the `past_key_values` cache if [`~PreTrainedModel.beam_search`] or + [`~PreTrainedModel.beam_sample`] is called. This is required to match `past_key_values` with the correct + beam_idx at every generation step. + + Output shares the same memory storage as `past`. + """ + + # Get a copy of `beam_idx` on all the devices where we need those indices. + device_to_beam_idx = { + past_state.device: beam_idx.to(past_state.device) for layer_past in past for past_state in layer_past + } + reordered_past = tuple( + ( + layer_past[0].index_select(0, device_to_beam_idx[layer_past[0].device]), + layer_past[1].index_select(0, device_to_beam_idx[layer_past[0].device]), + ) + for layer_past in past + ) + return reordered_past + + +@add_start_docstrings( + """ + The Falcon Model transformer with a sequence classification head on top (linear layer). + + [`FalconForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-1) 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). + """, + FALCON_START_DOCSTRING, +) +class FalconForSequenceClassification(FalconPreTrainedModel): + def __init__(self, config: FalconConfig): + super().__init__(config) + self.num_labels = config.num_labels + self.transformer = FalconModel(config) + self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(FALCON_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=SequenceClassifierOutputWithPast, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple[torch.Tensor], 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, + past_key_values=past_key_values, + attention_mask=attention_mask, + head_mask=head_mask, + 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: + sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(dim=-1) - 1 + else: + sequence_lengths = -1 + logger.warning( + f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " + "unexpected if using padding tokens in conjunction with `inputs_embeds.`" + ) + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] + + loss = None + if labels is not None: + 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, labels) + 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( + """ + Falcon Model 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. + """, + FALCON_START_DOCSTRING, +) +class FalconForTokenClassification(FalconPreTrainedModel): + def __init__(self, config: FalconConfig): + super().__init__(config) + self.num_labels = config.num_labels + + self.transformer = FalconModel(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.classifier = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(FALCON_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]] = None, + attention_mask: Optional[torch.Tensor] = None, + head_mask: Optional[torch.Tensor] = None, + inputs_embeds: Optional[torch.Tensor] = None, + labels: Optional[torch.Tensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple[torch.Tensor], 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 + + transformer_outputs = self.transformer( + input_ids, + past_key_values=past_key_values, + attention_mask=attention_mask, + head_mask=head_mask, + 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] + hidden_states = self.dropout(hidden_states) + logits = self.classifier(hidden_states) + + loss = None + if labels is not None: + batch_size, seq_length = labels.shape + loss_fct = CrossEntropyLoss() + loss = loss_fct( + logits.view(batch_size * seq_length, self.num_labels), labels.view(batch_size * seq_length) + ) + + if not return_dict: + output = (logits,) + transformer_outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@add_start_docstrings( + """ + The Falcon Model transformer with a span classification head on top for extractive question-answering tasks like + SQuAD (a linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + FALCON_START_DOCSTRING, +) +class FalconForQuestionAnswering(FalconPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.transformer = FalconModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, 2) + + # Initialize weights and apply final processing + self.post_init() + + @add_start_docstrings_to_model_forward(FALCON_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + head_mask: Optional[torch.FloatTensor] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, QuestionAnsweringModelOutput]: + r""" + start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the start of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the end of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + """ + 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, + head_mask=head_mask, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + sequence_output = outputs[0] + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + total_loss = None + if start_positions is not None and end_positions is not None: + # If we are on multi-GPU, split add a dimension + if len(start_positions.size()) > 1: + start_positions = start_positions.squeeze(-1) + if len(end_positions.size()) > 1: + end_positions = end_positions.squeeze(-1) + # sometimes the start/end positions are outside our model inputs, we ignore these terms + ignored_index = start_logits.size(1) + start_positions = start_positions.clamp(0, ignored_index) + end_positions = end_positions.clamp(0, ignored_index) + + loss_fct = CrossEntropyLoss(ignore_index=ignored_index) + start_loss = loss_fct(start_logits, start_positions) + end_loss = loss_fct(end_logits, end_positions) + total_loss = (start_loss + end_loss) / 2 + + if not return_dict: + output = (start_logits, end_logits) + outputs[2:] + return ((total_loss,) + output) if total_loss is not None else output + + return QuestionAnsweringModelOutput( + loss=total_loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) diff --git a/src/transformers/utils/dummy_pt_objects.py b/src/transformers/utils/dummy_pt_objects.py index a40deea2c0..9a41a8aa94 100644 --- a/src/transformers/utils/dummy_pt_objects.py +++ b/src/transformers/utils/dummy_pt_objects.py @@ -2876,6 +2876,51 @@ class EsmPreTrainedModel(metaclass=DummyObject): requires_backends(self, ["torch"]) +FALCON_PRETRAINED_MODEL_ARCHIVE_LIST = None + + +class FalconForCausalLM(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class FalconForQuestionAnswering(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class FalconForSequenceClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class FalconForTokenClassification(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class FalconModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + +class FalconPreTrainedModel(metaclass=DummyObject): + _backends = ["torch"] + + def __init__(self, *args, **kwargs): + requires_backends(self, ["torch"]) + + FLAUBERT_PRETRAINED_MODEL_ARCHIVE_LIST = None diff --git a/tests/models/falcon/__init__.py b/tests/models/falcon/__init__.py new file mode 100644 index 0000000000..e69de29bb2 diff --git a/tests/models/falcon/test_modeling_falcon.py b/tests/models/falcon/test_modeling_falcon.py new file mode 100644 index 0000000000..1c08f76bda --- /dev/null +++ b/tests/models/falcon/test_modeling_falcon.py @@ -0,0 +1,469 @@ +# coding=utf-8 +# Copyright 2023 The HuggingFace Inc. team. All rights reserved. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +""" Testing suite for the PyTorch Falcon model. """ + + +import unittest + +from transformers import AutoTokenizer, FalconConfig, is_torch_available +from transformers.testing_utils import require_torch, slow, torch_device + +from ...generation.test_utils import GenerationTesterMixin +from ...test_configuration_common import ConfigTester +from ...test_modeling_common import ModelTesterMixin, ids_tensor, random_attention_mask +from ...test_pipeline_mixin import PipelineTesterMixin + + +if is_torch_available(): + import torch + + from transformers import ( + FalconForCausalLM, + FalconForQuestionAnswering, + FalconForSequenceClassification, + FalconForTokenClassification, + FalconModel, + ) + + +class FalconModelTester: + def __init__( + self, + parent, + batch_size=3, + seq_length=7, + is_training=True, + use_input_mask=True, + use_token_type_ids=False, + use_labels=True, + vocab_size=99, + hidden_size=32, + num_hidden_layers=5, + 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, + 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.scope = scope + + 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 = random_attention_mask([self.batch_size, self.seq_length]) + + token_type_ids = None + + 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 FalconConfig( + 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, + is_decoder=False, + initializer_range=self.initializer_range, + pad_token_id=1, + new_decoder_architecture=True, + ) + + def create_and_check_model( + self, config, input_ids, token_type_ids, input_mask, sequence_labels, token_labels, choice_labels + ): + model = FalconModel(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)) + + 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 = FalconModel(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)) + + 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 = FalconForCausalLM(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)) + + 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 = FalconForCausalLM(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)) + + 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 FalconModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase): + all_model_classes = ( + ( + FalconModel, + FalconForCausalLM, + FalconForSequenceClassification, + FalconForTokenClassification, + FalconForQuestionAnswering, + ) + if is_torch_available() + else () + ) + all_generative_model_classes = (FalconForCausalLM,) if is_torch_available() else () + pipeline_model_mapping = ( + { + "feature-extraction": FalconModel, + "text-classification": FalconForSequenceClassification, + "text-generation": FalconForCausalLM, + "question-answering": FalconForQuestionAnswering, + "token-classification": FalconForTokenClassification, + "zero-shot": FalconForSequenceClassification, + } + if is_torch_available() + else {} + ) + test_headmasking = False + test_pruning = False + + def setUp(self): + self.model_tester = FalconModelTester(self) + self.config_tester = ConfigTester(self, config_class=FalconConfig, 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_position_embedding_types(self): + config, *inputs = self.model_tester.prepare_config_and_inputs() + for alibi in [True, False]: + config.alibi = alibi + self.model_tester.create_and_check_model(config, *inputs) + + def test_falcon_sequence_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) + sequence_labels = ids_tensor([self.model_tester.batch_size], self.model_tester.type_sequence_label_size) + model = FalconForSequenceClassification(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_falcon_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 = FalconForSequenceClassification(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_cache_conversions(self): + config, input_dict = self.model_tester.prepare_config_and_inputs_for_common() + input_ids = input_dict["input_ids"] + model = FalconForCausalLM(config) + model.to(torch_device) + model.eval() + result = model(input_ids, use_cache=True) + batch_size = input_ids.shape[0] + rw_cache = model._convert_to_rw_cache(result.past_key_values) + standard_cache = model._convert_cache_to_standard_format(rw_cache, batch_size) + for layer in range(len(rw_cache)): + for tensor_idx in range(2): + self.assertTrue(rw_cache[layer][tensor_idx].ndim == 3) + self.assertTrue(result.past_key_values[layer][tensor_idx].ndim == 4) + self.assertTrue( + torch.all(result.past_key_values[layer][tensor_idx] == standard_cache[layer][tensor_idx]) + ) + + def test_falcon_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 = FalconForSequenceClassification(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_past_key_values_format(self): + # Falcon can have different numbers of KV-heads than the number of query heads, so we need + # to override this test to use the right head counts. + for model_class in self.all_generative_model_classes: + config, inputs = self.model_tester.prepare_config_and_inputs_for_common() + + # If it doesn't support cache, pass the test + if not hasattr(config, "use_cache"): + return + + model = model_class(config).to(torch_device) + if "use_cache" not in inputs: + inputs["use_cache"] = True + outputs = model(**inputs) + + # If "past_key_values" is not returned, pass the test (e.g. RWKV uses a different cache name and format) + if "past_key_values" not in outputs: + return + + num_hidden_layers = ( + getattr(config, "decoder_layers", None) + or getattr(config, "num_decoder_layers", None) + or config.num_hidden_layers + ) + num_attention_heads = getattr(config, "num_kv_heads", config.num_attention_heads) + embed_dim = getattr(config, "d_model", config.hidden_size) + per_head_embed_dim = embed_dim // num_attention_heads + + past_kv = outputs["past_key_values"] + self.assertEqual(len(past_kv), num_hidden_layers) + + batch_size, seq_length = inputs["input_ids"].shape + for i in range(num_hidden_layers): + if config.new_decoder_architecture: + num_attention_heads = config.num_attention_heads + elif config.multi_query: + num_attention_heads = 1 + self.assertEqual(len(past_kv[0]), 2) # K V for the decoder = 2 + self.assertEqual( + past_kv[i][0].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim) + ) + self.assertEqual( + past_kv[i][1].shape, (batch_size, num_attention_heads, seq_length, per_head_embed_dim) + ) + + +@require_torch +class FalconLanguageGenerationTest(unittest.TestCase): + @slow + def test_lm_generate_falcon(self): + tokenizer = AutoTokenizer.from_pretrained("Rocketknight1/falcon-rw-1b") + model = FalconForCausalLM.from_pretrained("Rocketknight1/falcon-rw-1b") + model.eval() + model.to(torch_device) + inputs = tokenizer("My favorite food is", return_tensors="pt").to(torch_device) + + EXPECTED_OUTPUT = ( + "My favorite food is pizza. I love it so much that I have a pizza party every year for my birthday." + ) + + output_ids = model.generate(**inputs, do_sample=False, max_new_tokens=19) + output_str = tokenizer.batch_decode(output_ids)[0] + + self.assertEqual(output_str, EXPECTED_OUTPUT) + + @slow + def test_lm_generation_big_models(self): + # The big models are way too big for the CI, so we use tiny random models that resemble their + # architectures but with much smaller and fewer layers + for repo in ["Rocketknight1/tiny-random-falcon-7b", "Rocketknight1/tiny-random-falcon-40b"]: + tokenizer = AutoTokenizer.from_pretrained(repo) + model = FalconForCausalLM.from_pretrained(repo) + model.eval() + model.to(torch_device) + inputs = tokenizer("My favorite food is", return_tensors="pt").to(torch_device) + + # We just test that these run without errors - the models are randomly initialized + # and so the actual text outputs will be garbage + model.generate(**inputs, do_sample=False, max_new_tokens=4) + model.generate(**inputs, do_sample=True, max_new_tokens=4) + model.generate(**inputs, num_beams=2, max_new_tokens=4) + + @slow + def test_lm_generation_use_cache(self): + # The big models are way too big for the CI, so we use tiny random models that resemble their + # architectures but with much smaller and fewer layers + with torch.no_grad(): + for repo in [ + "Rocketknight1/falcon-rw-1b", + "Rocketknight1/tiny-random-falcon-7b", + "Rocketknight1/tiny-random-falcon-40b", + ]: + tokenizer = AutoTokenizer.from_pretrained(repo) + model = FalconForCausalLM.from_pretrained(repo) + model.eval() + model.to(device=torch_device) + inputs = tokenizer("My favorite food is", return_tensors="pt").to(torch_device) + + # Test results are the same with and without cache + outputs_no_cache = model.generate(**inputs, do_sample=False, max_new_tokens=20, use_cache=False) + outputs_cache = model.generate(**inputs, do_sample=False, max_new_tokens=20, use_cache=True) + self.assertTrue((outputs_cache - outputs_no_cache).sum().item() == 0)