Add modeling file

redirects should theoretically work, but still updating those repo references for clarity

.circleci/create_circleci_config.py

@@ -32,7 +32,8 @@ COMMON_ENV_VARIABLES = {
     "RUN_PT_TF_CROSS_TESTS": False,
     "RUN_PT_FLAX_CROSS_TESTS": False,
 }
-COMMON_PYTEST_OPTIONS = {"max-worker-restart": 0, "dist": "loadfile", "s": None}
+# Disable the use of {"s": None} as the output is way too long, causing the navigation on CircleCI impractical
+COMMON_PYTEST_OPTIONS = {"max-worker-restart": 0, "dist": "loadfile"}
 DEFAULT_DOCKER_IMAGE = [{"image": "cimg/python:3.8.12"}]
 
 
@@ -222,18 +223,27 @@ class CircleCIJob:
             # failure.
             test_command = f"({test_command}) || true"
         else:
-            test_command += " | tee tests_output.txt"
+            test_command += " || true"
         steps.append({"run": {"name": "Run tests", "command": test_command}})
 
+        check_test_command = f'if [ -s reports/{self.job_name}/failures_short.txt ]; '
+        check_test_command += 'then echo "Some test failed!"; echo ""; '
+        check_test_command += f'cat reports/{self.job_name}/failures_short.txt; '
+        check_test_command += 'echo ""; echo ""; '
+
+        py_command = f'import os; fp = open("reports/{self.job_name}/summary_short.txt"); failed = os.linesep.join([x for x in fp.read().split(os.linesep) if x.startswith(("FAILED ", "ERROR "))]); fp.close(); fp = open("summary_short.txt", "w"); fp.write(failed); fp.close()'
+        check_test_command += f"$(python3 -c '{py_command}'); "
+
+        check_test_command += f'cat summary_short.txt; echo ""; exit -1; '
+        check_test_command += f'elif [ -s reports/{self.job_name}/stats.txt ]; then echo "All tests pass!"; '
+
         # return code `124` means the previous (pytest run) step is timeout
         if self.name == "pr_documentation_tests":
-            checkout_doctest_command = 'if [ -s reports/tests_pr_documentation_tests/failures_short.txt ]; '
-            checkout_doctest_command += 'then echo "some test failed"; '
-            checkout_doctest_command += 'cat reports/tests_pr_documentation_tests/failures_short.txt; '
-            checkout_doctest_command += 'cat reports/tests_pr_documentation_tests/summary_short.txt; exit -1; '
-            checkout_doctest_command += 'elif [ -s reports/tests_pr_documentation_tests/stats.txt ]; then echo "All tests pass!"; '
-            checkout_doctest_command += 'elif [ -f 124.txt ]; then echo "doctest timeout!"; else echo "other fatal error)"; exit -1; fi;'
-            steps.append({"run": {"name": "Check doctest results", "command": checkout_doctest_command}})
+            check_test_command += 'elif [ -f 124.txt ]; then echo "doctest timeout!"; '
+
+        check_test_command += 'else echo "other fatal error"; echo ""; exit -1; fi;'
+
+        steps.append({"run": {"name": "Check test results", "command": check_test_command}})
 
         steps.append({"store_artifacts": {"path": "~/transformers/tests_output.txt"}})
         steps.append({"store_artifacts": {"path": "~/transformers/reports"}})
@@ -285,7 +295,7 @@ torch_job = CircleCIJob(
         "pip install -U --upgrade-strategy eager git+https://github.com/huggingface/accelerate",
     ],
     parallelism=1,
-    pytest_num_workers=3,
+    pytest_num_workers=8,
 )
 
 
@@ -298,8 +308,6 @@ tf_job = CircleCIJob(
         "pip install -U --upgrade-strategy eager tensorflow_probability",
     ],
     parallelism=1,
-    pytest_num_workers=6,
-    pytest_options={"rA": None},
 )
 
 
@@ -311,7 +319,6 @@ flax_job = CircleCIJob(
         "pip install -U --upgrade-strategy eager .[flax,testing,sentencepiece,flax-speech,vision]",
     ],
     parallelism=1,
-    pytest_options={"rA": None},
 )
 
 
@@ -323,7 +330,6 @@ pipelines_torch_job = CircleCIJob(
         "pip install --upgrade --upgrade-strategy eager pip",
         "pip install -U --upgrade-strategy eager .[sklearn,torch,testing,sentencepiece,torch-speech,vision,timm,video]",
     ],
-    pytest_options={"rA": None},
     marker="is_pipeline_test",
 )
 
@@ -337,7 +343,6 @@ pipelines_tf_job = CircleCIJob(
         "pip install -U --upgrade-strategy eager .[sklearn,tf-cpu,testing,sentencepiece,vision]",
         "pip install -U --upgrade-strategy eager tensorflow_probability",
     ],
-    pytest_options={"rA": None},
     marker="is_pipeline_test",
 )
 
@@ -598,7 +603,7 @@ def create_circleci_config(folder=None):
                 job.tests_to_run = [f"examples/{framework}"]
             else:
                 job.tests_to_run = [f for f in example_tests.split(" ") if f.startswith(f"examples/{framework}")]
-            
+
             if len(job.tests_to_run) > 0:
                 jobs.append(job)
 

.github/ISSUE_TEMPLATE/bug-report.yml

@@ -37,15 +37,16 @@ body:
           - pipelines: @Narsil
           - tensorflow: @gante and @Rocketknight1
           - tokenizers: @ArthurZucker
-          - trainer: @sgugger
+          - trainer: @muellerz and @pacman100
         
         Integrations:
         
           - deepspeed: HF Trainer/Accelerate: @pacman100
           - ray/raytune: @richardliaw, @amogkam
-          - Big Model Inference: @sgugger @muellerzr
+          - Big Model Inference: @SunMarc
+          - quantization (bitsandbytes, autogpt): @SunMarc and @younesbelkada
         
-        Documentation: @sgugger, @stevhliu and @MKhalusova
+        Documentation: @stevhliu and @MKhalusova
         
         Model hub:
 
@@ -61,7 +62,7 @@ body:
         Maintained examples (not research project or legacy):
         
           - Flax: @sanchit-gandhi
-          - PyTorch: @sgugger
+          - PyTorch: See Models above and tag the person corresponding to the modality of the example.
           - TensorFlow: @Rocketknight1
 
         Research projects are not maintained and should be taken as is.

.github/ISSUE_TEMPLATE/i18n.md

@@ -23,7 +23,7 @@ Some notes:
 * Please translate in a gender-neutral way.
 * Add your translations to the folder called `<languageCode>` inside the [source folder](https://github.com/huggingface/transformers/tree/main/docs/source).
 * Register your translation in `<languageCode>/_toctree.yml`; please follow the order of the [English version](https://github.com/huggingface/transformers/blob/main/docs/source/en/_toctree.yml).
-* Once you're finished, open a pull request and tag this issue by including #issue-number in the description, where issue-number is the number of this issue. Please ping @ArthurZucker, @sgugger for review.
+* Once you're finished, open a pull request and tag this issue by including #issue-number in the description, where issue-number is the number of this issue. Please ping @stevhliu and @MKhalusova for review.
 * 🙋 If you'd like others to help you with the translation, you can also post in the 🤗 [forums](https://discuss.huggingface.co/).
 
 ## Get Started section

.github/PULL_REQUEST_TEMPLATE.md

@@ -51,14 +51,16 @@ Library:
 - pipelines: @Narsil
 - tensorflow: @gante and @Rocketknight1
 - tokenizers: @ArthurZucker
-- trainer: @sgugger
+- trainer: @muellerz and @pacman100
 
 Integrations:
 
 - deepspeed: HF Trainer/Accelerate: @pacman100
 - ray/raytune: @richardliaw, @amogkam
+- Big Model Inference: @SunMarc
+- quantization (bitsandbytes, autogpt): @SunMarc and @younesbelkada
 
-Documentation: @sgugger, @stevhliu and @MKhalusova
+Documentation: @stevhliu and @MKhalusova
 
 HF projects:
 
@@ -70,7 +72,7 @@ HF projects:
 Maintained examples (not research project or legacy):
 
 - Flax: @sanchit-gandhi
-- PyTorch: @sgugger
+- PyTorch: See Models above and tag the person corresponding to the modality of the example.
 - TensorFlow: @Rocketknight1
 
  -->

.github/workflows/doctests.yml

@@ -34,7 +34,7 @@ jobs:
           nvidia-smi
 
       - name: Install transformers in edit mode
-        run: python3 -m pip install -e .
+        run: python3 -m pip install -e .[flax]
 
       - name: GPU visibility
         run: |
@@ -43,9 +43,13 @@ jobs:
       - name: Show installed libraries and their versions
         run: pip freeze
 
+      - name: Get doctest files
+        run: |
+          $(python3 -c 'from utils.tests_fetcher import get_all_doctest_files; to_test = get_all_doctest_files(); to_test = " ".join(to_test); fp = open("doc_tests.txt", "w"); fp.write(to_test); fp.close()')
+
       - name: Run doctests
         run: |
-          python3 -m pytest -v --make-reports doc_tests_gpu --doctest-modules $(cat utils/documentation_tests.txt) -sv --doctest-continue-on-failure --doctest-glob="*.md"
+          python3 -m pytest -v --make-reports doc_tests_gpu --doctest-modules $(cat doc_tests.txt) -sv --doctest-continue-on-failure --doctest-glob="*.md"
 
       - name: Failure short reports
         if: ${{ failure() }}

.github/workflows/stale.yml

@@ -2,7 +2,7 @@ name: Stale Bot
 
 on:
   schedule:
-    - cron: "0 15 * * *"
+    - cron: "0 8 * * *"
 
 jobs:
   close_stale_issues:

.github/workflows/update_metdata.yml

@@ -19,9 +19,9 @@ jobs:
       - name: Setup environment
         run: |
           pip install --upgrade pip
-          pip install datasets pandas
+          pip install datasets pandas==2.0.3
           pip install .[torch,tf,flax]
 
       - name: Update metadata
         run: |
-          python utils/update_metadata.py --token ${{ secrets.SYLVAIN_HF_TOKEN }} --commit_sha ${{ github.sha }}
+          python utils/update_metadata.py --token ${{ secrets.LYSANDRE_HF_TOKEN }} --commit_sha ${{ github.sha }}

Makefile

@@ -80,6 +80,7 @@ fix-copies:
 	python utils/check_copies.py --fix_and_overwrite
 	python utils/check_table.py --fix_and_overwrite
 	python utils/check_dummies.py --fix_and_overwrite
+	python utils/check_doctest_list.py --fix_and_overwrite
 	python utils/check_task_guides.py --fix_and_overwrite
 
 # Run tests for the library

README.md

@@ -310,6 +310,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -318,6 +319,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (from MetaAI) released with the paper [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
@@ -337,7 +339,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
@@ -372,8 +374,10 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh.
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
@@ -429,12 +433,14 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (from ADEPT) released in a [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
@@ -483,8 +489,10 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.

README_es.md

@@ -287,6 +287,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -295,6 +296,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (from MetaAI) released with the paper [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
@@ -314,7 +316,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER), released together with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
@@ -349,8 +351,10 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
@@ -406,12 +410,14 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (from ADEPT) released with the paper [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
@@ -460,8 +466,10 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.

README_hd.md

@@ -259,6 +259,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (एलेक्सा से) कागज के साथ [बीईआरटी के लिए ऑप्टिमल सबआर्किटेक्चर एक्सट्रैक्शन](https://arxiv.org/abs/ 2010.10499) एड्रियन डी विंटर और डैनियल जे पेरी द्वारा।
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (हरबिन इंस्टिट्यूट ऑफ़ टेक्नोलॉजी/माइक्रोसॉफ्ट रिसर्च एशिया/इंटेल लैब्स से) कागज के साथ [ब्रिजटॉवर: विजन-लैंग्वेज रिप्रेजेंटेशन लर्निंग में एनकोडर्स के बीच ब्रिज बनाना](<https://arxiv.org/abs/2206.08657>) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA से) Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park. द्वाराअनुसंधान पत्र [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) के साथ जारी किया गया
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google अनुसंधान से) साथ में कागज [ByT5: पूर्व-प्रशिक्षित बाइट-टू-बाइट मॉडल के साथ एक टोकन-मुक्त भविष्य की ओर] (https://arxiv.org/abs/2105.13626) Linting Xue, Aditya Barua, Noah Constant, रामी अल-रफू, शरण नारंग, मिहिर काले, एडम रॉबर्ट्स, कॉलिन रैफेल द्वारा पोस्ट किया गया।
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (इनरिया/फेसबुक/सोरबोन से) साथ में कागज [CamemBERT: एक टेस्टी फ्रेंच लैंग्वेज मॉडल](https:// arxiv.org/abs/1911.03894) लुई मार्टिन*, बेंजामिन मुलर*, पेड्रो जेवियर ऑर्टिज़ सुआरेज़*, योआन ड्यूपॉन्ट, लॉरेंट रोमरी, एरिक विलेमोन्टे डे ला क्लर्जरी, जैमे सेडाह और बेनोइट सगोट द्वारा।
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google रिसर्च से) साथ में दिया गया पेपर [कैनाइन: प्री-ट्रेनिंग ए एफिशिएंट टोकनाइजेशन-फ्री एनकोडर फॉर लैंग्वेज रिप्रेजेंटेशन]( https://arxiv.org/abs/2103.06874) जोनाथन एच क्लार्क, डैन गैरेट, यूलिया टर्क, जॉन विएटिंग द्वारा।
@@ -267,6 +268,7 @@ conda install -c huggingface transformers
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (OpenAI से) साथ वाला पेपर [लर्निंग ट्रांसफरेबल विजुअल मॉडल फ्रॉम नेचुरल लैंग्वेज सुपरविजन](https://arxiv.org /abs/2103.00020) एलेक रैडफोर्ड, जोंग वूक किम, क्रिस हैलासी, आदित्य रमेश, गेब्रियल गोह, संध्या अग्रवाल, गिरीश शास्त्री, अमांडा एस्केल, पामेला मिश्किन, जैक क्लार्क, ग्रेचेन क्रुएगर, इल्या सुत्स्केवर द्वारा।
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (सेल्सफोर्स से) साथ में पेपर [प्रोग्राम सिंथेसिस के लिए एक संवादात्मक प्रतिमान](https://arxiv.org/abs/2203.13474) एरिक निजकैंप, बो पैंग, हिरोआकी हयाशी, लिफू तू, हुआन वांग, यिंगबो झोउ, सिल्वियो सावरेस, कैमिंग जिओंग रिलीज।
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (MetaAI से) Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve. द्वाराअनुसंधान पत्र [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) के साथ जारी किया गया
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (माइक्रोसॉफ्ट रिसर्च एशिया से) कागज के साथ [फास्ट ट्रेनिंग कन्वर्जेंस के लिए सशर्त डीईटीआर](https://arxiv. org/abs/2108.06152) डेपू मेंग, ज़ियाओकांग चेन, ज़ेजिया फैन, गैंग ज़ेंग, होउकियांग ली, युहुई युआन, लेई सन, जिंगडोंग वांग द्वारा।
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech से) साथ में कागज [ConvBERT: स्पैन-आधारित डायनेमिक कनवल्शन के साथ BERT में सुधार](https://arxiv .org/abs/2008.02496) जिहांग जियांग, वीहाओ यू, डाकान झोउ, युनपेंग चेन, जियाशी फेंग, शुइचेंग यान द्वारा।
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI से) साथ वाला पेपर [A ConvNet for the 2020s](https://arxiv.org/abs /2201.03545) ज़ुआंग लियू, हेंज़ी माओ, चाओ-युआन वू, क्रिस्टोफ़ फीचटेनहोफ़र, ट्रेवर डेरेल, सैनिंग ज़ी द्वारा।
@@ -286,7 +288,7 @@ conda install -c huggingface transformers
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (फेसबुक से) साथ में कागज [ट्रांसफॉर्मर्स के साथ एंड-टू-एंड ऑब्जेक्ट डिटेक्शन](https://arxiv. org/abs/2005.12872) निकोलस कैरियन, फ़्रांसिस्को मस्सा, गेब्रियल सिनेव, निकोलस उसुनियर, अलेक्जेंडर किरिलोव, सर्गेई ज़ागोरुयको द्वारा।
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (माइक्रोसॉफ्ट रिसर्च से) कागज के साथ [DialoGPT: बड़े पैमाने पर जनरेटिव प्री-ट्रेनिंग फॉर कन्वर्सेशनल रिस्पांस जेनरेशन](https ://arxiv.org/abs/1911.00536) यिज़े झांग, सिकी सन, मिशेल गैली, येन-चुन चेन, क्रिस ब्रोकेट, जियांग गाओ, जियानफेंग गाओ, जिंगजिंग लियू, बिल डोलन द्वारा।
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI से) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. द्वाराअनुसंधान पत्र [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) के साथ जारी किया गया
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (Meta AI से) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. द्वाराअनुसंधान पत्र [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) के साथ जारी किया गया
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (हगिंगफेस से), साथ में कागज [डिस्टिलबर्ट, बीईआरटी का डिस्टिल्ड वर्जन: छोटा, तेज, सस्ता और हल्का] (https://arxiv.org/abs/1910.01108) विक्टर सनह, लिसांड्रे डेब्यू और थॉमस वुल्फ द्वारा पोस्ट किया गया। यही तरीका GPT-2 को [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERta से [DistilRoBERta](https://github.com) पर कंप्रेस करने के लिए भी लागू किया जाता है। / हगिंगफेस/ट्रांसफॉर्मर्स/ट्री/मेन/उदाहरण/डिस्टिलेशन), बहुभाषी BERT से [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) और डिस्टिलबर्ट का जर्मन संस्करण।
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (माइक्रोसॉफ्ट रिसर्च से) साथ में पेपर [DiT: सेल्फ सुपरवाइज्ड प्री-ट्रेनिंग फॉर डॉक्यूमेंट इमेज ट्रांसफॉर्मर](https://arxiv.org/abs/2203.02378) जुनलॉन्ग ली, यिहेंग जू, टेंगचाओ लव, लेई कुई, चा झांग द्वारा फुरु वेई द्वारा पोस्ट किया गया।
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER से) साथ में कागज [OCR-मुक्त डॉक्यूमेंट अंडरस्टैंडिंग ट्रांसफॉर्मर](https://arxiv.org/abs /2111.15664) गीवूक किम, टीकग्यू होंग, मूनबिन यिम, जियोंग्योन नाम, जिनयॉन्ग पार्क, जिनयॉन्ग यिम, वोनसेओक ह्वांग, सांगडू यूं, डोंगयून हान, सेउंग्युन पार्क द्वारा।
@@ -321,8 +323,10 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA से) साथ में कागज [GroupViT: टेक्स्ट सुपरविजन से सिमेंटिक सेगमेंटेशन इमर्जेस](https://arxiv .org/abs/2202.11094) जियारुई जू, शालिनी डी मेलो, सिफ़ी लियू, वोनमिन बायन, थॉमस ब्रेउएल, जान कौट्ज़, ज़ियाओलोंग वांग द्वारा।
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology से) Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik. द्वाराअनुसंधान पत्र [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) के साथ जारी किया गया
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (फेसबुक से) साथ में पेपर [ह्यूबर्ट: सेल्फ सुपरवाइज्ड स्पीच रिप्रेजेंटेशन लर्निंग बाय मास्क्ड प्रेडिक्शन ऑफ हिडन यूनिट्स](https ://arxiv.org/abs/2106.07447) वेई-निंग सू, बेंजामिन बोल्टे, याओ-हंग ह्यूबर्ट त्साई, कुशाल लखोटिया, रुस्लान सालाखुतदीनोव, अब्देलरहमान मोहम्मद द्वारा।
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (बर्कले से) साथ में कागज [I-BERT: Integer-only BERT Quantization](https:// arxiv.org/abs/2101.01321) सेहून किम, अमीर घोलमी, ज़ेवेई याओ, माइकल डब्ल्यू महोनी, कर्ट केटज़र द्वारा।
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce से) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. द्वाराअनुसंधान पत्र [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) के साथ जारी किया गया
@@ -378,12 +382,14 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google की ओर से) साथ में दिया गया पेपर [लंबे इनपुट सारांश के लिए ट्रांसफ़ॉर्मरों को बेहतर तरीके से एक्सटेंड करना](https://arxiv .org/abs/2208.04347) जेसन फांग, याओ झाओ, पीटर जे लियू द्वारा।
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (दीपमाइंड से) साथ में पेपर [पर्सीवर आईओ: संरचित इनपुट और आउटपुट के लिए एक सामान्य वास्तुकला] (https://arxiv.org/abs/2107.14795) एंड्रयू जेगल, सेबेस्टियन बोरग्यूड, जीन-बैप्टिस्ट अलायराक, कार्ल डोर्श, कैटलिन इओनेस्कु, डेविड द्वारा डिंग, स्कंद कोप्पुला, डैनियल ज़ोरान, एंड्रयू ब्रॉक, इवान शेलहैमर, ओलिवियर हेनाफ, मैथ्यू एम। बोट्विनिक, एंड्रयू ज़िसरमैन, ओरिओल विनियल्स, जोआओ कैरेरा द्वारा पोस्ट किया गया।
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (ADEPT से) Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani. द्वाराअनुसंधान पत्र [blog post](https://www.adept.ai/blog/persimmon-8b) के साथ जारी किया गया
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research से) कागज के साथ [PhoBERT: वियतनामी के लिए पूर्व-प्रशिक्षित भाषा मॉडल](https://www .aclweb.org/anthology/2020.findings-emnlp.92/) डैट क्वोक गुयेन और अन्ह तुआन गुयेन द्वारा पोस्ट किया गया।
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google से) Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova. द्वाराअनुसंधान पत्र [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) के साथ जारी किया गया
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP से) साथ वाला पेपर [प्रोग्राम अंडरस्टैंडिंग एंड जेनरेशन के लिए यूनिफाइड प्री-ट्रेनिंग](https://arxiv .org/abs/2103.06333) वसी उद्दीन अहमद, सैकत चक्रवर्ती, बैशाखी रे, काई-वेई चांग द्वारा।
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (माइक्रोसॉफ्ट रिसर्च से) साथ में पेपर [ProphetNet: प्रेडिक्टिंग फ्यूचर एन-ग्राम फॉर सीक्वेंस-टू-सीक्वेंस प्री-ट्रेनिंग ](https://arxiv.org/abs/2001.04063) यू यान, वीज़ेन क्यूई, येयुन गोंग, दयाहेंग लियू, नान डुआन, जिउशेंग चेन, रुओफ़ेई झांग और मिंग झोउ द्वारा पोस्ट किया गया।
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. से) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. द्वाराअनुसंधान पत्र [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) के साथ जारी किया गया
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. से) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. द्वाराअनुसंधान पत्र [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) के साथ जारी किया गया
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA से) साथ वाला पेपर [डीप लर्निंग इंफ़ेक्शन के लिए इंटीजर क्वांटिज़ेशन: प्रिंसिपल्स एंड एम्पिरिकल इवैल्यूएशन](https:// arxiv.org/abs/2004.09602) हाओ वू, पैट्रिक जुड, जिआओजी झांग, मिखाइल इसेव और पॉलियस माइकेविसियस द्वारा।
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (फेसबुक से) साथ में कागज [रिट्रीवल-ऑगमेंटेड जेनरेशन फॉर नॉलेज-इंटेंसिव एनएलपी टास्क](https://arxiv .org/abs/2005.11401) पैट्रिक लुईस, एथन पेरेज़, अलेक्जेंड्रा पिक्टस, फैबियो पेट्रोनी, व्लादिमीर कारपुखिन, नमन गोयल, हेनरिक कुटलर, माइक लुईस, वेन-ताउ यिह, टिम रॉकटाशेल, सेबस्टियन रिडेल, डौवे कीला द्वारा।
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google अनुसंधान से) केल्विन गु, केंटन ली, ज़ोरा तुंग, पानुपोंग पसुपत और मिंग-वेई चांग द्वारा साथ में दिया गया पेपर [REALM: रिट्रीवल-ऑगमेंटेड लैंग्वेज मॉडल प्री-ट्रेनिंग](https://arxiv.org/abs/2002.08909)।
@@ -432,8 +438,10 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (गूगल एआई से) कागज के साथ [एक इमेज इज़ वर्थ 16x16 वर्ड्स: ट्रांसफॉर्मर्स फॉर इमेज रिकॉग्निशन एट स्केल](https://arxiv.org/abs/2010.11929) एलेक्सी डोसोवित्स्की, लुकास बेयर, अलेक्जेंडर कोलेसनिकोव, डिर्क वीसेनबोर्न, शियाओहुआ झाई, थॉमस अनटरथिनर, मुस्तफा देहघानी, मैथियास मिंडरर, जॉर्ज हेगोल्ड, सिल्वेन गेली, जैकब उस्ज़कोरेइट द्वारा हॉल्सबी द्वारा पोस्ट किया गया।
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP से) साथ वाला पेपर [VisualBERT: A Simple and Performant Baseline for Vision and Language](https:/ /arxiv.org/pdf/1908.03557) लियुनियन हेरोल्ड ली, मार्क यात्स्कर, दा यिन, चो-जुई हसीह, काई-वेई चांग द्वारा।
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI से) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. द्वाराअनुसंधान पत्र [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) के साथ जारी किया गया
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (मेटा एआई से) साथ में कागज [मास्कड ऑटोएन्कोडर स्केलेबल विजन लर्नर्स हैं](https://arxiv.org/ एब्स/2111.06377) कैमिंग हे, ज़िनेली चेन, सेनिंग ज़ी, यांगहो ली, पिओट्र डॉलर, रॉस गिर्शिक द्वारा।
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (मेटा एआई से) साथ में कागज [लेबल-कुशल सीखने के लिए मास्क्ड स्याम देश के नेटवर्क](https://arxiv. org/abs/2204.07141) महमूद असरान, मथिल्डे कैरन, ईशान मिश्रा, पियोट्र बोजानोवस्की, फ्लोरियन बोर्डेस, पास्कल विंसेंट, आर्मंड जौलिन, माइकल रब्बत, निकोलस बल्लास द्वारा।
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise से) Jaehyeon Kim, Jungil Kong, Juhee Son. द्वाराअनुसंधान पत्र [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) के साथ जारी किया गया
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (फेसबुक एआई से) साथ में पेपर [wav2vec 2.0: ए फ्रेमवर्क फॉर सेल्फ-सुपरवाइज्ड लर्निंग ऑफ स्पीच रिप्रेजेंटेशन] (https://arxiv.org/abs/2006.11477) एलेक्सी बेवस्की, हेनरी झोउ, अब्देलरहमान मोहम्मद, माइकल औली द्वारा।
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI से) साथ वाला पेपर [FAIRSEQ S2T: FAIRSEQ के साथ फास्ट स्पीच-टू-टेक्स्ट मॉडलिंग ](https://arxiv.org/abs/2010.05171) चांगहान वांग, यूं तांग, जुताई मा, ऐनी वू, सरव्या पोपुरी, दिमित्रो ओखोनको, जुआन पिनो द्वारा पोस्ट किया गया।

README_ja.md

@@ -321,6 +321,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (BigScience workshop から) [BigScience Workshop](https://bigscience.huggingface.co/) から公開されました.
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa から) Adrian de Wynter and Daniel J. Perry から公開された研究論文: [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499)
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (Harbin Institute of Technology/Microsoft Research Asia/Intel Labs から) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA から) Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park. から公開された研究論文 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539)
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google Research から) Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel から公開された研究論文: [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626)
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (Inria/Facebook/Sorbonne から) Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot から公開された研究論文: [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894)
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google Research から) Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting から公開された研究論文: [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874)
@@ -329,6 +330,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (OpenAI から) Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever から公開された研究論文: [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020)
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (University of Göttingen から) Timo Lüddecke and Alexander Ecker から公開された研究論文: [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003)
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (Salesforce から) Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong から公開された研究論文: [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474)
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (MetaAI から) Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve. から公開された研究論文 [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/)
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (Microsoft Research Asia から) Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang から公開された研究論文: [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152)
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech から) Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan から公開された研究論文: [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496)
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI から) Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie から公開された研究論文: [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545)
@@ -348,7 +350,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (Facebook から) Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko から公開された研究論文: [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872)
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (Microsoft Research から) Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan から公開された研究論文: [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536)
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (SHI Labs から) Ali Hassani and Humphrey Shi から公開された研究論文: [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001)
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI から) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. から公開された研究論文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (Meta AI から) Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski. から公開された研究論文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (HuggingFace から), Victor Sanh, Lysandre Debut and Thomas Wolf. 同じ手法で GPT2, RoBERTa と Multilingual BERT の圧縮を行いました.圧縮されたモデルはそれぞれ [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation)、[DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation)、[DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/research_projects/distillation) と名付けられました. 公開された研究論文: [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108)
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (Microsoft Research から) Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei から公開された研究論文: [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378)
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER から), Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park から公開された研究論文: [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664)
@@ -383,8 +385,10 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) 坂本俊之(tanreinama)からリリースされました.
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (Microsoft から) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu から公開された研究論文: [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234).
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA から) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang から公開された研究論文: [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094)
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology から) Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik. から公開された研究論文 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf)
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook から) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed から公開された研究論文: [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447)
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley から) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer から公開された研究論文: [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321)
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI から) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever から公開された研究論文: [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/)
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce から) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. から公開された研究論文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)
@@ -440,12 +444,14 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (Google から) Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu から公開された研究論文: [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777)
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google から) Jason Phang, Yao Zhao, and Peter J. Liu から公開された研究論文: [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347)
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (Deepmind から) Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira から公開された研究論文: [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795)
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (ADEPT から) Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani. から公開された研究論文 [blog post](https://www.adept.ai/blog/persimmon-8b)
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research から) Dat Quoc Nguyen and Anh Tuan Nguyen から公開された研究論文: [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/)
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google から) Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova. から公開された研究論文 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347)
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP から) Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang から公開された研究論文: [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333)
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (Sea AI Labs から) Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng から公開された研究論文: [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418)
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research から) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou から公開された研究論文: [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063)
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. から) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. から公開された研究論文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. から) Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao. から公開された研究論文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA から) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius から公開された研究論文: [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602)
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (Facebook から) Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela から公開された研究論文: [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401)
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google Research から) Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang から公開された研究論文: [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909)
@@ -494,8 +500,10 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP から) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang から公開された研究論文: [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557)
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI から) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby から公開された研究論文: [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929)
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI から) Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He. から公開された研究論文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI から) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick から公開された研究論文: [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377)
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI から) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas から公開された研究論文: [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141)
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise から) Jaehyeon Kim, Jungil Kong, Juhee Son. から公開された研究論文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI から) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli から公開された研究論文: [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477)
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI から) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino から公開された研究論文: [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171)

README_ko.md

@@ -236,6 +236,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (Alexa 에서) Adrian de Wynter and Daniel J. Perry 의 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 논문과 함께 발표했습니다.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (NAVER CLOVA 에서 제공)은 Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.의 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539)논문과 함께 발표했습니다.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (Google Research 에서) Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel 의 [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) 논문과 함께 발표했습니다.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (Inria/Facebook/Sorbonne 에서) Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot 의 [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) 논문과 함께 발표했습니다.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (Google Research 에서) Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 의 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 논문과 함께 발표했습니다.
@@ -244,6 +245,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (OpenAI 에서) Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever 의 [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) 논문과 함께 발표했습니다.
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (University of Göttingen 에서) Timo Lüddecke and Alexander Ecker 의 [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) 논문과 함께 발표했습니다.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (Salesforce 에서) Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong 의 [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) 논문과 함께 발표했습니다.
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (MetaAI 에서 제공)은 Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.의 [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/)논문과 함께 발표했습니다.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (Microsoft Research Asia 에서) Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 의 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 논문과 함께 발표했습니다.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (YituTech 에서) Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 의 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 논문과 함께 발표했습니다.
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (Facebook AI 에서) Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 의 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 논문과 함께 발표했습니다.
@@ -263,7 +265,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (Facebook 에서) Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko 의 [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) 논문과 함께 발표했습니다.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (Microsoft Research 에서) Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan 의 [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) 논문과 함께 발표했습니다.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (SHI Labs 에서) Ali Hassani and Humphrey Shi 의 [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) 논문과 함께 발표했습니다.
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (Meta AI 에서 제공)은 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.의 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)논문과 함께 발표했습니다.
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (Meta AI 에서 제공)은 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.의 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193)논문과 함께 발표했습니다.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (HuggingFace 에서) Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT 의 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 논문과 함께 발표했습니다.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (Microsoft Research 에서) Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei 의 [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) 논문과 함께 발표했습니다.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (NAVER 에서) Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park 의 [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) 논문과 함께 발표했습니다.
@@ -298,8 +300,10 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu  의 [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234)  논문과 함께 발표했습니다.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA 에서) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 의 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 논문과 함께 발표했습니다.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (Allegro.pl, AGH University of Science and Technology 에서 제공)은 Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.의 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf)논문과 함께 발표했습니다.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (Facebook 에서) Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 의 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 논문과 함께 발표했습니다.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (Berkeley 에서) Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 의 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 논문과 함께 발표했습니다.
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI 에서) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 의 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 논문과 함께 발표했습니다.
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce 에서 제공)은 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.의 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)논문과 함께 발표했습니다.
@@ -355,12 +359,14 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (Google 에서) Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 의 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 논문과 함께 발표했습니다.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (Google 에서) Jason Phang, Yao Zhao, Peter J. Liu 의 [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) 논문과 함께 발표했습니다.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (Deepmind 에서) Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 의 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 논문과 함께 발표했습니다.
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (ADEPT 에서 제공)은 Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.의 [blog post](https://www.adept.ai/blog/persimmon-8b)논문과 함께 발표했습니다.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (VinAI Research 에서) Dat Quoc Nguyen and Anh Tuan Nguyen 의 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 논문과 함께 발표했습니다.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (Google 에서 제공)은 Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.의 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347)논문과 함께 발표했습니다.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (UCLA NLP 에서) Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang 의 [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) 논문과 함께 발표했습니다.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (Sea AI Labs 에서) Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng 의 [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) 논문과 함께 발표했습니다.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (Microsoft Research 에서) Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 의 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 논문과 함께 발표했습니다.
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. 에서 제공)은 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.의 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)논문과 함께 발표했습니다.
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (Nanjing University, The University of Hong Kong etc. 에서 제공)은 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.의 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf)논문과 함께 발표했습니다.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (NVIDIA 에서) Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 의 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 논문과 함께 발표했습니다.
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (Facebook 에서) Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela 의 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 논문과 함께 발표했습니다.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (Google Research 에서) Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 의 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 논문과 함께 발표했습니다.
@@ -409,8 +415,10 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (UCLA NLP 에서) Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 의 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 논문과 함께 발표했습니다.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (Google AI 에서) Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 의 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 논문과 함께 발표했습니다.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (Meta AI 에서 제공)은 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.의 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527)논문과 함께 발표했습니다.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (Meta AI 에서) Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 의 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 논문과 함께 발표했습니다.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (Meta AI 에서) Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 의 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) 논문과 함께 발표했습니다.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (Kakao Enterprise 에서 제공)은 Jaehyeon Kim, Jungil Kong, Juhee Son.의 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103)논문과 함께 발표했습니다.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (Facebook AI 에서) Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 의 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 논문과 함께 발표했습니다.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (Facebook AI 에서) Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 의 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 논문과 함께 발표했습니다.
@@ -429,7 +437,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (Facebook AI 에서) Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli 의 [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) 논문과 함께 발표했습니다.
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (Facebook AI 에서) Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli 의 [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) 논문과 함께 발표했습니다.
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (Huazhong University of Science & Technology 에서) Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu 의 [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) 논문과 함께 발표했습니다.
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (the University of Wisconsin - Madison 에서) Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 의 [You Only Sample (Almost) 논문과 함께 발표했습니다.
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (the University of Wisconsin - Madison 에서) Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 의 [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) 논문과 함께 발표했습니다.
 1. 새로운 모델을 올리고 싶나요? 우리가 **상세한 가이드와 템플릿** 으로 새로운 모델을 올리도록 도와드릴게요. 가이드와 템플릿은 이 저장소의 [`templates`](./templates) 폴더에서 확인하실 수 있습니다. [컨트리뷰션 가이드라인](./CONTRIBUTING.md)을 꼭 확인해주시고, PR을 올리기 전에 메인테이너에게 연락하거나 이슈를 오픈해 피드백을 받으시길 바랍니다.
 
 각 모델이 Flax, PyTorch, TensorFlow으로 구현되었는지 또는 🤗 Tokenizers 라이브러리가 지원하는 토크나이저를 사용하는지 확인하려면, [이 표](https://huggingface.co/docs/transformers/index#supported-frameworks)를 확인하세요.

README_zh-hans.md

@@ -260,6 +260,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (来自 Alexa) 伴随论文 [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) 由 Adrian de Wynter and Daniel J. Perry 发布。
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (来自 NAVER CLOVA) 伴随论文 [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) 由 Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park 发布。
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (来自 Google Research) 伴随论文 [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) 由 Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel 发布。
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (来自 Inria/Facebook/Sorbonne) 伴随论文 [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) 由 Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot 发布。
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (来自 Google Research) 伴随论文 [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) 由 Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting 发布。
@@ -268,6 +269,7 @@ conda install -c huggingface transformers
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (来自 OpenAI) 伴随论文 [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) 由 Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever 发布。
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (来自 University of Göttingen) 伴随论文 [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) 由 Timo Lüddecke and Alexander Ecker 发布。
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (来自 Salesforce) 伴随论文 [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) 由 Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong 发布。
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (来自 MetaAI) 伴随论文 [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) 由 Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve 发布。
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (来自 Microsoft Research Asia) 伴随论文 [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) 由 Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang 发布。
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (来自 YituTech) 伴随论文 [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) 由 Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan 发布。
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (来自 Facebook AI) 伴随论文 [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) 由 Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie 发布。
@@ -287,7 +289,7 @@ conda install -c huggingface transformers
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (来自 Facebook) 伴随论文 [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) 由 Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko 发布。
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (来自 Microsoft Research) 伴随论文 [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) 由 Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan 发布。
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (来自 SHI Labs) 伴随论文 [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) 由 Ali Hassani and Humphrey Shi 发布。
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (来自 Meta AI) 伴随论文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) 由 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski 发布。
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (来自 Meta AI) 伴随论文 [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) 由 Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski 发布。
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (来自 HuggingFace), 伴随论文 [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) 由 Victor Sanh, Lysandre Debut and Thomas Wolf 发布。 同样的方法也应用于压缩 GPT-2 到 [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa 到 [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT 到 [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) 和德语版 DistilBERT。
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (来自 Microsoft Research) 伴随论文 [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) 由 Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei 发布。
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (来自 NAVER) 伴随论文 [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) 由 Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park 发布。
@@ -322,8 +324,10 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (来自 UCSD, NVIDIA) 伴随论文 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 由 Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 发布。
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (来自 Allegro.pl, AGH University of Science and Technology) 伴随论文 [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) 由 Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik 发布。
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (来自 Facebook) 伴随论文 [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) 由 Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed 发布。
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (来自 Berkeley) 伴随论文 [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) 由 Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer 发布。
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (来自 OpenAI) 伴随论文 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 由 Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 发布。
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (来自 Salesforce) 伴随论文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) 由 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi 发布。
@@ -379,12 +383,14 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (来自 Google) 伴随论文 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 由 Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 发布。
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (来自 Google) 伴随论文 [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) 由 Jason Phang, Yao Zhao, Peter J. Liu 发布。
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (来自 Deepmind) 伴随论文 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 由 Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 发布。
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (来自 ADEPT) 伴随论文 [blog post](https://www.adept.ai/blog/persimmon-8b) 由 Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani 发布。
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (来自 VinAI Research) 伴随论文 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 由 Dat Quoc Nguyen and Anh Tuan Nguyen 发布。
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (来自 Google) 伴随论文 [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) 由 Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova 发布。
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (来自 UCLA NLP) 伴随论文 [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) 由 Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang 发布。
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (来自 Sea AI Labs) 伴随论文 [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) 由 Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng 发布。
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (来自 Microsoft Research) 伴随论文 [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) 由 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou 发布。
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (来自 Nanjing University, The University of Hong Kong etc.) 伴随论文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) 由 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao 发布。
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (来自 Nanjing University, The University of Hong Kong etc.) 伴随论文 [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) 由 Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao 发布。
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (来自 NVIDIA) 伴随论文 [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) 由 Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius 发布。
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (来自 Facebook) 伴随论文 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 由 Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela 发布。
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (来自 Google Research) 伴随论文 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 由 Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 发布。
@@ -433,8 +439,10 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (来自 UCLA NLP) 伴随论文 [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) 由 Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang 发布。
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (来自 Google AI) 伴随论文 [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) 由 Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby 发布。
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (来自 Meta AI) 伴随论文 [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) 由 Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He 发布。
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (来自 Meta AI) 伴随论文 [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) 由 Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick 发布。
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (来自 Meta AI) 伴随论文 [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas 发布.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (来自 Kakao Enterprise) 伴随论文 [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) 由 Jaehyeon Kim, Jungil Kong, Juhee Son 发布。
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (来自 Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) 由 Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (来自 Facebook AI) 伴随论文 [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) 由 Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (来自 Facebook AI) 伴随论文 [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) 由 Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino 发布。
@@ -453,7 +461,7 @@ conda install -c huggingface transformers
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (来自 Facebook AI) 伴随论文 [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) 由 Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli 发布。
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (来自 Facebook AI) 伴随论文 [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) 由 Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli 发布。
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (来自 Huazhong University of Science & Technology) 伴随论文 [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) 由 Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu 发布。
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (来自 the University of Wisconsin - Madison) 伴随论文 [You Only Sample (Almost) 由 Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 发布。
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (来自 the University of Wisconsin - Madison) 伴随论文 [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) 由 Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh 发布。
 1. 想要贡献新的模型？我们这里有一份**详细指引和模板**来引导你添加新的模型。你可以在 [`templates`](./templates) 目录中找到他们。记得查看 [贡献指南](./CONTRIBUTING.md) 并在开始写 PR 前联系维护人员或开一个新的 issue 来获得反馈。
 
 要检查某个模型是否已有 Flax、PyTorch 或 TensorFlow 的实现，或其是否在 🤗 Tokenizers 库中有对应词符化器（tokenizer），敬请参阅[此表](https://huggingface.co/docs/transformers/index#supported-frameworks)。

README_zh-hant.md

@@ -272,6 +272,7 @@ conda install -c huggingface transformers
 1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](https://huggingface.co/docs/transformers/model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](https://huggingface.co/docs/transformers/model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](https://huggingface.co/docs/transformers/model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](https://huggingface.co/docs/transformers/model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](https://huggingface.co/docs/transformers/model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](https://huggingface.co/docs/transformers/model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -280,6 +281,7 @@ conda install -c huggingface transformers
 1. **[CLIP](https://huggingface.co/docs/transformers/model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
 1. **[CLIPSeg](https://huggingface.co/docs/transformers/model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](https://huggingface.co/docs/transformers/model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
+1. **[CodeLlama](https://huggingface.co/docs/transformers/model_doc/llama_code)** (from MetaAI) released with the paper [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
 1. **[Conditional DETR](https://huggingface.co/docs/transformers/model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](https://huggingface.co/docs/transformers/model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[ConvNeXT](https://huggingface.co/docs/transformers/model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
@@ -299,7 +301,7 @@ conda install -c huggingface transformers
 1. **[DETR](https://huggingface.co/docs/transformers/model_doc/detr)** (from Facebook) released with the paper [End-to-End Object Detection with Transformers](https://arxiv.org/abs/2005.12872) by Nicolas Carion, Francisco Massa, Gabriel Synnaeve, Nicolas Usunier, Alexander Kirillov, Sergey Zagoruyko.
 1. **[DialoGPT](https://huggingface.co/docs/transformers/model_doc/dialogpt)** (from Microsoft Research) released with the paper [DialoGPT: Large-Scale Generative Pre-training for Conversational Response Generation](https://arxiv.org/abs/1911.00536) by Yizhe Zhang, Siqi Sun, Michel Galley, Yen-Chun Chen, Chris Brockett, Xiang Gao, Jianfeng Gao, Jingjing Liu, Bill Dolan.
 1. **[DiNAT](https://huggingface.co/docs/transformers/model_doc/dinat)** (from SHI Labs) released with the paper [Dilated Neighborhood Attention Transformer](https://arxiv.org/abs/2209.15001) by Ali Hassani and Humphrey Shi.
-1. **[DINOv2](https://huggingface.co/docs/transformers/main/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
+1. **[DINOv2](https://huggingface.co/docs/transformers/model_doc/dinov2)** (from Meta AI) released with the paper [DINOv2: Learning Robust Visual Features without Supervision](https://arxiv.org/abs/2304.07193) by Maxime Oquab, Timothée Darcet, Théo Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, Mahmoud Assran, Nicolas Ballas, Wojciech Galuba, Russell Howes, Po-Yao Huang, Shang-Wen Li, Ishan Misra, Michael Rabbat, Vasu Sharma, Gabriel Synnaeve, Hu Xu, Hervé Jegou, Julien Mairal, Patrick Labatut, Armand Joulin, Piotr Bojanowski.
 1. **[DistilBERT](https://huggingface.co/docs/transformers/model_doc/distilbert)** (from HuggingFace), released together with the paper [DistilBERT, a distilled version of BERT: smaller, faster, cheaper and lighter](https://arxiv.org/abs/1910.01108) by Victor Sanh, Lysandre Debut and Thomas Wolf. The same method has been applied to compress GPT2 into [DistilGPT2](https://github.com/huggingface/transformers/tree/main/examples/distillation), RoBERTa into [DistilRoBERTa](https://github.com/huggingface/transformers/tree/main/examples/distillation), Multilingual BERT into [DistilmBERT](https://github.com/huggingface/transformers/tree/main/examples/distillation) and a German version of DistilBERT.
 1. **[DiT](https://huggingface.co/docs/transformers/model_doc/dit)** (from Microsoft Research) released with the paper [DiT: Self-supervised Pre-training for Document Image Transformer](https://arxiv.org/abs/2203.02378) by Junlong Li, Yiheng Xu, Tengchao Lv, Lei Cui, Cha Zhang, Furu Wei.
 1. **[Donut](https://huggingface.co/docs/transformers/model_doc/donut)** (from NAVER) released with the paper [OCR-free Document Understanding Transformer](https://arxiv.org/abs/2111.15664) by Geewook Kim, Teakgyu Hong, Moonbin Yim, Jeongyeon Nam, Jinyoung Park, Jinyeong Yim, Wonseok Hwang, Sangdoo Yun, Dongyoon Han, Seunghyun Park.
@@ -334,8 +336,10 @@ conda install -c huggingface transformers
 1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
 1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](https://huggingface.co/docs/transformers/model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](https://huggingface.co/docs/transformers/model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](https://huggingface.co/docs/transformers/model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
+1. **[IDEFICS](https://huggingface.co/docs/transformers/model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh. 
 1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
 1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
@@ -391,12 +395,14 @@ conda install -c huggingface transformers
 1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](https://huggingface.co/docs/transformers/model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, Peter J. Liu.
 1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](https://huggingface.co/docs/transformers/main/model_doc/persimmon)** (from ADEPT) released with the paper [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](https://huggingface.co/docs/transformers/model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](https://huggingface.co/docs/transformers/model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](https://huggingface.co/docs/transformers/model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](https://huggingface.co/docs/transformers/model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi, Kyogu Lee. 
 1. **[ProphetNet](https://huggingface.co/docs/transformers/model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
-1. **[PVT](https://huggingface.co/docs/transformers/main/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
+1. **[PVT](https://huggingface.co/docs/transformers/model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](https://huggingface.co/docs/transformers/model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
@@ -445,8 +451,10 @@ conda install -c huggingface transformers
 1. **[Vision Transformer (ViT)](https://huggingface.co/docs/transformers/model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](https://huggingface.co/docs/transformers/model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](https://huggingface.co/docs/transformers/model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](https://huggingface.co/docs/transformers/model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](https://huggingface.co/docs/transformers/model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](https://huggingface.co/docs/transformers/model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
+1. **[VITS](https://huggingface.co/docs/transformers/model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](https://huggingface.co/docs/transformers/model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](https://huggingface.co/docs/transformers/model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
@@ -465,7 +473,7 @@ conda install -c huggingface transformers
 1. **[XLS-R](https://huggingface.co/docs/transformers/model_doc/xls_r)** (from Facebook AI) released with the paper [XLS-R: Self-supervised Cross-lingual Speech Representation Learning at Scale](https://arxiv.org/abs/2111.09296) by Arun Babu, Changhan Wang, Andros Tjandra, Kushal Lakhotia, Qiantong Xu, Naman Goyal, Kritika Singh, Patrick von Platen, Yatharth Saraf, Juan Pino, Alexei Baevski, Alexis Conneau, Michael Auli.
 1. **[XLSR-Wav2Vec2](https://huggingface.co/docs/transformers/model_doc/xlsr_wav2vec2)** (from Facebook AI) released with the paper [Unsupervised Cross-Lingual Representation Learning For Speech Recognition](https://arxiv.org/abs/2006.13979) by Alexis Conneau, Alexei Baevski, Ronan Collobert, Abdelrahman Mohamed, Michael Auli.
 1. **[YOLOS](https://huggingface.co/docs/transformers/model_doc/yolos)** (from Huazhong University of Science & Technology) released with the paper [You Only Look at One Sequence: Rethinking Transformer in Vision through Object Detection](https://arxiv.org/abs/2106.00666) by Yuxin Fang, Bencheng Liao, Xinggang Wang, Jiemin Fang, Jiyang Qi, Rui Wu, Jianwei Niu, Wenyu Liu.
-1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
+1. **[YOSO](https://huggingface.co/docs/transformers/model_doc/yoso)** (from the University of Wisconsin - Madison) released with the paper [You Only Sample (Almost) Once: Linear Cost Self-Attention Via Bernoulli Sampling](https://arxiv.org/abs/2111.09714) by Zhanpeng Zeng, Yunyang Xiong, Sathya N. Ravi, Shailesh Acharya, Glenn Fung, Vikas Singh.
 1. 想要貢獻新的模型？我們這裡有一份**詳細指引和模板**來引導你加入新的模型。你可以在 [`templates`](./templates) 目錄中找到它們。記得查看[貢獻指引](./CONTRIBUTING.md)並在開始寫 PR 前聯繫維護人員或開一個新的 issue 來獲得 feedbacks。
 
 要檢查某個模型是否已有 Flax、PyTorch 或 TensorFlow 的實作，或其是否在🤗 Tokenizers 函式庫中有對應的 tokenizer，敬請參閱[此表](https://huggingface.co/docs/transformers/index#supported-frameworks)。

awesome-transformers.md

@@ -601,3 +601,9 @@ All Hugging Face models and pipelines can be seamlessly integrated into BentoML
 
 Keywords: BentoML, Framework, Deployment, AI Applications
 
+## [LLaMA-Efficient-Tuning](https://github.com/hiyouga/LLaMA-Efficient-Tuning)
+
+[LLaMA-Efficient-Tuning](https://github.com/hiyouga/LLaMA-Efficient-Tuning) offers a user-friendly fine-tuning framework that incorporates PEFT. The repository includes training(fine-tuning) and inference examples for LLaMA-2, BLOOM, Falcon, Baichuan, Qwen, and other LLMs. A ChatGLM version is also available in [ChatGLM-Efficient-Tuning](https://github.com/hiyouga/ChatGLM-Efficient-Tuning).
+
+Keywords: PEFT, fine-tuning, LLaMA-2, ChatGLM, Qwen
+

docker/transformers-all-latest-gpu/Dockerfile

@@ -22,7 +22,6 @@ RUN python3 -m pip install --no-cache-dir --upgrade pip
 
 ARG REF=main
 RUN git clone https://github.com/huggingface/transformers && cd transformers && git checkout $REF
-RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime]
 
 # TODO: Handle these in a python utility script
 RUN [ ${#PYTORCH} -gt 0 -a "$PYTORCH" != "pre" ] && VERSION='torch=='$PYTORCH'.*' ||  VERSION='torch'; echo "export VERSION='$VERSION'" >> ~/.profile
@@ -32,7 +31,10 @@ RUN echo torch=$VERSION
 # TODO: We might need to specify proper versions that work with a specific torch version (especially for past CI).
 RUN [ "$PYTORCH" != "pre" ] && python3 -m pip install --no-cache-dir -U $VERSION torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/$CUDA || python3 -m pip install --no-cache-dir -U --pre torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/nightly/$CUDA
 
-RUN python3 -m pip install --no-cache-dir -U tensorflow==2.12 protobuf==3.20.3 tensorflow_text tensorflow_probability
+RUN python3 -m pip install --no-cache-dir -U tensorflow==2.13 protobuf==3.20.3 tensorflow_text tensorflow_probability
+
+RUN python3 -m pip install --no-cache-dir -e ./transformers[dev,onnxruntime]
+
 RUN python3 -m pip uninstall -y flax jax
 
 RUN python3 -m pip install --no-cache-dir intel_extension_for_pytorch==$INTEL_TORCH_EXT+cpu -f https://developer.intel.com/ipex-whl-stable-cpu
@@ -42,11 +44,19 @@ RUN python3 -m pip install -U "itsdangerous<2.1.0"
 
 RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/accelerate@main#egg=accelerate
 
+RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/peft@main#egg=peft
+
 # Add bitsandbytes for mixed int8 testing
 RUN python3 -m pip install --no-cache-dir bitsandbytes
 
-# For bettertransformer
-RUN python3 -m pip install --no-cache-dir optimum
+# Add auto-gptq for gtpq quantization testing
+RUN python3 -m pip install --no-cache-dir auto-gptq --extra-index-url https://huggingface.github.io/autogptq-index/whl/cu118/
+
+# Add einops for additional model testing
+RUN python3 -m pip install --no-cache-dir einops
+
+# For bettertransformer + gptq 
+RUN python3 -m pip install --no-cache-dir git+https://github.com/huggingface/optimum@main#egg=optimum
 
 # For video model testing
 RUN python3 -m pip install --no-cache-dir decord av==9.2.0

docker/transformers-tensorflow-gpu/Dockerfile

@@ -12,7 +12,7 @@ RUN git clone https://github.com/huggingface/transformers && cd transformers &&
 RUN python3 -m pip install --no-cache-dir -e ./transformers[dev-tensorflow,testing]
 
 # If set to nothing, will install the latest version
-ARG TENSORFLOW='2.12'
+ARG TENSORFLOW='2.13'
 
 RUN [ ${#TENSORFLOW} -gt 0 ] && VERSION='tensorflow=='$TENSORFLOW'.*' ||  VERSION='tensorflow'; python3 -m pip install --no-cache-dir -U $VERSION
 RUN python3 -m pip uninstall -y torch flax

docs/README.md

@@ -81,10 +81,10 @@ The `preview` command only works with existing doc files. When you add a complet
 
 ## Adding a new element to the navigation bar
 
-Accepted files are Markdown (.md or .md).
+Accepted files are Markdown (.md).
 
 Create a file with its extension and put it in the source directory. You can then link it to the toc-tree by putting
-the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/transformers/blob/main/docs/source/_toctree.yml) file.
+the filename without the extension in the [`_toctree.yml`](https://github.com/huggingface/transformers/blob/main/docs/source/en/_toctree.yml) file.
 
 ## Renaming section headers and moving sections
 
@@ -147,7 +147,7 @@ When adding a new model:
 - Add the classes that should be linked in the model. This generally includes the configuration, the tokenizer, and
   every model of that class (the base model, alongside models with additional heads), both in PyTorch and TensorFlow.
   The order is generally:
-    - Configuration,
+    - Configuration
     - Tokenizer
     - PyTorch base model
     - PyTorch head models

docs/TRANSLATING.md

@@ -54,4 +54,4 @@ The fields you should add are `local` (with the name of the file containing the
 
 Once you have translated the `_toctree.yml` file, you can start translating the [MDX](https://mdxjs.com/) files associated with your docs chapter.
 
-> 🙋 If you'd like others to help you with the translation, you should [open an issue](https://github.com/huggingface/transformers/issues) and tag @sgugger.
+> 🙋 If you'd like others to help you with the translation, you should [open an issue](https://github.com/huggingface/transformers/issues) and tag @stevhliu and @MKhalusova.

docs/source/de/index.md

@@ -218,7 +218,7 @@ Flax), PyTorch, und/oder TensorFlow haben.
 |       BigBird-Pegasus       |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |         Blenderbot          |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |       BlenderbotSmall       |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
-|            BLOOM            |       ❌       |       ✅       |       ✅        |         ❌         |      ❌      |
+|            BLOOM            |       ❌       |       ✅       |       ✅        |         ❌         |      ✅      |
 |          CamemBERT          |       ✅       |       ✅       |       ✅        |         ✅         |      ❌      |
 |           CANINE            |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            CLIP             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/de/quicktour.md

@@ -68,11 +68,13 @@ Installieren Sie die folgenden Abhängigkeiten, falls Sie dies nicht bereits get
 
 <frameworkcontent>
 <pt>
+
 ```bash
 pip install torch
 ```
 </pt>
 <tf>
+
 ```bash
 pip install tensorflow
 ```
@@ -226,6 +228,7 @@ Genau wie die [`pipeline`] akzeptiert der Tokenizer eine Liste von Eingaben. Dar
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> pt_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -237,6 +240,7 @@ Genau wie die [`pipeline`] akzeptiert der Tokenizer eine Liste von Eingaben. Dar
 ```
 </pt>
 <tf>
+
 ```py
 >>> tf_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -375,6 +379,7 @@ Ein besonders cooles 🤗 Transformers-Feature ist die Möglichkeit, ein Modell
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import AutoModel
 
@@ -383,6 +388,7 @@ Ein besonders cooles 🤗 Transformers-Feature ist die Möglichkeit, ein Modell
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import TFAutoModel
 

docs/source/en/_toctree.yml

@@ -19,150 +19,162 @@
     title: Train with a script
   - local: accelerate
     title: Set up distributed training with 🤗 Accelerate
+  - local: peft
+    title: Load and train adapters with 🤗 PEFT
   - local: model_sharing
     title: Share your model
   - local: transformers_agents
     title: Agents
+  - local: llm_tutorial
+    title: Generation with LLMs
   title: Tutorials
 - sections:
-  - sections:
-      - local: tasks/sequence_classification
-        title: Text classification
-      - local: tasks/token_classification
-        title: Token classification
-      - local: tasks/question_answering
-        title: Question answering
-      - local: tasks/language_modeling
-        title: Causal language modeling
-      - local: tasks/masked_language_modeling
-        title: Masked language modeling
-      - local: tasks/translation
-        title: Translation
-      - local: tasks/summarization
-        title: Summarization
-      - local: tasks/multiple_choice
-        title: Multiple choice
+  - isExpanded: false
+    sections:
+    - local: tasks/sequence_classification
+      title: Text classification
+    - local: tasks/token_classification
+      title: Token classification
+    - local: tasks/question_answering
+      title: Question answering
+    - local: tasks/language_modeling
+      title: Causal language modeling
+    - local: tasks/masked_language_modeling
+      title: Masked language modeling
+    - local: tasks/translation
+      title: Translation
+    - local: tasks/summarization
+      title: Summarization
+    - local: tasks/multiple_choice
+      title: Multiple choice
     title: Natural Language Processing
-    isExpanded: false
-  - sections:
-      - local: tasks/audio_classification
-        title: Audio classification
-      - local: tasks/asr
-        title: Automatic speech recognition
+  - isExpanded: false
+    sections:
+    - local: tasks/audio_classification
+      title: Audio classification
+    - local: tasks/asr
+      title: Automatic speech recognition
     title: Audio
-    isExpanded: false
-  - sections:
-      - local: tasks/image_classification
-        title: Image classification
-      - local: tasks/semantic_segmentation
-        title: Semantic segmentation
-      - local: tasks/video_classification
-        title: Video classification
-      - local: tasks/object_detection
-        title: Object detection
-      - local: tasks/zero_shot_object_detection
-        title: Zero-shot object detection
-      - local: tasks/zero_shot_image_classification
-        title: Zero-shot image classification
-      - local: tasks/monocular_depth_estimation
-        title: Depth estimation
+  - isExpanded: false
+    sections:
+    - local: tasks/image_classification
+      title: Image classification
+    - local: tasks/semantic_segmentation
+      title: Semantic segmentation
+    - local: tasks/video_classification
+      title: Video classification
+    - local: tasks/object_detection
+      title: Object detection
+    - local: tasks/zero_shot_object_detection
+      title: Zero-shot object detection
+    - local: tasks/zero_shot_image_classification
+      title: Zero-shot image classification
+    - local: tasks/monocular_depth_estimation
+      title: Depth estimation
     title: Computer Vision
-    isExpanded: false
-  - sections:
-      - local: tasks/image_captioning
-        title: Image captioning
-      - local: tasks/document_question_answering
-        title: Document Question Answering
-      - local: tasks/text-to-speech
-        title: Text to speech
+  - isExpanded: false
+    sections:
+    - local: tasks/image_captioning
+      title: Image captioning
+    - local: tasks/document_question_answering
+      title: Document Question Answering
+    - local: tasks/visual_question_answering
+      title: Visual Question Answering
+    - local: tasks/text-to-speech
+      title: Text to speech
     title: Multimodal
-    isExpanded: false
+  - isExpanded: false
+    sections:
+    - local: generation_strategies
+      title: Customize the generation strategy
+    title: Generation
   title: Task Guides
 - sections:
-    - local: fast_tokenizers
-      title: Use fast tokenizers from 🤗 Tokenizers
-    - local: multilingual
-      title: Run inference with multilingual models
-    - local: generation_strategies
-      title: Customize text generation strategy
-    - local: create_a_model
-      title: Use model-specific APIs
-    - local: custom_models
-      title: Share a custom model
-    - local: sagemaker
-      title: Run training on Amazon SageMaker
-    - local: serialization
-      title: Export to ONNX
-    - local: tflite
-      title: Export to TFLite
-    - local: torchscript
-      title: Export to TorchScript
-    - local: benchmarks
-      title: Benchmarks
-    - local: notebooks
-      title: Notebooks with examples
-    - local: community
-      title: Community resources
-    - local: custom_tools
-      title: Custom Tools and Prompts
-    - local: troubleshooting
-      title: Troubleshoot
+  - local: fast_tokenizers
+    title: Use fast tokenizers from 🤗 Tokenizers
+  - local: multilingual
+    title: Run inference with multilingual models
+  - local: create_a_model
+    title: Use model-specific APIs
+  - local: custom_models
+    title: Share a custom model
+  - local: chat_templating
+    title: Templates for chat models
+  - local: sagemaker
+    title: Run training on Amazon SageMaker
+  - local: serialization
+    title: Export to ONNX
+  - local: tflite
+    title: Export to TFLite
+  - local: torchscript
+    title: Export to TorchScript
+  - local: benchmarks
+    title: Benchmarks
+  - local: notebooks
+    title: Notebooks with examples
+  - local: community
+    title: Community resources
+  - local: custom_tools
+    title: Custom Tools and Prompts
+  - local: troubleshooting
+    title: Troubleshoot
   title: Developer guides
 - sections:
-    - local: performance
-      title: Overview
-    - sections:
-        - local: perf_train_gpu_one
-          title: Methods and tools for efficient training on a single GPU
-        - local: perf_train_gpu_many
-          title: Multiple GPUs and parallelism
-        - local: perf_train_cpu
-          title: Efficient training on CPU
-        - local: perf_train_cpu_many
-          title: Distributed CPU training
-        - local: perf_train_tpu
-          title: Training on TPUs
-        - local: perf_train_tpu_tf
-          title: Training on TPU with TensorFlow
-        - local: perf_train_special
-          title: Training on Specialized Hardware
-        - local: perf_hardware
-          title: Custom hardware for training
-        - local: hpo_train
-          title: Hyperparameter Search using Trainer API
-      title: Efficient training techniques
-    - sections:
-        - local: perf_infer_cpu
-          title: Inference on CPU
-        - local: perf_infer_gpu_one
-          title: Inference on one GPU
-        - local: perf_infer_gpu_many
-          title: Inference on many GPUs
-        - local: perf_infer_special
-          title: Inference on Specialized Hardware
-      title: Optimizing inference
-    - local: big_models
-      title: Instantiating a big model
-    - local: debugging
-      title: Troubleshooting
-    - local: tf_xla
-      title: XLA Integration for TensorFlow Models
+  - local: performance
+    title: Overview
+  - sections:
+    - local: perf_train_gpu_one
+      title: Methods and tools for efficient training on a single GPU
+    - local: perf_train_gpu_many
+      title: Multiple GPUs and parallelism
+    - local: perf_train_cpu
+      title: Efficient training on CPU
+    - local: perf_train_cpu_many
+      title: Distributed CPU training
+    - local: perf_train_tpu
+      title: Training on TPUs
+    - local: perf_train_tpu_tf
+      title: Training on TPU with TensorFlow
+    - local: perf_train_special
+      title: Training on Specialized Hardware
+    - local: perf_hardware
+      title: Custom hardware for training
+    - local: hpo_train
+      title: Hyperparameter Search using Trainer API
+    title: Efficient training techniques
+  - sections:
+    - local: perf_infer_cpu
+      title: Inference on CPU
+    - local: perf_infer_gpu_one
+      title: Inference on one GPU
+    - local: perf_infer_gpu_many
+      title: Inference on many GPUs
+    - local: perf_infer_special
+      title: Inference on Specialized Hardware
+    title: Optimizing inference
+  - local: big_models
+    title: Instantiating a big model
+  - local: debugging
+    title: Troubleshooting
+  - local: tf_xla
+    title: XLA Integration for TensorFlow Models
+  - local: perf_torch_compile
+    title: Optimize inference using `torch.compile()`
   title: Performance and scalability
 - sections:
-    - local: contributing
-      title: How to contribute to transformers?
-    - local: add_new_model
-      title: How to add a model to 🤗 Transformers?
-    - local: add_tensorflow_model
-      title: How to convert a 🤗 Transformers model to TensorFlow?
-    - local: add_new_pipeline
-      title: How to add a pipeline to 🤗 Transformers?
-    - local: testing
-      title: Testing
-    - local: pr_checks
-      title: Checks on a Pull Request
+  - local: contributing
+    title: How to contribute to transformers?
+  - local: add_new_model
+    title: How to add a model to 🤗 Transformers?
+  - local: add_tensorflow_model
+    title: How to convert a 🤗 Transformers model to TensorFlow?
+  - local: add_new_pipeline
+    title: How to add a pipeline to 🤗 Transformers?
+  - local: testing
+    title: Testing
+  - local: pr_checks
+    title: Checks on a Pull Request
   title: Contribute
-
 - sections:
   - local: philosophy
     title: Philosophy
@@ -273,6 +285,8 @@
         title: CANINE
       - local: model_doc/codegen
         title: CodeGen
+      - local: model_doc/code_llama
+        title: CodeLlama
       - local: model_doc/convbert
         title: ConvBERT
       - local: model_doc/cpm
@@ -301,6 +315,8 @@
         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
@@ -393,6 +409,8 @@
         title: Pegasus
       - local: model_doc/pegasus_x
         title: PEGASUS-X
+      - local: model_doc/persimmon
+        title: Persimmon
       - local: model_doc/phobert
         title: PhoBERT
       - local: model_doc/plbart
@@ -546,6 +564,8 @@
         title: Vision Transformer (ViT)
       - local: model_doc/vit_hybrid
         title: ViT Hybrid
+      - local: model_doc/vitdet
+        title: ViTDet
       - local: model_doc/vit_mae
         title: ViTMAE
       - local: model_doc/vit_msn
@@ -573,6 +593,8 @@
         title: MMS
       - local: model_doc/musicgen
         title: MusicGen
+      - local: model_doc/pop2piano
+        title: Pop2Piano
       - local: model_doc/sew
         title: SEW
       - local: model_doc/sew-d
@@ -587,6 +609,8 @@
         title: UniSpeech
       - local: model_doc/unispeech-sat
         title: UniSpeech-SAT
+      - local: model_doc/vits
+        title: VITS
       - local: model_doc/wav2vec2
         title: Wav2Vec2
       - local: model_doc/wav2vec2-conformer
@@ -614,6 +638,8 @@
         title: BLIP-2
       - local: model_doc/bridgetower
         title: BridgeTower
+      - local: model_doc/bros
+        title: BROS
       - local: model_doc/chinese_clip
         title: Chinese-CLIP
       - local: model_doc/clip
@@ -632,6 +658,8 @@
         title: GIT
       - local: model_doc/groupvit
         title: GroupViT
+      - local: model_doc/idefics
+        title: IDEFICS
       - local: model_doc/instructblip
         title: InstructBLIP
       - local: model_doc/layoutlm

docs/source/en/accelerate.md

@@ -133,4 +133,4 @@ accelerate launch train.py
 >>> notebook_launcher(training_function)
 ```
 
-For more information about 🤗 Accelerate and it's rich features, refer to the [documentation](https://huggingface.co/docs/accelerate).
+For more information about 🤗 Accelerate and its rich features, refer to the [documentation](https://huggingface.co/docs/accelerate).

docs/source/en/add_new_model.md

@@ -101,7 +101,7 @@ own regarding how code should be written :-)
 1. The forward pass of your model should be fully written in the modeling file while being fully independent of other
    models in the library. If you want to reuse a block from another model, copy the code and paste it with a
    `# Copied from` comment on top (see [here](https://github.com/huggingface/transformers/blob/v4.17.0/src/transformers/models/roberta/modeling_roberta.py#L160)
-   for a good example).
+   for a good example and [there](pr_checks#check-copies) for more documentation on Copied from). 
 2. The code should be fully understandable, even by a non-native English speaker. This means you should pick
    descriptive variable names and avoid abbreviations. As an example, `activation` is preferred to `act`.
    One-letter variable names are strongly discouraged unless it's an index in a for loop.
@@ -361,7 +361,7 @@ We expect that every model added to 🤗 Transformers passes a couple of integra
 model and the reimplemented version in 🤗 Transformers have to give the exact same output up to a precision of 0.001!
 Since it is normal that the exact same model written in different libraries can give a slightly different output
 depending on the library framework, we accept an error tolerance of 1e-3 (0.001). It is not enough if the model gives
-nearly the same output, they have to be the almost identical. Therefore, you will certainly compare the intermediate
+nearly the same output, they have to be almost identical. Therefore, you will certainly compare the intermediate
 outputs of the 🤗 Transformers version multiple times against the intermediate outputs of the original implementation of
 *brand_new_bert* in which case an **efficient** debugging environment of the original repository is absolutely
 important. Here is some advice is to make your debugging environment as efficient as possible.

docs/source/en/add_tensorflow_model.md

@@ -56,7 +56,7 @@ you might recall from our [general overview of 🤗 Transformers](add_new_model#
 that we are an opinionated bunch - the ease of use of 🤗 Transformers relies on consistent design choices. From
 experience, we can tell you a few important things about adding TensorFlow models:
 
-- Don't reinvent the wheel! More often that not, there are at least two reference implementations you should check: the
+- Don't reinvent the wheel! More often than not, there are at least two reference implementations you should check: the
 PyTorch equivalent of the model you are implementing and other TensorFlow models for the same class of problems.
 - Great model implementations survive the test of time. This doesn't happen because the code is pretty, but rather
 because the code is clear, easy to debug and build upon. If you make the life of the maintainers easy with your
@@ -101,7 +101,7 @@ TensorFlow-related pull request.
 
 **2. Prepare transformers dev environment**
 
-Having selected the model architecture, open an draft PR to signal your intention to work on it. Follow the
+Having selected the model architecture, open a draft PR to signal your intention to work on it. Follow the
 instructions below to set up your environment and open a draft PR.
 
 1. Fork the [repository](https://github.com/huggingface/transformers) by clicking on the 'Fork' button on the
@@ -328,7 +328,7 @@ That's it! 🎉
 ## Debugging mismatches across ML frameworks 🐛
 
 At some point, when adding a new architecture or when creating TensorFlow weights for an existing architecture, you
-might come across errors compaining about mismatches between PyTorch and TensorFlow. You might even decide to open the
+might come across errors complaining about mismatches between PyTorch and TensorFlow. You might even decide to open the
 model architecture code for the two frameworks, and find that they look identical. What's going on? 🤔
 
 First of all, let's talk about why understanding these mismatches matters. Many community members will use 🤗
@@ -351,7 +351,7 @@ ingredient here is patience. Here is our suggested workflow for when you come ac
    that you'll have to venture into the source implementation of said instruction. In some cases, you might find an
    issue with a reference implementation - don't abstain from opening an issue in the upstream repository.
 
-In some cases, in dicussion with the 🤗 Transformers team, we might find that the fixing the mismatch is infeasible.
+In some cases, in discussion with the 🤗 Transformers team, we might find that fixing the mismatch is infeasible.
 When the mismatch is very small in the output layers of the model (but potentially large in the hidden states), we
 might decide to ignore it in favor of distributing the model. The `pt-to-tf` CLI mentioned above has a `--max-error`
 flag to override the error message at weight conversion time.

docs/source/en/autoclass_tutorial.md

@@ -16,7 +16,7 @@ rendered properly in your Markdown viewer.
 
 # Load pretrained instances with an AutoClass
 
-With so many different Transformer architectures, it can be challenging to create one for your checkpoint. As a part of 🤗 Transformers core philosophy to make the library easy, simple and flexible to use, an `AutoClass` automatically infer and load the correct architecture from a given checkpoint. The `from_pretrained()` method lets you quickly load a pretrained model for any architecture so you don't have to devote time and resources to train a model from scratch. Producing this type of checkpoint-agnostic code means if your code works for one checkpoint, it will work with another checkpoint - as long as it was trained for a similar task - even if the architecture is different.
+With so many different Transformer architectures, it can be challenging to create one for your checkpoint. As a part of 🤗 Transformers core philosophy to make the library easy, simple and flexible to use, an `AutoClass` automatically infers and loads the correct architecture from a given checkpoint. The `from_pretrained()` method lets you quickly load a pretrained model for any architecture so you don't have to devote time and resources to train a model from scratch. Producing this type of checkpoint-agnostic code means if your code works for one checkpoint, it will work with another checkpoint - as long as it was trained for a similar task - even if the architecture is different.
 
 <Tip>
 

docs/source/en/big_models.md

@@ -23,11 +23,11 @@ from PyTorch is:
 2. Load your pretrained weights.
 3. Put those pretrained weights in your random model.
 
-Step 1 and 2 both require a full version of the model in memory, which is not a problem in most cases, but if your model starts weighing several GigaBytes, those two copies can make you got our of RAM. Even worse, if you are using `torch.distributed` to launch a distributed training, each process will load the pretrained model and store these two copies in RAM.
+Step 1 and 2 both require a full version of the model in memory, which is not a problem in most cases, but if your model starts weighing several GigaBytes, those two copies can make you get out of RAM. Even worse, if you are using `torch.distributed` to launch a distributed training, each process will load the pretrained model and store these two copies in RAM.
 
 <Tip>
 
-Note that the randomly created model is initialized with "empty" tensors, which take the space in memory without filling it (thus the random values are whatever was in this chunk of memory at a given time). The random initialization following the appropriate distribution for the kind of model/parameters instatiated (like a normal distribution for instance) is only performed after step 3 on the non-initialized weights, to be as fast as possible! 
+Note that the randomly created model is initialized with "empty" tensors, which take the space in memory without filling it (thus the random values are whatever was in this chunk of memory at a given time). The random initialization following the appropriate distribution for the kind of model/parameters instantiated (like a normal distribution for instance) is only performed after step 3 on the non-initialized weights, to be as fast as possible! 
 
 </Tip>
 
@@ -120,4 +120,4 @@ If you want to directly load such a sharded checkpoint inside a model without us
 
 Sharded checkpoints reduce the memory usage during step 2 of the workflow mentioned above, but in order to use that model in a low memory setting, we recommend leveraging our tools based on the Accelerate library.
 
-Please read the following guide for more information: [Large model loading using Accelerate](./main_classes/model#large-model-loading)
+Please read the following guide for more information: [Large model loading using Accelerate](./main_classes/model#large-model-loading)

docs/source/en/chat_templating.md

@@ -0,0 +1,255 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Templates for Chat Models
+
+## Introduction
+
+An increasingly common use case for LLMs is **chat**. In a chat context, rather than continuing a single string
+of text (as is the case with a standard language model), the model instead continues a conversation that consists
+of one or more **messages**, each of which includes a **role** as well as message text.
+
+Most commonly, these roles are "user" for messages sent by the user, and "assistant" for messages sent by the model.
+Some models also support a "system" role. System messages are usually sent at the beginning of the conversation
+and include directives about how the model should behave in the subsequent chat.
+
+All language models, including models fine-tuned for chat, operate on linear sequences of tokens and do not intrinsically
+have special handling for roles. This means that role information is usually injected by adding control tokens
+between messages, to indicate both the message boundary and the relevant roles.
+
+Unfortunately, there isn't (yet!) a standard for which tokens to use, and so different models have been trained
+with wildly different formatting and control tokens for chat. This can be a real problem for users - if you use the
+wrong format, then the model will be confused by your input, and your performance will be a lot worse than it should be.
+This is the problem that **chat templates** aim to resolve. 
+
+Chat conversations are typically represented as a list of dictionaries, where each dictionary contains `role`
+and `content` keys, and represents a single chat message. Chat templates are strings containing a Jinja template that
+specifies how to format a conversation for a given model into a single tokenizable sequence. By storing this information
+with the tokenizer, we can ensure that models get input data in the format they expect.
+
+Let's make this concrete with a quick example using the `BlenderBot` model. BlenderBot has an extremely simple default 
+template, which mostly just adds whitespace between rounds of dialogue:
+
+```python
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+>>> chat = [
+...   {"role": "user", "content": "Hello, how are you?"},
+...   {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
+...   {"role": "user", "content": "I'd like to show off how chat templating works!"},
+... ]
+
+>>> tokenizer.apply_chat_template(chat, tokenize=False)
+" Hello, how are you?  I'm doing great. How can I help you today?   I'd like to show off how chat templating works!</s>"
+```
+
+Notice how the entire chat is condensed into a single string. If we use `tokenize=True`, which is the default setting,
+that string will also be tokenized for us. To see a more complex template in action, though, let's use the 
+`meta-llama/Llama-2-7b-chat-hf` model. Note that this model has gated access, so you will have to
+[request access on the repo](https://huggingface.co/meta-llama/Llama-2-7b-chat-hf) if you want to run this code yourself:
+
+```python
+>> from transformers import AutoTokenizer
+>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-chat-hf")
+
+>> chat = [
+...   {"role": "user", "content": "Hello, how are you?"},
+...   {"role": "assistant", "content": "I'm doing great. How can I help you today?"},
+...   {"role": "user", "content": "I'd like to show off how chat templating works!"},
+... ]
+
+>> tokenizer.use_default_system_prompt = False
+>> tokenizer.apply_chat_template(chat, tokenize=False)
+"<s>[INST] Hello, how are you? [/INST] I'm doing great. How can I help you today? </s><s>[INST] I'd like to show off how chat templating works! [/INST]"
+```
+
+Note that this time, the tokenizer has added the control tokens [INST] and [/INST] to indicate the start and end of 
+user messages (but not assistant messages!)
+
+## How do chat templates work?
+
+The chat template for a model is stored on the `tokenizer.chat_template` attribute. If no chat template is set, the
+default template for that model class is used instead. Let's take a look at the template for `BlenderBot`:
+
+```python
+
+>>> from transformers import AutoTokenizer
+>>> tokenizer = AutoTokenizer.from_pretrained("facebook/blenderbot-400M-distill")
+
+>>> tokenizer.default_chat_template
+"{% for message in messages %}{% if message['role'] == 'user' %}{{ ' ' }}{% endif %}{{ message['content'] }}{% if not loop.last %}{{ '  ' }}{% endif %}{% endfor %}{{ eos_token }}"
+```
+
+That's kind of intimidating. Let's add some newlines and indentation to make it more readable. Note that
+we remove the first newline after each block as well as any preceding whitespace before a block by default, using the 
+Jinja `trim_blocks` and `lstrip_blocks` flags. This means that you can write your templates with indentations and 
+newlines and still have them function correctly!
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ ' ' }}
+    {% endif %}
+    {{ message['content'] }}
+    {% if not loop.last %}
+        {{ '  ' }}
+    {% endif %}
+{% endfor %}
+{{ eos_token }}
+```
+
+If you've never seen one of these before, this is a [Jinja template](https://jinja.palletsprojects.com/en/3.1.x/templates/).
+Jinja is a templating language that allows you to write simple code that generates text. In many ways, the code and
+syntax resembles Python. In pure Python, this template would look something like this:
+
+```python
+for idx, message in enumerate(messages):
+    if message['role'] == 'user':
+        print(' ')
+    print(message['content'])
+    if not idx == len(messages) - 1:  # Check for the last message in the conversation
+        print('  ')
+print(eos_token)
+```
+
+Effectively, the template does three things:
+1. For each message, if the message is a user message, add a blank space before it, otherwise print nothing.
+2. Add the message content
+3. If the message is not the last message, add two spaces after it. After the final message, print the EOS token.
+
+This is a pretty simple template - it doesn't add any control tokens, and it doesn't support "system" messages, which 
+are a common way to give the model directives about how it should behave in the subsequent conversation.
+But Jinja gives you a lot of flexibility to do those things! Let's see a Jinja template that can format inputs
+similarly to the way LLaMA formats them (note that the real LLaMA template includes handling for default system
+messages and slightly different system message handling in general - don't use this one in your actual code!)
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ bos_token + '[INST] ' + message['content'] + ' [/INST]' }}
+    {% elif message['role'] == 'system' %}
+        {{ '<<SYS>>\\n' + message['content'] + '\\n<</SYS>>\\n\\n' }}
+    {% elif message['role'] == 'assistant' %}
+        {{ ' '  + message['content'] + ' ' + eos_token }}
+    {% endif %}
+{% endfor %}
+```
+
+Hopefully if you stare at this for a little bit you can see what this template is doing - it adds specific tokens based
+on the "role" of each message, which represents who sent it. User, assistant and system messages are clearly
+distinguishable to the model because of the tokens they're wrapped in.
+
+## How do I create a chat template?
+
+Simple, just write a jinja template and set `tokenizer.chat_template`. You may find it easier to start with an 
+existing template from another model and simply edit it for your needs! For example, we could take the LLaMA template
+above and add "[ASST]" and "[/ASST]" to assistant messages:
+
+```
+{% for message in messages %}
+    {% if message['role'] == 'user' %}
+        {{ bos_token + '[INST] ' + message['content'].strip() + ' [/INST]' }}
+    {% elif message['role'] == 'system' %}
+        {{ '<<SYS>>\\n' + message['content'].strip() + '\\n<</SYS>>\\n\\n' }}
+    {% elif message['role'] == 'assistant' %}
+        {{ '[ASST] '  + message['content'] + ' [/ASST]' + eos_token }}
+    {% endif %}
+{% endfor %}
+```
+
+Now, simply set the `tokenizer.chat_template` attribute. Next time you use [`~PreTrainedTokenizer.apply_chat_template`], it will
+use your new template! This attribute will be saved in the `tokenizer_config.json` file, so you can use
+[`~utils.PushToHubMixin.push_to_hub`] to upload your new template to the Hub and make sure everyone's using the right
+template for your model!
+
+```python
+template = tokenizer.chat_template
+template = template.replace("SYS", "SYSTEM")  # Change the system token
+tokenizer.chat_template = template  # Set the new template
+tokenizer.push_to_hub("model_name")  # Upload your new template to the Hub!
+```
+
+The method [`~PreTrainedTokenizer.apply_chat_template`] which uses your chat template is called by the [`ConversationalPipeline`] class, so 
+once you set the correct chat template, your model will automatically become compatible with [`ConversationalPipeline`].
+
+## What are "default" templates?
+
+Before the introduction of chat templates, chat handling was hardcoded at the model class level. For backwards 
+compatibility, we have retained this class-specific handling as default templates, also set at the class level. If a
+model does not have a chat template set, but there is a default template for its model class, the `ConversationalPipeline`
+class and methods like `apply_chat_template` will use the class template instead. You can find out what the default
+template for your tokenizer is by checking the `tokenizer.default_chat_template` attribute.
+
+This is something we do purely for backward compatibility reasons, to avoid breaking any existing workflows. Even when
+the class template is appropriate for your model, we strongly recommend overriding the default template by
+setting the `chat_template` attribute explicitly to make it clear to users that your model has been correctly configured
+for chat, and to future-proof in case the default templates are ever altered or deprecated.
+
+## What template should I use?
+
+When setting the template for a model that's already been trained for chat, you should ensure that the template
+exactly matches the message formatting that the model saw during training, or else you will probably experience
+performance degradation. This is true even if you're training the model further - you will probably get the best 
+performance if you keep the chat tokens constant. This is very analogous to tokenization - you generally get the
+best performance for inference or fine-tuning when you precisely match the tokenization used during training.
+
+If you're training a model from scratch, or fine-tuning a base language model for chat, on the other hand,
+you have a lot of freedom to choose an appropriate template! LLMs are smart enough to learn to handle lots of different
+input formats. Our default template for models that don't have a class-specific template follows the 
+[ChatML format](https://github.com/openai/openai-python/blob/main/chatml.md), and this is a good, flexible choice for many use-cases. It looks like this:
+
+```
+{% for message in messages %}
+    {{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}
+{% endfor %}
+```
+
+If you like this one, here it is in one-liner form, ready to copy into your code:
+
+```
+tokenizer.chat_template = "{% for message in messages %}{{'<|im_start|>' + message['role'] + '\n' + message['content'] + '<|im_end|>' + '\n'}}{% endfor %}"
+```
+
+This template wraps each message in `<|im_start|>` and `<|im_end|>` tokens, and simply writes the role as a string, which
+allows for flexibility in the roles you train with. The output looks like this:
+
+```
+<|im_start|>system
+You are a helpful chatbot that will do its best not to say anything so stupid that people tweet about it.<|im_end|>
+<|im_start|>user
+How are you?<|im_end|>
+<|im_start|>assistant
+I'm doing great!<|im_end|>
+```
+
+The "user", "system" and "assistant" roles are the standard for chat, and we recommend using them when it makes sense,
+particularly if you want your model to operate well with [`ConversationalPipeline`]. However, you are not limited
+to these roles - templating is extremely flexible, and any string can be a role.
+
+## I want to use chat templates! How should I get started?
+
+If you have any chat models, you should set their `tokenizer.chat_template` attribute and test it using
+[`~PreTrainedTokenizer.apply_chat_template`]. This applies even if you're not the model owner - if you're using a model
+with an empty chat template, or one that's still using the default class template, please open a [pull request](https://huggingface.co/docs/hub/repositories-pull-requests-discussions) to
+the model repository so that this attribute can be set properly!
+
+Once the attribute is set, that's it, you're done! `tokenizer.apply_chat_template` will now work correctly for that
+model, which means it is also automatically supported in places like `ConversationalPipeline`!
+
+By ensuring that models have this attribute, we can make sure that the whole community gets to use the full power of
+open-source models. Formatting mismatches have been haunting the field and silently harming performance for too long - 
+it's time to put an end to them!

docs/source/en/community.md

@@ -1,4 +1,4 @@
-<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+<!--⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
 rendered properly in your Markdown viewer.
 -->
 
@@ -10,7 +10,7 @@ This page regroups resources around 🤗 Transformers developed by the community
 
 | Resource     |      Description      |      Author      |
 |:----------|:-------------|------:|
-| [Hugging Face Transformers Glossary Flashcards](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | A set of flashcards based on the [Transformers Docs Glossary](glossary) that has been put into a form which can be easily learnt/revised using [Anki ](https://apps.ankiweb.net/) an open source, cross platform app specifically designed for long term knowledge retention. See this [Introductory video on how to use the flashcards](https://www.youtube.com/watch?v=Dji_h7PILrw). | [Darigov Research](https://www.darigovresearch.com/) |
+| [Hugging Face Transformers Glossary Flashcards](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | A set of flashcards based on the [Transformers Docs Glossary](glossary) that has been put into a form which can be easily learned/revised using [Anki ](https://apps.ankiweb.net/) an open source, cross platform app specifically designed for long term knowledge retention. See this [Introductory video on how to use the flashcards](https://www.youtube.com/watch?v=Dji_h7PILrw). | [Darigov Research](https://www.darigovresearch.com/) |
 
 ## Community notebooks:
 
@@ -35,7 +35,7 @@ This page regroups resources around 🤗 Transformers developed by the community
 |[Speed up Fine-Tuning in Transformers with Dynamic Padding / Bucketing](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)|How to speed up fine-tuning by a factor of 2 using dynamic padding / bucketing|[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|
 |[Pretrain Reformer for Masked Language Modeling](https://github.com/patrickvonplaten/notebooks/blob/master/Reformer_For_Masked_LM.ipynb)| How to train a Reformer model with bi-directional self-attention layers | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tzzh0i8PgDQGV3SMFUGxM7_gGae3K-uW?usp=sharing)|
 |[Expand and Fine Tune Sci-BERT](https://github.com/lordtt13/word-embeddings/blob/master/COVID-19%20Research%20Data/COVID-SciBERT.ipynb)| How to increase vocabulary of a pretrained SciBERT model from AllenAI on the CORD dataset and pipeline it. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1rqAR40goxbAfez1xvF3hBJphSCsvXmh8)|
-|[Fine Tune BlenderBotSmall for Summarization using the Trainer API](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/fine-tune-blenderbot_small-for-summarization.ipynb)| How to fine tune BlenderBotSmall for summarization on a custom dataset, using the Trainer API. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/19Wmupuls7mykSGyRN_Qo6lPQhgp56ymq?usp=sharing)|
+|[Fine Tune BlenderBotSmall for Summarization using the Trainer API](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/fine-tune-blenderbot_small-for-summarization.ipynb)| How to fine-tune BlenderBotSmall for summarization on a custom dataset, using the Trainer API. | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/19Wmupuls7mykSGyRN_Qo6lPQhgp56ymq?usp=sharing)|
 |[Fine-tune Electra and interpret with Integrated Gradients](https://github.com/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb) | How to fine-tune Electra for sentiment analysis and interpret predictions with Captum Integrated Gradients | [Eliza Szczechla](https://elsanns.github.io) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb)|
 |[fine-tune a non-English GPT-2 Model with Trainer class](https://github.com/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb) | How to fine-tune a non-English GPT-2 Model with Trainer class | [Philipp Schmid](https://www.philschmid.de) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb)|
 |[Fine-tune a DistilBERT Model for Multi Label Classification task](https://github.com/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb) | How to fine-tune a DistilBERT Model for Multi Label Classification task | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb)|

docs/source/en/create_a_model.md

@@ -209,7 +209,7 @@ Easily reuse this checkpoint for another task by switching to a different model
 The last base class you need before using a model for textual data is a [tokenizer](main_classes/tokenizer) to convert raw text to tensors. There are two types of tokenizers you can use with 🤗 Transformers:
 
 - [`PreTrainedTokenizer`]: a Python implementation of a tokenizer.
-- [`PreTrainedTokenizerFast`]: a tokenizer from our Rust-based [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) library. This tokenizer type is significantly faster - especially during batch tokenization - due to it's Rust implementation. The fast tokenizer also offers additional methods like *offset mapping* which maps tokens to their original words or characters.
+- [`PreTrainedTokenizerFast`]: a tokenizer from our Rust-based [🤗 Tokenizer](https://huggingface.co/docs/tokenizers/python/latest/) library. This tokenizer type is significantly faster - especially during batch tokenization - due to its Rust implementation. The fast tokenizer also offers additional methods like *offset mapping* which maps tokens to their original words or characters.
 
 Both tokenizers support common methods such as encoding and decoding, adding new tokens, and managing special tokens.
 

docs/source/en/custom_models.md

@@ -341,7 +341,7 @@ model. This is different from pushing the code to the Hub in the sense that user
 get the custom models (contrarily to automatically downloading the model code from the Hub).
 
 As long as your config has a `model_type` attribute that is different from existing model types, and that your model
-classes have the right `config_class` attributes, you can just add them to the auto classes likes this:
+classes have the right `config_class` attributes, you can just add them to the auto classes like this:
 
 ```py
 from transformers import AutoConfig, AutoModel, AutoModelForImageClassification

docs/source/en/custom_tools.md

@@ -25,7 +25,7 @@ If you are not aware of what tools and agents are in the context of transformers
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API that is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API that is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>

docs/source/en/generation_strategies.md

@@ -57,10 +57,8 @@ When you load a model explicitly, you can inspect the generation configuration t
 >>> model = AutoModelForCausalLM.from_pretrained("distilgpt2")
 >>> model.generation_config
 GenerationConfig {
-    "_from_model_config": true,
     "bos_token_id": 50256,
     "eos_token_id": 50256,
-    "transformers_version": "4.26.0.dev0"
 }
 ```
 
@@ -77,7 +75,7 @@ producing highly repetitive results.
 You can override any `generation_config` by passing the parameters and their values directly to the [`generate`] method:
 
 ```python
->>> my_model.generate(**inputs, num_beams=4, do_sample=True)
+>>> my_model.generate(**inputs, num_beams=4, do_sample=True)  # doctest: +SKIP
 ```
 
 Even if the default decoding strategy mostly works for your task, you can still tweak a few things. Some of the
@@ -92,7 +90,7 @@ sequences that start with a lower probability initial tokens and would've been i
 - `do_sample`: if set to `True`, this parameter enables decoding strategies such as multinomial sampling, beam-search
 multinomial sampling, Top-K sampling and Top-p sampling. All these strategies select the next token from the probability
 distribution over the entire vocabulary with various strategy-specific adjustments.
-- `num_return_sequences`: the number of sequence candidates to return for each input. This options is only available for
+- `num_return_sequences`: the number of sequence candidates to return for each input. This option is only available for
 the decoding strategies that support multiple sequence candidates, e.g. variations of beam search and sampling. Decoding
 strategies like greedy search and contrastive search return a single output sequence.
 
@@ -107,11 +105,11 @@ If you would like to share your fine-tuned model with a specific generation conf
 ```python
 >>> from transformers import AutoModelForCausalLM, GenerationConfig
 
->>> model = AutoModelForCausalLM.from_pretrained("my_account/my_model")
+>>> model = AutoModelForCausalLM.from_pretrained("my_account/my_model")  # doctest: +SKIP
 >>> generation_config = GenerationConfig(
 ...     max_new_tokens=50, do_sample=True, top_k=50, eos_token_id=model.config.eos_token_id
 ... )
->>> generation_config.save_pretrained("my_account/my_model", push_to_hub=True)
+>>> generation_config.save_pretrained("my_account/my_model", push_to_hub=True)  # doctest: +SKIP
 ```
 
 You can also store several generation configurations in a single directory, making use of the `config_file_name`
@@ -133,19 +131,20 @@ one for summarization with beam search). You must have the right Hub permissions
 ...     pad_token=model.config.pad_token_id,
 ... )
 
->>> translation_generation_config.save_pretrained("t5-small", "translation_generation_config.json", push_to_hub=True)
+>>> # Tip: add `push_to_hub=True` to push to the Hub
+>>> translation_generation_config.save_pretrained("/tmp", "translation_generation_config.json")
 
 >>> # You could then use the named generation config file to parameterize generation
->>> generation_config = GenerationConfig.from_pretrained("t5-small", "translation_generation_config.json")
+>>> generation_config = GenerationConfig.from_pretrained("/tmp", "translation_generation_config.json")
 >>> inputs = tokenizer("translate English to French: Configuration files are easy to use!", return_tensors="pt")
 >>> outputs = model.generate(**inputs, generation_config=generation_config)
 >>> print(tokenizer.batch_decode(outputs, skip_special_tokens=True))
-['Les fichiers de configuration sont faciles à utiliser !']
+['Les fichiers de configuration sont faciles à utiliser!']
 ```
 
 ## Streaming
 
-The `generate()` supports streaming, through its `streamer` input. The `streamer` input is compatible any instance
+The `generate()` supports streaming, through its `streamer` input. The `streamer` input is compatible with any instance
 from a class that has the following methods: `put()` and `end()`. Internally, `put()` is used to push new tokens and
 `end()` is used to flag the end of text generation.
 
@@ -217,10 +216,9 @@ The two main parameters that enable and control the behavior of contrastive sear
 
 >>> outputs = model.generate(**inputs, penalty_alpha=0.6, top_k=4, max_new_tokens=100)
 >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
-['Hugging Face Company is a family owned and operated business. \
-We pride ourselves on being the best in the business and our customer service is second to none.\
-\n\nIf you have any questions about our products or services, feel free to contact us at any time.\
- We look forward to hearing from you!']
+['Hugging Face Company is a family owned and operated business. We pride ourselves on being the best
+in the business and our customer service is second to none.\n\nIf you have any questions about our
+products or services, feel free to contact us at any time. We look forward to hearing from you!']
 ```
 
 ### Multinomial sampling
@@ -233,7 +231,8 @@ risk of repetition.
 To enable multinomial sampling set `do_sample=True` and `num_beams=1`.
 
 ```python
->>> from transformers import AutoTokenizer, AutoModelForCausalLM
+>>> from transformers import AutoTokenizer, AutoModelForCausalLM, set_seed
+>>> set_seed(0)  # For reproducibility
 
 >>> checkpoint = "gpt2-large"
 >>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
@@ -244,11 +243,8 @@ To enable multinomial sampling set `do_sample=True` and `num_beams=1`.
 
 >>> outputs = model.generate(**inputs, do_sample=True, num_beams=1, max_new_tokens=100)
 >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
-['Today was an amazing day because we are now in the final stages of our trip to New York City which was very tough. \
-It is a difficult schedule and a challenging part of the year but still worth it. I have been taking things easier and \
-I feel stronger and more motivated to be out there on their tour. Hopefully, that experience is going to help them with \
-their upcoming events which are currently scheduled in Australia.\n\nWe love that they are here. They want to make a \
-name for themselves and become famous for what they']
+['Today was an amazing day because when you go to the World Cup and you don\'t, or when you don\'t get invited,
+that\'s a terrible feeling."']
 ```
 
 ### Beam-search decoding
@@ -272,7 +268,7 @@ To enable this decoding strategy, specify the `num_beams` (aka number of hypothe
 
 >>> outputs = model.generate(**inputs, num_beams=5, max_new_tokens=50)
 >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
-['It is astonishing how one can have such a profound impact on the lives of so many people in such a short period of \
+['It is astonishing how one can have such a profound impact on the lives of so many people in such a short period of
 time."\n\nHe added: "I am very proud of the work I have been able to do in the last few years.\n\n"I have']
 ```
 
@@ -282,7 +278,8 @@ As the name implies, this decoding strategy combines beam search with multinomia
 the `num_beams` greater than 1, and set `do_sample=True` to use this decoding strategy.
 
 ```python
->>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
+>>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM, set_seed
+>>> set_seed(0)  # For reproducibility
 
 >>> prompt = "translate English to German: The house is wonderful."
 >>> checkpoint = "t5-small"
@@ -302,27 +299,29 @@ the `num_beams` greater than 1, and set `do_sample=True` to use this decoding st
 The diverse beam search decoding strategy is an extension of the beam search strategy that allows for generating a more diverse
 set of beam sequences to choose from. To learn how it works, refer to [Diverse Beam Search: Decoding Diverse Solutions from Neural Sequence Models](https://arxiv.org/pdf/1610.02424.pdf).
 This approach has three main parameters: `num_beams`, `num_beam_groups`, and `diversity_penalty`.
-The diversily penalty ensures the outputs are distinct across groups, and beam search is used within each group.
+The diversity penalty ensures the outputs are distinct across groups, and beam search is used within each group.
 
 
 ```python
 >>> from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 
 >>> checkpoint = "google/pegasus-xsum"
->>> prompt = "The Permaculture Design Principles are a set of universal design principles \
->>> that can be applied to any location, climate and culture, and they allow us to design \
->>> the most efficient and sustainable human habitation and food production systems. \
->>> Permaculture is a design system that encompasses a wide variety of disciplines, such \
->>> as ecology, landscape design, environmental science and energy conservation, and the \
->>> Permaculture design principles are drawn from these various disciplines. Each individual \
->>> design principle itself embodies a complete conceptual framework based on sound \
->>> scientific principles. When we bring all these separate  principles together, we can \
->>> create a design system that both looks at whole systems, the parts that these systems \
->>> consist of, and how those parts interact with each other to create a complex, dynamic, \
->>> living system. Each design principle serves as a tool that allows us to integrate all \
->>> the separate parts of a design, referred to as elements, into a functional, synergistic, \
->>> whole system, where the elements harmoniously interact and work together in the most \
->>> efficient way possible."
+>>> prompt = (
+...     "The Permaculture Design Principles are a set of universal design principles "
+...     "that can be applied to any location, climate and culture, and they allow us to design "
+...     "the most efficient and sustainable human habitation and food production systems. "
+...     "Permaculture is a design system that encompasses a wide variety of disciplines, such "
+...     "as ecology, landscape design, environmental science and energy conservation, and the "
+...     "Permaculture design principles are drawn from these various disciplines. Each individual "
+...     "design principle itself embodies a complete conceptual framework based on sound "
+...     "scientific principles. When we bring all these separate  principles together, we can "
+...     "create a design system that both looks at whole systems, the parts that these systems "
+...     "consist of, and how those parts interact with each other to create a complex, dynamic, "
+...     "living system. Each design principle serves as a tool that allows us to integrate all "
+...     "the separate parts of a design, referred to as elements, into a functional, synergistic, "
+...     "whole system, where the elements harmoniously interact and work together in the most "
+...     "efficient way possible."
+... )
 
 >>> tokenizer = AutoTokenizer.from_pretrained(checkpoint)
 >>> inputs = tokenizer(prompt, return_tensors="pt")
@@ -331,7 +330,8 @@ The diversily penalty ensures the outputs are distinct across groups, and beam s
 
 >>> outputs = model.generate(**inputs, num_beams=5, num_beam_groups=5, max_new_tokens=30, diversity_penalty=1.0)
 >>> tokenizer.decode(outputs[0], skip_special_tokens=True)
-'The aim of this project is to create a new type of living system, one that is more sustainable and efficient than the current one.'
+'The Design Principles are a set of universal design principles that can be applied to any location, climate and
+culture, and they allow us to design the'
 ```
 
 This guide illustrates the main parameters that enable various decoding strategies. More advanced parameters exist for the
@@ -365,11 +365,12 @@ To enable assisted decoding, set the `assistant_model` argument with a model.
 ['Alice and Bob are sitting in a bar. Alice is drinking a beer and Bob is drinking a']
 ```
 
-When using assisted decoding with sampling methods, you can use the `temperarure` argument to control the randomness
+When using assisted decoding with sampling methods, you can use the `temperature` argument to control the randomness
 just like in multinomial sampling. However, in assisted decoding, reducing the temperature will help improving latency.
 
 ```python
->>> from transformers import AutoModelForCausalLM, AutoTokenizer
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer, set_seed
+>>> set_seed(42)  # For reproducibility
 
 >>> prompt = "Alice and Bob"
 >>> checkpoint = "EleutherAI/pythia-1.4b-deduped"
@@ -382,5 +383,5 @@ just like in multinomial sampling. However, in assisted decoding, reducing the t
 >>> assistant_model = AutoModelForCausalLM.from_pretrained(assistant_checkpoint)
 >>> outputs = model.generate(**inputs, assistant_model=assistant_model, do_sample=True, temperature=0.5)
 >>> tokenizer.batch_decode(outputs, skip_special_tokens=True)
-["Alice and Bob are sitting on the sofa. Alice says, 'I'm going to my room"]
+['Alice and Bob are going to the same party. It is a small party, in a small']
 ```

docs/source/en/glossary.md

@@ -187,7 +187,7 @@ The model head refers to the last layer of a neural network that accepts the raw
 
 ### image patch
 
-Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the `patch_size` - or resolution - of the model in it's configuration.
+Vision-based Transformers models split an image into smaller patches which are linearly embedded, and then passed as a sequence to the model. You can find the `patch_size` - or resolution - of the model in its configuration.
 
 ### inference
 

docs/source/en/hpo_train.md

@@ -54,6 +54,18 @@ For optuna, see optuna [object_parameter](https://optuna.readthedocs.io/en/stabl
 ...     }
 ```
 
+Optuna provides multi-objective HPO. You can pass `direction` in `hyperparameter_search` and define your own compute_objective to return multiple objective values. The Pareto Front (`List[BestRun]`) will be returned in hyperparameter_search, you should refer to the test case `TrainerHyperParameterMultiObjectOptunaIntegrationTest` in [test_trainer](https://github.com/huggingface/transformers/blob/main/tests/trainer/test_trainer.py). It's like following
+
+```py
+>>> best_trials = trainer.hyperparameter_search(
+...     direction=["minimize", "maximize"],
+...     backend="optuna",
+...     hp_space=optuna_hp_space,
+...     n_trials=20,
+...     compute_objective=compute_objective,
+... )
+```
+
 For raytune, see raytune [object_parameter](https://docs.ray.io/en/latest/tune/api/search_space.html), it's like following:
 
 ```py

docs/source/en/index.md

@@ -76,6 +76,7 @@ The documentation is organized into five sections:
 1. **[BLOOM](model_doc/bloom)** (from BigScience workshop) released by the [BigScience Workshop](https://bigscience.huggingface.co/).
 1. **[BORT](model_doc/bort)** (from Alexa) released with the paper [Optimal Subarchitecture Extraction For BERT](https://arxiv.org/abs/2010.10499) by Adrian de Wynter and Daniel J. Perry.
 1. **[BridgeTower](model_doc/bridgetower)** (from Harbin Institute of Technology/Microsoft Research Asia/Intel Labs) released with the paper [BridgeTower: Building Bridges Between Encoders in Vision-Language Representation Learning](https://arxiv.org/abs/2206.08657) by Xiao Xu, Chenfei Wu, Shachar Rosenman, Vasudev Lal, Wanxiang Che, Nan Duan.
+1. **[BROS](model_doc/bros)** (from NAVER CLOVA) released with the paper [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
 1. **[ByT5](model_doc/byt5)** (from Google Research) released with the paper [ByT5: Towards a token-free future with pre-trained byte-to-byte models](https://arxiv.org/abs/2105.13626) by Linting Xue, Aditya Barua, Noah Constant, Rami Al-Rfou, Sharan Narang, Mihir Kale, Adam Roberts, Colin Raffel.
 1. **[CamemBERT](model_doc/camembert)** (from Inria/Facebook/Sorbonne) released with the paper [CamemBERT: a Tasty French Language Model](https://arxiv.org/abs/1911.03894) by Louis Martin*, Benjamin Muller*, Pedro Javier Ortiz Suárez*, Yoann Dupont, Laurent Romary, Éric Villemonte de la Clergerie, Djamé Seddah and Benoît Sagot.
 1. **[CANINE](model_doc/canine)** (from Google Research) released with the paper [CANINE: Pre-training an Efficient Tokenization-Free Encoder for Language Representation](https://arxiv.org/abs/2103.06874) by Jonathan H. Clark, Dan Garrette, Iulia Turc, John Wieting.
@@ -84,6 +85,7 @@ The documentation is organized into five sections:
 1. **[CLIP](model_doc/clip)** (from OpenAI) released with the paper [Learning Transferable Visual Models From Natural Language Supervision](https://arxiv.org/abs/2103.00020) by Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, Ilya Sutskever.
 1. **[CLIPSeg](model_doc/clipseg)** (from University of Göttingen) released with the paper [Image Segmentation Using Text and Image Prompts](https://arxiv.org/abs/2112.10003) by Timo Lüddecke and Alexander Ecker.
 1. **[CodeGen](model_doc/codegen)** (from Salesforce) released with the paper [A Conversational Paradigm for Program Synthesis](https://arxiv.org/abs/2203.13474) by Erik Nijkamp, Bo Pang, Hiroaki Hayashi, Lifu Tu, Huan Wang, Yingbo Zhou, Silvio Savarese, Caiming Xiong.
+1. **[CodeLlama](model_doc/llama_code)** (from MetaAI) released with the paper [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
 1. **[Conditional DETR](model_doc/conditional_detr)** (from Microsoft Research Asia) released with the paper [Conditional DETR for Fast Training Convergence](https://arxiv.org/abs/2108.06152) by Depu Meng, Xiaokang Chen, Zejia Fan, Gang Zeng, Houqiang Li, Yuhui Yuan, Lei Sun, Jingdong Wang.
 1. **[ConvBERT](model_doc/convbert)** (from YituTech) released with the paper [ConvBERT: Improving BERT with Span-based Dynamic Convolution](https://arxiv.org/abs/2008.02496) by Zihang Jiang, Weihao Yu, Daquan Zhou, Yunpeng Chen, Jiashi Feng, Shuicheng Yan.
 1. **[ConvNeXT](model_doc/convnext)** (from Facebook AI) released with the paper [A ConvNet for the 2020s](https://arxiv.org/abs/2201.03545) by Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, Saining Xie.
@@ -138,8 +140,10 @@ The documentation is organized into five sections:
 1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
 1. **[Graphormer](model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
 1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
+1. **[HerBERT](model_doc/herbert)** (from Allegro.pl, AGH University of Science and Technology) released with the paper [KLEJ: Comprehensive Benchmark for Polish Language Understanding](https://www.aclweb.org/anthology/2020.acl-main.111.pdf) by Piotr Rybak, Robert Mroczkowski, Janusz Tracz, Ireneusz Gawlik.
 1. **[Hubert](model_doc/hubert)** (from Facebook) released with the paper [HuBERT: Self-Supervised Speech Representation Learning by Masked Prediction of Hidden Units](https://arxiv.org/abs/2106.07447) by Wei-Ning Hsu, Benjamin Bolte, Yao-Hung Hubert Tsai, Kushal Lakhotia, Ruslan Salakhutdinov, Abdelrahman Mohamed.
 1. **[I-BERT](model_doc/ibert)** (from Berkeley) released with the paper [I-BERT: Integer-only BERT Quantization](https://arxiv.org/abs/2101.01321) by Sehoon Kim, Amir Gholami, Zhewei Yao, Michael W. Mahoney, Kurt Keutzer.
+1. **[IDEFICS](model_doc/idefics)** (from HuggingFace) released with the paper [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents](https://huggingface.co/papers/2306.16527) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh.
 1. **[ImageGPT](model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
 1. **[Informer](model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
 1. **[InstructBLIP](model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
@@ -195,10 +199,12 @@ The documentation is organized into five sections:
 1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
 1. **[PEGASUS-X](model_doc/pegasus_x)** (from Google) released with the paper [Investigating Efficiently Extending Transformers for Long Input Summarization](https://arxiv.org/abs/2208.04347) by Jason Phang, Yao Zhao, and Peter J. Liu.
 1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
+1. **[Persimmon](model_doc/persimmon)** (from ADEPT) released in a [blog post](https://www.adept.ai/blog/persimmon-8b) by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
 1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
 1. **[Pix2Struct](model_doc/pix2struct)** (from Google) released with the paper [Pix2Struct: Screenshot Parsing as Pretraining for Visual Language Understanding](https://arxiv.org/abs/2210.03347) by Kenton Lee, Mandar Joshi, Iulia Turc, Hexiang Hu, Fangyu Liu, Julian Eisenschlos, Urvashi Khandelwal, Peter Shaw, Ming-Wei Chang, Kristina Toutanova.
 1. **[PLBart](model_doc/plbart)** (from UCLA NLP) released with the paper [Unified Pre-training for Program Understanding and Generation](https://arxiv.org/abs/2103.06333) by Wasi Uddin Ahmad, Saikat Chakraborty, Baishakhi Ray, Kai-Wei Chang.
 1. **[PoolFormer](model_doc/poolformer)** (from Sea AI Labs) released with the paper [MetaFormer is Actually What You Need for Vision](https://arxiv.org/abs/2111.11418) by Yu, Weihao and Luo, Mi and Zhou, Pan and Si, Chenyang and Zhou, Yichen and Wang, Xinchao and Feng, Jiashi and Yan, Shuicheng.
+1. **[Pop2Piano](model_doc/pop2piano)** released with the paper [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
 1. **[ProphetNet](model_doc/prophetnet)** (from Microsoft Research) released with the paper [ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training](https://arxiv.org/abs/2001.04063) by Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang and Ming Zhou.
 1. **[PVT](model_doc/pvt)** (from Nanjing University, The University of Hong Kong etc.) released with the paper [Pyramid Vision Transformer: A Versatile Backbone for Dense Prediction without Convolutions](https://arxiv.org/pdf/2102.12122.pdf) by Wenhai Wang, Enze Xie, Xiang Li, Deng-Ping Fan, Kaitao Song, Ding Liang, Tong Lu, Ping Luo, Ling Shao.
 1. **[QDQBert](model_doc/qdqbert)** (from NVIDIA) released with the paper [Integer Quantization for Deep Learning Inference: Principles and Empirical Evaluation](https://arxiv.org/abs/2004.09602) by Hao Wu, Patrick Judd, Xiaojie Zhang, Mikhail Isaev and Paulius Micikevicius.
@@ -249,8 +255,10 @@ The documentation is organized into five sections:
 1. **[Vision Transformer (ViT)](model_doc/vit)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
 1. **[VisualBERT](model_doc/visual_bert)** (from UCLA NLP) released with the paper [VisualBERT: A Simple and Performant Baseline for Vision and Language](https://arxiv.org/pdf/1908.03557) by Liunian Harold Li, Mark Yatskar, Da Yin, Cho-Jui Hsieh, Kai-Wei Chang.
 1. **[ViT Hybrid](model_doc/vit_hybrid)** (from Google AI) released with the paper [An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale](https://arxiv.org/abs/2010.11929) by Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, Neil Houlsby.
+1. **[VitDet](model_doc/vitdet)** (from Meta AI) released with the paper [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
 1. **[ViTMAE](model_doc/vit_mae)** (from Meta AI) released with the paper [Masked Autoencoders Are Scalable Vision Learners](https://arxiv.org/abs/2111.06377) by Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, Ross Girshick.
 1. **[ViTMSN](model_doc/vit_msn)** (from Meta AI) released with the paper [Masked Siamese Networks for Label-Efficient Learning](https://arxiv.org/abs/2204.07141) by Mahmoud Assran, Mathilde Caron, Ishan Misra, Piotr Bojanowski, Florian Bordes, Pascal Vincent, Armand Joulin, Michael Rabbat, Nicolas Ballas.
+1. **[VITS](model_doc/vits)** (from Kakao Enterprise) released with the paper [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
 1. **[ViViT](model_doc/vivit)** (from Google Research) released with the paper [ViViT: A Video Vision Transformer](https://arxiv.org/abs/2103.15691) by Anurag Arnab, Mostafa Dehghani, Georg Heigold, Chen Sun, Mario Lučić, Cordelia Schmid.
 1. **[Wav2Vec2](model_doc/wav2vec2)** (from Facebook AI) released with the paper [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations](https://arxiv.org/abs/2006.11477) by Alexei Baevski, Henry Zhou, Abdelrahman Mohamed, Michael Auli.
 1. **[Wav2Vec2-Conformer](model_doc/wav2vec2-conformer)** (from Facebook AI) released with the paper [FAIRSEQ S2T: Fast Speech-to-Text Modeling with FAIRSEQ](https://arxiv.org/abs/2010.05171) by Changhan Wang, Yun Tang, Xutai Ma, Anne Wu, Sravya Popuri, Dmytro Okhonko, Juan Pino.
@@ -300,8 +308,9 @@ Flax), PyTorch, and/or TensorFlow.
 |        BlenderbotSmall        |       ✅        |         ✅         |      ✅      |
 |             BLIP              |       ✅        |         ✅         |      ❌      |
 |            BLIP-2             |       ✅        |         ❌         |      ❌      |
-|             BLOOM             |       ✅        |         ❌         |      ❌      |
+|             BLOOM             |       ✅        |         ❌         |      ✅      |
 |          BridgeTower          |       ✅        |         ❌         |      ❌      |
+|             BROS              |       ✅        |         ❌         |      ❌      |
 |           CamemBERT           |       ✅        |         ✅         |      ❌      |
 |            CANINE             |       ✅        |         ❌         |      ❌      |
 |         Chinese-CLIP          |       ✅        |         ❌         |      ❌      |
@@ -309,6 +318,7 @@ Flax), PyTorch, and/or TensorFlow.
 |             CLIP              |       ✅        |         ✅         |      ✅      |
 |            CLIPSeg            |       ✅        |         ❌         |      ❌      |
 |            CodeGen            |       ✅        |         ❌         |      ❌      |
+|           CodeLlama           |       ✅        |         ❌         |      ❌      |
 |       Conditional DETR        |       ✅        |         ❌         |      ❌      |
 |           ConvBERT            |       ✅        |         ✅         |      ❌      |
 |           ConvNeXT            |       ✅        |         ✅         |      ❌      |
@@ -360,6 +370,7 @@ Flax), PyTorch, and/or TensorFlow.
 |           GroupViT            |       ✅        |         ✅         |      ❌      |
 |            Hubert             |       ✅        |         ✅         |      ❌      |
 |            I-BERT             |       ✅        |         ❌         |      ❌      |
+|            IDEFICS            |       ✅        |         ❌         |      ❌      |
 |           ImageGPT            |       ✅        |         ❌         |      ❌      |
 |           Informer            |       ✅        |         ❌         |      ❌      |
 |         InstructBLIP          |       ✅        |         ❌         |      ❌      |
@@ -410,9 +421,11 @@ Flax), PyTorch, and/or TensorFlow.
 |            Pegasus            |       ✅        |         ✅         |      ✅      |
 |           PEGASUS-X           |       ✅        |         ❌         |      ❌      |
 |           Perceiver           |       ✅        |         ❌         |      ❌      |
+|           Persimmon           |       ✅        |         ❌         |      ❌      |
 |          Pix2Struct           |       ✅        |         ❌         |      ❌      |
 |            PLBart             |       ✅        |         ❌         |      ❌      |
 |          PoolFormer           |       ✅        |         ❌         |      ❌      |
+|           Pop2Piano           |       ✅        |         ❌         |      ❌      |
 |          ProphetNet           |       ✅        |         ❌         |      ❌      |
 |              PVT              |       ✅        |         ❌         |      ❌      |
 |            QDQBert            |       ✅        |         ❌         |      ❌      |
@@ -465,8 +478,10 @@ Flax), PyTorch, and/or TensorFlow.
 |          VisualBERT           |       ✅        |         ❌         |      ❌      |
 |              ViT              |       ✅        |         ✅         |      ✅      |
 |          ViT Hybrid           |       ✅        |         ❌         |      ❌      |
+|            VitDet             |       ✅        |         ❌         |      ❌      |
 |            ViTMAE             |       ✅        |         ✅         |      ❌      |
 |            ViTMSN             |       ✅        |         ❌         |      ❌      |
+|             VITS              |       ✅        |         ❌         |      ❌      |
 |             ViViT             |       ✅        |         ❌         |      ❌      |
 |           Wav2Vec2            |       ✅        |         ✅         |      ✅      |
 |      Wav2Vec2-Conformer       |       ✅        |         ❌         |      ❌      |

docs/source/en/internal/generation_utils.md

@@ -75,39 +75,104 @@ values. Here, for instance, it has two keys that are `sequences` and `scores`.
 We document here all output types.
 
 
-### GreedySearchOutput
-
-[[autodoc]] generation.GreedySearchDecoderOnlyOutput
+### PyTorch
 
 [[autodoc]] generation.GreedySearchEncoderDecoderOutput
 
-[[autodoc]] generation.FlaxGreedySearchOutput
-
-### SampleOutput
-
-[[autodoc]] generation.SampleDecoderOnlyOutput
+[[autodoc]] generation.GreedySearchDecoderOnlyOutput
 
 [[autodoc]] generation.SampleEncoderDecoderOutput
 
-[[autodoc]] generation.FlaxSampleOutput
-
-### BeamSearchOutput
-
-[[autodoc]] generation.BeamSearchDecoderOnlyOutput
+[[autodoc]] generation.SampleDecoderOnlyOutput
 
 [[autodoc]] generation.BeamSearchEncoderDecoderOutput
 
-### BeamSampleOutput
+[[autodoc]] generation.BeamSearchDecoderOnlyOutput
+
+[[autodoc]] generation.BeamSampleEncoderDecoderOutput
 
 [[autodoc]] generation.BeamSampleDecoderOnlyOutput
 
-[[autodoc]] generation.BeamSampleEncoderDecoderOutput
+[[autodoc]] generation.ContrastiveSearchEncoderDecoderOutput
+
+[[autodoc]] generation.ContrastiveSearchDecoderOnlyOutput
+
+### TensorFlow
+
+[[autodoc]] generation.TFGreedySearchEncoderDecoderOutput
+
+[[autodoc]] generation.TFGreedySearchDecoderOnlyOutput
+
+[[autodoc]] generation.TFSampleEncoderDecoderOutput
+
+[[autodoc]] generation.TFSampleDecoderOnlyOutput
+
+[[autodoc]] generation.TFBeamSearchEncoderDecoderOutput
+
+[[autodoc]] generation.TFBeamSearchDecoderOnlyOutput
+
+[[autodoc]] generation.TFBeamSampleEncoderDecoderOutput
+
+[[autodoc]] generation.TFBeamSampleDecoderOnlyOutput
+
+[[autodoc]] generation.TFContrastiveSearchEncoderDecoderOutput
+
+[[autodoc]] generation.TFContrastiveSearchDecoderOnlyOutput
+
+### FLAX
+
+[[autodoc]] generation.FlaxSampleOutput
+
+[[autodoc]] generation.FlaxGreedySearchOutput
+
+[[autodoc]] generation.FlaxBeamSearchOutput
 
 ## LogitsProcessor
 
 A [`LogitsProcessor`] can be used to modify the prediction scores of a language model head for
 generation.
 
+### PyTorch
+
+[[autodoc]] AlternatingCodebooksLogitsProcessor
+    - __call__
+
+[[autodoc]] ClassifierFreeGuidanceLogitsProcessor
+    - __call__
+
+[[autodoc]] EncoderNoRepeatNGramLogitsProcessor
+    - __call__
+
+[[autodoc]] EncoderRepetitionPenaltyLogitsProcessor
+    - __call__
+
+[[autodoc]] EpsilonLogitsWarper
+    - __call__
+
+[[autodoc]] EtaLogitsWarper
+    - __call__
+
+[[autodoc]] ExponentialDecayLengthPenalty
+    - __call__
+
+[[autodoc]] ForcedBOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] ForcedEOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] ForceTokensLogitsProcessor
+    - __call__
+
+[[autodoc]] HammingDiversityLogitsProcessor
+    - __call__
+
+[[autodoc]] InfNanRemoveLogitsProcessor
+    - __call__
+
+[[autodoc]] LogitNormalization
+    - __call__
+
 [[autodoc]] LogitsProcessor
     - __call__
 
@@ -123,43 +188,54 @@ generation.
 [[autodoc]] MinNewTokensLengthLogitsProcessor
     - __call__
 
-[[autodoc]] TemperatureLogitsWarper
-    - __call__
-
-[[autodoc]] RepetitionPenaltyLogitsProcessor
-    - __call__
-
-[[autodoc]] TopPLogitsWarper
-    - __call__
-
-[[autodoc]] TopKLogitsWarper
-    - __call__
-
-[[autodoc]] TypicalLogitsWarper
+[[autodoc]] NoBadWordsLogitsProcessor
     - __call__
 
 [[autodoc]] NoRepeatNGramLogitsProcessor
     - __call__
 
-[[autodoc]] SequenceBiasLogitsProcessor
-    - __call__
-
-[[autodoc]] NoBadWordsLogitsProcessor
-    - __call__
-
 [[autodoc]] PrefixConstrainedLogitsProcessor
     - __call__
 
-[[autodoc]] HammingDiversityLogitsProcessor
+[[autodoc]] RepetitionPenaltyLogitsProcessor
     - __call__
 
-[[autodoc]] ForcedBOSTokenLogitsProcessor
+[[autodoc]] SequenceBiasLogitsProcessor
     - __call__
 
-[[autodoc]] ForcedEOSTokenLogitsProcessor
+[[autodoc]] SuppressTokensAtBeginLogitsProcessor
     - __call__
 
-[[autodoc]] InfNanRemoveLogitsProcessor
+[[autodoc]] SuppressTokensLogitsProcessor
+    - __call__
+
+[[autodoc]] TemperatureLogitsWarper
+    - __call__
+
+[[autodoc]] TopKLogitsWarper
+    - __call__
+
+[[autodoc]] TopPLogitsWarper
+    - __call__
+
+[[autodoc]] TypicalLogitsWarper
+    - __call__
+
+[[autodoc]] UnbatchedClassifierFreeGuidanceLogitsProcessor
+    - __call__
+
+[[autodoc]] WhisperTimeStampLogitsProcessor
+    - __call__
+
+### TensorFlow
+
+[[autodoc]] TFForcedBOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] TFForcedEOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] TFForceTokensLogitsProcessor
     - __call__
 
 [[autodoc]] TFLogitsProcessor
@@ -171,15 +247,6 @@ generation.
 [[autodoc]] TFLogitsWarper
     - __call__
 
-[[autodoc]] TFTemperatureLogitsWarper
-    - __call__
-
-[[autodoc]] TFTopPLogitsWarper
-    - __call__
-
-[[autodoc]] TFTopKLogitsWarper
-    - __call__
-
 [[autodoc]] TFMinLengthLogitsProcessor
     - __call__
 
@@ -192,10 +259,30 @@ generation.
 [[autodoc]] TFRepetitionPenaltyLogitsProcessor
     - __call__
 
-[[autodoc]] TFForcedBOSTokenLogitsProcessor
+[[autodoc]] TFSuppressTokensAtBeginLogitsProcessor
     - __call__
 
-[[autodoc]] TFForcedEOSTokenLogitsProcessor
+[[autodoc]] TFSuppressTokensLogitsProcessor
+    - __call__
+
+[[autodoc]] TFTemperatureLogitsWarper
+    - __call__
+
+[[autodoc]] TFTopKLogitsWarper
+    - __call__
+
+[[autodoc]] TFTopPLogitsWarper
+    - __call__
+
+### FLAX
+
+[[autodoc]] FlaxForcedBOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] FlaxForcedEOSTokenLogitsProcessor
+    - __call__
+
+[[autodoc]] FlaxForceTokensLogitsProcessor
     - __call__
 
 [[autodoc]] FlaxLogitsProcessor
@@ -207,27 +294,30 @@ generation.
 [[autodoc]] FlaxLogitsWarper
     - __call__
 
-[[autodoc]] FlaxTemperatureLogitsWarper
+[[autodoc]] FlaxMinLengthLogitsProcessor
     - __call__
 
-[[autodoc]] FlaxTopPLogitsWarper
+[[autodoc]] FlaxSuppressTokensAtBeginLogitsProcessor
+    - __call__
+
+[[autodoc]] FlaxSuppressTokensLogitsProcessor
+    - __call__
+
+[[autodoc]] FlaxTemperatureLogitsWarper
     - __call__
 
 [[autodoc]] FlaxTopKLogitsWarper
     - __call__
 
-[[autodoc]] FlaxForcedBOSTokenLogitsProcessor
+[[autodoc]] FlaxTopPLogitsWarper
     - __call__
 
-[[autodoc]] FlaxForcedEOSTokenLogitsProcessor
-    - __call__
-
-[[autodoc]] FlaxMinLengthLogitsProcessor
+[[autodoc]] FlaxWhisperTimeStampLogitsProcessor
     - __call__
 
 ## StoppingCriteria
 
-A [`StoppingCriteria`] can be used to change when to stop generation (other than EOS token).
+A [`StoppingCriteria`] can be used to change when to stop generation (other than EOS token). Please note that this is exclusivelly available to our PyTorch implementations.
 
 [[autodoc]] StoppingCriteria
     - __call__
@@ -243,7 +333,7 @@ A [`StoppingCriteria`] can be used to change when to stop generation (other than
 
 ## Constraints
 
-A [`Constraint`] can be used to force the generation to include specific tokens or sequences in the output.
+A [`Constraint`] can be used to force the generation to include specific tokens or sequences in the output. Please note that this is exclusivelly available to our PyTorch implementations.
 
 [[autodoc]] Constraint
 

docs/source/en/llm_tutorial.md

@@ -0,0 +1,221 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+
+# Generation with LLMs
+
+[[open-in-colab]]
+
+LLMs, or Large Language Models, are the key component behind text generation. In a nutshell, they consist of large pretrained transformer models trained to predict the next word (or, more precisely, token) given some input text. Since they predict one token at a time, you need to do something more elaborate to generate new sentences other than just calling the model -- you need to do autoregressive generation.
+
+Autoregressive generation is the inference-time procedure of iteratively calling a model with its own generated outputs, given a few initial inputs. In 🤗 Transformers, this is handled by the [`~generation.GenerationMixin.generate`] method, which is available to all models with generative capabilities.
+
+This tutorial will show you how to:
+
+* Generate text with an LLM
+* Avoid common pitfalls
+* Next steps to help you get the most out of your LLM
+
+Before you begin, make sure you have all the necessary libraries installed:
+
+```bash
+pip install transformers bitsandbytes>=0.39.0 -q
+```
+
+
+## Generate text
+
+A language model trained for [causal language modeling](tasks/language_modeling) takes a sequence of text tokens as input and returns the probability distribution for the next token.
+
+<!-- [GIF 1 -- FWD PASS] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_1_1080p.mov"
+    ></video>
+    <figcaption>"Forward pass of an LLM"</figcaption>
+</figure>
+
+A critical aspect of autoregressive generation with LLMs is how to select the next token from this probability distribution. Anything goes in this step as long as you end up with a token for the next iteration. This means it can be as simple as selecting the most likely token from the probability distribution or as complex as applying a dozen transformations before sampling from the resulting distribution.
+
+<!-- [GIF 2 -- TEXT GENERATION] -->
+<figure class="image table text-center m-0 w-full">
+    <video
+        style="max-width: 90%; margin: auto;"
+        autoplay loop muted playsinline
+        src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/blog/assisted-generation/gif_2_1080p.mov"
+    ></video>
+    <figcaption>"Autoregressive generation iteratively selects the next token from a probability distribution to generate text"</figcaption>
+</figure>
+
+The process depicted above is repeated iteratively until some stopping condition is reached. Ideally, the stopping condition is dictated by the model, which should learn when to output an end-of-sequence (`EOS`) token. If this is not the case, generation stops when some predefined maximum length is reached.
+
+Properly setting up the token selection step and the stopping condition is essential to make your model behave as you'd expect on your task. That is why we have a [`~generation.GenerationConfig`] file associated with each model, which contains a good default generative parameterization and is loaded alongside your model.
+
+Let's talk code!
+
+<Tip>
+
+If you're interested in basic LLM usage, our high-level [`Pipeline`](pipeline_tutorial) interface is a great starting point. However, LLMs often require advanced features like quantization and fine control of the token selection step, which is best done through [`~generation.GenerationMixin.generate`]. Autoregressive generation with LLMs is also resource-intensive and should be executed on a GPU for adequate throughput.
+
+</Tip>
+
+<!-- TODO: update example to llama 2 (or a newer popular baseline) when it becomes ungated -->
+First, you need to load the model.
+
+```py
+>>> from transformers import AutoModelForCausalLM
+
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+You'll notice two flags in the `from_pretrained` call:
+
+ - `device_map` ensures the model is moved to your GPU(s)
+ - `load_in_4bit` applies [4-bit dynamic quantization](main_classes/quantization) to massively reduce the resource requirements
+
+There are other ways to initialize a model, but this is a good baseline to begin with an LLM.
+
+Next, you need to preprocess your text input with a [tokenizer](tokenizer_summary).
+
+```py
+>>> from transformers import AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> model_inputs = tokenizer(["A list of colors: red, blue"], return_tensors="pt").to("cuda")
+```
+
+The `model_inputs` variable holds the tokenized text input, as well as the attention mask. While [`~generation.GenerationMixin.generate`] does its best effort to infer the attention mask when it is not passed, we recommend passing it whenever possible for optimal results.
+
+Finally, call the [`~generation.GenerationMixin.generate`] method to returns the generated tokens, which should be converted to text before printing.
+
+```py
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A list of colors: red, blue, green, yellow, black, white, and brown'
+```
+
+And that's it! In a few lines of code, you can harness the power of an LLM.
+
+
+## Common pitfalls
+
+There are many [generation strategies](generation_strategies), and sometimes the default values may not be appropriate for your use case. If your outputs aren't aligned with what you're expecting, we've created a list of the most common pitfalls and how to avoid them.
+
+```py
+>>> from transformers import AutoModelForCausalLM, AutoTokenizer
+
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model = AutoModelForCausalLM.from_pretrained(
+...     "openlm-research/open_llama_7b", device_map="auto", load_in_4bit=True
+... )
+```
+
+### Generated output is too short/long
+
+If not specified in the [`~generation.GenerationConfig`] file, `generate` returns up to 20 tokens by default. We highly recommend manually setting `max_new_tokens` in your `generate` call to control the maximum number of new tokens it can return. Keep in mind LLMs (more precisely, [decoder-only models](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt)) also return the input prompt as part of the output.
+
+
+```py
+>>> model_inputs = tokenizer(["A sequence of numbers: 1, 2"], return_tensors="pt").to("cuda")
+
+>>> # By default, the output will contain up to 20 tokens
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5'
+
+>>> # Setting `max_new_tokens` allows you to control the maximum length
+>>> generated_ids = model.generate(**model_inputs, max_new_tokens=50)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'A sequence of numbers: 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,'
+```
+
+### Incorrect generation mode
+
+By default, and unless specified in the [`~generation.GenerationConfig`] file, `generate` selects the most likely token at each iteration (greedy decoding). Depending on your task, this may be undesirable; creative tasks like chatbots or writing an essay benefit from sampling. On the other hand, input-grounded tasks like audio transcription or translation benefit from greedy decoding. Enable sampling with `do_sample=True`, and you can learn more about this topic in this [blog post](https://huggingface.co/blog/how-to-generate).
+
+```py
+>>> # Set seed or reproducibility -- you don't need this unless you want full reproducibility
+>>> from transformers import set_seed
+>>> set_seed(0)
+
+>>> model_inputs = tokenizer(["I am a cat."], return_tensors="pt").to("cuda")
+
+>>> # LLM + greedy decoding = repetitive, boring output
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat. I am a cat. I am a cat. I am a cat'
+
+>>> # With sampling, the output becomes more creative!
+>>> generated_ids = model.generate(**model_inputs, do_sample=True)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'I am a cat.\nI just need to be. I am always.\nEvery time'
+```
+
+### Wrong padding side
+
+LLMs are [decoder-only](https://huggingface.co/learn/nlp-course/chapter1/6?fw=pt) architectures, meaning they continue to iterate on your input prompt. If your inputs do not have the same length, they need to be padded. Since LLMs are not trained to continue from pad tokens, your input needs to be left-padded. Make sure you also don't forget to pass the attention mask to generate!
+
+```py
+>>> # The tokenizer initialized above has right-padding active by default: the 1st sequence,
+>>> # which is shorter, has padding on the right side. Generation fails.
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids[0], skip_special_tokens=True)[0]
+''
+
+>>> # With left-padding, it works as expected!
+>>> tokenizer = AutoTokenizer.from_pretrained("openlm-research/open_llama_7b", padding_side="left")
+>>> tokenizer.pad_token = tokenizer.eos_token  # Llama has no pad token by default
+>>> model_inputs = tokenizer(
+...     ["1, 2, 3", "A, B, C, D, E"], padding=True, return_tensors="pt"
+... ).to("cuda")
+>>> generated_ids = model.generate(**model_inputs)
+>>> tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
+'1, 2, 3, 4, 5, 6,'
+```
+
+<!-- TODO: when the prompting guide is ready, mention the importance of setting the right prompt in this section -->
+
+## Further resources
+
+While the autoregressive generation process is relatively straightforward, making the most out of your LLM can be a challenging endeavor because there are many moving parts. For your next steps to help you dive deeper into LLM usage and understanding:
+
+<!-- TODO: complete with new guides -->
+### Advanced generate usage
+
+1. [Guide](generation_strategies) on how to control different generation methods, how to set up the generation configuration file, and how to stream the output;
+2. API reference on [`~generation.GenerationConfig`], [`~generation.GenerationMixin.generate`], and [generate-related classes](internal/generation_utils).
+
+### LLM leaderboards
+
+1. [Open LLM Leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard), which focuses on the quality of the open-source models;
+2. [Open LLM-Perf Leaderboard](https://huggingface.co/spaces/optimum/llm-perf-leaderboard), which focuses on LLM throughput.
+
+### Latency and throughput
+
+1. [Guide](main_classes/quantization) on dynamic quantization, which shows you how to drastically reduce your memory requirements.
+
+### Related libraries
+
+1. [`text-generation-inference`](https://github.com/huggingface/text-generation-inference), a production-ready server for LLMs;
+2. [`optimum`](https://github.com/huggingface/optimum), an extension of 🤗 Transformers that optimizes for specific hardware devices.

docs/source/en/main_classes/agent.md

@@ -18,7 +18,7 @@ rendered properly in your Markdown viewer.
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API which is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API which is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>

docs/source/en/main_classes/deepspeed.md

@@ -168,6 +168,8 @@ If after trying everything suggested you still encounter build issues, please, p
 
 To deploy the DeepSpeed integration adjust the [`Trainer`] command line arguments to include a new argument `--deepspeed ds_config.json`, where `ds_config.json` is the DeepSpeed configuration file as
    documented [here](https://www.deepspeed.ai/docs/config-json/). The file naming is up to you.
+   It's recommended to use DeepSpeed's `add_config_arguments` utility to add the necessary command line arguments to your code.
+   For more information please see [DeepSpeed's Argument Parsing](https://deepspeed.readthedocs.io/en/latest/initialize.html#argument-parsing) doc.
 
 You can use a launcher of your choice here. You can continue using the pytorch launcher:
 
@@ -1410,7 +1412,7 @@ the full fp32 mode, by explicitly disabling the otherwise default fp16 mixed pre
 ```json
 {
     "fp16": {
-        "enabled": "false",
+        "enabled": false,
     }
 }
 ```
@@ -2063,20 +2065,20 @@ In this case you usually need to raise the value of `initial_scale_power`. Setti
 
 ## Non-Trainer Deepspeed Integration
 
-The [`~deepspeed.HfDeepSpeedConfig`] is used to integrate Deepspeed into the 🤗 Transformers core
+The [`~integrations.HfDeepSpeedConfig`] is used to integrate Deepspeed into the 🤗 Transformers core
 functionality, when [`Trainer`] is not used. The only thing that it does is handling Deepspeed ZeRO-3 param gathering and automatically splitting the model onto multiple gpus during `from_pretrained` call. Everything else you have to do by yourself.
 
 When using [`Trainer`] everything is automatically taken care of.
 
 When not using [`Trainer`], to efficiently deploy DeepSpeed ZeRO-3, you must instantiate the
-[`~deepspeed.HfDeepSpeedConfig`] object before instantiating the model and keep that object alive.
+[`~integrations.HfDeepSpeedConfig`] object before instantiating the model and keep that object alive.
 
 If you're using Deepspeed ZeRO-1 or ZeRO-2 you don't need to use `HfDeepSpeedConfig` at all.
 
 For example for a pretrained model:
 
 ```python
-from transformers.deepspeed import HfDeepSpeedConfig
+from transformers.integrations import HfDeepSpeedConfig
 from transformers import AutoModel
 import deepspeed
 
@@ -2090,7 +2092,7 @@ engine = deepspeed.initialize(model=model, config_params=ds_config, ...)
 or for non-pretrained model:
 
 ```python
-from transformers.deepspeed import HfDeepSpeedConfig
+from transformers.integrations import HfDeepSpeedConfig
 from transformers import AutoModel, AutoConfig
 import deepspeed
 
@@ -2106,7 +2108,7 @@ Please note that if you're not using the [`Trainer`] integration, you're complet
 
 ## HfDeepSpeedConfig
 
-[[autodoc]] deepspeed.HfDeepSpeedConfig
+[[autodoc]] integrations.HfDeepSpeedConfig
     - all
 
 ### Custom DeepSpeed ZeRO Inference
@@ -2159,7 +2161,7 @@ Make sure to:
 
 
 from transformers import AutoTokenizer, AutoConfig, AutoModelForSeq2SeqLM
-from transformers.deepspeed import HfDeepSpeedConfig
+from transformers.integrations import HfDeepSpeedConfig
 import deepspeed
 import os
 import torch

docs/source/en/main_classes/output.md

@@ -40,6 +40,13 @@ an optional `attentions` attribute. Here we have the `loss` since we passed alon
 `hidden_states` and `attentions` because we didn't pass `output_hidden_states=True` or
 `output_attentions=True`.
 
+<Tip>
+
+When passing `output_hidden_states=True` you may expect the `outputs.hidden_states[-1]` to match `outputs.last_hidden_states` exactly.
+However, this is not always the case. Some models apply normalization or subsequent process to the last hidden state when it's returned.
+</Tip>
+
+
 You can access each attribute as you would usually do, and if that attribute has not been returned by the model, you
 will get `None`. Here for instance `outputs.loss` is the loss computed by the model, and `outputs.attentions` is
 `None`.

docs/source/en/main_classes/pipelines.md

@@ -318,6 +318,13 @@ Pipelines available for audio tasks include the following.
     - __call__
     - all
 
+### TextToAudioPipeline
+
+[[autodoc]] TextToAudioPipeline
+    - __call__
+    - all
+
+
 ### ZeroShotAudioClassificationPipeline
 
 [[autodoc]] ZeroShotAudioClassificationPipeline

docs/source/en/main_classes/quantization.md

@@ -16,6 +16,138 @@ rendered properly in your Markdown viewer.
 
 # Quantize 🤗 Transformers models
 
+## `AutoGPTQ` Integration
+
+🤗 Transformers has integrated `optimum` API to perform GPTQ quantization on language models. You can load and quantize your model in 8, 4, 3 or even 2 bits without a big drop of performance and faster inference speed! This is supported by most GPU hardwares.
+
+To learn more about the the quantization model, check out: 
+- the [GPTQ](https://arxiv.org/pdf/2210.17323.pdf) paper
+- the `optimum` [guide](https://huggingface.co/docs/optimum/llm_quantization/usage_guides/quantization) on GPTQ quantization
+- the [`AutoGPTQ`](https://github.com/PanQiWei/AutoGPTQ) library used as the backend
+
+### Requirements
+
+You need to have the following requirements installed to run the code below: 
+
+- Install latest `AutoGPTQ` library
+`pip install auto-gptq`
+
+- Install latest `optimum` from source 
+`pip install git+https://github.com/huggingface/optimum.git`
+
+- Install latest `transformers` from source 
+`pip install git+https://github.com/huggingface/transformers.git`
+
+- Install latest `accelerate` library 
+`pip install --upgrade accelerate`
+
+Note that GPTQ integration supports for now only text models and you may encounter unexpected behaviour for vision, speech or multi-modal models.
+
+### Load and quantize a model
+
+GPTQ is a quantization method that requires weights calibration before using the quantized models. If you want to quantize transformers model from scratch, it might take some time before producing the quantized model (~5 min on a Google colab for `facebook/opt-350m` model). 
+
+Hence, there are two different scenarios where you want to use GPTQ-quantized models. The first use case would be to load models that has been already quantized by other users that are available on the Hub, the second use case would be to quantize your model from scratch and save it or push it on the Hub so that other users can also use it.
+#### GPTQ Configuration
+
+In order to load and quantize a model, you need to create a [`GPTQConfig`]. You need to pass the number of `bits`, a `dataset` in order to calibrate the quantization and the `tokenizer` of the model in order prepare the dataset.
+
+```python 
+model_id = "facebook/opt-125m"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+gptq_config = GPTQConfig(bits=4, dataset = "c4", tokenizer=tokenizer)
+```
+
+Note that you can pass your own dataset as a list of string. However, it is highly recommended to use the dataset from the GPTQ paper. 
+```python
+dataset = ["auto-gptq is an easy-to-use model quantization library with user-friendly apis, based on GPTQ algorithm."]
+quantization = GPTQConfig(bits=4, dataset = dataset, tokenizer=tokenizer)
+```
+
+#### Quantization
+
+You can quantize a model by using `from_pretrained` and setting the `quantization_config`. 
+
+```python
+from transformers import AutoModelForCausalLM
+model = AutoModelForCausalLM.from_pretrained(model_id, quantization_config=gptq_config)
+```
+Note that you will need a GPU to quantize a model. We will put the model in the cpu and move the modules back and forth to the gpu in order to quantize them.
+
+If you want to maximize your gpus usage while using cpu offload, you can set `device_map = "auto"`.
+```python
+from transformers import AutoModelForCausalLM
+model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", quantization_config=gptq_config)
+```
+Note that disk offload is not supported. Furthermore, if you are out of memory because of the dataset, you may have to pass `max_memory` in `from_pretained`. Checkout this [guide](https://huggingface.co/docs/accelerate/usage_guides/big_modeling#designing-a-device-map) to learn more about `device_map` and `max_memory`.
+
+<Tip warning={true}>
+GPTQ quantization only works for text model for now. Futhermore, the quantization process can a lot of time depending on one's hardware (175B model = 4 gpu hours using NVIDIA A100). Please check on the hub if there is not a GPTQ quantized version of the model. If not, you can submit a demand on github. 
+</Tip>
+
+### Push quantized model to 🤗 Hub
+
+You can push the quantized model like any 🤗 model to Hub with `push_to_hub`. The quantization config will be saved and pushed along the model. 
+
+```python
+quantized_model.push_to_hub("opt-125m-gptq")
+tokenizer.push_to_hub("opt-125m-gptq")
+```
+
+If you want to save your quantized model on your local machine, you can also do it with `save_pretrained`: 
+```python
+quantized_model.save_pretrained("opt-125m-gptq")
+tokenizer.save_pretrained("opt-125m-gptq")
+```
+
+Note that if you have quantized your model with a `device_map`, make sure to move the entire model to one of your gpus or the `cpu` before saving it. 
+```python
+quantized_model.to("cpu")
+quantized_model.save_pretrained("opt-125m-gptq")
+```
+
+### Load a quantized model from the 🤗 Hub
+
+You can load a quantized model from the Hub by using `from_pretrained`.
+Make sure that the pushed weights are quantized, by checking that the attribute `quantization_config` is present in the model configuration object.
+
+```python
+from transformers import AutoModelForCausalLM
+model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq")
+```
+
+If you want to load a model faster and without allocating more memory than needed, the `device_map` argument also works with quantized model. Make sure that you have `accelerate` library installed.
+```python
+from transformers import AutoModelForCausalLM
+model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto")
+```
+
+### Exllama kernels for faster inference
+
+For 4-bit model, you can use the exllama kernels in order to a faster inference speed. It is activated by default. You can change that behavior by passing `disable_exllama` in [`GPTQConfig`]. This will overwrite the quantization config stored in the config. Note that you will only be able to overwrite the attributes related to the kernels. Furthermore, you need to have the entire model on gpus if you want to use exllama kernels. 
+
+```py
+import torch
+gptq_config = GPTQConfig(bits=4, disable_exllama=False)
+model = AutoModelForCausalLM.from_pretrained("{your_username}/opt-125m-gptq", device_map="auto", quantization_config = gptq_config)
+```
+
+Note that only 4-bit models are supported for now. Furthermore, it is recommended to deactivate the exllama kernels if you are finetuning a quantized model with peft. 
+
+#### Fine-tune a quantized model 
+
+With the official support of adapters in the Hugging Face ecosystem, you can fine-tune models that have been quantized with GPTQ. 
+Please have a look at [`peft`](https://github.com/huggingface/peft) library for more details.
+
+### Example demo
+
+Check out the Google Colab [notebook](https://colab.research.google.com/drive/1_TIrmuKOFhuRRiTWN94iLKUFu6ZX4ceb?usp=sharing) to learn how to quantize your model with GPTQ and how finetune the quantized model with peft. 
+
+### GPTQConfig
+
+[[autodoc]] GPTQConfig
+
+
 ## `bitsandbytes` Integration
 
 🤗 Transformers is closely integrated with most used modules on `bitsandbytes`. You can load your model in 8-bit precision with few lines of code.
@@ -29,6 +161,29 @@ If you want to quantize your own pytorch model, check out this [documentation](h
 
 Here are the things you can do using `bitsandbytes` integration
 
+### General usage
+
+You can quantize a model by using the `load_in_8bit` or `load_in_4bit` argument when calling the [`~PreTrainedModel.from_pretrained`] method as long as your model supports loading with 🤗 Accelerate and contains `torch.nn.Linear` layers. This should work for any modality as well.
+
+```python
+from transformers import AutoModelForCausalLM
+
+model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True)
+model_4bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_4bit=True)
+```
+
+By default all other modules (e.g. `torch.nn.LayerNorm`) will be converted in `torch.float16`, but if you want to change their `dtype` you can overwrite the `torch_dtype` argument:
+
+```python
+>>> import torch
+>>> from transformers import AutoModelForCausalLM
+
+>>> model_8bit = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", load_in_8bit=True, torch_dtype=torch.float32)
+>>> model_8bit.model.decoder.layers[-1].final_layer_norm.weight.dtype
+torch.float32
+```
+
+
 ### FP4 quantization 
 
 #### Requirements
@@ -41,7 +196,7 @@ Make sure that you have installed the requirements below before running any of t
 - Install latest `accelerate`
 `pip install --upgrade accelerate`
 
-- Install latest `transformers` from source 
+- Install latest `transformers`
 `pip install --upgrade transformers`
 
 #### Tips and best practices
@@ -106,9 +261,9 @@ Note also that `device_map` is optional but setting `device_map = 'auto'` is pre
 
 </Tip>
 
-#### Advanced usecases
+#### Advanced use cases
 
-Here we will cover some advanced usecases you can perform with FP4 quantization 
+Here we will cover some advanced use cases you can perform with FP4 quantization 
 
 ##### Change the compute dtype
 
@@ -184,15 +339,15 @@ model = AutoModelForCausalLM.from_pretrained("{your_username}/bloom-560m-8bit",
 Note that in this case, you don't need to specify the arguments `load_in_8bit=True`, but you need to make sure that `bitsandbytes` and `accelerate` are installed.
 Note also that `device_map` is optional but setting `device_map = 'auto'` is prefered for inference as it will dispatch efficiently the model on the available ressources.
 
-### Advanced usecases
+### Advanced use cases
 
 This section is intended to advanced users, that want to explore what it is possible to do beyond loading and running 8-bit models.
 
 #### Offload between `cpu` and `gpu`
 
-One of the advanced usecase of this is being able to load a model and dispatch the weights between `CPU` and `GPU`. Note that the weights that will be dispatched on CPU **will not** be converted in 8-bit, thus kept in `float32`. This feature is intended for users that want to fit a very large model and dispatch the model between GPU and CPU.
+One of the advanced use case of this is being able to load a model and dispatch the weights between `CPU` and `GPU`. Note that the weights that will be dispatched on CPU **will not** be converted in 8-bit, thus kept in `float32`. This feature is intended for users that want to fit a very large model and dispatch the model between GPU and CPU.
 
-First, load a `BitsAndBytesConfig` from `transformers` and set the attribute `llm_int8_enable_fp32_cpu_offload` to `True`:
+First, load a [`BitsAndBytesConfig`] from `transformers` and set the attribute `llm_int8_enable_fp32_cpu_offload` to `True`:
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
@@ -226,7 +381,7 @@ And that's it! Enjoy your model!
 
 You can play with the `llm_int8_threshold` argument to change the threshold of the outliers. An "outlier" is a hidden state value that is greater than a certain threshold. 
 This corresponds to the outlier threshold for outlier detection as described in `LLM.int8()` paper. Any hidden states value that is above this threshold will be considered an outlier and the operation on those values will be done in fp16. Values are usually normally distributed, that is, most values are in the range [-3.5, 3.5], but there are some exceptional systematic outliers that are very differently distributed for large models. These outliers are often in the interval [-60, -6] or [6, 60]. Int8 quantization works well for values of magnitude ~5, but beyond that, there is a significant performance penalty. A good default threshold is 6, but a lower threshold might be needed for more unstable models (small models, fine-tuning).
-This argument can impact the inference speed of the model. We suggest to play with this parameter to find which one is the best for your usecase.
+This argument can impact the inference speed of the model. We suggest to play with this parameter to find which one is the best for your use case.
 
 ```python
 from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
@@ -280,4 +435,4 @@ Note that you don't need to pass `device_map` when loading the model for trainin
 
 ## Quantization with 🤗 `optimum` 
 
-Please have a look at [Optimum documentation](https://huggingface.co/docs/optimum/index) to learn more about quantization methods that are supported by `optimum` and see if these are applicable for your usecase.
+Please have a look at [Optimum documentation](https://huggingface.co/docs/optimum/index) to learn more about quantization methods that are supported by `optimum` and see if these are applicable for your use case.

docs/source/en/main_classes/tokenizer.md

@@ -55,6 +55,7 @@ to a given token).
 
 [[autodoc]] PreTrainedTokenizer
     - __call__
+    - apply_chat_template
     - batch_decode
     - decode
     - encode
@@ -68,6 +69,7 @@ loaded very simply into 🤗 transformers. Take a look at the [Using tokenizers
 
 [[autodoc]] PreTrainedTokenizerFast
     - __call__
+    - apply_chat_template
     - batch_decode
     - decode
     - encode

docs/source/en/main_classes/trainer.md

@@ -60,7 +60,7 @@ from transformers import Trainer
 
 class CustomTrainer(Trainer):
     def compute_loss(self, model, inputs, return_outputs=False):
-        labels = inputs.get("labels")
+        labels = inputs.pop("labels")
         # forward pass
         outputs = model(**inputs)
         logits = outputs.get("logits")
@@ -456,6 +456,10 @@ as the model saving with FSDP activated is only available with recent fixes.
     If `"True"`, FSDP explicitly prefetches the next upcoming all-gather while executing in the forward pass. 
   - `limit_all_gathers` can be specified in the config file. 
     If `"True"`, FSDP explicitly synchronizes the CPU thread to prevent too many in-flight all-gathers.
+  - `activation_checkpointing` can be specified in the config file.
+    If `"True"`, FSDP activation checkpointing is a technique to reduce memory usage by clearing activations of
+    certain layers and recomputing them during a backward pass. Effectively, this trades extra computation time
+    for reduced memory usage.
 
 **Few caveats to be aware of**
 - it is incompatible with `generate`, thus is incompatible with `--predict_with_generate` 

docs/source/en/model_doc/auto.md

@@ -330,6 +330,14 @@ The following auto classes are available for the following audio tasks.
 
 [[autodoc]] AutoModelForAudioXVector
 
+### AutoModelForTextToSpectrogram
+
+[[autodoc]] AutoModelForTextToSpectrogram
+
+### AutoModelForTextToWaveform
+
+[[autodoc]] AutoModelForTextToWaveform
+
 ## Multimodal
 
 The following auto classes are available for the following multimodal tasks.

docs/source/en/model_doc/bark.md

@@ -26,8 +26,67 @@ Bark is made of 4 main models:
 
 It should be noted that each of the first three modules can support conditional speaker embeddings to condition the output sound according to specific predefined voice.
 
+### Optimizing Bark
 
-### Tips:
+Bark can be optimized with just a few extra lines of code, which **significantly reduces its memory footprint** and **accelerates inference**.
+
+#### Using half-precision
+
+You can speed up inference and reduce memory footprint by 50% simply by loading the model in half-precision.
+
+```python
+from transformers import BarkModel
+import torch
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+model = BarkModel.from_pretrained("suno/bark-small", torch_dtype=torch.float16).to(device)
+```
+
+#### Using 🤗 Better Transformer
+
+Better Transformer is an 🤗 Optimum feature that performs kernel fusion under the hood. You can gain 20% to 30% in speed with zero performance degradation. It only requires one line of code to export the model to 🤗 Better Transformer:
+
+```python
+model =  model.to_bettertransformer()
+```
+
+Note that 🤗 Optimum must be installed before using this feature. [Here's how to install it.](https://huggingface.co/docs/optimum/installation)
+
+#### Using CPU offload
+
+As mentioned above, Bark is made up of 4 sub-models, which are called up sequentially during audio generation. In other words, while one sub-model is in use, the other sub-models are idle.
+
+If you're using a CUDA device, a simple solution to benefit from an 80% reduction in memory footprint is to offload the GPU's submodels when they're idle. This operation is called CPU offloading. You can use it with one line of code.
+
+```python
+model.enable_cpu_offload()
+```
+
+Note that 🤗 Accelerate must be installed before using this feature. [Here's how to install it.](https://huggingface.co/docs/accelerate/basic_tutorials/install)
+
+#### Combining optimizaton techniques
+
+You can combine optimization techniques, and use CPU offload, half-precision and 🤗 Better Transformer all at once.
+
+```python
+from transformers import BarkModel
+import torch
+
+device = "cuda" if torch.cuda.is_available() else "cpu"
+
+# load in fp16
+model = BarkModel.from_pretrained("suno/bark-small", torch_dtype=torch.float16).to(device)
+
+# convert to bettertransformer
+model = BetterTransformer.transform(model, keep_original_model=False)
+
+# enable CPU offload
+model.enable_cpu_offload()
+```
+
+Find out more on inference optimization techniques [here](https://huggingface.co/docs/transformers/perf_infer_gpu_one).
+
+### Tips
 
 Suno offers a library of voice presets in a number of languages [here](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c).
 These presets are also uploaded in the hub [here](https://huggingface.co/suno/bark-small/tree/main/speaker_embeddings) or [here](https://huggingface.co/suno/bark/tree/main/speaker_embeddings).
@@ -103,6 +162,7 @@ The original code can be found [here](https://github.com/suno-ai/bark).
 
 [[autodoc]] BarkModel
     - generate
+    - enable_cpu_offload
 
 ## BarkSemanticModel
 

docs/source/en/model_doc/blip.md

@@ -30,7 +30,7 @@ The abstract from the paper is the following:
 *Vision-Language Pre-training (VLP) has advanced the performance for many vision-language tasks. 
 However, most existing pre-trained models only excel in either understanding-based tasks or generation-based tasks. Furthermore, performance improvement has been largely achieved by scaling up the dataset with noisy image-text pairs collected from the web, which is a suboptimal source of supervision. In this paper, we propose BLIP, a new VLP framework which transfers flexibly to both vision-language understanding and generation tasks. BLIP effectively utilizes the noisy web data by bootstrapping the captions, where a captioner generates synthetic captions and a filter removes the noisy ones. We achieve state-of-the-art results on a wide range of vision-language tasks, such as image-text retrieval (+2.7% in average recall@1), image captioning (+2.8% in CIDEr), and VQA (+1.6% in VQA score). BLIP also demonstrates strong generalization ability when directly transferred to videolanguage tasks in a zero-shot manner. Code, models, and datasets are released.*
 
-![BLIP.gif](https://s3.amazonaws.com/moonup/production/uploads/1670928184033-62441d1d9fdefb55a0b7d12c.gif)
+![BLIP.gif](https://cdn-uploads.huggingface.co/production/uploads/1670928184033-62441d1d9fdefb55a0b7d12c.gif)
 
 This model was contributed by [ybelkada](https://huggingface.co/ybelkada).
 The original code can be found [here](https://github.com/salesforce/BLIP).

docs/source/en/model_doc/bloom.md

@@ -85,3 +85,13 @@ See also:
 
 [[autodoc]] BloomForQuestionAnswering
     - forward
+
+## FlaxBloomModel
+
+[[autodoc]] FlaxBloomModel
+    - __call__
+
+## FlaxBloomForCausalLM
+
+[[autodoc]] FlaxBloomForCausalLM
+    - __call__

docs/source/en/model_doc/bros.md

@@ -0,0 +1,115 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# BROS
+
+## Overview
+
+The BROS model was proposed in [BROS: A Pre-trained Language Model Focusing on Text and Layout for Better Key Information Extraction from Documents](https://arxiv.org/abs/2108.04539) by Teakgyu Hong, Donghyun Kim, Mingi Ji, Wonseok Hwang, Daehyun Nam, Sungrae Park.
+
+BROS stands for *BERT Relying On Spatiality*. It is an encoder-only Transformer model that takes a sequence of tokens and their bounding boxes as inputs and outputs a sequence of hidden states. BROS encode relative spatial information instead of using absolute spatial information.
+
+It is pre-trained with two objectives: a token-masked language modeling objective (TMLM) used in BERT, and a novel area-masked language modeling objective (AMLM)
+In TMLM, tokens are randomly masked, and the model predicts the masked tokens using spatial information and other unmasked tokens.
+AMLM is a 2D version of TMLM. It randomly masks text tokens and predicts with the same information as TMLM, but it masks text blocks (areas).
+
+`BrosForTokenClassification` has a simple linear layer on top of BrosModel. It predicts the label of each token.
+`BrosSpadeEEForTokenClassification` has an `initial_token_classifier` and `subsequent_token_classifier` on top of BrosModel. `initial_token_classifier` is used to predict the first token of each entity, and `subsequent_token_classifier` is used to predict the next token of within entity. `BrosSpadeELForTokenClassification` has an `entity_linker` on top of BrosModel. `entity_linker` is used to predict the relation between two entities.
+
+`BrosForTokenClassification` and `BrosSpadeEEForTokenClassification` essentially perform the same job. However, `BrosForTokenClassification` assumes input tokens are perfectly serialized (which is very challenging task since they exist in a 2D space), while `BrosSpadeEEForTokenClassification` allows for more flexibility in handling serialization errors as it predicts next connection tokens from one token.
+
+`BrosSpadeELForTokenClassification` perform the intra-entity linking task. It predicts relation from one token (of one entity) to another token (of another entity) if these two entities share some relation.
+
+BROS achieves comparable or better result on Key Information Extraction (KIE) benchmarks such as FUNSD, SROIE, CORD and SciTSR, without relying on explicit visual features.
+
+
+The abstract from the paper is the following:
+
+*Key information extraction (KIE) from document images requires understanding the contextual and spatial semantics of texts in two-dimensional (2D) space. Many recent studies try to solve the task by developing pre-trained language models focusing on combining visual features from document images with texts and their layout. On the other hand, this paper tackles the problem by going back to the basic: effective combination of text and layout. Specifically, we propose a pre-trained language model, named BROS (BERT Relying On Spatiality), that encodes relative positions of texts in 2D space and learns from unlabeled documents with area-masking strategy. With this optimized training scheme for understanding texts in 2D space, BROS shows comparable or better performance compared to previous methods on four KIE benchmarks (FUNSD, SROIE*, CORD, and SciTSR) without relying on visual features. This paper also reveals two real-world challenges in KIE tasks-(1) minimizing the error from incorrect text ordering and (2) efficient learning from fewer downstream examples-and demonstrates the superiority of BROS over previous methods.*
+
+Tips:
+
+- [`~transformers.BrosModel.forward`] requires `input_ids` and `bbox` (bounding box). Each bounding box should be in (x0, y0, x1, y1) format (top-left corner, bottom-right corner). Obtaining of Bounding boxes depends on external OCR system. The `x` coordinate should be normalized by document image width, and the `y` coordinate should be normalized by document image height.
+
+```python
+def expand_and_normalize_bbox(bboxes, doc_width, doc_height):
+    # here, bboxes are numpy array
+
+    # Normalize bbox -> 0 ~ 1
+    bboxes[:, [0, 2]] = bboxes[:, [0, 2]] / width
+    bboxes[:, [1, 3]] = bboxes[:, [1, 3]] / height
+```
+
+- [`~transformers.BrosForTokenClassification.forward`, `~transformers.BrosSpadeEEForTokenClassification.forward`, `~transformers.BrosSpadeEEForTokenClassification.forward`] require not only `input_ids` and `bbox` but also `box_first_token_mask` for loss calculation. It is a mask to filter out non-first tokens of each box. You can obtain this mask by saving start token indices of bounding boxes when creating `input_ids` from words. You can make `box_first_token_mask` with following code,
+
+
+```python
+def make_box_first_token_mask(bboxes, words, tokenizer, max_seq_length=512):
+
+    box_first_token_mask = np.zeros(max_seq_length, dtype=np.bool_)
+
+    # encode(tokenize) each word from words (List[str])
+    input_ids_list: List[List[int]] = [tokenizer.encode(e, add_special_tokens=False) for e in words]
+
+    # get the length of each box
+    tokens_length_list: List[int] = [len(l) for l in input_ids_list]
+
+    box_end_token_indices = np.array(list(itertools.accumulate(tokens_length_list)))
+    box_start_token_indices = box_end_token_indices - np.array(tokens_length_list)
+
+    # filter out the indices that are out of max_seq_length
+    box_end_token_indices = box_end_token_indices[box_end_token_indices < max_seq_length - 1]
+    if len(box_start_token_indices) > len(box_end_token_indices):
+        box_start_token_indices = box_start_token_indices[: len(box_end_token_indices)]
+
+    # set box_start_token_indices to True
+    box_first_token_mask[box_start_token_indices] = True
+
+    return box_first_token_mask
+
+```
+
+- Demo scripts can be found [here](https://github.com/clovaai/bros).
+
+This model was contributed by [jinho8345](https://huggingface.co/jinho8345). The original code can be found [here](https://github.com/clovaai/bros).
+
+## BrosConfig
+
+[[autodoc]] BrosConfig
+
+## BrosProcessor
+
+[[autodoc]] BrosProcessor
+    - __call__
+
+## BrosModel
+
+[[autodoc]] BrosModel
+    - forward
+
+
+## BrosForTokenClassification
+
+[[autodoc]] BrosForTokenClassification
+    - forward
+
+
+## BrosSpadeEEForTokenClassification
+
+[[autodoc]] BrosSpadeEEForTokenClassification
+    - forward
+
+
+## BrosSpadeELForTokenClassification
+
+[[autodoc]] BrosSpadeELForTokenClassification
+    - forward

docs/source/en/model_doc/clip.md

@@ -184,6 +184,11 @@ The resource should ideally demonstrate something new instead of duplicating an
 [[autodoc]] FlaxCLIPTextModel
     - __call__
 
+## FlaxCLIPTextModelWithProjection
+
+[[autodoc]] FlaxCLIPTextModelWithProjection
+    - __call__
+
 ## FlaxCLIPVisionModel
 
 [[autodoc]] FlaxCLIPVisionModel

docs/source/en/model_doc/code_llama.md

@@ -0,0 +1,118 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contains specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# CodeLlama
+
+## Overview
+
+The Code Llama model was proposed in [Code Llama: Open Foundation Models for Code](https://ai.meta.com/research/publications/code-llama-open-foundation-models-for-code/) by Baptiste Rozière, Jonas Gehring, Fabian Gloeckle, Sten Sootla, Itai Gat, Xiaoqing Ellen Tan, Yossi Adi, Jingyu Liu, Tal Remez, Jérémy Rapin, Artyom Kozhevnikov, Ivan Evtimov, Joanna Bitton, Manish Bhatt, Cristian Canton Ferrer, Aaron Grattafiori, Wenhan Xiong, Alexandre Défossez, Jade Copet, Faisal Azhar, Hugo Touvron, Louis Martin, Nicolas Usunier, Thomas Scialom, Gabriel Synnaeve.
+
+The abstract from the paper is the following:
+
+*We release Code Llama, a family of large language models for code based on Llama 2 providing state-of-the-art performance among open models, infilling capabilities, support for large input contexts, and zero-shot instruction following ability for programming tasks. We provide multiple flavors to cover a wide range of applications: foundation models (Code Llama), Python specializations (Code Llama - Python), and instruction-following models (Code Llama - Instruct) with 7B, 13B and 34B parameters each. All models are trained on sequences of 16k tokens and show improvements on inputs with up to 100k tokens. 7B and 13B Code Llama and Code Llama - Instruct variants support infilling based on surrounding content. Code Llama reaches state-of-the-art performance among open models on several code benchmarks, with scores of up to 53% and 55% on HumanEval and MBPP, respectively. Notably, Code Llama - Python 7B outperforms Llama 2 70B on HumanEval and MBPP, and all our models outperform every other publicly available model on MultiPL-E. We release Code Llama under a permissive license that allows for both research and commercial use.*
+
+Check out all Code Llama models [here](https://huggingface.co/models?search=code_llama) and the officially released ones in the [codellama org](https://huggingface.co/codellama).
+
+<Tip warning={true}>
+
+The `Llama2` family models, on which Code Llama is based, were trained using `bfloat16`, but the original inference uses `float16`. Let's look at the different precisions:
+
+* `float32`: PyTorch convention on model initialization is to load models in `float32`, no matter with which `dtype` the model weights were stored. `transformers` also follows this convention for consistency with PyTorch. This will be picked by default. If you want the `AutoModel` API to cast the load the checkpoints with the storage weights type, you must specify `torch_dtype="auto"`, e.g. `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`.
+* `bfloat16`: Code Llama was trained with this precision, so we recommend using it for further training or fine-tuning.
+* `float16`: We recommend running inference using this precision, as it's usually faster than `bfloat16`, and evaluation metrics show no discernible degradation with respect to `bfloat16`. You can also run inference using `bfloat16`, and we recommend you check inference results with both `float16` and `bfloat16` after fine-tuning.
+
+As mentioned above, the `dtype` of the storage weights is mostly irrelevant unless you are using `torch_dtype="auto"` when initializing a model using. The reason is that the model will first be downloaded (using the `dtype` of the checkpoints online) and then will be casted to the default `dtype` of `torch` (becomes `torch.float32`). If there is a specified `torch_dtype`, it will be used instead.
+
+</Tip>
+
+Tips:
+
+- These models have the same architecture as the `Llama2` models
+- The infilling task is supported out of the box. You should be using the `tokenizer.fill_token` where you want your input to be filled.
+- The model conversion script is the same as for the `Llama2` family:
+
+Here is a sample usage
+```bash
+python src/transformers/models/llama/convert_llama_weights_to_hf.py \
+    --input_dir /path/to/downloaded/llama/weights --model_size 7B --output_dir /output/path
+```
+Note that executing the script requires enough CPU RAM to host the whole model in float16 precision (even if the biggest versions
+come in several checkpoints they each contain a part of each weight of the model, so we need to load them all in RAM).
+
+- After conversion, the model and tokenizer can be loaded via:
+
+```python
+>>> from transformers import LlamaForCausalLM, CodeLlamaTokenizer
+
+>>> tokenizer = CodeLlamaTokenizer.from_pretrained("codellama/CodeLlama-7b-hf")
+>>> model = LlamaForCausalLM.from_pretrained("codellama/CodeLlama-7b-hf")
+>>> PROMPT = '''def remove_non_ascii(s: str) -> str:
+    """ <FILL_ME>
+    return result
+'''
+>>> input_ids = tokenizer(PROMPT, return_tensors="pt")["input_ids"]
+>>> generated_ids = model.generate(input_ids, max_new_tokens=128)
+
+>>> filling = tokenizer.batch_decode(generated_ids[:, input_ids.shape[1]:], skip_special_tokens = True)[0]
+>>> print(PROMPT.replace("<FILL_ME>", filling))
+def remove_non_ascii(s: str) -> str:
+    """ Remove non-ASCII characters from a string.
+
+    Args:
+        s: The string to remove non-ASCII characters from.
+
+    Returns:
+        The string with non-ASCII characters removed.
+    """
+    result = ""
+    for c in s:
+        if ord(c) < 128:
+            result += c
+    return result
+```
+
+If you only want the infilled part:
+```python
+>>> from transformers import pipeline
+>>> import torch
+
+>>> generator = pipeline("text-generation",model="codellama/CodeLlama-7b-hf",torch_dtype=torch.float16, device_map="auto")
+>>> generator('def remove_non_ascii(s: str) -> str:\n    """ <FILL_ME>\n    return result', max_new_tokens = 128, return_type = 1)
+```
+
+Under the hood, the tokenizer [automatically splits by `<FILL_ME>`](https://huggingface.co/docs/transformers/main/model_doc/code_llama#transformers.CodeLlamaTokenizer.fill_token) to create a formatted input string that follows [the original training pattern](https://github.com/facebookresearch/codellama/blob/cb51c14ec761370ba2e2bc351374a79265d0465e/llama/generation.py#L402). This is more robust than preparing the pattern yourself: it avoids pitfalls, such as token glueing, that are very hard to debug.  To see how much CPU and GPU memory you need for this model or others, try [this calculator](https://huggingface.co/spaces/hf-accelerate/model-memory-usage) which can help determine that value.
+
+- The LLaMA tokenizer is a BPE model based on [sentencepiece](https://github.com/google/sentencepiece). One quirk of sentencepiece is that when decoding a sequence, if the first token is the start of the word (e.g. "Banana"), the tokenizer does not prepend the prefix space to the string.
+
+This model was contributed by [ArthurZucker](https://huggingface.co/ArthurZ). The original code of the authors can be found [here](https://github.com/facebookresearch/llama).
+
+
+## CodeLlamaTokenizer
+
+[[autodoc]] CodeLlamaTokenizer
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - save_vocabulary
+
+## CodeLlamaTokenizerFast
+
+[[autodoc]] CodeLlamaTokenizerFast
+    - build_inputs_with_special_tokens
+    - get_special_tokens_mask
+    - create_token_type_ids_from_sequences
+    - update_post_processor
+    - save_vocabulary

docs/source/en/model_doc/deberta-v2.md

@@ -152,3 +152,8 @@ contributed by [kamalkraj](https://huggingface.co/kamalkraj). The original code
 
 [[autodoc]] TFDebertaV2ForQuestionAnswering
     - call
+
+## TFDebertaV2ForMultipleChoice
+
+[[autodoc]] TFDebertaV2ForMultipleChoice
+    - call

docs/source/en/model_doc/donut.md

@@ -80,11 +80,9 @@ into a single instance to both extract the input features and decode the predict
 ...     pixel_values.to(device),
 ...     decoder_input_ids=decoder_input_ids.to(device),
 ...     max_length=model.decoder.config.max_position_embeddings,
-...     early_stopping=True,
 ...     pad_token_id=processor.tokenizer.pad_token_id,
 ...     eos_token_id=processor.tokenizer.eos_token_id,
 ...     use_cache=True,
-...     num_beams=1,
 ...     bad_words_ids=[[processor.tokenizer.unk_token_id]],
 ...     return_dict_in_generate=True,
 ... )
@@ -125,11 +123,9 @@ into a single instance to both extract the input features and decode the predict
 ...     pixel_values.to(device),
 ...     decoder_input_ids=decoder_input_ids.to(device),
 ...     max_length=model.decoder.config.max_position_embeddings,
-...     early_stopping=True,
 ...     pad_token_id=processor.tokenizer.pad_token_id,
 ...     eos_token_id=processor.tokenizer.eos_token_id,
 ...     use_cache=True,
-...     num_beams=1,
 ...     bad_words_ids=[[processor.tokenizer.unk_token_id]],
 ...     return_dict_in_generate=True,
 ... )
@@ -172,11 +168,9 @@ into a single instance to both extract the input features and decode the predict
 ...     pixel_values.to(device),
 ...     decoder_input_ids=decoder_input_ids.to(device),
 ...     max_length=model.decoder.config.max_position_embeddings,
-...     early_stopping=True,
 ...     pad_token_id=processor.tokenizer.pad_token_id,
 ...     eos_token_id=processor.tokenizer.eos_token_id,
 ...     use_cache=True,
-...     num_beams=1,
 ...     bad_words_ids=[[processor.tokenizer.unk_token_id]],
 ...     return_dict_in_generate=True,
 ... )

docs/source/en/model_doc/falcon.md

@@ -0,0 +1,84 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Falcon
+
+## Overview
+
+Falcon is a class of causal decoder-only models built by [TII](https://www.tii.ae/). The largest Falcon checkpoints
+have been trained on >=1T tokens of text, with a particular emphasis on the [RefinedWeb](https://arxiv.org/abs/2306.01116)
+corpus. They are made available under the Apache 2.0 license.
+
+
+Falcon's architecture is modern and optimized for inference, with multi-query attention and support for efficient
+attention variants like `FlashAttention`. Both 'base' models trained only as causal language models as well as
+'instruct' models that have received further fine-tuning are available.
+
+
+Falcon models are (as of 2023) some of the largest and most powerful open-source language models,
+and consistently rank highly in the [OpenLLM leaderboard](https://huggingface.co/spaces/HuggingFaceH4/open_llm_leaderboard).
+
+## Converting custom checkpoints 
+
+<Tip>
+
+Falcon models were initially added to the Hugging Face Hub as custom code checkpoints. However, Falcon is now fully
+supported in the Transformers library. If you fine-tuned a model from a custom code checkpoint, we recommend converting
+your checkpoint to the new in-library format, as this should give significant improvements to stability and
+performance, especially for generation, as well as removing the need to use `trust_remote_code=True`!
+
+</Tip>
+
+You can convert custom code checkpoints to full Transformers checkpoints using the `convert_custom_code_checkpoint.py` 
+script located in the
+[Falcon model directory](https://github.com/huggingface/transformers/tree/main/src/transformers/models/falcon)
+of the Transformers library. To use this script, simply call it with 
+`python convert_custom_code_checkpoint.py --checkpoint_dir my_model`. This will convert your checkpoint in-place, and
+you can immediately load it from the directory afterwards with e.g. `from_pretrained()`. If your model hasn't been
+uploaded to the Hub, we recommend making a backup before attempting the conversion, just in case!
+
+
+## FalconConfig
+
+[[autodoc]] FalconConfig
+    - all
+
+## FalconModel
+
+[[autodoc]] FalconModel
+    - forward
+
+## FalconForCausalLM
+
+[[autodoc]] FalconForCausalLM
+    - forward
+
+## FalconForSequenceClassification
+
+[[autodoc]] FalconForSequenceClassification
+    - forward
+
+## FalconForTokenClassification
+
+[[autodoc]] FalconForTokenClassification
+    - forward
+
+## FalconForQuestionAnswering
+
+[[autodoc]] FalconForQuestionAnswering
+    - forward
+
+

docs/source/en/model_doc/idefics.md

@@ -0,0 +1,63 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# IDEFICS
+
+## Overview
+
+The IDEFICS model was proposed in [OBELICS: An Open Web-Scale Filtered Dataset of Interleaved Image-Text Documents
+](https://huggingface.co/papers/2306.16527
+) by Hugo Laurençon, Lucile Saulnier, Léo Tronchon, Stas Bekman, Amanpreet Singh, Anton Lozhkov, Thomas Wang, Siddharth Karamcheti, Alexander M. Rush, Douwe Kiela, Matthieu Cord, Victor Sanh
+
+The abstract from the paper is the following:
+
+*Large multimodal models trained on natural documents, which interleave images and text, outperform models trained on image-text pairs on various multimodal benchmarks that require reasoning over one or multiple images to generate a text. However, the datasets used to train these models have not been released, and the collection process has not been fully specified. We introduce the OBELICS dataset, an open web-scale filtered dataset of interleaved image-text documents comprising 141 million web pages extracted from Common Crawl, 353 million associated images, and 115 billion text tokens. We describe the dataset creation process, present comprehensive filtering rules, and provide an analysis of the dataset's content. To show the viability of OBELISC, we train an 80 billion parameters vision and language model on the dataset and obtain competitive performance on various multimodal benchmarks. We release the code to reproduce the dataset along with the dataset itself.*
+
+This model was contributed by [HuggingFaceM4](https://huggingface.co/HuggingFaceM4). The original code can be found [here](<INSERT LINK TO GITHUB REPO HERE>). (TODO: don't have a public link yet).
+
+
+<Tip warning={true}>
+
+Idefics modeling code in Transformers is for finetuning and inferencing the pre-trained Idefics models.
+
+To train a new Idefics model from scratch use the m4 codebase (a link will be provided once it's made public)
+
+</Tip>
+
+
+## IdeficsConfig
+
+[[autodoc]] IdeficsConfig
+
+## IdeficsModel
+
+[[autodoc]] IdeficsModel
+    - forward
+
+## IdeficsForVisionText2Text
+
+[[autodoc]] IdeficsForVisionText2Text
+    - forward
+
+## IdeficsImageProcessor
+
+[[autodoc]] IdeficsImageProcessor
+    - preprocess
+
+## IdeficsProcessor
+
+[[autodoc]] IdeficsProcessor
+    - __call__

docs/source/en/model_doc/llama.md

@@ -55,6 +55,28 @@ Based on the original LLaMA model, Meta AI has released some follow-up works:
 
 - **Llama2**: Llama2 is an improved version of Llama with some architectural tweaks (Grouped Query Attention), and is pre-trained on 2Trillion tokens. Refer to the documentation of Llama2 which can be found [here](llama2).
 
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with LLaMA. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+<PipelineTag pipeline="text-classification"/>
+
+- A [notebook](https://colab.research.google.com/github/bigscience-workshop/petals/blob/main/examples/prompt-tuning-sst2.ipynb#scrollTo=f04ba4d2) on how to use prompt tuning to adapt the LLaMA model for text classification task. 🌎
+
+<PipelineTag pipeline="question-answering"/>
+
+- [StackLLaMA: A hands-on guide to train LLaMA with RLHF](https://huggingface.co/blog/stackllama#stackllama-a-hands-on-guide-to-train-llama-with-rlhf), a blog post about how to train LLaMA to answer questions on [Stack Exchange](https://stackexchange.com/) with RLHF.
+
+⚗️ Optimization
+- A [notebook](https://colab.research.google.com/drive/1SQUXq1AMZPSLD4mk3A3swUIc6Y2dclme?usp=sharing) on how to fine-tune LLaMA model using xturing library on GPU which has limited memory. 🌎 
+
+⚡️ Inference
+- A [notebook](https://colab.research.google.com/github/DominguesM/alpaca-lora-ptbr-7b/blob/main/notebooks/02%20-%20Evaluate.ipynb) on how to run the LLaMA Model using PeftModel from the 🤗 PEFT library. 🌎 
+- A [notebook](https://colab.research.google.com/drive/1l2GiSSPbajVyp2Nk3CFT4t3uH6-5TiBe?usp=sharing) on how to load a PEFT adapter LLaMA model with LangChain. 🌎
+
+🚀 Deploy
+- A [notebook](https://colab.research.google.com/github/lxe/simple-llama-finetuner/blob/master/Simple_LLaMA_FineTuner.ipynb#scrollTo=3PM_DilAZD8T) on how to fine-tune LLaMA model using LoRA method via the 🤗 PEFT library with intuitive UI. 🌎 
+- A [notebook](https://github.com/aws/amazon-sagemaker-examples/blob/main/introduction_to_amazon_algorithms/jumpstart-foundation-models/text-generation-open-llama.ipynb) on how to deploy Open-LLaMA model for text generation on Amazon SageMaker. 🌎 
 
 ## LlamaConfig
 

docs/source/en/model_doc/llama2.md

@@ -26,10 +26,21 @@ The abstract from the paper is the following:
 
 Checkout all Llama2 models [here](https://huggingface.co/models?search=llama2)
 
+<Tip warning={true}>
+
+The `Llama2` models were trained using `bfloat16`, but the original inference uses `float16. The checkpoints uploaded on the hub use `torch_dtype = 'float16'` which will be
+used by the `AutoModel` API to cast the checkpoints from `torch.float32` to `torch.float16`. 
+
+The `dtype` of the online weights is mostly irrelevant, unless you are using `torch_dtype="auto"` when initializing a model using `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`. The reason is that the model will first be downloaded ( using the `dtype` of the checkpoints online) then it will be casted to the default `dtype` of `torch` (becomes `torch.float32`) and finally, if there is a `torch_dtype` provided in the config, it will be used. 
+
+Training the model in `float16` is not recommended and known to produce `nan`, as such the model should be trained in `bfloat16`.
+
+</Tip>
+
 Tips:
 
 - Weights for the Llama2 models can be obtained by filling out [this form](https://ai.meta.com/resources/models-and-libraries/llama-downloads/)
-- The architecture is very similar to the first Llama, with the addition of Groupe Query Attention (GQA) following this [paper](https://arxiv.org/pdf/2305.13245.pdf)
+- The architecture is very similar to the first Llama, with the addition of Grouped Query Attention (GQA) following this [paper](https://arxiv.org/pdf/2305.13245.pdf)
 - Setting `config.pretraining_tp` to a value different than 1 will activate the more accurate but slower computation of the linear layers, which should better match the original logits.
 - The original model uses `pad_id = -1` which means that there is no padding token. We can't have the same logic, make sure to add a padding token using `tokenizer.add_special_tokens({"pad_token":"<pad>"})` and resize the token embedding accordingly. You should also set the `model.config.pad_token_id`. The `embed_tokens` layer of the model is initialized with `self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.config.padding_idx)`, which makes sure that encoding the padding token will output zeros, so passing it when initializing is recommended.
 - After filling out the form and gaining access to the model checkpoints, you should be able to use the already converted checkpoints. Otherwise, if you are converting your own model, feel free to use the [conversion script](https://github.com/huggingface/transformers/blob/main/src/transformers/models/llama/convert_llama_weights_to_hf.py). The script can be called with the following (example) command:
@@ -55,6 +66,35 @@ come in several checkpoints they each contain a part of each weight of the model
 
 This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ) with contributions from [Lysandre Debut](https://huggingface.co/lysandre). The code of the implementation in Hugging Face is based on GPT-NeoX [here](https://github.com/EleutherAI/gpt-neox). The original code of the authors can be found [here](https://github.com/facebookresearch/llama).
 
+## Resources
+
+A list of official Hugging Face and community (indicated by 🌎) resources to help you get started with LLaMA2. If you're interested in submitting a resource to be included here, please feel free to open a Pull Request and we'll review it! The resource should ideally demonstrate something new instead of duplicating an existing resource.
+
+- [Llama 2 is here - get it on Hugging Face](https://huggingface.co/blog/llama2), a blog post about Llama 2 and how to use it with 🤗 Transformers and 🤗 PEFT.
+- [LLaMA 2 - Every Resource you need](https://www.philschmid.de/llama-2), a compilation of relevant resources to learn about LLaMA 2 and how to get started quickly.
+
+<PipelineTag pipeline="text-generation"/>
+
+- A [notebook](https://colab.research.google.com/drive/1PEQyJO1-f6j0S_XJ8DV50NkpzasXkrzd?usp=sharing) on how to fine-tune Llama 2 in Google Colab using QLoRA and 4-bit precision. 🌎
+- A [notebook](https://colab.research.google.com/drive/134o_cXcMe_lsvl15ZE_4Y75Kstepsntu?usp=sharing) on how to fine-tune the "Llama-v2-7b-guanaco" model with 4-bit QLoRA and generate Q&A datasets from PDFs. 🌎
+
+<PipelineTag pipeline="text-classification"/>
+
+- A [notebook](https://colab.research.google.com/drive/1ggaa2oRFphdBmqIjSEbnb_HGkcIRC2ZB?usp=sharing) on how to fine-tune the Llama 2 model with QLoRa, TRL, and Korean text classification dataset. 🌎🇰🇷
+
+⚗️ Optimization
+- [Fine-tune Llama 2 with DPO](https://huggingface.co/blog/dpo-trl), a guide to using the TRL library's DPO method to fine tune Llama 2 on a specific dataset.
+- [Extended Guide: Instruction-tune Llama 2](https://www.philschmid.de/instruction-tune-llama-2), a guide to training Llama 2 to generate instructions from inputs, transforming the model from instruction-following to instruction-giving.
+- A [notebook](https://colab.research.google.com/drive/1SYpgFpcmtIUzdE7pxqknrM4ArCASfkFQ?usp=sharing) on how to fine-tune the Llama 2 model on a personal computer using QLoRa and TRL. 🌎
+
+⚡️ Inference
+- A [notebook](https://colab.research.google.com/drive/1TC56ArKerXUpbgRy5vM3woRsbTEVNq7h?usp=sharing) on how to quantize the Llama 2 model using GPTQ from the AutoGPTQ library. 🌎
+- A [notebook](https://colab.research.google.com/drive/1X1z9Q6domMKl2CnEM0QGHNwidLfR4dW2?usp=sharing) on how to run the Llama 2 Chat Model with 4-bit quantization on a local computer or Google Colab. 🌎
+
+🚀 Deploy
+- [Fine-tune LLaMA 2 (7-70B) on Amazon SageMaker](https://www.philschmid.de/sagemaker-llama2-qlora), a complete guide from setup to QLoRA fine-tuning and deployment on Amazon SageMaker.
+- [Deploy Llama 2 7B/13B/70B on Amazon SageMaker](https://www.philschmid.de/sagemaker-llama-llm), a guide on using Hugging Face's LLM DLC container for secure and scalable deployment.
+
 
 ## LlamaConfig
 

docs/source/en/model_doc/mms.md

@@ -165,7 +165,147 @@ To further improve performance from ASR models, language model decoding can be u
 
 ### Speech Synthesis (TTS)
 
-Individual TTS models are available for each of the 1100+ languages. The models and inference documentation can be found [here](https://huggingface.co/facebook/mms-tts).
+MMS-TTS uses the same model architecture as VITS, which was added to 🤗 Transformers in v4.33. MMS trains a separate 
+model checkpoint for each of the 1100+ languages in the project. All available checkpoints can be found on the Hugging 
+Face Hub: [facebook/mms-tts](https://huggingface.co/models?sort=trending&search=facebook%2Fmms-tts), and the inference 
+documentation under [VITS](https://huggingface.co/docs/transformers/main/en/model_doc/vits).
+
+#### Inference
+
+To use the MMS model, first update to the latest version of the Transformers library:
+
+```bash
+pip install --upgrade transformers accelerate
+```
+
+Since the flow-based model in VITS is non-deterministic, it is good practice to set a seed to ensure reproducibility of 
+the outputs. 
+
+- For languages with a Roman alphabet, such as English or French, the tokenizer can be used directly to 
+pre-process the text inputs. The following code example runs a forward pass using the MMS-TTS English checkpoint:
+
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+model = VitsModel.from_pretrained("facebook/mms-tts-eng")
+
+inputs = tokenizer(text="Hello - my dog is cute", return_tensors="pt")
+
+set_seed(555)  # make deterministic
+
+with torch.no_grad():
+   outputs = model(**inputs)
+
+waveform = outputs.waveform[0]
+```
+
+The resulting waveform can be saved as a `.wav` file:
+
+```python
+import scipy
+
+scipy.io.wavfile.write("synthesized_speech.wav", rate=model.config.sampling_rate, data=waveform)
+```
+
+Or displayed in a Jupyter Notebook / Google Colab:
+
+```python
+from IPython.display import Audio
+
+Audio(waveform, rate=model.config.sampling_rate)
+```
+
+For certain languages with non-Roman alphabets, such as Arabic, Mandarin or Hindi, the [`uroman`](https://github.com/isi-nlp/uroman) 
+perl package is required to pre-process the text inputs to the Roman alphabet.
+
+You can check whether you require the `uroman` package for your language by inspecting the `is_uroman` attribute of 
+the pre-trained `tokenizer`:
+
+```python
+from transformers import VitsTokenizer
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+print(tokenizer.is_uroman)
+```
+
+If required, you should apply the uroman package to your text inputs **prior** to passing them to the `VitsTokenizer`, 
+since currently the tokenizer does not support performing the pre-processing itself.
+
+To do this, first clone the uroman repository to your local machine and set the bash variable `UROMAN` to the local path:
+
+```bash
+git clone https://github.com/isi-nlp/uroman.git
+cd uroman
+export UROMAN=$(pwd)
+```
+
+You can then pre-process the text input using the following code snippet. You can either rely on using the bash variable 
+`UROMAN` to point to the uroman repository, or you can pass the uroman directory as an argument to the `uromaize` function:
+
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+import os
+import subprocess
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
+model = VitsModel.from_pretrained("facebook/mms-tts-kor")
+
+def uromanize(input_string, uroman_path):
+    """Convert non-Roman strings to Roman using the `uroman` perl package."""
+    script_path = os.path.join(uroman_path, "bin", "uroman.pl")
+
+    command = ["perl", script_path]
+
+    process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    # Execute the perl command
+    stdout, stderr = process.communicate(input=input_string.encode())
+
+    if process.returncode != 0:
+        raise ValueError(f"Error {process.returncode}: {stderr.decode()}")
+
+    # Return the output as a string and skip the new-line character at the end
+    return stdout.decode()[:-1]
+
+text = "이봐 무슨 일이야"
+uromaized_text = uromanize(text, uroman_path=os.environ["UROMAN"])
+
+inputs = tokenizer(text=uromaized_text, return_tensors="pt")
+
+set_seed(555)  # make deterministic
+with torch.no_grad():
+   outputs = model(inputs["input_ids"])
+
+waveform = outputs.waveform[0]
+```
+
+**Tips:**
+
+* The MMS-TTS checkpoints are trained on lower-cased, un-punctuated text. By default, the `VitsTokenizer` *normalizes* the inputs by removing any casing and punctuation, to avoid passing out-of-vocabulary characters to the model. Hence, the model is agnostic to casing and punctuation, so these should be avoided in the text prompt. You can disable normalisation by setting `noramlize=False` in the call to the tokenizer, but this will lead to un-expected behaviour and is discouraged.
+* The speaking rate can be varied by setting the attribute `model.speaking_rate` to a chosen value. Likewise, the randomness of the noise is controlled by `model.noise_scale`:
+
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+model = VitsModel.from_pretrained("facebook/mms-tts-eng")
+
+inputs = tokenizer(text="Hello - my dog is cute", return_tensors="pt")
+
+# make deterministic
+set_seed(555)  
+
+# make speech faster and more noisy
+model.speaking_rate = 1.5
+model.noise_scale = 0.8
+
+with torch.no_grad():
+   outputs = model(**inputs)
+```
+
 
 ### Language Identification (LID)
 
@@ -173,11 +313,12 @@ Different LID models are available based on the number of languages they can rec
 
 #### Inference
 First, we install transformers and some other libraries
-```
-pip install torch accelerate torchaudio datasets
+
+```bash
+pip install torch accelerate datasets[audio]
 pip install --upgrade transformers
 ````
-pip install torch datasets[audio]
+
 Next, we load a couple of audio samples via `datasets`. Make sure that the audio data is sampled to 16000 kHz.
 
 ```py

docs/source/en/model_doc/musicgen.md

@@ -53,6 +53,10 @@ better results than greedy, thus we encourage sampling mode to be used where pos
 and can be explicitly specified by setting `do_sample=True` in the call to [`MusicgenForConditionalGeneration.generate`],
 or by overriding the model's generation config (see below).
 
+Generation is limited by the sinusoidal positional embeddings to 30 second inputs. Meaning, MusicGen cannot generate more
+than 30 seconds of audio (1503 tokens), and input audio passed by Audio-Prompted Generation contributes to this limit so,
+given an input of 20 seconds of audio, MusicGen cannot generate more than 10 seconds of additional audio.
+
 ### Unconditional Generation
 
 The inputs for unconditional (or 'null') generation can be obtained through the method
@@ -210,28 +214,7 @@ The MusicGen model can be de-composed into three distinct stages:
 
 Thus, the MusicGen model can either be used as a standalone decoder model, corresponding to the class [`MusicgenForCausalLM`],
 or as a composite model that includes the text encoder and audio encoder/decoder, corresponding to the class
-[`MusicgenForConditionalGeneration`].
-
-Since the text encoder and audio encoder/decoder models are frozen during training, the MusicGen decoder [`MusicgenForCausalLM`]
-can be trained standalone on a dataset of encoder hidden-states and audio codes. For inference, the trained decoder can
-be combined with the frozen text encoder and audio encoder/decoders to recover the composite [`MusicgenForConditionalGeneration`]
-model.
-
-Below, we demonstrate how to construct the composite [`MusicgenForConditionalGeneration`] model from its three constituent
-parts, as would typically be done following training of the MusicGen decoder LM:
-
-```python
->>> from transformers import AutoConfig, AutoModelForTextEncoding, AutoModel, MusicgenForCausalLM, MusicgenForConditionalGeneration
-
->>> text_encoder = AutoModelForTextEncoding.from_pretrained("t5-base")
->>> audio_encoder = AutoModel.from_pretrained("facebook/encodec_32khz")
->>> decoder_config = AutoConfig.from_pretrained("facebook/musicgen-small").decoder
->>> decoder = MusicgenForCausalLM.from_pretrained("facebook/musicgen-small", **decoder_config)
-
->>> model = MusicgenForConditionalGeneration.from_sub_models_pretrained(text_encoder, audio_encoder, decoder)
-```
-
-If only the decoder needs to be loaded from the pre-trained checkpoint for the composite model, it can be loaded by first 
+[`MusicgenForConditionalGeneration`]. If only the decoder needs to be loaded from the pre-trained checkpoint, it can be loaded by first 
 specifying the correct config, or be accessed through the `.decoder` attribute of the composite model:
 
 ```python
@@ -245,6 +228,11 @@ specifying the correct config, or be accessed through the `.decoder` attribute o
 >>> decoder = MusicgenForConditionalGeneration.from_pretrained("facebook/musicgen-small").decoder
 ```
 
+Since the text encoder and audio encoder/decoder models are frozen during training, the MusicGen decoder [`MusicgenForCausalLM`]
+can be trained standalone on a dataset of encoder hidden-states and audio codes. For inference, the trained decoder can
+be combined with the frozen text encoder and audio encoder/decoders to recover the composite [`MusicgenForConditionalGeneration`]
+model.
+
 Tips:
 * MusicGen is trained on the 32kHz checkpoint of Encodec. You should ensure you use a compatible version of the Encodec model.
 * Sampling mode tends to deliver better results than greedy - you can toggle sampling with the variable `do_sample` in the call to [`MusicgenForConditionalGeneration.generate`]

docs/source/en/model_doc/persimmon.md

@@ -0,0 +1,96 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Persimmon
+
+## Overview
+
+The Persimmon model was created by [ADEPT](https://www.adept.ai/blog/persimmon-8b), and authored by Erich Elsen, Augustus Odena, Maxwell Nye, Sağnak Taşırlar, Tri Dao, Curtis Hawthorne, Deepak Moparthi, Arushi Somani.
+
+The authors introduced Persimmon-8B, a decoder model based on the classic transformers architecture, with query and key normalization. Persimmon-8B is a fully permissively-licensed model with approximately 8 billion parameters, released under the Apache license.  Some of the key attributes of Persimmon-8B are long context size (16K), performance, and capabilities for multimodal extensions.
+
+The authors showcase their approach to model evaluation, focusing on practical text generation, mirroring how users interact with language models. The work also includes a comparative analysis, pitting Persimmon-8B against other prominent models (MPT 7B Instruct and Llama 2 Base 7B 1-Shot), across various evaluation tasks. The results demonstrate Persimmon-8B's competitive performance, even with limited training data.
+
+In terms of model details, the work outlines the architecture and training methodology of Persimmon-8B, providing insights into its design choices, sequence length, and dataset composition. The authors present a fast inference code that outperforms traditional implementations through operator fusion and CUDA graph utilization while maintaining code coherence. They express their anticipation of how the community will leverage this contribution to drive innovation, hinting at further upcoming releases as part of an ongoing series of developments.
+
+
+<Tip warning={true}>
+
+The `Persimmon` models were trained using `bfloat16`, but the original inference uses `float16` The checkpoints uploaded on the hub use `torch_dtype = 'float16'` which will be
+used by the `AutoModel` API to cast the checkpoints from `torch.float32` to `torch.float16`. 
+
+The `dtype` of the online weights is mostly irrelevant, unless you are using `torch_dtype="auto"` when initializing a model using `model = AutoModelForCausalLM.from_pretrained("path", torch_dtype = "auto")`. The reason is that the model will first be downloaded ( using the `dtype` of the checkpoints online) then it will be cast to the default `dtype` of `torch` (becomes `torch.float32`). Users should specify the `torch_dtype` they want, and if they don't it will be `torch.float32`.
+
+Finetuning the model in `float16` is not recommended and known to produce `nan`, as such the model should be fine-tuned in `bfloat16`.
+
+</Tip>
+
+
+Tips:
+
+- To convert the model, you need to clone the original repository using `git clone https://github.com/persimmon-ai-labs/adept-inference`, then get the checkpoints:
+
+```bash
+git clone https://github.com/persimmon-ai-labs/adept-inference
+wget https://axtkn4xl5cip.objectstorage.us-phoenix-1.oci.customer-oci.com/n/axtkn4xl5cip/b/adept-public-data/o/8b_base_model_release.tar
+tar -xvf 8b_base_model_release.tar
+python src/transformers/models/persimmon/convert_persimmon_weights_to_hf.py  --input_dir /path/to/downloaded/persimmon/weights/ --output_dir /output/path \
+    --pt_model_path /path/to/8b_chat_model_release/iter_0001251/mp_rank_00/model_optim_rng.pt
+    --ada_lib_path /path/to/adept-inference
+```
+
+For the chat model:
+```bash
+wget https://axtkn4xl5cip.objectstorage.us-phoenix-1.oci.customer-oci.com/n/axtkn4xl5cip/b/adept-public-data/o/8b_chat_model_release.tar
+tar -xvf 8b_base_model_release.tar
+```
+
+Thereafter, models can be loaded via:
+
+```py
+from transformers import PersimmonForCausalLM, PersimmonTokenizer
+
+model = PersimmonForCausalLM.from_pretrained("/output/path")
+tokenizer = PersimmonTokenizer.from_pretrained("/output/path")
+```
+
+This model was contributed by [ArthurZ](https://huggingface.co/ArthurZ).
+The original code can be found [here](https://github.com/persimmon-ai-labs/adept-inference).
+
+- Perismmon uses a `sentencepiece` based tokenizer, with a `Unigram` model. It supports bytefallback, which is only available in `tokenizers==0.14.0` for the fast tokenizer.
+The `LlamaTokenizer` is used as it is a standard wrapper around sentencepiece. The `chat` template will be updated with the templating functions in a follow up PR!
+
+- The authors suggest to use the following prompt format for the chat mode: `f"human: {prompt}\n\nadept:"`
+
+
+## PersimmonConfig
+
+[[autodoc]] PersimmonConfig
+
+## PersimmonModel
+
+[[autodoc]] PersimmonModel
+    - forward
+
+## PersimmonForCausalLM
+
+[[autodoc]] PersimmonForCausalLM
+    - forward
+
+## PersimmonForSequenceClassification
+
+[[autodoc]] PersimmonForSequenceClassification
+    - forward

docs/source/en/model_doc/pop2piano.md

@@ -0,0 +1,196 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Pop2Piano
+
+<div class="flex flex-wrap space-x-1">
+<a href="https://huggingface.co/spaces/sweetcocoa/pop2piano">
+<img alt="Spaces" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue">
+</a>
+</div>
+
+## Overview
+
+The Pop2Piano model was proposed in [Pop2Piano : Pop Audio-based Piano Cover Generation](https://arxiv.org/abs/2211.00895) by Jongho Choi and Kyogu Lee.
+
+Piano covers of pop music are widely enjoyed, but generating them from music is not a trivial task. It requires great 
+expertise with playing piano as well as knowing different characteristics and melodies of a song. With Pop2Piano you 
+can directly generate a cover from a song's audio waveform. It is the first model to directly generate a piano cover 
+from pop audio without melody and chord extraction modules. 
+
+Pop2Piano is an encoder-decoder Transformer model based on [T5](https://arxiv.org/pdf/1910.10683.pdf). The input audio 
+is transformed to its waveform and passed to the encoder, which transforms it to a latent representation. The decoder 
+uses these latent representations to generate token ids in an autoregressive way. Each token id corresponds to one of four 
+different token types: time, velocity, note and 'special'. The token ids are then decoded to their equivalent MIDI file.
+
+
+The abstract from the paper is the following:
+
+*Piano covers of pop music are enjoyed by many people. However, the
+task of automatically generating piano covers of pop music is still
+understudied. This is partly due to the lack of synchronized
+{Pop, Piano Cover} data pairs, which made it challenging to apply
+the latest data-intensive deep learning-based methods. To leverage
+the power of the data-driven approach, we make a large amount of
+paired and synchronized {Pop, Piano Cover} data using an automated
+pipeline. In this paper, we present Pop2Piano, a Transformer network
+that generates piano covers given waveforms of pop music. To the best
+of our knowledge, this is the first model to generate a piano cover
+directly from pop audio without using melody and chord extraction
+modules. We show that Pop2Piano, trained with our dataset, is capable
+of producing plausible piano covers.*
+
+
+Tips:
+
+1. To use Pop2Piano, you will need to install the 🤗 Transformers library, as well as the following third party modules:  
+```
+pip install pretty-midi==0.2.9 essentia==2.1b6.dev1034 librosa scipy
+```
+Please note that you may need to restart your runtime after installation.
+2. Pop2Piano is an Encoder-Decoder based model like T5.
+3. Pop2Piano can be used to generate midi-audio files for a given audio sequence.
+4. Choosing different composers in `Pop2PianoForConditionalGeneration.generate()` can lead to variety of different results.
+5. Setting the sampling rate to 44.1 kHz when loading the audio file can give good performance.
+6. Though Pop2Piano was mainly trained on Korean Pop music, it also does pretty well on other Western Pop or Hip Hop songs.
+
+This model was contributed by [Susnato Dhar](https://huggingface.co/susnato).
+The original code can be found [here](https://github.com/sweetcocoa/pop2piano).
+
+## Examples
+
+- Example using HuggingFace Dataset:
+
+```python
+>>> from datasets import load_dataset
+>>> from transformers import Pop2PianoForConditionalGeneration, Pop2PianoProcessor
+
+>>> model = Pop2PianoForConditionalGeneration.from_pretrained("sweetcocoa/pop2piano")
+>>> processor = Pop2PianoProcessor.from_pretrained("sweetcocoa/pop2piano")
+>>> ds = load_dataset("sweetcocoa/pop2piano_ci", split="test")
+
+>>> inputs = processor(
+...     audio=ds["audio"][0]["array"], sampling_rate=ds["audio"][0]["sampling_rate"], return_tensors="pt"
+... )
+>>> model_output = model.generate(input_features=inputs["input_features"], composer="composer1")
+>>> tokenizer_output = processor.batch_decode(
+...     token_ids=model_output, feature_extractor_output=inputs
+... )["pretty_midi_objects"][0]
+>>> tokenizer_output.write("./Outputs/midi_output.mid")
+```
+
+- Example using your own audio file:
+
+```python
+>>> import librosa
+>>> from transformers import Pop2PianoForConditionalGeneration, Pop2PianoProcessor
+
+>>> audio, sr = librosa.load("<your_audio_file_here>", sr=44100)  # feel free to change the sr to a suitable value.
+>>> model = Pop2PianoForConditionalGeneration.from_pretrained("sweetcocoa/pop2piano")
+>>> processor = Pop2PianoProcessor.from_pretrained("sweetcocoa/pop2piano")
+
+>>> inputs = processor(audio=audio, sampling_rate=sr, return_tensors="pt")
+>>> model_output = model.generate(input_features=inputs["input_features"], composer="composer1")
+>>> tokenizer_output = processor.batch_decode(
+...     token_ids=model_output, feature_extractor_output=inputs
+... )["pretty_midi_objects"][0]
+>>> tokenizer_output.write("./Outputs/midi_output.mid")
+```
+
+- Example of processing multiple audio files in batch:
+
+```python
+>>> import librosa
+>>> from transformers import Pop2PianoForConditionalGeneration, Pop2PianoProcessor
+
+>>> # feel free to change the sr to a suitable value.
+>>> audio1, sr1 = librosa.load("<your_first_audio_file_here>", sr=44100)  
+>>> audio2, sr2 = librosa.load("<your_second_audio_file_here>", sr=44100)
+>>> model = Pop2PianoForConditionalGeneration.from_pretrained("sweetcocoa/pop2piano")
+>>> processor = Pop2PianoProcessor.from_pretrained("sweetcocoa/pop2piano")
+
+>>> inputs = processor(audio=[audio1, audio2], sampling_rate=[sr1, sr2], return_attention_mask=True, return_tensors="pt")
+>>> # Since we now generating in batch(2 audios) we must pass the attention_mask
+>>> model_output = model.generate(
+...     input_features=inputs["input_features"],
+...     attention_mask=inputs["attention_mask"],
+...     composer="composer1",
+... )
+>>> tokenizer_output = processor.batch_decode(
+...     token_ids=model_output, feature_extractor_output=inputs
+... )["pretty_midi_objects"]
+
+>>> # Since we now have 2 generated MIDI files
+>>> tokenizer_output[0].write("./Outputs/midi_output1.mid")
+>>> tokenizer_output[1].write("./Outputs/midi_output2.mid")
+```
+
+
+- Example of processing multiple audio files in batch (Using `Pop2PianoFeatureExtractor` and `Pop2PianoTokenizer`):
+
+```python
+>>> import librosa
+>>> from transformers import Pop2PianoForConditionalGeneration, Pop2PianoFeatureExtractor, Pop2PianoTokenizer
+
+>>> # feel free to change the sr to a suitable value.
+>>> audio1, sr1 = librosa.load("<your_first_audio_file_here>", sr=44100)  
+>>> audio2, sr2 = librosa.load("<your_second_audio_file_here>", sr=44100)
+>>> model = Pop2PianoForConditionalGeneration.from_pretrained("sweetcocoa/pop2piano")
+>>> feature_extractor = Pop2PianoFeatureExtractor.from_pretrained("sweetcocoa/pop2piano")
+>>> tokenizer = Pop2PianoTokenizer.from_pretrained("sweetcocoa/pop2piano")
+
+>>> inputs = feature_extractor(
+...     audio=[audio1, audio2], 
+...     sampling_rate=[sr1, sr2], 
+...     return_attention_mask=True, 
+...     return_tensors="pt",
+... )
+>>> # Since we now generating in batch(2 audios) we must pass the attention_mask
+>>> model_output = model.generate(
+...     input_features=inputs["input_features"],
+...     attention_mask=inputs["attention_mask"],
+...     composer="composer1",
+... )
+>>> tokenizer_output = tokenizer.batch_decode(
+...     token_ids=model_output, feature_extractor_output=inputs
+... )["pretty_midi_objects"]
+
+>>> # Since we now have 2 generated MIDI files
+>>> tokenizer_output[0].write("./Outputs/midi_output1.mid")
+>>> tokenizer_output[1].write("./Outputs/midi_output2.mid")
+```
+
+
+## Pop2PianoConfig
+
+[[autodoc]] Pop2PianoConfig
+
+## Pop2PianoFeatureExtractor
+
+[[autodoc]] Pop2PianoFeatureExtractor
+    - __call__
+
+## Pop2PianoForConditionalGeneration
+
+[[autodoc]] Pop2PianoForConditionalGeneration
+    - forward
+    - generate
+
+## Pop2PianoTokenizer
+
+[[autodoc]] Pop2PianoTokenizer
+    - __call__
+
+## Pop2PianoProcessor
+
+[[autodoc]] Pop2PianoProcessor
+    - __call__

docs/source/en/model_doc/speech-encoder-decoder.md

@@ -111,7 +111,7 @@ speech inputs) and `labels` (which are the `input_ids` of the encoded target seq
 >>> labels = tokenizer(ds[0]["text"], return_tensors="pt").input_ids
 
 >>> # the forward function automatically creates the correct decoder_input_ids
->>> loss = model(**input_features).loss
+>>> loss = model(input_values=input_values, labels=labels).loss
 >>> loss.backward()
 ```
 
@@ -129,4 +129,4 @@ speech inputs) and `labels` (which are the `input_ids` of the encoded target seq
 
 [[autodoc]] FlaxSpeechEncoderDecoderModel
     - __call__
-    - from_encoder_decoder_pretrained
+    - from_encoder_decoder_pretrained

docs/source/en/model_doc/speecht5.md

@@ -71,7 +71,7 @@ This model was contributed by [Matthijs](https://huggingface.co/Matthijs). The o
 
 [[autodoc]] SpeechT5ForTextToSpeech
     - forward
-    - generate_speech
+    - generate
 
 ## SpeechT5ForSpeechToSpeech
 

docs/source/en/model_doc/vitdet.md

@@ -0,0 +1,39 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# ViTDet
+
+## Overview
+
+The ViTDet model was proposed in [Exploring Plain Vision Transformer Backbones for Object Detection](https://arxiv.org/abs/2203.16527) by Yanghao Li, Hanzi Mao, Ross Girshick, Kaiming He.
+VitDet leverages the plain [Vision Transformer](vit) for the task of object detection.
+
+The abstract from the paper is the following:
+
+*We explore the plain, non-hierarchical Vision Transformer (ViT) as a backbone network for object detection. This design enables the original ViT architecture to be fine-tuned for object detection without needing to redesign a hierarchical backbone for pre-training. With minimal adaptations for fine-tuning, our plain-backbone detector can achieve competitive results. Surprisingly, we observe: (i) it is sufficient to build a simple feature pyramid from a single-scale feature map (without the common FPN design) and (ii) it is sufficient to use window attention (without shifting) aided with very few cross-window propagation blocks. With plain ViT backbones pre-trained as Masked Autoencoders (MAE), our detector, named ViTDet, can compete with the previous leading methods that were all based on hierarchical backbones, reaching up to 61.3 AP_box on the COCO dataset using only ImageNet-1K pre-training. We hope our study will draw attention to research on plain-backbone detectors.*
+
+Tips:
+
+- For the moment, only the backbone is available.
+
+This model was contributed by [nielsr](https://huggingface.co/nielsr).
+The original code can be found [here](https://github.com/facebookresearch/detectron2/tree/main/projects/ViTDet).
+
+
+## VitDetConfig
+
+[[autodoc]] VitDetConfig
+
+## VitDetModel
+
+[[autodoc]] VitDetModel
+    - forward

docs/source/en/model_doc/vits.md

@@ -0,0 +1,162 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# VITS
+
+## Overview
+
+The VITS model was proposed in [Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech](https://arxiv.org/abs/2106.06103) by Jaehyeon Kim, Jungil Kong, Juhee Son.
+
+
+VITS (**V**ariational **I**nference with adversarial learning for end-to-end **T**ext-to-**S**peech) is an end-to-end 
+speech synthesis model that predicts a speech waveform conditional on an input text sequence. It is a conditional variational 
+autoencoder (VAE) comprised of a posterior encoder, decoder, and conditional prior.
+
+A set of spectrogram-based acoustic features are predicted by the flow-based module, which is formed of a Transformer-based
+text encoder and multiple coupling layers. The spectrogram is decoded using a stack of transposed convolutional layers,
+much in the same style as the HiFi-GAN vocoder. Motivated by the one-to-many nature of the TTS problem, where the same text 
+input can be spoken in multiple ways, the model also includes a stochastic duration predictor, which allows the model to 
+synthesise speech with different rhythms from the same input text. 
+
+The model is trained end-to-end with a combination of losses derived from variational lower bound and adversarial training. 
+To improve the expressiveness of the model, normalizing flows are applied to the conditional prior distribution. During 
+inference, the text encodings are up-sampled based on the duration prediction module, and then mapped into the 
+waveform using a cascade of the flow module and HiFi-GAN decoder. Due to the stochastic nature of the duration predictor,
+the model is non-deterministic, and thus requires a fixed seed to generate the same speech waveform.
+
+The abstract from the paper is the following:
+
+*Several recent end-to-end text-to-speech (TTS) models enabling single-stage training and parallel sampling have been proposed, but their sample quality does not match that of two-stage TTS systems. In this work, we present a parallel end-to-end TTS method that generates more natural sounding audio than current two-stage models. Our method adopts variational inference augmented with normalizing flows and an adversarial training process, which improves the expressive power of generative modeling. We also propose a stochastic duration predictor to synthesize speech with diverse rhythms from input text. With the uncertainty modeling over latent variables and the stochastic duration predictor, our method expresses the natural one-to-many relationship in which a text input can be spoken in multiple ways with different pitches and rhythms. A subjective human evaluation (mean opinion score, or MOS) on the LJ Speech, a single speaker dataset, shows that our method outperforms the best publicly available TTS systems and achieves a MOS comparable to ground truth.*
+
+This model can also be used with TTS checkpoints from [Massively Multilingual Speech (MMS)](https://arxiv.org/abs/2305.13516) 
+as these checkpoints use the same architecture and a slightly modified tokenizer.
+
+This model was contributed by [Matthijs](https://huggingface.co/Matthijs) and [sanchit-gandhi](https://huggingface.co/sanchit-gandhi). The original code can be found [here](https://github.com/jaywalnut310/vits).
+
+## Model Usage
+
+Both the VITS and MMS-TTS checkpoints can be used with the same API. Since the flow-based model is non-deterministic, it 
+is good practice to set a seed to ensure reproducibility of the outputs. For languages with a Roman alphabet, 
+such as English or French, the tokenizer can be used directly to pre-process the text inputs. The following code example 
+runs a forward pass using the MMS-TTS English checkpoint:
+
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+model = VitsModel.from_pretrained("facebook/mms-tts-eng")
+
+inputs = tokenizer(text="Hello - my dog is cute", return_tensors="pt")
+
+set_seed(555)  # make deterministic
+
+with torch.no_grad():
+   outputs = model(**inputs)
+
+waveform = outputs.waveform[0]
+```
+
+The resulting waveform can be saved as a `.wav` file:
+
+```python
+import scipy
+
+scipy.io.wavfile.write("techno.wav", rate=model.config.sampling_rate, data=waveform)
+```
+
+Or displayed in a Jupyter Notebook / Google Colab:
+
+```python
+from IPython.display import Audio
+
+Audio(waveform, rate=model.config.sampling_rate)
+```
+
+For certain languages with a non-Roman alphabet, such as Arabic, Mandarin or Hindi, the [`uroman`](https://github.com/isi-nlp/uroman) 
+perl package is required to pre-process the text inputs to the Roman alphabet.
+
+You can check whether you require the `uroman` package for your language by inspecting the `is_uroman` attribute of 
+the pre-trained `tokenizer`:
+
+```python
+from transformers import VitsTokenizer
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-eng")
+print(tokenizer.is_uroman)
+```
+
+If required, you should apply the uroman package to your text inputs **prior** to passing them to the `VitsTokenizer`, 
+since currently the tokenizer does not support performing the pre-processing itself.  
+
+To do this, first clone the uroman repository to your local machine and set the bash variable `UROMAN` to the local path:
+
+```bash
+git clone https://github.com/isi-nlp/uroman.git
+cd uroman
+export UROMAN=$(pwd)
+```
+
+You can then pre-process the text input using the following code snippet. You can either rely on using the bash variable 
+`UROMAN` to point to the uroman repository, or you can pass the uroman directory as an argument to the `uromaize` function:
+
+```python
+import torch
+from transformers import VitsTokenizer, VitsModel, set_seed
+import os
+import subprocess
+
+tokenizer = VitsTokenizer.from_pretrained("facebook/mms-tts-kor")
+model = VitsModel.from_pretrained("facebook/mms-tts-kor")
+
+def uromanize(input_string, uroman_path):
+    """Convert non-Roman strings to Roman using the `uroman` perl package."""
+    script_path = os.path.join(uroman_path, "bin", "uroman.pl")
+
+    command = ["perl", script_path]
+
+    process = subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
+    # Execute the perl command
+    stdout, stderr = process.communicate(input=input_string.encode())
+
+    if process.returncode != 0:
+        raise ValueError(f"Error {process.returncode}: {stderr.decode()}")
+
+    # Return the output as a string and skip the new-line character at the end
+    return stdout.decode()[:-1]
+
+text = "이봐 무슨 일이야"
+uromaized_text = uromanize(text, uroman_path=os.environ["UROMAN"])
+
+inputs = tokenizer(text=uromaized_text, return_tensors="pt")
+
+set_seed(555)  # make deterministic
+with torch.no_grad():
+   outputs = model(inputs["input_ids"])
+
+waveform = outputs.waveform[0]
+```
+
+## VitsConfig
+
+[[autodoc]] VitsConfig
+
+## VitsTokenizer
+
+[[autodoc]] VitsTokenizer
+    - __call__
+    - save_vocabulary
+
+## VitsModel
+
+[[autodoc]] VitsModel
+    - forward

docs/source/en/model_memory_anatomy.md

@@ -76,7 +76,7 @@ GPU memory occupied: 0 MB.
 
 That looks good: the GPU memory is not occupied as we would expect before we load any models. If that's not the case on 
 your machine make sure to stop all processes that are using GPU memory. However, not all free GPU memory can be used by 
-the user. When a model is loaded to the GPU also the kernels are loaded which can take up 1-2GB of memory. To see how 
+the user. When a model is loaded to the GPU the kernels are also loaded, which can take up 1-2GB of memory. To see how 
 much it is we load a tiny tensor into the GPU which triggers the kernels to be loaded as well.
 
 ```py
@@ -105,7 +105,7 @@ how much space just the weights use.
 GPU memory occupied: 2631 MB.
 ```
 
-We can see that the model weights alone take up 1.3 GB of the GPU memory. The exact number depends on the specific 
+We can see that the model weights alone take up 1.3 GB of GPU memory. The exact number depends on the specific 
 GPU you are using. Note that on newer GPUs a model can sometimes take up more space since the weights are loaded in an 
 optimized fashion that speeds up the usage of the model. Now we can also quickly check if we get the same result 
 as with `nvidia-smi` CLI:
@@ -184,7 +184,7 @@ GPU memory occupied: 14949 MB.
 We see that already a relatively small batch size almost fills up our GPU's entire memory. However, a larger batch size 
 can often result in faster model convergence or better end performance. So ideally we want to tune the batch size to our
 model's needs and not to the GPU limitations. What's interesting is that we use much more memory than the size of the model. 
-To understand a bit better why this is the case let's have look at a model's operations and memory needs.
+To understand a bit better why this is the case let's have a look at a model's operations and memory needs.
 
 ## Anatomy of Model's Operations
 

docs/source/en/model_sharing.md

@@ -95,7 +95,7 @@ Specify `from_pt=True` to convert a checkpoint from PyTorch to TensorFlow:
 >>> tf_model = TFDistilBertForSequenceClassification.from_pretrained("path/to/awesome-name-you-picked", from_pt=True)
 ```
 
-Then you can save your new TensorFlow model with it's new checkpoint:
+Then you can save your new TensorFlow model with its new checkpoint:
 
 ```py
 >>> tf_model.save_pretrained("path/to/awesome-name-you-picked")
@@ -201,7 +201,7 @@ Or perhaps you'd like to add the TensorFlow version of your fine-tuned PyTorch m
 >>> tf_model.push_to_hub("my-awesome-model")
 ```
 
-Now when you navigate to the your Hugging Face profile, you should see your newly created model repository. Clicking on the **Files** tab will display all the files you've uploaded to the repository.
+Now when you navigate to your Hugging Face profile, you should see your newly created model repository. Clicking on the **Files** tab will display all the files you've uploaded to the repository.
 
 For more details on how to create and upload files to a repository, refer to the Hub documentation [here](https://huggingface.co/docs/hub/how-to-upstream).
 

docs/source/en/multilingual.md

@@ -171,9 +171,9 @@ Tokenize the text:
 MBart forces the target language id as the first generated token to translate to the target language. Set the `forced_bos_token_id` to `en` in the `generate` method to translate to English:
 
 ```py
->>> generated_tokens = model.generate(**encoded_en, forced_bos_token_id=tokenizer.lang_code_to_id("en_XX"))
+>>> generated_tokens = model.generate(**encoded_en, forced_bos_token_id=tokenizer.lang_code_to_id["en_XX"])
 >>> tokenizer.batch_decode(generated_tokens, skip_special_tokens=True)
 "Don't interfere with the wizard's affairs, because they are subtle, will soon get angry."
 ```
 
-If you are using the `facebook/mbart-large-50-many-to-one-mmt` checkpoint, you don't need to force the target language id as the first generated token otherwise the usage is the same.
+If you are using the `facebook/mbart-large-50-many-to-one-mmt` checkpoint, you don't need to force the target language id as the first generated token otherwise the usage is the same.

docs/source/en/peft.md

@@ -0,0 +1,216 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+http://www.apache.org/licenses/LICENSE-2.0
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# Load adapters with 🤗 PEFT
+
+[[open-in-colab]]
+
+[Parameter-Efficient Fine Tuning (PEFT)](https://huggingface.co/blog/peft) methods freeze the pretrained model parameters during fine-tuning and add a small number of trainable parameters (the adapters) on top of it. The adapters are trained to learn task-specific information. This approach has been shown to be very memory-efficient with lower compute usage while producing results comparable to a fully fine-tuned model. 
+
+Adapters trained with PEFT are also usually an order of magnitude smaller than the full model, making it convenient to share, store, and load them.
+
+<div class="flex flex-col justify-center">
+  <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/peft/PEFT-hub-screenshot.png"/>
+  <figcaption class="text-center">The adapter weights for a OPTForCausalLM model stored on the Hub are only ~6MB compared to the full size of the model weights, which can be ~700MB.</figcaption>
+</div>
+
+If you're interested in learning more about the 🤗 PEFT library, check out the [documentation](https://huggingface.co/docs/peft/index).
+
+## Setup
+
+Get started by installing 🤗 PEFT:
+
+```bash
+pip install peft
+```
+
+If you want to try out the brand new features, you might be interested in installing the library from source:
+
+```bash
+pip install git+https://github.com/huggingface/peft.git
+```
+
+## Supported PEFT models
+
+🤗 Transformers natively supports some PEFT methods, meaning you can load adapter weights stored locally or on the Hub and easily run or train them with a few lines of code. The following methods are supported:
+
+- [Low Rank Adapters](https://huggingface.co/docs/peft/conceptual_guides/lora)
+- [IA3](https://huggingface.co/docs/peft/conceptual_guides/ia3)
+- [AdaLoRA](https://arxiv.org/abs/2303.10512)
+
+If you want to use other PEFT methods, such as prompt learning or prompt tuning, or about the 🤗 PEFT library in general, please refer to the [documentation](https://huggingface.co/docs/peft/index).
+
+
+## Load a PEFT adapter
+
+To load and use a PEFT adapter model from 🤗 Transformers, make sure the Hub repository or local directory contains an `adapter_config.json` file and the adapter weights, as shown in the example image above. Then you can load the PEFT adapter model using the `AutoModelFor` class. For example, to load a PEFT adapter model for causal language modeling:
+
+1. specify the PEFT model id
+2. pass it to the [`AutoModelForCausalLM`] class
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+peft_model_id = "ybelkada/opt-350m-lora"
+model = AutoModelForCausalLM.from_pretrained(peft_model_id)
+```
+
+<Tip>
+
+You can load a PEFT adapter with either an `AutoModelFor` class or the base model class like `OPTForCausalLM` or `LlamaForCausalLM`.
+
+</Tip>
+
+You can also load a PEFT adapter by calling the `load_adapter` method:
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+model_id = "facebook/opt-350m"
+peft_model_id = "ybelkada/opt-350m-lora"
+
+model = AutoModelForCausalLM.from_pretrained(model_id)
+model.load_adapter(peft_model_id)
+```
+
+## Load in 8bit or 4bit
+
+The `bitsandbytes` integration supports 8bit and 4bit precision data types, which are useful for loading large models because it saves memory (see the `bitsandbytes` integration [guide](./quantization#bitsandbytes-integration) to learn more). Add the `load_in_8bit` or `load_in_4bit` parameters to [`~PreTrainedModel.from_pretrained`] and set `device_map="auto"` to effectively distribute the model to your hardware:
+
+```py
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+peft_model_id = "ybelkada/opt-350m-lora"
+model = AutoModelForCausalLM.from_pretrained(peft_model_id, device_map="auto", load_in_8bit=True)
+```
+
+## Add a new adapter
+
+You can use [`~peft.PeftModel.add_adapter`] to add a new adapter to a model with an existing adapter as long as the new adapter is the same type as the current one. For example, if you have an existing LoRA adapter attached to a model:
+
+```py
+from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
+from peft import PeftConfig
+
+model_id = "facebook/opt-350m"
+model = AutoModelForCausalLM.from_pretrained(model_id)
+
+lora_config = LoraConfig(
+    target_modules=["q_proj", "k_proj"],
+    init_lora_weights=False
+)
+
+model.add_adapter(lora_config, adapter_name="adapter_1")
+```
+
+To add a new adapter:
+
+```py
+# attach new adapter with same config
+model.add_adapter(lora_config, adapter_name="adapter_2")
+```
+
+Now you can use [`~peft.PeftModel.set_adapter`] to set which adapter to use:
+
+```py
+# use adapter_1
+model.set_adapter("adapter_1")
+output = model.generate(**inputs)
+print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
+
+# use adapter_2
+model.set_adapter("adapter_2")
+output_enabled = model.generate(**inputs)
+print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
+```
+
+## Enable and disable adapters
+
+Once you've added an adapter to a model, you can enable or disable the adapter module. To enable the adapter module:
+
+```py
+from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
+from peft import PeftConfig
+
+model_id = "facebook/opt-350m"
+adapter_model_id = "ybelkada/opt-350m-lora"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+text = "Hello"
+inputs = tokenizer(text, return_tensors="pt")
+
+model = AutoModelForCausalLM.from_pretrained(model_id)
+peft_config = PeftConfig.from_pretrained(adapter_model_id)
+
+# to initiate with random weights
+peft_config.init_lora_weights = False
+
+model.add_adapter(peft_config)
+model.enable_adapters()
+output = model.generate(**inputs)
+```
+
+To disable the adapter module:
+
+```py
+model.disable_adapters()
+output = model.generate(**inputs)
+```
+
+## Train a PEFT adapter
+
+PEFT adapters are supported by the [`Trainer`] class so that you can train an adapter for your specific use case. It only requires adding a few more lines of code. For example, to train a LoRA adapter:
+
+<Tip>
+
+If you aren't familiar with fine-tuning a model with [`Trainer`], take a look at the [Fine-tune a pretrained model](training) tutorial.
+
+</Tip>
+
+1. Define your adapter configuration with the task type and hyperparameters (see [`~peft.LoraConfig`] for more details about what the hyperparameters do).
+
+```py
+from peft import LoraConfig
+
+peft_config = LoraConfig(
+    lora_alpha=16,
+    lora_dropout=0.1,
+    r=64,
+    bias="none",
+    task_type="CAUSAL_LM",
+)
+```
+
+2. Add adapter to the model.
+
+```py
+model.add_adapter(peft_config)
+```
+
+3. Now you can pass the model to [`Trainer`]!
+
+```py
+trainer = Trainer(model=model, ...)
+trainer.train()
+```
+
+To save your trained adapter and load it back:
+
+```py
+model.save_pretrained(save_dir)
+model = AutoModelForCausalLM.from_pretrained(save_dir)
+```
+
+<!--
+TODO: (@younesbelkada @stevhliu)
+-   Link to PEFT docs for further details
+-   Trainer  
+-   8-bit / 4-bit examples ?
+-->

docs/source/en/perf_infer_gpu_many.md

@@ -22,6 +22,99 @@ Note: A multi GPU setup can use the majority of the strategies described in the
 
 </Tip>
 
-## `BetterTransformer` for faster inference
+## BetterTransformer
 
-We have recently integrated `BetterTransformer` for faster inference on multi-GPU for text, image and audio models. Check the documentation about this integration [here](https://huggingface.co/docs/optimum/bettertransformer/overview) for more details.
+[BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview) converts 🤗 Transformers models to use the PyTorch-native fastpath execution, which calls optimized kernels like Flash Attention under the hood.  
+
+BetterTransformer is also supported for faster inference on single and multi-GPU for text, image, and audio models.
+
+<Tip>
+
+Flash Attention can only be used for models using fp16 or bf16 dtype. Make sure to cast your model to the appropriate dtype before using BetterTransformer.
+  
+</Tip>
+
+### Decoder models
+
+For text models, especially decoder-based models (GPT, T5, Llama, etc.), the BetterTransformer API converts all attention operations to use the [`torch.nn.functional.scaled_dot_product_attention` operator](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) (SDPA) that is only available in PyTorch 2.0 and onwards. 
+
+To convert a model to BetterTransformer:
+
+```python
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+# convert the model to BetterTransformer
+model.to_bettertransformer()
+
+# Use it for training or inference
+```
+
+SDPA can also call [Flash Attention](https://arxiv.org/abs/2205.14135) kernels under the hood. To enable Flash Attention or to check that it is available in a given setting (hardware, problem size), use [`torch.backends.cuda.sdp_kernel`](https://pytorch.org/docs/master/backends.html#torch.backends.cuda.sdp_kernel) as a context manager:
+
+
+```diff
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m").to("cuda")
+# convert the model to BetterTransformer
+model.to_bettertransformer()
+
+input_text = "Hello my dog is cute and"
+inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
+
++ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    outputs = model.generate(**inputs)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+If you see a bug with a traceback saying 
+
+```bash
+RuntimeError: No available kernel.  Aborting execution.
+```
+
+try using the PyTorch nightly version, which may have a broader coverage for Flash Attention:
+
+```bash
+pip3 install -U --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
+```
+
+Have a look at this [blog post](https://pytorch.org/blog/out-of-the-box-acceleration/) to learn more about what is possible with the BetterTransformer + SDPA API.
+
+### Encoder models
+
+For encoder models during inference, BetterTransformer dispatches the forward call of encoder layers to an equivalent of [`torch.nn.TransformerEncoderLayer`](https://pytorch.org/docs/stable/generated/torch.nn.TransformerEncoderLayer.html) that will execute the fastpath implementation of the encoder layers.
+
+Because `torch.nn.TransformerEncoderLayer` fastpath does not support training, it is dispatched to `torch.nn.functional.scaled_dot_product_attention` instead, which does not leverage nested tensors but can use Flash Attention or Memory-Efficient Attention fused kernels.
+
+More details about BetterTransformer performance can be found in this [blog post](https://medium.com/pytorch/bettertransformer-out-of-the-box-performance-for-huggingface-transformers-3fbe27d50ab2), and you can learn more about BetterTransformer for encoder models in this [blog](https://pytorch.org/blog/a-better-transformer-for-fast-transformer-encoder-inference/).
+
+
+## Advanced usage: mixing FP4 (or Int8) and BetterTransformer
+
+You can combine the different methods described above to get the best performance for your model. For example, you can use BetterTransformer with FP4 mixed-precision inference + flash attention:
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.float16
+)
+
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", quantization_config=quantization_config)
+
+input_text = "Hello my dog is cute and"
+inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    outputs = model.generate(**inputs)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```

docs/source/en/perf_infer_gpu_one.md

@@ -17,7 +17,19 @@ rendered properly in your Markdown viewer.
 
 In addition to this guide, relevant information can be found as well in [the guide for training on a single GPU](perf_train_gpu_one) and [the guide for inference on CPUs](perf_infer_cpu).
 
-## Better Transformer: PyTorch-native transformer fastpath
+## BetterTransformer
+
+[BetterTransformer](https://huggingface.co/docs/optimum/bettertransformer/overview) converts 🤗 Transformers models to use the PyTorch-native fastpath execution, which calls optimized kernels like Flash Attention under the hood.  
+
+BetterTransformer is also supported for faster inference on single and multi-GPU for text, image, and audio models.
+
+<Tip>
+
+Flash Attention can only be used for models using fp16 or bf16 dtype. Make sure to cast your model to the appropriate dtype before using BetterTransformer.
+  
+</Tip>
+
+### Encoder models
 
 PyTorch-native [`nn.MultiHeadAttention`](https://pytorch.org/blog/a-better-transformer-for-fast-transformer-encoder-inference/) attention fastpath, called BetterTransformer, can be used with Transformers through the integration in the [🤗 Optimum library](https://huggingface.co/docs/optimum/bettertransformer/overview).
 
@@ -36,7 +48,61 @@ model = model.reverse_bettertransformer()
 model.save_pretrained("saved_model")
 ```
 
-As of PyTorch 2.0, the attention fastpath is supported for both encoders and decoders. The list of supported architectures can be found [here](https://huggingface.co/docs/optimum/bettertransformer/overview#supported-models).
+Have a look at this [blog post](https://medium.com/pytorch/bettertransformer-out-of-the-box-performance-for-huggingface-transformers-3fbe27d50ab2) to learn more about what is possible to do with `BetterTransformer` API for encoder models.
+
+### Decoder models
+
+For text models, especially decoder-based models (GPT, T5, Llama, etc.), the BetterTransformer API converts all attention operations to use the [`torch.nn.functional.scaled_dot_product_attention` operator](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) (SDPA) that is only available in PyTorch 2.0 and onwards. 
+
+To convert a model to BetterTransformer:
+
+```python
+from transformers import AutoModelForCausalLM
+
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m")
+# convert the model to BetterTransformer
+model.to_bettertransformer()
+
+# Use it for training or inference
+```
+
+SDPA can also call [Flash Attention](https://arxiv.org/abs/2205.14135) kernels under the hood. To enable Flash Attention or to check that it is available in a given setting (hardware, problem size), use [`torch.backends.cuda.sdp_kernel`](https://pytorch.org/docs/master/backends.html#torch.backends.cuda.sdp_kernel) as a context manager:
+
+
+```diff
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", torch_dtype=torch.float16).to("cuda")
+# convert the model to BetterTransformer
+model.to_bettertransformer()
+
+input_text = "Hello my dog is cute and"
+inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
+
++ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    outputs = model.generate(**inputs)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```
+
+If you see a bug with a traceback saying 
+
+```bash
+RuntimeError: No available kernel.  Aborting execution.
+```
+
+try using the PyTorch nightly version, which may have a broader coverage for Flash Attention:
+
+```bash
+pip3 install -U --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
+```
+
+Or make sure your model is correctly casted in float16 or bfloat16
+
+
+Have a look at [this detailed blogpost](https://pytorch.org/blog/out-of-the-box-acceleration/) to read more about what is possible to do with `BetterTransformer` + SDPA API.
 
 ## `bitsandbytes` integration for FP4 mixed-precision inference
 
@@ -104,12 +170,12 @@ Note that this feature can also be used in a multi GPU setup.
 From the paper [`LLM.int8() : 8-bit Matrix Multiplication for Transformers at Scale`](https://arxiv.org/abs/2208.07339), we support Hugging Face integration for all models in the Hub with a few lines of code.
 The method reduces `nn.Linear` size by 2 for `float16` and `bfloat16` weights and by 4 for `float32` weights, with close to no impact to the quality by operating on the outliers in half-precision.
 
-![HFxbitsandbytes.png](https://s3.amazonaws.com/moonup/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png)
+![HFxbitsandbytes.png](https://cdn-uploads.huggingface.co/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png)
 
 Int8 mixed-precision matrix decomposition works by separating a matrix multiplication into two streams: (1) a systematic feature outlier stream matrix multiplied in fp16 (0.01%), (2) a regular stream of int8 matrix multiplication (99.9%). With this method, int8 inference with no predictive degradation is possible for very large models.
 For more details regarding the method, check out the [paper](https://arxiv.org/abs/2208.07339) or our [blogpost about the integration](https://huggingface.co/blog/hf-bitsandbytes-integration).
 
-![MixedInt8.gif](https://s3.amazonaws.com/moonup/production/uploads/1660567469965-62441d1d9fdefb55a0b7d12c.gif)
+![MixedInt8.gif](https://cdn-uploads.huggingface.co/production/uploads/1660567469965-62441d1d9fdefb55a0b7d12c.gif)
 
 Note, that you would require a GPU to run mixed-8bit models as the kernels have been compiled for GPUs only. Make sure that you have enough GPU memory to store the quarter (or half if your model weights are in half precision) of the model before using this feature.
 Below are some notes to help you use this module, or follow the demos on [Google colab](#colab-demos).
@@ -182,3 +248,28 @@ Check out the demo for running T5-11b (42GB in fp32)! Using 8-bit quantization o
 Or this demo for BLOOM-3B:
 
 [![Open In Colab: BLOOM-3b demo](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1qOjXfQIAULfKvZqwCen8-MoWKGdSatZ4?usp=sharing)
+
+## Advanced usage: mixing FP4 (or Int8) and BetterTransformer
+
+You can combine the different methods described above to get the best performance for your model. For example, you can use BetterTransformer with FP4 mixed-precision inference + flash attention:
+
+```py
+import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
+
+quantization_config = BitsAndBytesConfig(
+    load_in_4bit=True,
+    bnb_4bit_compute_dtype=torch.float16
+)
+
+tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
+model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", quantization_config=quantization_config)
+
+input_text = "Hello my dog is cute and"
+inputs = tokenizer(input_text, return_tensors="pt").to("cuda")
+
+with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
+    outputs = model.generate(**inputs)
+
+print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+```

docs/source/en/perf_torch_compile.md

@@ -0,0 +1,359 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Optimize inference using torch.compile()
+
+This guide aims to provide a benchmark on the inference speed-ups introduced with [`torch.compile()`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) for [computer vision models in 🤗 Transformers](https://huggingface.co/models?pipeline_tag=image-classification&library=transformers&sort=trending).
+
+## Benefits of torch.compile
+   
+Depending on the model and the GPU, `torch.compile()` yields up to 30% speed-up during inference. To use `torch.compile()`, simply install any version of `torch` above 2.0. 
+
+Compiling a model takes time, so it's useful if you are compiling the model only once instead of every time you infer.
+To compile any computer vision model of your choice, call `torch.compile()` on the model as shown below:
+
+```diff
+from transformers import AutoModelForImageClassification
+
+model = AutoModelForImageClassification.from_pretrained(MODEL_ID).to("cuda")
++ model = torch.compile(model)
+```
+
+`compile()` comes with multiple modes for compiling, which essentially differ in compilation time and inference overhead. `max-autotune` takes longer than `reduce-overhead` but results in faster inference. Default mode is fastest for compilation but is not as efficient compared to `reduce-overhead` for inference time. In this guide, we used the default mode. You can learn more about it [here](https://pytorch.org/get-started/pytorch-2.0/#user-experience).
+
+We benchmarked `torch.compile` with different computer vision models, tasks, types of hardware, and batch sizes on `torch` version 2.0.1.
+
+## Benchmarking code 
+
+Below you can find the benchmarking code for each task. We warm up the GPU before inference and take the mean time of 300 inferences, using the same image each time.
+
+### Image Classification with ViT
+
+```python 
+import torch
+from PIL import Image
+import requests
+import numpy as np
+from transformers import AutoImageProcessor, AutoModelForImageClassification
+
+url = 'http://images.cocodataset.org/val2017/000000039769.jpg'
+image = Image.open(requests.get(url, stream=True).raw)
+
+processor = AutoImageProcessor.from_pretrained("google/vit-base-patch16-224")
+model = AutoModelForImageClassification.from_pretrained("google/vit-base-patch16-224").to("cuda")
+model = torch.compile(model)
+
+processed_input = processor(image, return_tensors='pt').to(device="cuda")
+
+with torch.no_grad():
+    _ = model(**processed_input)
+
+```
+
+#### Object Detection with DETR
+
+```python 
+from transformers import AutoImageProcessor, AutoModelForObjectDetection
+
+processor = AutoImageProcessor.from_pretrained("facebook/detr-resnet-50")
+model = AutoModelForObjectDetection.from_pretrained("facebook/detr-resnet-50").to("cuda")
+model = torch.compile(model)
+
+texts = ["a photo of a cat", "a photo of a dog"]
+inputs = processor(text=texts, images=image, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    _ = model(**inputs)
+```
+
+#### Image Segmentation with Segformer
+
+```python 
+from transformers import SegformerImageProcessor, SegformerForSemanticSegmentation
+
+processor = SegformerImageProcessor.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512")
+model = SegformerForSemanticSegmentation.from_pretrained("nvidia/segformer-b0-finetuned-ade-512-512").to("cuda")
+model = torch.compile(model)
+seg_inputs = processor(images=image, return_tensors="pt").to("cuda")
+
+with torch.no_grad():
+    _ = model(**seg_inputs)
+```
+
+Below you can find the list of the models we benchmarked.
+
+**Image Classification** 
+- [google/vit-base-patch16-224](https://huggingface.co/google/vit-base-patch16-224)
+- [microsoft/beit-base-patch16-224-pt22k-ft22k](https://huggingface.co/microsoft/beit-base-patch16-224-pt22k-ft22k)
+- [facebook/convnext-large-224](https://huggingface.co/facebook/convnext-large-224)
+- [microsoft/resnet-50](https://huggingface.co/)
+
+**Image Segmentation** 
+- [nvidia/segformer-b0-finetuned-ade-512-512](https://huggingface.co/nvidia/segformer-b0-finetuned-ade-512-512)
+- [facebook/mask2former-swin-tiny-coco-panoptic](https://huggingface.co/facebook/mask2former-swin-tiny-coco-panoptic)
+- [facebook/maskformer-swin-base-ade](https://huggingface.co/facebook/maskformer-swin-base-ade)
+- [google/deeplabv3_mobilenet_v2_1.0_513](https://huggingface.co/google/deeplabv3_mobilenet_v2_1.0_513)
+
+**Object Detection** 
+- [google/owlvit-base-patch32](https://huggingface.co/google/owlvit-base-patch32)
+- [facebook/detr-resnet-101](https://huggingface.co/facebook/detr-resnet-101)
+- [microsoft/conditional-detr-resnet-50](https://huggingface.co/microsoft/conditional-detr-resnet-50)
+
+Below you can find visualization of inference durations with and without `torch.compile()` and percentage improvements for each model in different hardware and batch sizes. 
+
+<div class="flex">
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/a100_batch_comp.png" />
+  </div>
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/v100_batch_comp.png" />
+  </div>
+   <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/t4_batch_comp.png" />
+  </div>
+</div>
+
+<div class="flex">
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/A100_1_duration.png" />
+  </div>
+  <div>
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/A100_1_percentage.png" />
+  </div>
+</div>
+
+
+![Duration Comparison on V100 with Batch Size of 1](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/v100_1_duration.png)
+
+![Percentage Improvement on T4 with Batch Size of 4](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/torch_compile/T4_4_percentage.png)
+
+Below you can find inference durations in milliseconds for each model with and without `compile()`. Note that OwlViT results in OOM in larger batch sizes.
+
+### A100 (batch size: 1)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 9.325 | 7.584 | 
+| Image Segmentation/Segformer | 11.759 | 10.500 |
+| Object Detection/OwlViT | 24.978 | 18.420 |
+| Image Classification/BeiT | 11.282 | 8.448 | 
+| Object Detection/DETR | 34.619 | 19.040 |
+| Image Classification/ConvNeXT | 10.410 | 10.208 | 
+| Image Classification/ResNet | 6.531 | 4.124 |
+| Image Segmentation/Mask2former | 60.188 | 49.117 |
+| Image Segmentation/Maskformer | 75.764 | 59.487 | 
+| Image Segmentation/MobileNet | 8.583 | 3.974 |
+| Object Detection/Resnet-101 | 36.276 | 18.197 |
+| Object Detection/Conditional-DETR | 31.219 | 17.993 |
+
+
+### A100 (batch size: 4)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 14.832 | 14.499 | 
+| Image Segmentation/Segformer | 18.838 | 16.476 |
+| Image Classification/BeiT | 13.205 | 13.048 | 
+| Object Detection/DETR | 48.657 | 32.418|
+| Image Classification/ConvNeXT | 22.940 | 21.631 | 
+| Image Classification/ResNet | 6.657 | 4.268 |
+| Image Segmentation/Mask2former | 74.277 | 61.781 |
+| Image Segmentation/Maskformer | 180.700 | 159.116 | 
+| Image Segmentation/MobileNet | 14.174 | 8.515 |
+| Object Detection/Resnet-101 | 68.101 | 44.998 |
+| Object Detection/Conditional-DETR | 56.470 | 35.552 |
+
+### A100 (batch size: 16)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 40.944 | 40.010 | 
+| Image Segmentation/Segformer | 37.005 | 31.144 |
+| Image Classification/BeiT | 41.854 | 41.048 | 
+| Object Detection/DETR | 164.382 | 161.902 |
+| Image Classification/ConvNeXT | 82.258 | 75.561 | 
+| Image Classification/ResNet | 7.018 | 5.024 |
+| Image Segmentation/Mask2former | 178.945 | 154.814 |
+| Image Segmentation/Maskformer | 638.570 | 579.826 | 
+| Image Segmentation/MobileNet | 51.693 | 30.310 |
+| Object Detection/Resnet-101 | 232.887 | 155.021 |
+| Object Detection/Conditional-DETR | 180.491 | 124.032 |
+
+### V100 (batch size: 1)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 10.495 | 6.00 | 
+| Image Segmentation/Segformer | 13.321 | 5.862 | 
+| Object Detection/OwlViT | 25.769 | 22.395 | 
+| Image Classification/BeiT | 11.347 | 7.234 | 
+| Object Detection/DETR | 33.951 | 19.388 |
+| Image Classification/ConvNeXT | 11.623 | 10.412 | 
+| Image Classification/ResNet | 6.484 | 3.820 |
+| Image Segmentation/Mask2former | 64.640 | 49.873 |
+| Image Segmentation/Maskformer | 95.532 | 72.207 | 
+| Image Segmentation/MobileNet | 9.217 | 4.753 |
+| Object Detection/Resnet-101 | 52.818 | 28.367 |
+| Object Detection/Conditional-DETR | 39.512 | 20.816 |
+
+### V100 (batch size: 4)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 15.181 | 14.501 | 
+| Image Segmentation/Segformer | 16.787 | 16.188 |
+| Image Classification/BeiT | 15.171 | 14.753 | 
+| Object Detection/DETR | 88.529 | 64.195 |
+| Image Classification/ConvNeXT | 29.574 | 27.085 | 
+| Image Classification/ResNet | 6.109 | 4.731 |
+| Image Segmentation/Mask2former | 90.402 | 76.926 |
+| Image Segmentation/Maskformer | 234.261 | 205.456 | 
+| Image Segmentation/MobileNet | 24.623 | 14.816 |
+| Object Detection/Resnet-101 | 134.672 | 101.304 |
+| Object Detection/Conditional-DETR | 97.464 | 69.739 |
+
+### V100 (batch size: 16)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 52.209 | 51.633 | 
+| Image Segmentation/Segformer | 61.013 | 55.499 |
+| Image Classification/BeiT | 53.938 | 53.581  |
+| Object Detection/DETR | OOM | OOM |
+| Image Classification/ConvNeXT | 109.682 | 100.771 | 
+| Image Classification/ResNet | 14.857 | 12.089 |
+| Image Segmentation/Mask2former | 249.605 | 222.801 |
+| Image Segmentation/Maskformer | 831.142 | 743.645 | 
+| Image Segmentation/MobileNet | 93.129 | 55.365 |
+| Object Detection/Resnet-101 | 482.425 | 361.843 |
+| Object Detection/Conditional-DETR | 344.661 | 255.298 |
+
+### T4 (batch size: 1)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 16.520 | 15.786 | 
+| Image Segmentation/Segformer | 16.116 | 14.205 |
+| Object Detection/OwlViT | 53.634 | 51.105 |
+| Image Classification/BeiT | 16.464 | 15.710 | 
+| Object Detection/DETR | 73.100 | 53.99 |
+| Image Classification/ConvNeXT | 32.932 | 30.845 | 
+| Image Classification/ResNet | 6.031 | 4.321 |
+| Image Segmentation/Mask2former | 79.192 | 66.815 |
+| Image Segmentation/Maskformer | 200.026 | 188.268 | 
+| Image Segmentation/MobileNet | 18.908 | 11.997 |
+| Object Detection/Resnet-101 | 106.622 | 82.566 |
+| Object Detection/Conditional-DETR | 77.594 | 56.984 |
+
+### T4 (batch size: 4)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 43.653 | 43.626 | 
+| Image Segmentation/Segformer | 45.327 | 42.445 |
+| Image Classification/BeiT | 52.007 | 51.354 | 
+| Object Detection/DETR | 277.850 | 268.003 |
+| Image Classification/ConvNeXT | 119.259 | 105.580 | 
+| Image Classification/ResNet | 13.039 | 11.388 |
+| Image Segmentation/Mask2former | 201.540 | 184.670 |
+| Image Segmentation/Maskformer | 764.052 | 711.280 | 
+| Image Segmentation/MobileNet | 74.289 | 48.677 |
+| Object Detection/Resnet-101 | 421.859 | 357.614 |
+| Object Detection/Conditional-DETR | 289.002 | 226.945 |
+
+### T4 (batch size: 16)
+
+| **Task/Model** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|
+| Image Classification/ViT | 163.914 | 160.907 | 
+| Image Segmentation/Segformer | 192.412 | 163.620 |
+| Image Classification/BeiT | 188.978 | 187.976 | 
+| Object Detection/DETR | OOM | OOM |
+| Image Classification/ConvNeXT | 422.886 | 388.078 | 
+| Image Classification/ResNet | 44.114 | 37.604 |
+| Image Segmentation/Mask2former | 756.337 | 695.291 |
+| Image Segmentation/Maskformer | 2842.940 | 2656.88 | 
+| Image Segmentation/MobileNet | 299.003 | 201.942 |
+| Object Detection/Resnet-101 |  1619.505 | 1262.758 | 
+| Object Detection/Conditional-DETR | 1137.513 | 897.390|
+
+## PyTorch Nightly
+We also benchmarked on PyTorch nightly (2.1.0dev, find the wheel [here](https://download.pytorch.org/whl/nightly/cu118)) and observed improvement in latency both for uncompiled and compiled models. 
+
+### A100
+
+| **Task/Model** | **Batch Size** | **torch 2.0 - no compile** | **torch 2.0 -<br> compile** |
+|:---:|:---:|:---:|:---:|
+| Image Classification/BeiT | Unbatched | 12.462 | 6.954 | 
+| Image Classification/BeiT | 4 | 14.109 | 12.851 | 
+| Image Classification/BeiT | 16 | 42.179 | 42.147 | 
+| Object Detection/DETR | Unbatched | 30.484 | 15.221 |
+| Object Detection/DETR | 4 | 46.816 | 30.942 |
+| Object Detection/DETR | 16 | 163.749 | 163.706  |
+
+### T4
+
+| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|:---:|
+| Image Classification/BeiT | Unbatched | 14.408 | 14.052 | 
+| Image Classification/BeiT | 4 | 47.381 | 46.604 | 
+| Image Classification/BeiT | 16 | 42.179 | 42.147  | 
+| Object Detection/DETR | Unbatched | 68.382 | 53.481 |
+| Object Detection/DETR | 4 | 269.615 | 204.785 |
+| Object Detection/DETR | 16 | OOM | OOM   |
+
+### V100
+
+| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|:---:|
+| Image Classification/BeiT | Unbatched | 13.477 | 7.926 | 
+| Image Classification/BeiT | 4 | 15.103 | 14.378 | 
+| Image Classification/BeiT | 16 | 52.517 | 51.691  | 
+| Object Detection/DETR | Unbatched | 28.706 | 19.077 |
+| Object Detection/DETR | 4 | 88.402 | 62.949|
+| Object Detection/DETR | 16 | OOM | OOM  |
+
+
+## Reduce Overhead
+We benchmarked `reduce-overhead` compilation mode for A100 and T4 in Nightly.
+
+### A100
+
+| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** |
+|:---:|:---:|:---:|:---:|
+| Image Classification/ConvNeXT | Unbatched | 11.758 | 7.335 | 
+| Image Classification/ConvNeXT | 4 | 23.171 | 21.490 | 
+| Image Classification/ResNet | Unbatched | 7.435 | 3.801 | 
+| Image Classification/ResNet | 4 | 7.261 | 2.187 | 
+| Object Detection/Conditional-DETR | Unbatched | 32.823 | 11.627  | 
+| Object Detection/Conditional-DETR | 4 | 50.622 | 33.831  | 
+| Image Segmentation/MobileNet | Unbatched | 9.869 | 4.244 |
+| Image Segmentation/MobileNet | 4 | 14.385 | 7.946 |
+
+
+### T4
+
+| **Task/Model** | **Batch Size** | **torch 2.0 - <br>no compile** | **torch 2.0 - <br>compile** | 
+|:---:|:---:|:---:|:---:|
+| Image Classification/ConvNeXT | Unbatched | 32.137 | 31.84 | 
+| Image Classification/ConvNeXT | 4 | 120.944 | 110.209 | 
+| Image Classification/ResNet | Unbatched | 9.761 | 7.698 | 
+| Image Classification/ResNet | 4 | 15.215 | 13.871 | 
+| Object Detection/Conditional-DETR | Unbatched | 72.150 | 57.660  | 
+| Object Detection/Conditional-DETR | 4 | 301.494 | 247.543  | 
+| Image Segmentation/MobileNet | Unbatched | 22.266 | 19.339  |
+| Image Segmentation/MobileNet | 4 | 78.311 | 50.983 |
+
+

docs/source/en/perf_train_gpu_one.md

@@ -237,10 +237,11 @@ For example if you have [NVIDIA/apex](https://github.com/NVIDIA/apex) installed,
 fastest training experience among all supported AdamW optimizers.
 
 [`Trainer`] integrates a variety of optimizers that can be used out of box: `adamw_hf`, `adamw_torch`, `adamw_torch_fused`, 
-`adamw_apex_fused`, `adamw_anyprecision` or `adafactor`. More optimizers can be plugged in via a third-party implementation.
+`adamw_apex_fused`, `adamw_anyprecision`, `adafactor`, or `adamw_bnb_8bit`. More optimizers can be plugged in via a third-party implementation.
 
-Let's take a closer look at two alternatives to AdamW optimizer - Adafactor (available in Trainer), and 8bit BNB quantized 
-optimizer (third-party implementation).
+Let's take a closer look at two alternatives to AdamW optimizer:
+1. `adafactor` which is available in [`Trainer`]
+2. `adamw_bnb_8bit` is also available in Trainer, but a third-party integration is provided below for demonstration.
 
 For comparison, for a 3B-parameter model, like “t5-3b”: 
 * A standard AdamW optimizer will need 24GB of GPU memory because it uses 8 bytes for each parameter (8*3 => 24GB)
@@ -269,7 +270,13 @@ Instead of aggregating optimizer states like Adafactor, 8-bit Adam keeps the ful
 means that it stores the state with lower precision and dequantizes it only for the optimization. This is similar to the 
 idea behind mixed precision training.
 
-To use the 8-bit optimizer, you need to install it separately and then pass it as a custom optimizer to the [`Trainer`]. 
+To use `adamw_bnb_8bit`, you simply need to set `optim="adamw_bnb_8bit"` in [`TrainingArguments`]:
+
+```py
+training_args = TrainingArguments(per_device_train_batch_size=4, optim="adamw_bnb_8bit", **default_args)
+```
+
+However, we can also use a third-party implementation of the 8-bit optimizer for demonstration purposes to see how that can be integrated.
 
 First, follow the installation guide in the GitHub [repo](https://github.com/TimDettmers/bitsandbytes) to install the `bitsandbytes` library 
 that implements the 8-bit Adam optimizer.
@@ -311,13 +318,6 @@ adam_bnb_optim = bnb.optim.Adam8bit(
 )
 ```
 
-<Tip>
-
-To use the 8-bit optimizer with an existing pretrained model, you need to make a change to the embedding layer.
-Read [this issue](https://github.com/huggingface/transformers/issues/14819) for more information.
-
-</Tip>
-
 Finally, pass the custom optimizer as an argument to the `Trainer`:
 
 ```py
@@ -503,10 +503,10 @@ Most related papers and implementations are built around Tensorflow/TPUs:
 
 And for Pytorch DeepSpeed has built one as well: [DeepSpeed-MoE: Advancing Mixture-of-Experts Inference and Training to Power Next-Generation AI Scale](https://arxiv.org/abs/2201.05596), [Mixture of Experts](https://www.deepspeed.ai/tutorials/mixture-of-experts/) - blog posts:  [1](https://www.microsoft.com/en-us/research/blog/deepspeed-powers-8x-larger-moe-model-training-with-high-performance/), [2](https://www.microsoft.com/en-us/research/publication/scalable-and-efficient-moe-training-for-multitask-multilingual-models/) and specific deployment with large transformer-based natural language generation models: [blog post](https://www.deepspeed.ai/news/2021/12/09/deepspeed-moe-nlg.html), [Megatron-Deepspeed branch](Thttps://github.com/microsoft/Megatron-DeepSpeed/tree/moe-training).
 
-## Using PyTorch native attention
+## Using PyTorch native attention and Flash Attention
 
-PyTorch 2.0 released the native [`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) (SDPA), 
-that allows to use fused GPU kernels as [memory-efficient attention](https://arxiv.org/abs/2112.05682) and [flash attention](https://arxiv.org/abs/2205.14135).
+PyTorch 2.0 released a native [`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) (SDPA), 
+that allows using fused GPU kernels such as [memory-efficient attention](https://arxiv.org/abs/2112.05682) and [flash attention](https://arxiv.org/abs/2205.14135).
 
 After installing the [`optimum`](https://github.com/huggingface/optimum) package, the relevant internal modules can be 
 replaced to use PyTorch's native attention with:
@@ -516,3 +516,13 @@ model = model.to_bettertransformer()
 ```
 
 Once converted, train the model as usual.
+
+<Tip warning={true}>
+
+The PyTorch-native `scaled_dot_product_attention` operator can only dispatch to Flash Attention if no `attention_mask` is provided.
+
+By default, in training mode, the BetterTransformer integration **drops the mask support and can only be used for training that does not require a padding mask for batched training**. This is the case, for example, during masked language modeling or causal language modeling. BetterTransformer is not suited for fine-tuning models on tasks that require a padding mask. 
+
+</Tip>
+
+Check out this [blogpost](https://pytorch.org/blog/out-of-the-box-acceleration/) to learn more about acceleration and memory-savings with SDPA.

docs/source/en/pipeline_tutorial.md

@@ -204,7 +204,7 @@ page.
 
 Using a [`pipeline`] for vision tasks is practically identical.
 
-Specify your task and pass your image to the classifier. The image can be a link or a local path to the image. For example, what species of cat is shown below?
+Specify your task and pass your image to the classifier. The image can be a link, a local path or a base64-encoded image. For example, what species of cat is shown below?
 
 ![pipeline-cat-chonk](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/pipeline-cat-chonk.jpeg)
 

docs/source/en/pipeline_webserver.md

@@ -87,6 +87,13 @@ of the model on the webserver. This way, no unnecessary RAM is being used.
 Then the queuing mechanism allows you to do fancy stuff like maybe accumulating a few
 items before inferring to use dynamic batching:
 
+<Tip warning={true}>
+
+The code sample below is intentionally written like pseudo-code for readability.
+Do not run this without checking if it makes sense for your system resources!
+
+</Tip>
+
 ```py
 (string, rq) = await q.get()
 strings = []
@@ -104,11 +111,7 @@ for rq, out in zip(queues, outs):
     await rq.put(out)
 ```
 
-<Tip warning={true}>
-Do not activate this without checking it makes sense for your load!
-</Tip>
-
-The proposed code is optimized for readability, not for being the best code.
+Again, the proposed code is optimized for readability, not for being the best code.
 First of all, there's no batch size limit which is usually not a 
 great idea. Next, the timeout is reset on every queue fetch, meaning you could
 wait much more than 1ms before running the inference (delaying the first request 

docs/source/en/pr_checks.md

@@ -142,3 +142,58 @@ Additional checks concern PRs that add new models, mainly that:
 - All checkpoints used actually exist on the Hub
 
 -->
+
+### Check copies
+
+Since the Transformers library is very opinionated with respect to model code, and each model should fully be implemented in a single file without relying on other models, we have added a mechanism that checks whether a copy of the code of a layer of a given model stays consistent with the original. This way, when there is a bug fix, we can see all other impacted models and choose to trickle down the modification or break the copy.
+
+<Tip>
+
+If a file is a full copy of another file, you should register it in the constant `FULL_COPIES` of `utils/check_copies.py`.
+
+</Tip>
+
+This mechanism relies on comments of the form `# Copied from xxx`. The `xxx` should contain the whole path to the class of function which is being copied below. For instance, `RobertaSelfOutput` is a direct copy of the `BertSelfOutput` class, so you can see [here](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L289) it has a comment:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertSelfOutput
+```
+
+Note that instead of applying this to a whole class, you can apply it to the relevant methods that are copied from. For instance [here](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L598) you can see how `RobertaPreTrainedModel._init_weights` is copied from the same method in `BertPreTrainedModel` with the comment:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertPreTrainedModel._init_weights
+```
+
+Sometimes the copy is exactly the same except for names: for instance in `RobertaAttention`, we use `RobertaSelfAttention` insted of `BertSelfAttention` but other than that, the code is exactly the same. This is why `# Copied from` supports simple string replacements with the follwoing syntax: `Copied from xxx with foo->bar`. This means the code is copied with all instances of `foo` being replaced by `bar`. You can see how it used [here](https://github.com/huggingface/transformers/blob/2bd7a27a671fd1d98059124024f580f8f5c0f3b5/src/transformers/models/roberta/modeling_roberta.py#L304C1-L304C86) in `RobertaAttention` with the comment:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertAttention with Bert->Roberta
+```
+
+Note that there shouldn't be any spaces around the arrow (unless that space is part of the pattern to replace of course).
+
+You can add several patterns separated by a comma. For instance here `CamemberForMaskedLM` is a direct copy of `RobertaForMaskedLM` with two replacements: `Roberta` to `Camembert` and `ROBERTA` to `CAMEMBERT`. You can see [here](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/camembert/modeling_camembert.py#L929) this is done with the comment:
+
+```py
+# Copied from transformers.models.roberta.modeling_roberta.RobertaForMaskedLM with Roberta->Camembert, ROBERTA->CAMEMBERT
+```
+
+If the order matters (because one of the replacements might conflict with a previous one), the replacements are executed from left to right.
+
+<Tip>
+
+If the replacements change the formatting (if you replace a short name by a very long name for instance), the copy is checked after applying the auto-formatter.
+
+</Tip>
+
+Another way when the patterns are just different casings of the same replacement (with an uppercased and a lowercased variants) is just to add the option `all-casing`. [Here](https://github.com/huggingface/transformers/blob/15082a9dc6950ecae63a0d3e5060b2fc7f15050a/src/transformers/models/mobilebert/modeling_mobilebert.py#L1237) is an example in `MobileBertForSequenceClassification` with the comment:
+
+```py
+# Copied from transformers.models.bert.modeling_bert.BertForSequenceClassification with Bert->MobileBert all-casing
+```
+
+In this case, the code is copied from `BertForSequenceClassification` by replacing:
+- `Bert` by `MobileBert` (for instance when using `MobileBertModel` in the init)
+- `bert` by `mobilebert` (for instance when defining `self.mobilebert`)
+- `BERT` by `MOBILEBERT` (in the constant `MOBILEBERT_INPUTS_DOCSTRING`)

docs/source/en/quicktour.md

@@ -30,11 +30,13 @@ You'll also need to install your preferred machine learning framework:
 
 <frameworkcontent>
 <pt>
+
 ```bash
 pip install torch
 ```
 </pt>
 <tf>
+
 ```bash
 pip install tensorflow
 ```
@@ -208,6 +210,7 @@ A tokenizer can also accept a list of inputs, and pad and truncate the text to r
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> pt_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -219,6 +222,7 @@ A tokenizer can also accept a list of inputs, and pad and truncate the text to r
 ```
 </pt>
 <tf>
+
 ```py
 >>> tf_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -352,6 +356,7 @@ One particularly cool 🤗 Transformers feature is the ability to save a model a
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import AutoModel
 
@@ -360,6 +365,7 @@ One particularly cool 🤗 Transformers feature is the ability to save a model a
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import TFAutoModel
 

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:
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
-[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [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), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [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), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [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), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MVP](../model_doc/mvp), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [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)
 
 
 
@@ -154,7 +154,7 @@ This dataset contains the token sequences, but some of these are longer than the
 
 You can now use a second preprocessing function to
 - concatenate all the sequences
-- split the concatenated sequences into shorter chunks defined by `block_size`, which should be both shorter than the maximum input length and short enough for your GPU RAM. 
+- split the concatenated sequences into shorter chunks defined by `block_size`, which should be both shorter than the maximum input length and short enough for your GPU RAM.
 
 ```py
 >>> block_size = 128

docs/source/en/tasks/semantic_segmentation.md

@@ -221,6 +221,10 @@ logits first, and then reshaped to match the size of the labels before you can c
 <pt>
 
 ```py
+>>> import numpy as np
+>>> import torch
+>>> from torch import nn
+
 >>> def compute_metrics(eval_pred):
 ...     with torch.no_grad():
 ...         logits, labels = eval_pred

docs/source/en/tasks/sequence_classification.md

@@ -33,7 +33,7 @@ The task illustrated in this tutorial is supported by the following model archit
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
 
-[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [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), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
+[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
 
 
 

docs/source/en/tasks/text-to-speech.md

@@ -19,16 +19,40 @@ rendered properly in your Markdown viewer.
 [[open-in-colab]]
 
 Text-to-speech (TTS) is the task of creating natural-sounding speech from text, where the speech can be generated in multiple 
-languages and for multiple speakers. The only text-to-speech model currently available in 🤗 Transformers 
-is [SpeechT5](model_doc/speecht5), though more will be added in the future. SpeechT5 is pre-trained on a combination of 
+languages and for multiple speakers. Several text-to-speech models are currently available in 🤗 Transformers, such as 
+[Bark](../model_doc/bark), [MMS](../model_doc/mms), [VITS](../model_doc/vits) and [SpeechT5](../model_doc/speecht5). 
+
+You can easily generate audio using the `"text-to-audio"` pipeline (or its alias - `"text-to-speech"`). Some models, like Bark, 
+can also be conditioned to generate non-verbal communications such as laughing, sighing and crying, or even add music.
+Here's an example of how you would use the `"text-to-speech"` pipeline with Bark: 
+
+```py
+>>> from transformers import pipeline
+
+>>> pipe = pipeline("text-to-speech", model="suno/bark-small")
+>>> text = "[clears throat] This is a test ... and I just took a long pause."
+>>> output = pipe(text)
+```
+
+Here's a code snippet you can use to listen to the resulting audio in a notebook: 
+
+```python
+>>> from IPython.display import Audio
+>>> Audio(output["audio"], rate=output["sampling_rate"])
+```
+
+For more examples on what Bark and other pretrained TTS models can do, refer to our 
+[Audio course](https://huggingface.co/learn/audio-course/chapter6/pre-trained_models). 
+
+If you are looking to fine-tune a TTS model, you can currently fine-tune SpeechT5 only. SpeechT5 is pre-trained on a combination of 
 speech-to-text and text-to-speech data, allowing it to learn a unified space of hidden representations shared by both text 
 and speech. This means that the same pre-trained model can be fine-tuned for different tasks. Furthermore, SpeechT5 
 supports multiple speakers through x-vector speaker embeddings. 
 
-This guide illustrates how to:
+The remainder of this guide illustrates how to:
 
-1. Fine-tune [SpeechT5](model_doc/speecht5) that was originally trained on English speech on the Dutch (`nl`) language subset of the [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli) dataset.
-2. Use your fine-tuned model for inference.
+1. Fine-tune [SpeechT5](../model_doc/speecht5) that was originally trained on English speech on the Dutch (`nl`) language subset of the [VoxPopuli](https://huggingface.co/datasets/facebook/voxpopuli) dataset.
+2. Use your refined model for inference in one of two ways: using a pipeline or directly.
 
 Before you begin, make sure you have all the necessary libraries installed:
 
@@ -485,6 +509,12 @@ the `per_device_train_batch_size` incrementally by factors of 2 and increase `gr
 >>> trainer.train()
 ```
 
+To be able to use your checkpoint with a pipeline, make sure to save the processor with the checkpoint: 
+
+```py
+>>> processor.save_pretrained("YOUR_ACCOUNT_NAME/speecht5_finetuned_voxpopuli_nl")
+```
+
 Push the final model to the 🤗 Hub:
 
 ```py
@@ -493,29 +523,70 @@ Push the final model to the 🤗 Hub:
 
 ## Inference
 
+### Inference with a pipeline
+
 Great, now that you've fine-tuned a model, you can use it for inference!
-Load the model from the 🤗 Hub (make sure to use your account name in the following code snippet): 
+First, let's see how you can use it with a corresponding pipeline. Let's create a `"text-to-speech"` pipeline with your 
+checkpoint: 
+
+```py
+>>> from transformers import pipeline
+
+>>> pipe = pipeline("text-to-speech", model="YOUR_ACCOUNT_NAME/speecht5_finetuned_voxpopuli_nl")
+```
+
+Pick a piece of text in Dutch you'd like narrated, e.g.:
+
+```py
+>>> text = "hallo allemaal, ik praat nederlands. groetjes aan iedereen!"
+```
+
+To use SpeechT5 with the pipeline, you'll need a speaker embedding. Let's get it from an example in the test dataset: 
+
+```py
+>>> example = dataset["test"][304]
+>>> speaker_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
+```
+
+Now you can pass the text and speaker embeddings to the pipeline, and it will take care of the rest: 
+
+```py
+>>> forward_params = {"speaker_embeddings": speaker_embeddings}
+>>> output = pipe(text, forward_params=forward_params)
+>>> output
+{'audio': array([-6.82714235e-05, -4.26525949e-04,  1.06134125e-04, ...,
+        -1.22392643e-03, -7.76011671e-04,  3.29112721e-04], dtype=float32),
+ 'sampling_rate': 16000}
+```
+
+You can then listen to the result:
+
+```py
+>>> from IPython.display import Audio
+>>> Audio(output['audio'], rate=output['sampling_rate']) 
+```
+
+### Run inference manually
+
+You can achieve the same inference results without using the pipeline, however, more steps will be required. 
+
+Load the model from the 🤗 Hub: 
 
 ```py
 >>> model = SpeechT5ForTextToSpeech.from_pretrained("YOUR_ACCOUNT/speecht5_finetuned_voxpopuli_nl")
 ```
 
-Pick an example, here we'll take one from the test dataset. Obtain a speaker embedding. 
+Pick an example from the test dataset obtain a speaker embedding. 
 
 ```py 
 >>> example = dataset["test"][304]
 >>> speaker_embeddings = torch.tensor(example["speaker_embeddings"]).unsqueeze(0)
 ```
 
-Define some input text and tokenize it.
+Define the input text and tokenize it.
 
 ```py 
 >>> text = "hallo allemaal, ik praat nederlands. groetjes aan iedereen!"
-```
-
-Preprocess the input text: 
-
-```py
 >>> inputs = processor(text=text, return_tensors="pt")
 ```
 

docs/source/en/tasks/token_classification.md

@@ -32,8 +32,7 @@ The task illustrated in this tutorial is supported by the following model archit
 
 <!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
 
-[ALBERT](../model_doc/albert), [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), [MPT](../model_doc/mpt), [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), [BROS](../model_doc/bros), [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), [MPT](../model_doc/mpt), [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)
 
 <!--End of the generated tip-->
 

docs/source/en/tasks/visual_question_answering.md

@@ -0,0 +1,401 @@
+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Visual Question Answering
+
+[[open-in-colab]]
+
+Visual Question Answering (VQA) is the task of answering open-ended questions based on an image. 
+The input to models supporting this task is typically a combination of an image and a question, and the output is an 
+answer expressed in natural language.
+
+Some noteworthy use case examples for VQA include:
+* Accessibility applications for visually impaired individuals.
+* Education: posing questions about visual materials presented in lectures or textbooks. VQA can also be utilized in interactive museum exhibits or historical sites.
+* Customer service and e-commerce: VQA can enhance user experience by letting users ask questions about products. 
+* Image retrieval: VQA models can be used to retrieve images with specific characteristics. For example, the user can ask "Is there a dog?" to find all images with dogs from a set of images.
+
+In this guide you'll learn how to:
+
+- Fine-tune a classification VQA model, specifically [ViLT](../model_doc/vilt), on the [`Graphcore/vqa` dataset](https://huggingface.co/datasets/Graphcore/vqa).
+- Use your fine-tuned ViLT for inference.
+- Run zero-shot VQA inference with a generative model, like BLIP-2.
+
+## Fine-tuning ViLT
+
+ViLT model incorporates text embeddings into a Vision Transformer (ViT), allowing it to have a minimal design for 
+Vision-and-Language Pre-training (VLP). This model can be used for several downstream tasks. For the VQA task, a classifier 
+head is placed on top (a linear layer on top of the final hidden state of the `[CLS]` token) and randomly initialized. 
+Visual Question Answering is thus treated as a **classification problem**.
+
+More recent models, such as BLIP, BLIP-2, and InstructBLIP, treat VQA as a generative task. Later in this guide we 
+illustrate how to use them for zero-shot VQA inference. 
+
+Before you begin, make sure you have all the necessary libraries installed. 
+
+```bash
+pip install -q transformers datasets
+```
+
+We encourage you to share your model with the community. Log in to your Hugging Face account to upload it to the 🤗 Hub.
+When prompted, enter your token to log in:
+
+```py
+>>> from huggingface_hub import notebook_login
+
+>>> notebook_login()
+```
+
+Let's define the model checkpoint as a global variable.
+
+```py
+>>> model_checkpoint = "dandelin/vilt-b32-mlm"
+```
+
+## Load the data
+
+For illustration purposes, in this guide we use a very small sample of the annotated visual question answering `Graphcore/vqa` dataset. 
+You can find the full dataset on [🤗 Hub](https://huggingface.co/datasets/Graphcore/vqa).
+
+As an alternative to the [`Graphcore/vqa` dataset](https://huggingface.co/datasets/Graphcore/vqa), you can download the 
+same data manually from the official [VQA dataset page](https://visualqa.org/download.html). If you prefer to follow the 
+tutorial with your custom data, check out how to [Create an image dataset](https://huggingface.co/docs/datasets/image_dataset#loading-script)
+guide in the 🤗 Datasets documentation.  
+
+Let's load the first 200 examples from the validation split and explore the dataset's features:  
+
+```python
+>>> from datasets import load_dataset
+
+>>> dataset = load_dataset("Graphcore/vqa", split="validation[:200]")
+>>> dataset
+Dataset({
+    features: ['question', 'question_type', 'question_id', 'image_id', 'answer_type', 'label'],
+    num_rows: 200
+})
+```
+
+Let's take a look at an example to understand the dataset's features:
+
+```py
+>>> dataset[0]
+{'question': 'Where is he looking?',
+ 'question_type': 'none of the above',
+ 'question_id': 262148000,
+ 'image_id': '/root/.cache/huggingface/datasets/downloads/extracted/ca733e0e000fb2d7a09fbcc94dbfe7b5a30750681d0e965f8e0a23b1c2f98c75/val2014/COCO_val2014_000000262148.jpg',
+ 'answer_type': 'other',
+ 'label': {'ids': ['at table', 'down', 'skateboard', 'table'],
+  'weights': [0.30000001192092896,
+   1.0,
+   0.30000001192092896,
+   0.30000001192092896]}}
+```
+
+The features relevant to the task include: 
+* `question`: the question to be answered from the image
+* `image_id`: the path to the image the question refers to
+* `label`: the annotations
+
+We can remove the rest of the features as they won't be necessary: 
+
+```py 
+>>> dataset = dataset.remove_columns(['question_type', 'question_id', 'answer_type'])
+```
+
+As you can see, the `label` feature contains several answers to the same question (called `ids` here) collected by different human annotators. 
+This is because the answer to a question can be subjective. In this case, the question is "where is he looking?". Some people 
+annotated this with "down", others with "at table", another one with "skateboard", etc. 
+
+Take a look at the image and consider which answer would you give:
+
+```python
+>>> from PIL import Image
+
+>>> image = Image.open(dataset[0]['image_id'])
+>>> image
+```
+
+<div class="flex justify-center">
+     <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/vqa-example.png" alt="VQA Image Example"/>
+</div>
+
+Due to the questions' and answers' ambiguity, datasets like this are treated as a multi-label classification problem (as 
+multiple answers are possibly valid). Moreover, rather than just creating a one-hot encoded vector, one creates a 
+soft encoding, based on the number of times a certain answer appeared in the annotations.
+
+For instance, in the example above, because the answer "down" is selected way more often than other answers, it has a 
+score (called `weight` in the dataset) of 1.0, and the rest of the answers have scores < 1.0. 
+
+To later instantiate the model with an appropriate classification head, let's create two dictionaries: one that maps 
+the label name to an integer and vice versa:
+
+```py
+>>> import itertools
+
+>>> labels = [item['ids'] for item in dataset['label']]
+>>> flattened_labels = list(itertools.chain(*labels))
+>>> unique_labels = list(set(flattened_labels))
+
+>>> label2id = {label: idx for idx, label in enumerate(unique_labels)}
+>>> id2label = {idx: label for label, idx in label2id.items()} 
+```
+
+Now that we have the mappings, we can replace the string answers with their ids, and flatten the dataset for a more convenient further preprocessing. 
+
+```python
+>>> def replace_ids(inputs):
+...   inputs["label"]["ids"] = [label2id[x] for x in inputs["label"]["ids"]]
+...   return inputs
+
+
+>>> dataset = dataset.map(replace_ids)
+>>> flat_dataset = dataset.flatten()
+>>> flat_dataset.features
+{'question': Value(dtype='string', id=None),
+ 'image_id': Value(dtype='string', id=None),
+ 'label.ids': Sequence(feature=Value(dtype='int64', id=None), length=-1, id=None),
+ 'label.weights': Sequence(feature=Value(dtype='float64', id=None), length=-1, id=None)}
+```
+
+## Preprocessing data
+
+The next step is to load a ViLT processor to prepare the image and text data for the model. 
+[`ViltProcessor`] wraps a BERT tokenizer and ViLT image processor into a convenient single processor:
+
+```py 
+>>> from transformers import ViltProcessor
+
+>>> processor = ViltProcessor.from_pretrained(model_checkpoint)
+```
+
+To preprocess the data we need to encode the images and questions using the [`ViltProcessor`]. The processor will use 
+the [`BertTokenizerFast`] to tokenize the text and create `input_ids`, `attention_mask` and `token_type_ids` for the text data. 
+As for images, the processor will leverage [`ViltImageProcessor`] to resize and normalize the image, and create `pixel_values` and `pixel_mask`.
+
+All these preprocessing steps are done under the hood, we only need to call the `processor`. However, we still need to 
+prepare the target labels. In this representation, each element corresponds to a possible answer (label). For correct answers, the element holds 
+their respective score (weight), while the remaining elements are set to zero.
+
+The following function applies the `processor` to the images and questions and formats the labels as described above:
+
+```py
+>>> import torch
+
+>>> def preprocess_data(examples):
+...     image_paths = examples['image_id']
+...     images = [Image.open(image_path) for image_path in image_paths]
+...     texts = examples['question']    
+
+...     encoding = processor(images, texts, padding="max_length", truncation=True, return_tensors="pt")
+
+...     for k, v in encoding.items():
+...           encoding[k] = v.squeeze()
+    
+...     targets = []
+
+...     for labels, scores in zip(examples['label.ids'], examples['label.weights']):
+...         target = torch.zeros(len(id2label))
+
+...         for label, score in zip(labels, scores):
+...             target[label] = score
+      
+...         targets.append(target)
+
+...     encoding["labels"] = targets
+    
+...     return encoding
+```
+
+To apply the preprocessing function over the entire dataset, use 🤗 Datasets [`~datasets.map`] function. You can speed up `map` by 
+setting `batched=True` to process multiple elements of the dataset at once. At this point, feel free to remove the columns you don't need.
+
+```py
+>>> processed_dataset = flat_dataset.map(preprocess_data, batched=True, remove_columns=['question','question_type',  'question_id', 'image_id', 'answer_type', 'label.ids', 'label.weights'])
+>>> processed_dataset
+Dataset({
+    features: ['input_ids', 'token_type_ids', 'attention_mask', 'pixel_values', 'pixel_mask', 'labels'],
+    num_rows: 200
+})
+```
+
+As a final step, create a batch of examples using [`DefaultDataCollator`]:
+
+```py
+>>> from transformers import DefaultDataCollator
+
+>>> data_collator = DefaultDataCollator()
+```
+
+## Train the model
+
+You’re ready to start training your model now! Load ViLT with [`ViltForQuestionAnswering`]. Specify the number of labels 
+along with the label mappings:
+
+```py
+>>> from transformers import ViltForQuestionAnswering
+
+>>> model = ViltForQuestionAnswering.from_pretrained(model_checkpoint, num_labels=len(id2label), id2label=id2label, label2id=label2id)
+```
+
+At this point, only three steps remain:
+
+1. Define your training hyperparameters in [`TrainingArguments`]:
+
+```py
+>>> from transformers import TrainingArguments
+
+>>> repo_id = "MariaK/vilt_finetuned_200"
+
+>>> training_args = TrainingArguments(
+...     output_dir=repo_id,
+...     per_device_train_batch_size=4,
+...     num_train_epochs=20,
+...     save_steps=200,
+...     logging_steps=50,
+...     learning_rate=5e-5,
+...     save_total_limit=2,
+...     remove_unused_columns=False,
+...     push_to_hub=True,
+... )
+```
+
+2. Pass the training arguments to [`Trainer`] along with the model, dataset, processor, and data collator.
+
+```py
+>>> from transformers import Trainer
+
+>>> trainer = Trainer(
+...     model=model,
+...     args=training_args,
+...     data_collator=data_collator,
+...     train_dataset=processed_dataset,
+...     tokenizer=processor,
+... )
+```
+
+3. Call [`~Trainer.train`] to finetune your model.
+
+```py
+>>> trainer.train() 
+```
+
+Once training is completed, share your model to the Hub with the [`~Trainer.push_to_hub`] method to share your final model on the 🤗 Hub:
+
+```py
+>>> trainer.push_to_hub()
+```
+
+## Inference
+
+Now that you have fine-tuned a ViLT model, and uploaded it to the 🤗 Hub, you can use it for inference. The simplest
+way to try out your fine-tuned model for inference is to use it in a [`Pipeline`].
+
+```py
+>>> from transformers import pipeline
+
+>>> pipe = pipeline("visual-question-answering", model="MariaK/vilt_finetuned_200")
+```
+
+The model in this guide has only been trained on 200 examples, so don't expect a lot from it. Let's see if it at least 
+learned something from the data and take the first example from the dataset to illustrate inference:
+
+```py
+>>> example = dataset[0]
+>>> image = Image.open(example['image_id'])
+>>> question = example['question']
+>>> print(question)
+>>> pipe(image, question, top_k=1)
+"Where is he looking?"
+[{'score': 0.5498199462890625, 'answer': 'down'}]
+```
+
+Even though not very confident, the model indeed has learned something. With more examples and longer training, you'll get far better results!
+
+You can also manually replicate the results of the pipeline if you'd like:
+1. Take an image and a question, prepare them for the model using the processor from your model.
+2. Forward the result or preprocessing through the model.
+3. From the logits, get the most likely answer's id, and find the actual answer in the `id2label`.
+
+```py
+>>> processor = ViltProcessor.from_pretrained("MariaK/vilt_finetuned_200")
+
+>>> image = Image.open(example['image_id'])
+>>> question = example['question']
+
+>>> # prepare inputs
+>>> inputs = processor(image, question, return_tensors="pt")
+
+>>> model = ViltForQuestionAnswering.from_pretrained("MariaK/vilt_finetuned_200")
+
+>>> # forward pass
+>>> with torch.no_grad():
+...     outputs = model(**inputs)
+
+>>> logits = outputs.logits
+>>> idx = logits.argmax(-1).item()
+>>> print("Predicted answer:", model.config.id2label[idx])
+Predicted answer: down
+```
+
+## Zero-shot VQA
+
+The previous model treated VQA as a classification task. Some recent models, such as BLIP, BLIP-2, and InstructBLIP approach 
+VQA as a generative task. Let's take [BLIP-2](../model_doc/blip-2) as an example. It introduced a new visual-language pre-training 
+paradigm in which any combination of pre-trained vision encoder and LLM can be used (learn more in the [BLIP-2 blog post](https://huggingface.co/blog/blip-2)). 
+This enables achieving state-of-the-art results on multiple visual-language tasks including visual question answering. 
+
+Let's illustrate how you can use this model for VQA. First, let's load the model. Here we'll explicitly send the model to a 
+GPU, if available, which we didn't need to do earlier when training, as [`Trainer`] handles this automatically: 
+
+```py
+>>> from transformers import AutoProcessor, Blip2ForConditionalGeneration
+>>> import torch
+
+>>> processor = AutoProcessor.from_pretrained("Salesforce/blip2-opt-2.7b")
+>>> model = Blip2ForConditionalGeneration.from_pretrained("Salesforce/blip2-opt-2.7b", torch_dtype=torch.float16)
+>>> device = "cuda" if torch.cuda.is_available() else "cpu"
+>>> model.to(device)
+```
+
+The model takes image and text as input, so let's use the exact same image/question pair from the first example in the VQA dataset: 
+
+```py 
+>>> example = dataset[0]
+>>> image = Image.open(example['image_id'])
+>>> question = example['question']
+```
+
+To use BLIP-2 for visual question answering task, the textual prompt has to follow a specific format: `Question: {} Answer:`.
+
+```py
+>>> prompt = f"Question: {question} Answer:" 
+```
+
+Now we need to preprocess the image/prompt with the model's processor, pass the processed input through the model, and decode the output:
+
+```py
+>>> inputs = processor(image, text=prompt, return_tensors="pt").to(device, torch.float16)
+
+>>> generated_ids = model.generate(**inputs, max_new_tokens=10)
+>>> generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
+>>> print(generated_text)
+"He is looking at the crowd" 
+```
+
+As you can see, the model recognized the crowd, and the direction of the face (looking down), however, it seems to miss 
+the fact the crowd is behind the skater. Still, in cases where acquiring human-annotated datasets is not feasible, this 
+approach can quickly produce useful results.
+ 

docs/source/en/testing.md

@@ -112,7 +112,7 @@ pytest tests/test_optimization.py --collect-only -q
 To run an individual test module:
 
 ```bash
-pytest tests/test_logging.py
+pytest tests/utils/test_logging.py
 ```
 
 ### Run specific tests
@@ -432,14 +432,14 @@ pytest --instafail
 On a GPU-enabled setup, to test in CPU-only mode add `CUDA_VISIBLE_DEVICES=""`:
 
 ```bash
-CUDA_VISIBLE_DEVICES="" pytest tests/test_logging.py
+CUDA_VISIBLE_DEVICES="" pytest tests/utils/test_logging.py
 ```
 
 or if you have multiple gpus, you can specify which one is to be used by `pytest`. For example, to use only the
 second gpu if you have gpus `0` and `1`, you can run:
 
 ```bash
-CUDA_VISIBLE_DEVICES="1" pytest tests/test_logging.py
+CUDA_VISIBLE_DEVICES="1" pytest tests/utils/test_logging.py
 ```
 
 This is handy when you want to run different tasks on different GPUs.
@@ -511,6 +511,21 @@ from transformers.testing_utils import get_gpu_count
 n_gpu = get_gpu_count()  # works with torch and tf
 ```
 
+### Testing with a specific PyTorch backend or device
+
+To run the test suite on a specific torch device add `TRANSFORMERS_TEST_DEVICE="$device"` where `$device` is the target backend. For example, to test on CPU only:
+```bash
+TRANSFORMERS_TEST_DEVICE="cpu" pytest tests/utils/test_logging.py
+```
+
+This variable is useful for testing custom or less common PyTorch backends such as `mps`. It can also be used to achieve the same effect as `CUDA_VISIBLE_DEVICES` by targeting specific GPUs or testing in CPU-only mode.
+
+Certain devices will require an additional import after importing `torch` for the first time. This can be specified using the environment variable `TRANSFORMERS_TEST_BACKEND`:
+```bash
+TRANSFORMERS_TEST_BACKEND="torch_npu" pytest tests/utils/test_logging.py
+```
+
+
 ### Distributed training
 
 `pytest` can't deal with distributed training directly. If this is attempted - the sub-processes don't do the right
@@ -538,7 +553,7 @@ according captured output will usually be shown along with the failure traceback
 To disable output capturing and to get the `stdout` and `stderr` normally, use `-s` or `--capture=no`:
 
 ```bash
-pytest -s tests/test_logging.py
+pytest -s tests/utils/test_logging.py
 ```
 
 To send test results to JUnit format output:
@@ -552,7 +567,7 @@ py.test tests --junitxml=result.xml
 To have no color (e.g., yellow on white background is not readable):
 
 ```bash
-pytest --color=no tests/test_logging.py
+pytest --color=no tests/utils/test_logging.py
 ```
 
 ### Sending test report to online pastebin service
@@ -560,7 +575,7 @@ pytest --color=no tests/test_logging.py
 Creating a URL for each test failure:
 
 ```bash
-pytest --pastebin=failed tests/test_logging.py
+pytest --pastebin=failed tests/utils/test_logging.py
 ```
 
 This will submit test run information to a remote Paste service and provide a URL for each failure. You may select
@@ -569,7 +584,7 @@ tests as usual or add for example -x if you only want to send one particular fai
 Creating a URL for a whole test session log:
 
 ```bash
-pytest --pastebin=all tests/test_logging.py
+pytest --pastebin=all tests/utils/test_logging.py
 ```
 
 ## Writing tests
@@ -1199,7 +1214,7 @@ tf.random.set_seed(seed)
 To start a debugger at the point of the warning, do this:
 
 ```bash
-pytest tests/test_logging.py -W error::UserWarning --pdb
+pytest tests/utils/test_logging.py -W error::UserWarning --pdb
 ```
 
 ## Working with github actions workflows

docs/source/en/tokenizer_summary.md

@@ -141,7 +141,7 @@ on.
 
 Byte-Pair Encoding (BPE) was introduced in [Neural Machine Translation of Rare Words with Subword Units (Sennrich et
 al., 2015)](https://arxiv.org/abs/1508.07909). BPE relies on a pre-tokenizer that splits the training data into
-words. Pretokenization can be as simple as space tokenization, e.g. [GPT-2](model_doc/gpt2), [Roberta](model_doc/roberta). More advanced pre-tokenization include rule-based tokenization, e.g. [XLM](model_doc/xlm),
+words. Pretokenization can be as simple as space tokenization, e.g. [GPT-2](model_doc/gpt2), [RoBERTa](model_doc/roberta). More advanced pre-tokenization include rule-based tokenization, e.g. [XLM](model_doc/xlm),
 [FlauBERT](model_doc/flaubert) which uses Moses for most languages, or [GPT](model_doc/gpt) which uses
 Spacy and ftfy, to count the frequency of each word in the training corpus.
 

docs/source/en/transformers_agents.md

@@ -14,11 +14,11 @@ rendered properly in your Markdown viewer.
 
 -->
 
-# Transformers Agent
+# Transformers Agents
 
 <Tip warning={true}>
 
-Transformers Agent is an experimental API which is subject to change at any time. Results returned by the agents
+Transformers Agents is an experimental API which is subject to change at any time. Results returned by the agents
 can vary as the APIs or underlying models are prone to change.
 
 </Tip>
@@ -206,25 +206,13 @@ This method can also take arguments if you would like to pass non-text types or
 
 ### ⚠️ Remote execution
 
-For demonstration purposes and so that this can be used with all setups, we have created remote executors for several 
-of the default tools the agent has access. These are created using 
-[inference endpoints](https://huggingface.co/inference-endpoints). To see how to set up remote executors tools yourself,
+For demonstration purposes and so that it could be used with all setups, we had created remote executors for several 
+of the default tools the agent has access for the release. These are created using 
+[inference endpoints](https://huggingface.co/inference-endpoints).
+
+We have turned these off for now, but in order to see how to set up remote executors tools yourself,
 we recommend reading the [custom tool guide](./custom_tools).
 
-In order to run with remote tools, specifying `remote=True` to either [`~Agent.run`] or [`~Agent.chat`] is sufficient.
-
-For example, the following command could be run on any device efficiently, without needing significant RAM or GPU:
-
-```py
-agent.run("Draw me a picture of rivers and lakes", remote=True)
-```
-
-The same can be said for [`~Agent.chat`]:
-
-```py
-agent.chat("Draw me a picture of rivers and lakes", remote=True)
-```
-
 ### What's happening here? What are tools, and what are agents?
 
 <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/diagram.png">

docs/source/es/quicktour.md

@@ -68,11 +68,13 @@ Instala las siguientes dependencias si aún no lo has hecho:
 
 <frameworkcontent>
 <pt>
+
 ```bash
 pip install torch
 ```
 </pt>
 <tf>
+
 ```bash
 pip install tensorflow
 ```
@@ -224,6 +226,7 @@ Como con el [`pipeline`], el tokenizador aceptará una lista de inputs. Además,
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> pt_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -235,6 +238,7 @@ Como con el [`pipeline`], el tokenizador aceptará una lista de inputs. Además,
 ```
 </pt>
 <tf>
+
 ```py
 >>> tf_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -377,6 +381,7 @@ Una característica particularmente interesante de 🤗 Transformers es la habil
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import AutoModel
 
@@ -385,6 +390,7 @@ Una característica particularmente interesante de 🤗 Transformers es la habil
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import TFAutoModel
 

docs/source/es/tasks/language_modeling.md

@@ -122,7 +122,7 @@ Así es como puedes crear una función de preprocesamiento para convertir la lis
 ...     return tokenizer([" ".join(x) for x in examples["answers.text"]], truncation=True)
 ```
 
-Usa de 🤗 Datasets la función [`map`](https://huggingface.co/docs/datasets/package_reference/main_classes.html#datasets.Dataset.map) para aplicar la función de preprocesamiento sobre el dataset en su totalidad. Puedes acelerar la función `map` configurando el argumento `batched=True` para procesar múltiples elementos del dataset a la vez y aumentar la cantidad de procesos con `num_proc`. Elimina las columnas que no necesitas:
+Usa de 🤗 Datasets la función [`map`](https://huggingface.co/docs/datasets/process#map) para aplicar la función de preprocesamiento sobre el dataset en su totalidad. Puedes acelerar la función `map` configurando el argumento `batched=True` para procesar múltiples elementos del dataset a la vez y aumentar la cantidad de procesos con `num_proc`. Elimina las columnas que no necesitas:
 
 ```py
 >>> tokenized_eli5 = eli5.map(

docs/source/fr/index.md

@@ -25,7 +25,7 @@ Apprentissage automatique de pointe pour [PyTorch](https://pytorch.org/), [Tenso
 🗣️ **Audio**: reconnaissance automatique de la parole et classification audio.<br>
 🐙 **Multimodalité**: système de question-réponse avec des tableaux ou images, reconnaissance optique de caractères, extraction d'information depuis des documents scannés et classification de vidéo.
 
-🤗 Transformers prend en charge l'interopérabilité entre PyTorch, TensorFlow et JAX. Cela permet d'utiliser un framework différent à chaque étape de la vie d'un modèle, par example entraîner un modèle en trois lignes de code avec un framework, et le charger pour l'inférence avec un autre. Les modèles peuvent également être exportés dans un format comme ONNX et TorchScript pour être déployés dans des environnements de production.
+🤗 Transformers prend en charge l'interopérabilité entre PyTorch, TensorFlow et JAX. Cela permet d'utiliser un framework différent à chaque étape de la vie d'un modèle, par exemple entraîner un modèle en trois lignes de code avec un framework, et le charger pour l'inférence avec un autre. Les modèles peuvent également être exportés dans un format comme ONNX et TorchScript pour être déployés dans des environnements de production.
 
 Rejoignez la communauté grandissante sur le [Hub](https://huggingface.co/models), le [forum](https://discuss.huggingface.co/) ou [Discord](https://discord.com/invite/JfAtkvEtRb) dès aujourd'hui !
 
@@ -407,4 +407,4 @@ Le tableau ci-dessous représente la prise en charge actuelle dans la bibliothè
 |             YOLOS             |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             YOSO              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 
-<!-- End table-->
+<!-- End table-->

docs/source/fr/quicktour.md

@@ -30,11 +30,13 @@ Vous aurez aussi besoin d'installer votre bibliothèque d'apprentissage profond
 
 <frameworkcontent>
 <pt>
+
 ```bash
 pip install torch
 ```
 </pt>
 <tf>
+
 ```bash
 pip install tensorflow
 ```
@@ -58,7 +60,7 @@ Le [`pipeline`] est le moyen le plus simple d'utiliser un modèle pré-entraîn
 | Traduction                  | Traduit du texte d'un langage à un autre                                                                      | Texte                | pipeline(task="translation")                  |
 | Classification d'image       | Attribue une catégorie à une image                                                                           | Image                | pipeline(task="image-classification")         |
 | Segmentation d'image           | Attribue une catégorie à chaque pixel d'une image (supporte la segmentation sémantique, panoptique et d'instance) | Image                | pipeline(task="image-segmentation")           |
-| Détection d'objects             | Prédit les délimitations et catégories d'objects dans une image                                                | Image                | pipeline(task="object-detection")             |
+| Détection d'objets             | Prédit les délimitations et catégories d'objets dans une image                                                | Image                | pipeline(task="object-detection")             |
 | Classification d'audio       | Attribue une catégorie à un fichier audio                                                                    | Audio                | pipeline(task="audio-classification")         |
 | Reconnaissance automatique de la parole | Extrait le discours d'un fichier audio en texte                                                                  | Audio                | pipeline(task="automatic-speech-recognition") |
 | Question réponse visuels    | Etant données une image et une question, répond correctement à une question sur l'image                                   | Modalités multiples  | pipeline(task="vqa")                          |
@@ -97,7 +99,7 @@ Le [`pipeline`] peut aussi itérer sur un jeu de données entier pour n'importe
 >>> speech_recognizer = pipeline("automatic-speech-recognition", model="facebook/wav2vec2-base-960h")
 ```
 
-Chargez un jeu de données audio (voir le 🤗 Datasets [Quick Start](https://huggingface.co/docs/datasets/quickstart#audio) pour plus de détails) sur lequel vous souhaitez itérer. Pour cet example, nous chargons le jeu de données [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) :
+Chargez un jeu de données audio (voir le 🤗 Datasets [Quick Start](https://huggingface.co/docs/datasets/quickstart#audio) pour plus de détails) sur lequel vous souhaitez itérer. Pour cet exemple, nous chargeons le jeu de données [MInDS-14](https://huggingface.co/datasets/PolyAI/minds14) :
 
 ```py
 >>> from datasets import load_dataset, Audio
@@ -153,7 +155,7 @@ Utilisez [`TFAutoModelForSequenceClassification`] et [`AutoTokenizer`] pour char
 </tf>
 </frameworkcontent>
 
-Specifiez le modèle et le tokenizer dans le [`pipeline`], et utilisez le `classifier` sur le texte en français :
+Spécifiez le modèle et le tokenizer dans le [`pipeline`], et utilisez le `classifier` sur le texte en français :
 
 ```py
 >>> classifier = pipeline("sentiment-analysis", model=model, tokenizer=tokenizer)
@@ -203,6 +205,7 @@ Un tokenizer peut également accepter une liste de textes, et remplir et tronque
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> pt_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -214,6 +217,7 @@ Un tokenizer peut également accepter une liste de textes, et remplir et tronque
 ```
 </pt>
 <tf>
+
 ```py
 >>> tf_batch = tokenizer(
 ...     ["We are very happy to show you the 🤗 Transformers library.", "We hope you don't hate it."],
@@ -346,6 +350,7 @@ Une fonctionnalité particulièrement cool 🤗 Transformers est la possibilité
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import AutoModel
 
@@ -354,6 +359,7 @@ Une fonctionnalité particulièrement cool 🤗 Transformers est la possibilité
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import TFAutoModel
 
@@ -412,7 +418,7 @@ En fonction de votre tâche, vous passerez généralement les paramètres suivan
    >>> model = AutoModelForSequenceClassification.from_pretrained("distilbert-base-uncased")
    ```
 
-2. [`TrainingArguments`] contient les hyperparamètres du modèle que vous pouvez changer comme le taux d'apprentissage, la taille due l'échantillon, et le nombre d'époques pour s'entraîner. Les valeurs par défaut sont utilisées si vous ne spécifiez pas d'hyperparamètres d'apprentissage :
+2. [`TrainingArguments`] contient les hyperparamètres du modèle que vous pouvez changer comme le taux d'apprentissage, la taille de l'échantillon, et le nombre d'époques pour s'entraîner. Les valeurs par défaut sont utilisées si vous ne spécifiez pas d'hyperparamètres d'apprentissage :
 
    ```py
    >>> from transformers import TrainingArguments
@@ -541,4 +547,4 @@ Tous les modèles sont des modèles standard [`tf.keras.Model`](https://www.tens
 
 ## Et après ?
 
-Maintenant que vous avez terminé la visite rapide de 🤗 Transformers, consultez nos guides et apprenez à faire des choses plus spécifiques comme créer un modèle personnalisé, finetuner un modèle pour une tâche, et comment entraîner un modèle avec un script. Si vous souhaitez en savoir plus sur les concepts fondamentaux de 🤗 Transformers, jetez un œil à nos guides conceptuels !
+Maintenant que vous avez terminé la visite rapide de 🤗 Transformers, consultez nos guides et apprenez à faire des choses plus spécifiques comme créer un modèle personnalisé, finetuner un modèle pour une tâche, et comment entraîner un modèle avec un script. Si vous souhaitez en savoir plus sur les concepts fondamentaux de 🤗 Transformers, jetez un œil à nos guides conceptuels !

docs/source/it/perf_infer_gpu_one.md

@@ -32,12 +32,12 @@ Nota che questa funzione può essere utilizzata anche nelle configurazioni multi
 Dal paper [`LLM.int8() : 8-bit Matrix Multiplication for Transformers at Scale`](https://arxiv.org/abs/2208.07339), noi supportiamo l'integrazione di Hugging Face per tutti i modelli dell'Hub con poche righe di codice.
 Il metodo `nn.Linear` riduce la dimensione di 2 per i pesi `float16` e `bfloat16` e di 4 per i pesi `float32`, con un impatto quasi nullo sulla qualità, operando sugli outlier in half-precision.
 
-![HFxbitsandbytes.png](https://s3.amazonaws.com/moonup/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png)
+![HFxbitsandbytes.png](https://cdn-uploads.huggingface.co/production/uploads/1659861207959-62441d1d9fdefb55a0b7d12c.png)
 
 Il metodo Int8 mixed-precision matrix decomposition funziona separando la moltiplicazione tra matrici in due flussi: (1) una matrice di flusso di outlier di caratteristiche sistematiche moltiplicata in fp16, (2) in flusso regolare di moltiplicazione di matrici int8 (99,9%). Con questo metodo, è possibile effettutare inferenza int8 per modelli molto grandi senza degrado predittivo.
 Per maggiori dettagli sul metodo, consultare il [paper](https://arxiv.org/abs/2208.07339) o il nostro [blogpost sull'integrazione](https://huggingface.co/blog/hf-bitsandbytes-integration).
 
-![MixedInt8.gif](https://s3.amazonaws.com/moonup/production/uploads/1660567469965-62441d1d9fdefb55a0b7d12c.gif)
+![MixedInt8.gif](https://cdn-uploads.huggingface.co/production/uploads/1660567469965-62441d1d9fdefb55a0b7d12c.gif)
 
 Nota che è necessaria una GPU per eseguire modelli di tipo mixed-8bit, poiché i kernel sono stati compilati solo per le GPU. Prima di utilizzare questa funzione, assicurarsi di disporre di memoria sufficiente sulla GPU per memorizzare un quarto del modello (o la metà se i pesi del modello sono in mezza precisione).
 Di seguito sono riportate alcune note per aiutarvi a utilizzare questo modulo, oppure seguite le dimostrazioni su [Google colab](#colab-demos).

docs/source/it/quicktour.md

@@ -68,11 +68,13 @@ Installa le seguenti dipendenze se non lo hai già fatto:
 
 <frameworkcontent>
 <pt>
+
 ```bash
 pip install torch
 ```
 </pt>
 <tf>
+
 ```bash
 pip install tensorflow
 ```
@@ -379,6 +381,7 @@ Una caratteristica particolarmente interessante di 🤗 Transformers è la sua a
 
 <frameworkcontent>
 <pt>
+
 ```py
 >>> from transformers import AutoModel
 
@@ -387,6 +390,7 @@ Una caratteristica particolarmente interessante di 🤗 Transformers è la sua a
 ```
 </pt>
 <tf>
+
 ```py
 >>> from transformers import TFAutoModel
 

docs/source/ko/_toctree.yml

@@ -19,8 +19,14 @@
     title: 스크립트로 학습하기
   - local: accelerate
     title: 🤗 Accelerate로 분산 학습 구성하기
+  - local: peft
+    title: 🤗 PEFT로 어댑터 로드 및 학습하기
   - local: model_sharing
     title: 만든 모델 공유하기
+  - local: transformers_agents
+    title: 에이전트
+  - local: llm_tutorial
+    title: 대규모 언어 모델로 생성하기
   title: 튜토리얼
 - sections:
   - sections:
@@ -71,6 +77,8 @@
         title: 이미지 캡셔닝
       - local: tasks/document_question_answering
         title: 문서 질의 응답(Document Question Answering)
+      - local: tasks/visual_question_answering
+        title: 시각적 질의응답 (Visual Question Answering)
     title: 멀티모달
     isExpanded: false
   title: 태스크 가이드
@@ -97,8 +105,8 @@
       title: (번역중) Benchmarks
     - local: in_translation
       title: (번역중) Notebooks with examples
-    - local: in_translation
-      title: (번역중) Community resources
+    - local: community
+      title: 커뮤니티 리소스
     - local: custom_tools
       title: 사용자 정의 도구와 프롬프트
     - local: troubleshooting
@@ -113,20 +121,20 @@
       title: (번역중) Training on many GPUs
     - local: perf_train_cpu
       title: CPU에서 훈련
-    - local: in_translation
-      title: (번역중) Training on many CPUs
+    - local: perf_train_cpu_many
+      title: 다중 CPU에서 훈련하기
     - local: in_translation
       title: (번역중) Training on TPUs
-    - local: in_translation
-      title: (번역중) Training on TPU with TensorFlow
+    - local: perf_train_tpu_tf
+      title: TensorFlow로 TPU에서 훈련하기
     - local: in_translation
       title: (번역중) Training on Specialized Hardware
     - local: perf_infer_cpu
       title: CPU로 추론하기
-    - local: in_translation
-      title: (번역중) Inference on one GPU
-    - local: in_translation
-      title: (번역중) Inference on many GPUs
+    - local: perf_infer_gpu_one
+      title: 하나의 GPU를 활용한 추론
+    - local: perf_infer_gpu_many
+      title: 다중 GPU에서 추론
     - local: in_translation
       title: (번역중) Inference on Specialized Hardware
     - local: perf_hardware
@@ -141,31 +149,31 @@
       title: TensorFlow 모델을 위한 XLA 통합
   title: (번역중) 성능 및 확장성
 - sections:
-    - local: in_translation
-      title: (번역중) How to contribute to transformers?
-    - local: in_translation
-      title: (번역중) How to add a model to 🤗 Transformers?
-    - local: in_translation
-      title: (번역중) How to convert a 🤗 Transformers model to TensorFlow?
-    - local: in_translation
-      title: (번역중) How to add a pipeline to 🤗 Transformers?
+    - local: contributing
+      title: 🤗 Transformers에 기여하는 방법
+    - local: add_new_model
+      title: 🤗 Transformers에 새로운 모델을 추가하는 방법 
+    - local: add_tensorflow_model
+      title: 어떻게 🤗 Transformers 모델을 TensorFlow로 변환하나요?
+    - local: add_new_pipeline
+      title: 어떻게 🤗 Transformers에 파이프라인을 추가하나요?
     - local: testing
       title: 테스트
-    - local: in_translation
-      title: (번역중) Checks on a Pull Request
+    - local: pr_checks
+      title: Pull Request에 대한 검사
   title: (번역중) 기여하기
 
 - sections:
-  - local: in_translation
-    title: (번역중) Philosophy
+  - local: philosophy
+    title: 이념과 목표
   - local: in_translation
     title: (번역중) Glossary
   - local: task_summary
     title: 🤗 Transformers로 할 수 있는 작업
   - local: tasks_explained
     title: 🤗 Transformers로 작업을 해결하는 방법
-  - local: in_translation
-    title: (번역중) The Transformer model family
+  - local: model_summary
+    title: Transformer 모델군
   - local: in_translation
     title: (번역중) Summary of the tokenizers
   - local: attention
@@ -178,6 +186,8 @@
     title: 고정 길이 모델의 펄플렉서티(Perplexity)
   - local: pipeline_webserver
     title: 추론 웹 서버를 위한 파이프라인
+  - local: model_memory_anatomy
+    title: 모델 학습 해부하기
   title: (번역중) 개념 가이드
 - sections:
   - sections:
@@ -327,8 +337,10 @@
         title: (번역중) Jukebox
       - local: in_translation
         title: (번역중) LED
-      - local: in_translation
-        title: (번역중) LLaMA
+      - local: model_doc/llama
+        title: LLaMA
+      - local: model_doc/llama2
+        title: LLaMA2
       - local: in_translation
         title: (번역중) Longformer
       - local: in_translation

docs/source/ko/add_new_model.md

@@ -0,0 +1,630 @@
+<!--Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# Hugging Face Transformers를 추가하는 방법은 무엇인가요? [[how-to-add-a-model-to-transformers]]
+
+Hugging Face Transformers 라이브러리는 커뮤니티 기여자들 덕분에 새로운 모델을 제공할 수 있는 경우가 많습니다. 하지만 이는 도전적인 프로젝트이며 Hugging Face Transformers 라이브러리와 구현할 모델에 대한 깊은 이해가 필요합니다. Hugging Face에서는 더 많은 커뮤니티 멤버가 모델을 적극적으로 추가할 수 있도록 지원하고자 하며, 이 가이드를 통해 PyTorch 모델을 추가하는 과정을 안내하고 있습니다 (PyTorch가 설치되어 있는지 확인해주세요).
+
+<Tip>
+
+TensorFlow 모델을 구현하고자 하는 경우 [🤗 Transformers 모델을 TensorFlow로 변환하는 방법](add_tensorflow_model) 가이드를 살펴보세요!
+
+</Tip>
+
+이 과정을 진행하면 다음과 같은 내용을 이해하게 됩니다:
+
+- 오픈 소스의 모범 사례에 대한 통찰력을 얻습니다.
+- 가장 인기 있는 딥러닝 라이브러리의 설계 원칙을 이해합니다.
+- 대규모 모델을 효율적으로 테스트하는 방법을 배웁니다.
+- `black`, `ruff`, `make fix-copies`와 같은 Python 유틸리티를 통합하여 깔끔하고 가독성 있는 코드를 작성하는 방법을 배웁니다.
+
+Hugging Face 팀은 항상 도움을 줄 준비가 되어 있으므로 혼자가 아니라는 점을 기억하세요. 🤗 ❤️
+
+시작에 앞서 🤗 Transformers에 원하는 모델을 추가하기 위해 [New model addition](https://github.com/huggingface/transformers/issues/new?assignees=&labels=New+model&template=new-model-addition.yml) 이슈를 열어야 합니다. 특정 모델을 기여하는 데 특별히 까다로운 기준을 가지지 않는 경우 [New model label](https://github.com/huggingface/transformers/labels/New%20model)을 필터링하여 요청되지 않은 모델이 있는지 확인하고 작업할 수 있습니다.
+
+새로운 모델 요청을 열었다면 첫 번째 단계는 🤗 Transformers에 익숙해지는 것입니다!
+
+## 🤗 Transformers의 전반적인 개요  [[general-overview-of-transformers]]
+
+먼저 🤗 Transformers에 대한 전반적인 개요를 파악해야 합니다. 🤗 Transformers는 매우 주관적인 라이브러리이기 때문에 해당 라이브러리의 철학이나 설계 선택 사항에 동의하지 않을 수도 있습니다. 그러나 우리의 경험상 라이브러리의 기본적인 설계 선택과 철학은 🤗 Transformers의 규모를 효율적으로 확장하면서 유지 보수 비용을 합리적인 수준으로 유지하는 것입니다.
+
+[라이브러리의 철학에 대한 문서](philosophy)를 읽는 것이 라이브러리를 더 잘 이해하는 좋은 시작점입니다. 모든 모델에 적용하려는 몇 가지 작업 방식에 대한 선택 사항이 있습니다:
+
+- 일반적으로 추상화보다는 구성을 선호합니다.
+- 코드를 복제하는 것이 항상 나쁜 것은 아닙니다. 코드의 가독성이나 접근성을 크게 향상시킨다면 복제하는 것은 좋습니다.
+- 모델 파일은 가능한 한 독립적으로 유지되어야 합니다. 따라서 특정 모델의 코드를 읽을 때 해당 `modeling_....py` 파일만 확인하면 됩니다.
+
+우리는 라이브러리의 코드가 제품을 제공하는 수단뿐만 아니라 개선하고자 하는 제품이라고도 생각합니다. 따라서 모델을 추가할 때, 사용자는 모델을 사용할 사람뿐만 아니라 코드를 읽고 이해하고 필요한 경우 조정할 수 있는 모든 사람까지도 포함한다는 점을 기억해야 합니다.
+
+이를 염두에 두고 일반적인 라이브러리 설계에 대해 조금 더 자세히 알아보겠습니다.
+
+### 모델 개요 [[overview-of-models]]
+
+모델을 성공적으로 추가하려면 모델과 해당 구성인 [`PreTrainedModel`] 및 [`PretrainedConfig`] 간의 상호작용을 이해하는 것이 중요합니다. 예를 들어, 🤗 Transformers에 추가하려는 모델을 `BrandNewBert`라고 부르겠습니다.
+
+다음을 살펴보겠습니다:
+
+<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers_overview.png"/>
+
+보다시피, 🤗 Transformers에서는 상속을 사용하지만 추상화 수준을 최소한으로 유지합니다. 라이브러리의 어떤 모델에서도 두 수준 이상의 추상화가 존재하지 않습니다. `BrandNewBertModel`은 `BrandNewBertPreTrainedModel`에서 상속받고, 이 클래스는 [`PreTrainedModel`]에서 상속받습니다. 이로써 새로운 모델은 [`PreTrainedModel`]에만 의존하도록 하려고 합니다. 모든 새로운 모델에 자동으로 제공되는 중요한 기능은 [`~PreTrainedModel.from_pretrained`] 및 [`~PreTrainedModel.save_pretrained`]입니다. 이러한 기능 외에도 `BrandNewBertModel.forward`와 같은 다른 중요한 기능은 새로운 `modeling_brand_new_bert.py` 스크립트에서 완전히 정의되어야 합니다. 또한 `BrandNewBertForMaskedLM`과 같은 특정 헤드 레이어를 가진 모델은 `BrandNewBertModel`을 상속받지 않고 forward pass에서 호출할 수 있는 `BrandNewBertModel`을 사용하여 추상화 수준을 낮게 유지합니다. 모든 새로운 모델은 `BrandNewBertConfig`라는 구성 클래스를 필요로 합니다. 이 구성은 항상 [`PreTrainedModel`]의 속성으로 저장되며, 따라서 `BrandNewBertPreTrainedModel`을 상속받는 모든 클래스에서 `config` 속성을 통해 액세스할 수 있습니다:
+
+```python
+model = BrandNewBertModel.from_pretrained("brandy/brand_new_bert")
+model.config  # model has access to its config
+```
+
+모델과 마찬가지로 구성은 [`PretrainedConfig`]에서 기본 직렬화 및 역직렬화 기능을 상속받습니다. 구성과 모델은 항상 *pytorch_model.bin* 파일과 *config.json* 파일로 각각 별도로 직렬화됩니다. [`~PreTrainedModel.save_pretrained`]를 호출하면 자동으로 [`~PretrainedConfig.save_pretrained`]도 호출되므로 모델과 구성이 모두 저장됩니다.
+
+
+### 코드 스타일 [[code-style]]
+
+새로운 모델을 작성할 때, Transformers는 주관적인 라이브러리이며 몇 가지 독특한 코딩 스타일이 있습니다:
+
+1. 모델의 forward pass는 모델 파일에 완전히 작성되어야 합니다. 라이브러리의 다른 모델에서 블록을 재사용하려면 코드를 복사하여 위에 `# Copied from` 주석과 함께 붙여넣으면 됩니다 (예: [여기](https://github.com/huggingface/transformers/blob/v4.17.0/src/transformers/models/roberta/modeling_roberta.py#L160)를 참조하세요).
+2. 코드는 완전히 이해하기 쉬워야 합니다. 변수 이름을 명확하게 지정하고 약어를 사용하지 않는 것이 좋습니다. 예를 들어, `act`보다는 `activation`을 선호합니다. 한 글자 변수 이름은 루프의 인덱스인 경우를 제외하고 권장되지 않습니다.
+3. 더 일반적으로, 짧은 마법 같은 코드보다는 길고 명시적인 코드를 선호합니다.
+4. PyTorch에서 `nn.Sequential`을 하위 클래스로 만들지 말고 `nn.Module`을 하위 클래스로 만들고 forward pass를 작성하여 다른 사람이 코드를 빠르게 디버그할 수 있도록 합니다. print 문이나 중단점을 추가할 수 있습니다.
+5. 함수 시그니처에는 타입 주석을 사용해야 합니다. 그 외에는 타입 주석보다 변수 이름이 훨씬 읽기 쉽고 이해하기 쉽습니다.
+
+### 토크나이저 개요 [[overview-of-tokenizers]]
+ 
+아직 준비되지 않았습니다 :-( 이 섹션은 곧 추가될 예정입니다!
+
+## 🤗 Transformers에 모델 추가하는 단계별 방법  [[stepbystep-recipe-to-add-a-model-to-transformers]]
+
+각자 모델을 이식하는 방법에 대한 선호가 다르기 때문에 다른 기여자들이 Hugging Face에 모델을 이식하는 방법에 대한 요약을 살펴보는 것이 매우 유용할 수 있습니다. 다음은 모델을 이식하는 방법에 대한 커뮤니티 블로그 게시물 목록입니다:
+
+1. [GPT2 모델 이식하기](https://medium.com/huggingface/from-tensorflow-to-pytorch-265f40ef2a28) - [Thomas](https://huggingface.co/thomwolf)
+2. [WMT19 MT 모델 이식하기](https://huggingface.co/blog/porting-fsmt) - [Stas](https://huggingface.co/stas)
+
+경험상 모델을 추가할 때 주의해야 할 가장 중요한 사항은 다음과 같습니다:
+
+-  같은 일을 반복하지 마세요! 새로운 🤗 Transformers 모델을 위해 추가할 코드의 대부분은 이미 🤗 Transformers 어딘가에 존재합니다. 이미 존재하는 복사할 수 있는 유사한 모델과 토크나이저를 찾는데 시간을 투자하세요. [grep](https://www.gnu.org/software/grep/)와 [rg](https://github.com/BurntSushi/ripgrep)를 참고하세요. 모델의 토크나이저가 한 모델을 기반으로 하고 모델링 코드가 다른 모델을 기반으로 하는 경우가 존재할 수도 있습니다. 예를 들어 FSMT의 모델링 코드는 BART를 기반으로 하고 FSMT의 토크나이저 코드는 XLM을 기반으로 합니다.
+-  이것은 과학적인 도전보다는 공학적인 도전입니다. 논문의 모델의 모든 이론적 측면을 이해하려는 것보다 효율적인 디버깅 환경을 만드는 데 더 많은 시간을 소비해야 합니다.
+-  막힐 때 도움을 요청하세요! 모델은 🤗 Transformers의 핵심 구성 요소이므로 Hugging Face의 우리는 당신이 모델을 추가하는 각 단계에서 기꺼이 도움을 줄 준비가 되어 있습니다. 진전이 없다고 느끼면 주저하지 말고 도움을 요청하세요.
+
+다음에서는 모델을 🤗 Transformers로 이식하는 데 가장 유용한 일반적인 절차를 제공하려고 노력합니다.
+
+다음 목록은 모델을 추가하는 데 수행해야 할 모든 작업의 요약이며 To-Do 목록으로 사용할 수 있습니다:
+
+☐ (선택 사항) BrandNewBert의 이론적 측면 이해<br>
+☐ Hugging Face 개발 환경 준비<br>
+☐ 원본 리포지토리의 디버깅 환경 설정<br>
+☐ 원본 리포지토리와 체크포인트를 사용하여 `forward()` pass가 성공적으로 실행되는 스크립트 작성<br>
+☐ 🤗 Transformers에 모델 스켈레톤 성공적으로 추가<br>
+☐ 원본 체크포인트를 🤗 Transformers 체크포인트로 성공적으로 변환<br>
+☐ 🤗 Transformers에서 원본 체크포인트와 동일한 출력을 내주는 `forward()` pass 성공적으로 실행<br>
+☐ 🤗 Transformers에서 모델 테스트 완료<br>
+☐ 🤗 Transformers에 토크나이저 성공적으로 추가<br>
+☐ 종단 간 통합 테스트 실행<br>
+☐ 문서 작성 완료<br>
+☐ 모델 가중치를 허브에 업로드<br>
+☐ Pull request 제출<br>
+☐ (선택 사항) 데모 노트북 추가
+
+우선, 일반적으로는 `BrandNewBert`의 이론적인 이해로 시작하는 것을 권장합니다. 그러나 이론적 측면을 직접 이해하는 대신 *직접 해보면서* 모델의 이론적 측면을 이해하는 것을 선호하는 경우 바로 `BrandNewBert` 코드 베이스로 빠져드는 것도 괜찮습니다. 이 옵션은 엔지니어링 기술이 이론적 기술보다 더 뛰어난 경우, `BrandNewBert`의 논문을 이해하는 데 어려움이 있는 경우, 또는 과학적인 논문을 읽는 것보다 프로그래밍에 훨씬 더 흥미 있는 경우에 더 적합할 수 있습니다.
+
+### 1. (선택 사항) BrandNewBert의 이론적 측면 [[1-optional-theoretical-aspects-of-brandnewbert]]
+
+만약 그런 서술적인 작업이 존재한다면, *BrandNewBert*의 논문을 읽어보는 시간을 가져야 합니다. 이해하기 어려운 섹션이 많을 수 있습니다. 그렇더라도 걱정하지 마세요! 목표는 논문의 깊은 이론적 이해가 아니라 *BrandNewBert*를 🤗 Transformers에서 효과적으로 재구현하기 위해 필요한 정보를 추출하는 것입니다. 이를 위해 이론적 측면에 너무 많은 시간을 투자할 필요는 없지만 다음과 같은 실제적인 측면에 집중해야 합니다:
+
+- *BrandNewBert*는 어떤 유형의 모델인가요? BERT와 유사한 인코더 모델인가요? GPT2와 유사한 디코더 모델인가요? BART와 유사한 인코더-디코더 모델인가요? 이들 간의 차이점에 익숙하지 않은 경우[model_summary](model_summary)를 참조하세요.
+- *BrandNewBert*의 응용 분야는 무엇인가요? 텍스트 분류인가요? 텍스트 생성인가요? 요약과 같은 Seq2Seq 작업인가요?
+- *brand_new_bert*와 BERT/GPT-2/BART의 차이점은 무엇인가요?
+- *brand_new_bert*와 가장 유사한 [🤗 Transformers 모델](https://huggingface.co/transformers/#contents)은 무엇인가요?
+- 어떤 종류의 토크나이저가 사용되나요? Sentencepiece 토크나이저인가요? Word piece 토크나이저인가요? BERT 또는 BART에 사용되는 동일한 토크나이저인가요?
+
+모델의 아키텍처에 대해 충분히 이해했다는 생각이 든 후, 궁금한 사항이 있으면 Hugging Face 팀에 문의하십시오. 이는 모델의 아키텍처, 어텐션 레이어 등에 관한 질문을 포함할 수 있습니다. Hugging Face의 유지 관리자들은 보통 코드를 검토하는 것에 대해 매우 기뻐하므로 당신을 돕는 일을 매우 환영할 것입니다!
+
+### 2. 개발 환경 설정 [[2-next-prepare-your-environment]]
+
+1. 저장소 페이지에서 "Fork" 버튼을 클릭하여 저장소의 사본을 GitHub 사용자 계정으로 만듭니다.
+
+2. `transformers` fork를 로컬 디스크에 클론하고 베이스 저장소를 원격 저장소로 추가합니다:
+
+```bash
+git clone https://github.com/[your Github handle]/transformers.git
+cd transformers
+git remote add upstream https://github.com/huggingface/transformers.git
+```
+
+3. 개발 환경을 설정합니다. 다음 명령을 실행하여 개발 환경을 설정할 수 있습니다:
+
+```bash
+python -m venv .env
+source .env/bin/activate
+pip install -e ".[dev]"
+```
+
+각 운영 체제에 따라 Transformers의 선택적 의존성이 개수가 증가하면 이 명령이 실패할 수 있습니다. 그런 경우에는 작업 중인 딥 러닝 프레임워크 (PyTorch, TensorFlow 및/또는 Flax)을 설치한 후, 다음 명령을 수행하면 됩니다:
+
+```bash
+pip install -e ".[quality]"
+```
+
+대부분의 경우에는 이것으로 충분합니다. 그런 다음 상위 디렉토리로 돌아갑니다.
+
+```bash
+cd ..
+```
+
+4. Transformers에 *brand_new_bert*의 PyTorch 버전을 추가하는 것을 권장합니다. PyTorch를 설치하려면 다음 링크의 지침을 따르십시오: https://pytorch.org/get-started/locally/.
+
+**참고:** CUDA를 설치할 필요는 없습니다. 새로운 모델이 CPU에서 작동하도록 만드는 것으로 충분합니다.
+
+5. *brand_new_bert*를 이식하기 위해서는 해당 원본 저장소에 접근할 수 있어야 합니다:
+
+```bash
+git clone https://github.com/org_that_created_brand_new_bert_org/brand_new_bert.git
+cd brand_new_bert
+pip install -e .
+```
+
+이제 *brand_new_bert*를 🤗 Transformers로 이식하기 위한 개발 환경을 설정하였습니다.
+
+### 3.-4. 원본 저장소에서 사전 훈련된 체크포인트 실행하기 [[3.-4.-run-a-pretrained-checkpoint-using-the-original-repository]]
+
+먼저, 원본 *brand_new_bert* 저장소에서 작업을 시작합니다. 원본 구현은 보통 "연구용"으로 많이 사용됩니다. 즉, 문서화가 부족하고 코드가 이해하기 어려울 수 있습니다. 그러나 이것이 바로 *brand_new_bert*를 다시 구현하려는 동기가 되어야 합니다. Hugging Face에서의 주요 목표 중 하나는 **거인의 어깨 위에 서는 것**이며, 이는 여기에서 쉽게 해석되어 동작하는 모델을 가져와서 가능한 한 **접근 가능하고 사용자 친화적이며 아름답게** 만드는 것입니다. 이것은 🤗 Transformers에서 모델을 다시 구현하는 가장 중요한 동기입니다 - 새로운 복잡한 NLP 기술을 **모두에게** 접근 가능하게 만드는 것을 목표로 합니다.
+
+따라서 원본 저장소에 대해 자세히 살펴보는 것으로 시작해야 합니다.
+
+원본 저장소에서 공식 사전 훈련된 모델을 성공적으로 실행하는 것은 종종 **가장 어려운** 단계입니다. 우리의 경험에 따르면, 원본 코드 베이스에 익숙해지는 데 시간을 투자하는 것이 매우 중요합니다. 다음을 파악해야 합니다:
+
+- 사전 훈련된 가중치를 어디서 찾을 수 있는지?
+- 사전 훈련된 가중치를 해당 모델에로드하는 방법은?
+- 모델과 독립적으로 토크나이저를 실행하는 방법은?
+- 간단한 forward pass에 필요한 클래스와 함수를 파악하기 위해 forward pass를 한 번 추적해 보세요. 일반적으로 해당 함수들만 다시 구현하면 됩니다.
+- 모델의 중요한 구성 요소를 찾을 수 있어야 합니다. 모델 클래스는 어디에 있나요? 모델 하위 클래스(*EncoderModel*, *DecoderModel* 등)가 있나요? self-attention 레이어는 어디에 있나요? self-attention, cross-attention 등 여러 가지 다른 어텐션 레이어가 있나요?
+- 원본 환경에서 모델을 디버그할 수 있는 방법은 무엇인가요? *print* 문을 추가해야 하나요? *ipdb*와 같은 대화식 디버거를 사용할 수 있나요? PyCharm과 같은 효율적인 IDE를 사용해 모델을 디버그할 수 있나요?
+
+원본 저장소에서 코드를 이식하는 작업을 시작하기 전에 원본 저장소에서 코드를 **효율적으로** 디버그할 수 있어야 합니다! 또한, 오픈 소스 라이브러리로 작업하고 있다는 것을 기억해야 합니다. 따라서 원본 저장소에서 issue를 열거나 pull request를 열기를 주저하지 마십시오. 이 저장소의 유지 관리자들은 누군가가 자신들의 코드를 살펴본다는 것에 대해 매우 기뻐할 것입니다!
+
+현재 시점에서, 원래 모델을 디버깅하기 위해 어떤 디버깅 환경과 전략을 선호하는지는 당신에게 달렸습니다. 우리는 고가의 GPU 환경을 구축하는 것은 비추천합니다. 대신, 원래 저장소로 들어가서 작업을 시작할 때와 🤗 Transformers 모델의 구현을 시작할 때에도 CPU에서 작업하는 것이 좋습니다. 모델이 이미 🤗 Transformers로 성공적으로 이식되었을 때에만 모델이 GPU에서도 예상대로 작동하는지 확인해야합니다.
+
+일반적으로, 원래 모델을 실행하기 위한 두 가지 가능한 디버깅 환경이 있습니다.
+
+- [Jupyter 노트북](https://jupyter.org/) / [Google Colab](https://colab.research.google.com/notebooks/intro.ipynb)
+- 로컬 Python 스크립트
+
+Jupyter 노트북의 장점은 셀 단위로 실행할 수 있다는 것입니다. 이는 논리적인 구성 요소를 더 잘 분리하고 중간 결과를 저장할 수 있으므로 디버깅 사이클이 더 빨라질 수 있습니다. 또한, 노트북은 다른 기여자와 쉽게 공유할 수 있으므로 Hugging Face 팀의 도움을 요청하려는 경우 매우 유용할 수 있습니다. Jupyter 노트북에 익숙하다면 이를 사용하는 것을 강력히 추천합니다.
+
+Jupyter 노트북의 단점은 사용에 익숙하지 않은 경우 새로운 프로그래밍 환경에 적응하는 데 시간을 할애해야 하며, `ipdb`와 같은 알려진 디버깅 도구를 더 이상 사용할 수 없을 수도 있다는 것입니다.
+
+각 코드 베이스에 대해 좋은 첫 번째 단계는 항상 **작은** 사전 훈련된 체크포인트를 로드하고 더미 정수 벡터 입력을 사용하여 단일 forward pass를 재현하는 것입니다. 이와 같은 스크립트는 다음과 같을 수 있습니다(의사 코드로 작성):
+
+```python
+model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
+input_ids = [0, 4, 5, 2, 3, 7, 9]  # vector of input ids
+original_output = model.predict(input_ids)
+```
+
+다음으로, 디버깅 전략에 대해 일반적으로 다음과 같은 몇 가지 선택지가 있습니다:
+
+- 원본 모델을 많은 작은 테스트 가능한 구성 요소로 분해하고 각각에 대해 forward pass를 실행하여 검증합니다.
+- 원본 모델을 원본 *tokenizer*과 원본 *model*로만 분해하고 해당 부분에 대해 forward pass를 실행한 후 검증을 위해 중간 출력(print 문 또는 중단점)을 사용합니다.
+
+다시 말하지만, 어떤 전략을 선택할지는 당신에게 달려 있습니다. 원본 코드 베이스에 따라 하나 또는 다른 전략이 유리할 수 있습니다.
+
+원본 코드 베이스를 모델의 작은 하위 구성 요소로 분해할 수 있는지 여부, 예를 들어 원본 코드 베이스가 즉시 실행 모드에서 간단히 실행될 수 있는 경우, 그런 경우에는 그 노력이 가치가 있다는 것이 일반적입니다. 초기에 더 어려운 방법을 선택하는 것에는 몇 가지 중요한 장점이 있습니다.
+
+- 원본 모델을 🤗 Transformers 구현과 비교할 때 각 구성 요소가 일치하는지 자동으로 확인할 수 있습니다. 즉, 시각적인 비교(print 문을 통한 비교가 아닌) 대신 🤗 Transformers 구현과 그에 대응하는 원본 구성 요소가 일치하는지 확인할 수 있습니다.
+- 전체 모델을 모듈별로, 즉 작은 구성 요소로 분해함으로써 모델을 이식하는 큰 문제를 단순히 개별 구성 요소를 이식하는 작은 문제로 분해할 수 있으므로 작업을 더 잘 구조화할 수 있습니다.
+- 모델을 논리적으로 의미 있는 구성 요소로 분리하는 것은 모델의 설계에 대한 더 나은 개요를 얻고 모델을 더 잘 이해하는 데 도움이 됩니다.
+- 이러한 구성 요소별 테스트를 통해 코드를 변경하면서 회귀가 발생하지 않도록 보장할 수 있습니다.
+
+[Lysandre의 ELECTRA 통합 검사](https://gist.github.com/LysandreJik/db4c948f6b4483960de5cbac598ad4ed)는 이를 수행하는 좋은 예제입니다.
+
+그러나 원본 코드 베이스가 매우 복잡하거나 중간 구성 요소를 컴파일된 모드에서 실행하는 것만 허용하는 경우, 모델을 테스트 가능한 작은 하위 구성 요소로 분해하는 것이 시간이 많이 소요되거나 불가능할 수도 있습니다. [T5의 MeshTensorFlow](https://github.com/tensorflow/mesh/tree/master/mesh_tensorflow) 라이브러리는 매우 복잡하며 모델을 하위 구성 요소로 분해하는 간단한 방법을 제공하지 않습니다. 이러한 라이브러리의 경우, 보통 print 문을 통해 확인합니다.
+
+어떤 전략을 선택하더라도 권장되는 절차는 동일합니다. 먼저 시작 레이어를 디버그하고 마지막 레이어를 마지막에 디버그하는 것이 좋습니다.
+
+다음 순서로 각 레이어의 출력을 검색하는 것이 좋습니다:
+
+1. 모델에 전달된 입력 ID 가져오기
+2. 워드 임베딩 가져오기
+3. 첫 번째 Transformer 레이어의 입력 가져오기
+4. 첫 번째 Transformer 레이어의 출력 가져오기
+5. 다음 n-1개의 Transformer 레이어의 출력 가져오기
+6. BrandNewBert 모델의 출력 가져오기
+
+입력 ID는 정수 배열로 구성되며, 예를 들어 `input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]`와 같을 수 있습니다.
+
+다음 레이어의 출력은 종종 다차원 실수 배열로 구성되며, 다음과 같이 나타낼 수 있습니다:
+
+```
+[[
+ [-0.1465, -0.6501,  0.1993,  ...,  0.1451,  0.3430,  0.6024],
+ [-0.4417, -0.5920,  0.3450,  ..., -0.3062,  0.6182,  0.7132],
+ [-0.5009, -0.7122,  0.4548,  ..., -0.3662,  0.6091,  0.7648],
+ ...,
+ [-0.5613, -0.6332,  0.4324,  ..., -0.3792,  0.7372,  0.9288],
+ [-0.5416, -0.6345,  0.4180,  ..., -0.3564,  0.6992,  0.9191],
+ [-0.5334, -0.6403,  0.4271,  ..., -0.3339,  0.6533,  0.8694]]],
+```
+
+🤗 Transformers에 추가되는 모든 모델은 통합 테스트를 통과해야 합니다. 즉, 원본 모델과 🤗 Transformers의 재구현 버전이 0.001의 정밀도로 정확히 동일한 출력을 내야 합니다! 동일한 모델이 다른 라이브러리에서 작성되었을 때 라이브러리 프레임워크에 따라 약간 다른 출력을 얻는 것은 정상이므로 1e-3(0.001)의 오차는 허용합니다. 거의 동일한 출력을 내는 것만으로는 충분하지 않으며, 완벽히 일치하는 수준이어야 합니다. 따라서 🤗 Transformers 버전의 중간 출력을 *brand_new_bert*의 원래 구현의 중간 출력과 여러 번 비교해야 합니다. 이 경우 원본 저장소의 **효율적인** 디버깅 환경이 절대적으로 중요합니다. 디버깅 환경을 가능한 한 효율적으로 만드는 몇 가지 조언을 제시합니다.
+
+- 중간 결과를 디버그하는 가장 좋은 방법을 찾으세요. 원본 저장소가 PyTorch로 작성되었다면 원본 모델을 더 작은 하위 구성 요소로 분해하여 중간 값을 검색하는 긴 스크립트를 작성하는 것에 시간을 투자할 가치가 있습니다. 원본 저장소가 Tensorflow 1로 작성되었다면 [tf.print](https://www.tensorflow.org/api_docs/python/tf/print)와 같은 Tensorflow 출력 작업을 사용하여 중간 값을 출력해야 할 수도 있습니다. 원본 저장소가 Jax로 작성되었다면 forward pass를 실행할 때 모델이 **jit 되지 않도록** 해야 합니다. 예를 들어 [이 링크](https://github.com/google/jax/issues/196)를 확인해 보세요.
+- 사용 가능한 가장 작은 사전 훈련된 체크포인트를 사용하세요. 체크포인트가 작을수록 디버그 사이클이 더 빨라집니다. 전반적으로 forward pass에 10초 이상이 걸리는 경우 효율적이지 않습니다. 매우 큰 체크포인트만 사용할 수 있는 경우, 새 환경에서 임의로 초기화된 가중치로 더미 모델을 만들고 해당 가중치를 🤗 Transformers 버전과 비교하기 위해 저장하는 것이 더 의미가 있을 수 있습니다.
+- 디버깅 설정에서 가장 쉽게 forward pass를 호출하는 방법을 사용하세요. 원본 저장소에서 **단일** forward pass만 호출하는 함수를 찾는 것이 이상적입니다. 이 함수는 일반적으로 `predict`, `evaluate`, `forward`, `__call__`과 같이 호출됩니다. `autoregressive_sample`과 같은 텍스트 생성에서 `forward`를 여러 번 호출하여 텍스트를 생성하는 등의 작업을 수행하는 함수를 디버그하고 싶지 않을 것입니다.
+- 토큰화 과정을 모델의 *forward* pass와 분리하려고 노력하세요. 원본 저장소에서 입력 문자열을 입력해야 하는 예제가 있는 경우, 입력 문자열이 입력 ID로 변경되는 순간을 찾아서 시작하세요. 이 경우 직접 ID를 입력할 수 있도록 작은 스크립트를 작성하거나 원본 코드를 수정해야 할 수도 있습니다.
+- 디버깅 설정에서 모델이 훈련 모드가 아니라는 것을 확인하세요. 훈련 모드에서는 모델의 여러 드롭아웃 레이어 때문에 무작위 출력이 생성될 수 있습니다. 디버깅 환경에서 forward pass가 **결정론적**이도록 해야 합니다. 또는 동일한 프레임워크에 있는 경우 *transformers.utils.set_seed*를 사용하세요.
+
+다음 섹션에서는 *brand_new_bert*에 대해 이 작업을 수행하는 데 더 구체적인 세부 사항/팁을 제공합니다.
+
+### 5.-14. 🤗 Transformers에 BrandNewBert를 이식하기 [[5.-14.-port-brandnewbert-to-transformers]]
+
+이제, 마침내 🤗 Transformers에 새로운 코드를 추가할 수 있습니다. 🤗 Transformers 포크의 클론으로 이동하세요:
+
+```bash
+cd transformers
+```
+
+다음과 같이 이미 존재하는 모델의 모델 아키텍처와 정확히 일치하는 모델을 추가하는 특별한 경우에는 [이 섹션](#write-a-conversion-script)에 설명된대로 변환 스크립트만 추가하면 됩니다. 이 경우에는 이미 존재하는 모델의 전체 모델 아키텍처를 그대로 재사용할 수 있습니다.
+
+그렇지 않으면 새로운 모델 생성을 시작합시다. 여기에서 두 가지 선택지가 있습니다:
+
+- `transformers-cli add-new-model-like`를 사용하여 기존 모델과 유사한 새로운 모델 추가하기
+- `transformers-cli add-new-model`을 사용하여 템플릿을 기반으로 한 새로운 모델 추가하기 (선택한 모델 유형에 따라 BERT 또는 Bart와 유사한 모습일 것입니다)
+
+두 경우 모두, 모델의 기본 정보를 입력하는 설문조사가 제시됩니다. 두 번째 명령어는 `cookiecutter`를 설치해야 합니다. 자세한 정보는 [여기](https://github.com/huggingface/transformers/tree/main/templates/adding_a_new_model)에서 확인할 수 있습니다.
+
+**huggingface/transformers 메인 저장소에 Pull Request 열기**
+
+자동으로 생성된 코드를 수정하기 전에, 지금은 "작업 진행 중 (WIP)" 풀 리퀘스트를 열기 위한 시기입니다. 예를 들어, 🤗 Transformers에 "*brand_new_bert* 추가"라는 제목의 "[WIP] Add *brand_new_bert*" 풀 리퀘스트를 엽니다. 이렇게 하면 당신과 Hugging Face 팀이 🤗 Transformers에 모델을 통합하는 작업을 함께할 수 있습니다.
+
+다음을 수행해야 합니다:
+
+1. 메인 브랜치에서 작업을 잘 설명하는 이름으로 브랜치 생성
+
+```bash
+git checkout -b add_brand_new_bert
+```
+
+2. 자동으로 생성된 코드 커밋
+
+```bash
+git add .
+git commit
+```
+
+3. 현재 메인을 가져오고 리베이스
+
+```bash
+git fetch upstream
+git rebase upstream/main
+```
+
+4. 변경 사항을 계정에 푸시
+
+```bash
+git push -u origin a-descriptive-name-for-my-changes
+```
+
+5. 만족스럽다면, GitHub에서 자신의 포크한 웹 페이지로 이동합니다. "Pull request"를 클릭합니다. Hugging Face 팀의 일부 멤버의 GitHub 핸들을 리뷰어로 추가하여 Hugging Face 팀이 앞으로의 변경 사항에 대해 알림을 받을 수 있도록 합니다.
+
+6. GitHub 풀 리퀘스트 웹 페이지 오른쪽에 있는 "Convert to draft"를 클릭하여 PR을 초안으로 변경합니다.
+
+다음으로, 어떤 진전을 이루었다면 작업을 커밋하고 계정에 푸시하여 풀 리퀘스트에 표시되도록 해야 합니다. 또한, 다음과 같이 현재 메인과 작업을 업데이트해야 합니다:
+
+```bash
+git fetch upstream
+git merge upstream/main
+```
+
+일반적으로, 모델 또는 구현에 관한 모든 질문은 자신의 PR에서 해야 하며, PR에서 토론되고 해결되어야 합니다. 이렇게 하면 Hugging Face 팀이 새로운 코드를 커밋하거나 질문을 할 때 항상 알림을 받을 수 있습니다. Hugging Face 팀에게 문제 또는 질문을 효율적으로 이해할 수 있도록 추가한 코드를 명시하는 것이 도움이 될 때가 많습니다.
+
+이를 위해, 변경 사항을 모두 볼 수 있는 "Files changed" 탭으로 이동하여 질문하고자 하는 줄로 이동한 다음 "+" 기호를 클릭하여 코멘트를 추가할 수 있습니다. 질문이나 문제가 해결되면, 생성된 코멘트의 "Resolve" 버튼을 클릭할 수 있습니다.
+
+마찬가지로, Hugging Face 팀은 코드를 리뷰할 때 코멘트를 남길 것입니다. 우리는 PR에서 대부분의 질문을 GitHub에서 묻는 것을 권장합니다. 공개에 크게 도움이 되지 않는 매우 일반적인 질문의 경우, Slack이나 이메일을 통해 Hugging Face 팀에게 문의할 수 있습니다.
+
+**5. brand_new_bert에 대해 생성된 모델 코드를 적용하기**
+
+먼저, 우리는 모델 자체에만 초점을 맞추고 토크나이저에 대해서는 신경 쓰지 않을 것입니다. 모든 관련 코드는 다음의 생성된 파일에서 찾을 수 있습니다: `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py` 및 `src/transformers/models/brand_new_bert/configuration_brand_new_bert.py`.
+
+이제 마침내 코딩을 시작할 수 있습니다 :). `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`의 생성된 코드는 인코더 전용 모델인 경우 BERT와 동일한 아키텍처를 가지거나, 인코더-디코더 모델인 경우 BART와 동일한 아키텍처를 가질 것입니다. 이 시점에서, 모델의 이론적 측면에 대해 배운 내용을 다시 상기해야 합니다: *모델이 BERT 또는 BART와 어떻게 다른가요?*. 자주 변경해야 하는 것은 *self-attention* 레이어, 정규화 레이어의 순서 등을 변경하는 것입니다. 다시 말하지만, 자신의 모델을 구현하는 데 도움이 되도록 Transformers에서 이미 존재하는 모델의 유사한 아키텍처를 살펴보는 것이 유용할 수 있습니다.
+
+**참고로** 이 시점에서, 코드가 완전히 정확하거나 깨끗하다고 확신할 필요는 없습니다. 오히려 처음에는 원본 코드의 첫 번째 *불완전하고* 복사된 버전을 `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`에 추가하는 것이 좋습니다. 필요한 모든 코드가 추가될 때까지 이러한 작업을 진행한 후, 다음 섹션에서 설명한 변환 스크립트를 사용하여 코드를 점진적으로 개선하고 수정하는 것이 훨씬 효율적입니다. 이 시점에서 작동해야 하는 유일한 것은 다음 명령이 작동하는 것입니다:
+
+```python
+from transformers import BrandNewBertModel, BrandNewBertConfig
+
+model = BrandNewBertModel(BrandNewBertConfig())
+```
+
+위의 명령은 `BrandNewBertConfig()`에 정의된 기본 매개변수에 따라 무작위 가중치로 모델을 생성하며, 이로써 모든 구성 요소의 `init()` 메서드가 작동함을 보장합니다.
+
+모든 무작위 초기화는 `BrandnewBertPreTrainedModel` 클래스의 `_init_weights` 메서드에서 수행되어야 합니다. 이 메서드는 구성 설정 변수에 따라 모든 리프 모듈을 초기화해야 합니다. BERT의 `_init_weights` 메서드 예제는 다음과 같습니다:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstance(module, nn.Linear):
+        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, nn.LayerNorm):
+        module.bias.data.zero_()
+        module.weight.data.fill_(1.0)
+```
+
+몇 가지 모듈에 대해 특별한 초기화가 필요한 경우 사용자 정의 방식을 사용할 수도 있습니다. 예를 들어, `Wav2Vec2ForPreTraining`에서 마지막 두 개의 선형 레이어는 일반적인 PyTorch `nn.Linear`의 초기화를 가져야 하지만, 다른 모든 레이어는 위와 같은 초기화를 사용해야 합니다. 이는 다음과 같이 코드화됩니다:
+
+```py
+def _init_weights(self, module):
+    """Initialize the weights"""
+    if isinstnace(module, Wav2Vec2ForPreTraining):
+        module.project_hid.reset_parameters()
+        module.project_q.reset_parameters()
+        module.project_hid._is_hf_initialized = True
+        module.project_q._is_hf_initialized = True
+    elif isinstance(module, nn.Linear):
+        module.weight.data.normal_(mean=0.0, std=self.config.initializer_range)
+        if module.bias is not None:
+            module.bias.data.zero_()
+```
+
+`_is_hf_initialized` 플래그는 서브모듈을 한 번만 초기화하도록 내부적으로 사용됩니다. `module.project_q` 및 `module.project_hid`에 대해 `True`로 설정함으로써, 우리가 수행한 사용자 정의 초기화가 이후에 덮어쓰이지 않도록 합니다. 즉, `_init_weights` 함수가 이들에게 적용되지 않습니다.
+
+**6. 변환 스크립트 작성하기**
+
+다음으로, 디버그에 사용한 체크포인트를 기존 저장소에서 만든 🤗 Transformers 구현과 호환되는 체크포인트로 변환할 수 있는 변환 스크립트를 작성해야 합니다. 변환 스크립트를 처음부터 작성하는 것보다는 *brand_new_bert*와 동일한 프레임워크로 작성된 유사한 모델을 변환한 기존 변환 스크립트를 찾아보는 것이 좋습니다. 일반적으로 기존 변환 스크립트를 복사하여 사용 사례에 맞게 약간 수정하는 것으로 충분합니다. 모델에 대해 유사한 기존 변환 스크립트를 어디에서 찾을 수 있는지 Hugging Face 팀에게 문의하는 것을 망설이지 마세요.
+
+- TensorFlow에서 PyTorch로 모델을 이전하는 경우, 좋은 참고 자료로 BERT의 변환 스크립트 [여기](https://github.com/huggingface/transformers/blob/7acfa95afb8194f8f9c1f4d2c6028224dbed35a2/src/transformers/models/bert/modeling_bert.py#L91)를 참조할 수 있습니다.
+- PyTorch에서 PyTorch로 모델을 이전하는 경우, 좋은 참고 자료로 BART의 변환 스크립트 [여기](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bart/convert_bart_original_pytorch_checkpoint_to_pytorch.py)를 참조할 수 있습니다.
+
+다음에서는 PyTorch 모델이 레이어 가중치를 저장하고 레이어 이름을 정의하는 방법에 대해 간단히 설명하겠습니다. PyTorch에서 레이어의 이름은 레이어에 지정한 클래스 속성의 이름으로 정의됩니다. 다음과 같이 PyTorch에서 `SimpleModel`이라는 더미 모델을 정의해 봅시다:
+
+```python
+from torch import nn
+
+
+class SimpleModel(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.dense = nn.Linear(10, 10)
+        self.intermediate = nn.Linear(10, 10)
+        self.layer_norm = nn.LayerNorm(10)
+```
+
+이제 이 모델 정의의 인스턴스를 생성할 수 있으며 `dense`, `intermediate`, `layer_norm` 등의 가중치가 랜덤하게 할당됩니다. 모델을 출력하여 아키텍처를 확인할 수 있습니다.
+
+```python
+model = SimpleModel()
+
+print(model)
+```
+
+이는 다음과 같이 출력됩니다:
+
+```
+SimpleModel(
+  (dense): Linear(in_features=10, out_features=10, bias=True)
+  (intermediate): Linear(in_features=10, out_features=10, bias=True)
+  (layer_norm): LayerNorm((10,), eps=1e-05, elementwise_affine=True)
+)
+```
+
+우리는 레이어의 이름이 PyTorch에서 클래스 속성의 이름으로 정의되어 있는 것을 볼 수 있습니다. 특정 레이어의 가중치 값을 출력하여 확인할 수 있습니다:
+
+```python
+print(model.dense.weight.data)
+```
+
+가중치가 무작위로 초기화되었음을 확인할 수 있습니다.
+
+```
+tensor([[-0.0818,  0.2207, -0.0749, -0.0030,  0.0045, -0.1569, -0.1598,  0.0212,
+         -0.2077,  0.2157],
+        [ 0.1044,  0.0201,  0.0990,  0.2482,  0.3116,  0.2509,  0.2866, -0.2190,
+          0.2166, -0.0212],
+        [-0.2000,  0.1107, -0.1999, -0.3119,  0.1559,  0.0993,  0.1776, -0.1950,
+         -0.1023, -0.0447],
+        [-0.0888, -0.1092,  0.2281,  0.0336,  0.1817, -0.0115,  0.2096,  0.1415,
+         -0.1876, -0.2467],
+        [ 0.2208, -0.2352, -0.1426, -0.2636, -0.2889, -0.2061, -0.2849, -0.0465,
+          0.2577,  0.0402],
+        [ 0.1502,  0.2465,  0.2566,  0.0693,  0.2352, -0.0530,  0.1859, -0.0604,
+          0.2132,  0.1680],
+        [ 0.1733, -0.2407, -0.1721,  0.1484,  0.0358, -0.0633, -0.0721, -0.0090,
+          0.2707, -0.2509],
+        [-0.1173,  0.1561,  0.2945,  0.0595, -0.1996,  0.2988, -0.0802,  0.0407,
+          0.1829, -0.1568],
+        [-0.1164, -0.2228, -0.0403,  0.0428,  0.1339,  0.0047,  0.1967,  0.2923,
+          0.0333, -0.0536],
+        [-0.1492, -0.1616,  0.1057,  0.1950, -0.2807, -0.2710, -0.1586,  0.0739,
+          0.2220,  0.2358]]).
+```
+
+변환 스크립트에서는 이러한 무작위로 초기화된 가중치를 체크포인트의 해당 레이어의 정확한 가중치로 채워야 합니다. 예를 들면 다음과 같습니다:
+
+```python
+# retrieve matching layer weights, e.g. by
+# recursive algorithm
+layer_name = "dense"
+pretrained_weight = array_of_dense_layer
+
+model_pointer = getattr(model, "dense")
+
+model_pointer.weight.data = torch.from_numpy(pretrained_weight)
+```
+
+이렇게 하면 PyTorch 모델의 무작위로 초기화된 각 가중치와 해당 체크포인트 가중치가 **모양과 이름** 모두에서 정확히 일치하는지 확인해야 합니다. 이를 위해 모양에 대한 assert 문을 추가하고 체크포인트 가중치의 이름을 출력해야 합니다. 예를 들어 다음과 같은 문장을 추가해야 합니다:
+
+```python
+assert (
+    model_pointer.weight.shape == pretrained_weight.shape
+), f"Pointer shape of random weight {model_pointer.shape} and array shape of checkpoint weight {pretrained_weight.shape} mismatched"
+```
+
+또한 두 가중치의 이름을 출력하여 일치하는지 확인해야 합니다. *예시*:
+
+```python
+logger.info(f"Initialize PyTorch weight {layer_name} from {pretrained_weight.name}")
+```
+
+모양 또는 이름이 일치하지 않는 경우, 랜덤으로 초기화된 레이어에 잘못된 체크포인트 가중치를 할당한 것으로 추측됩니다.
+
+잘못된 모양은 `BrandNewBertConfig()`의 구성 매개변수 설정이 변환하려는 체크포인트에 사용된 설정과 정확히 일치하지 않기 때문일 가능성이 가장 큽니다. 그러나 PyTorch의 레이어 구현 자체에서 가중치를 전치해야 할 수도 있습니다.
+
+마지막으로, **모든** 필요한 가중치가 초기화되었는지 확인하고 초기화에 사용되지 않은 모든 체크포인트 가중치를 출력하여 모델이 올바르게 변환되었는지 확인해야 합니다. 잘못된 모양 문장이나 잘못된 이름 할당으로 인해 변환 시도가 실패하는 것은 완전히 정상입니다. 이는 `BrandNewBertConfig()`에서 잘못된 매개변수를 사용하거나 🤗 Transformers 구현에서 잘못된 아키텍처, 🤗 Transformers 구현의 구성 요소 중 하나의 `init()` 함수에 버그가 있는 경우이거나 체크포인트 가중치 중 하나를 전치해야 하는 경우일 가능성이 가장 높습니다.
+
+이 단계는 이전 단계와 함께 반복되어야 하며 모든 체크포인트의 가중치가 Transformers 모델에 올바르게 로드되었을 때까지 계속되어야 합니다. 🤗 Transformers 구현에 체크포인트를 올바르게 로드한 후에는 `/path/to/converted/checkpoint/folder`와 같은 원하는 폴더에 모델을 저장할 수 있어야 합니다. 해당 폴더에는 `pytorch_model.bin` 파일과 `config.json` 파일이 모두 포함되어야 합니다.
+
+```python
+model.save_pretrained("/path/to/converted/checkpoint/folder")
+```
+
+**7. 순방향 패스 구현하기**
+
+🤗 Transformers 구현에 사전 훈련된 가중치를 정확하게 로드한 후에는 순방향 패스가 올바르게 구현되었는지 확인해야 합니다. [원본 저장소에 익숙해지기](#34-run-a-pretrained-checkpoint-using-the-original-repository)에서 이미 원본 저장소를 사용하여 모델의 순방향 패스를 실행하는 스크립트를 만들었습니다. 이제 원본 대신 🤗 Transformers 구현을 사용하는 유사한 스크립트를 작성해야 합니다. 다음과 같이 작성되어야 합니다:
+
+```python
+model = BrandNewBertModel.from_pretrained("/path/to/converted/checkpoint/folder")
+input_ids = [0, 4, 4, 3, 2, 4, 1, 7, 19]
+output = model(input_ids).last_hidden_states
+```
+
+🤗 Transformers 구현과 원본 모델 구현이 처음부터 정확히 동일한 출력을 제공하지 않거나 순방향 패스에서 오류가 발생할 가능성이 매우 높습니다. 실망하지 마세요. 예상된 일입니다! 먼저, 순방향 패스에서 오류가 발생하지 않도록 해야 합니다. 종종 잘못된 차원이 사용되어 *차원 불일치* 오류가 발생하거나 잘못된 데이터 유형 개체가 사용되는 경우가 있습니다. 예를 들면 `torch.long` 대신에 `torch.float32`가 사용된 경우입니다. 해결할 수 없는 오류가 발생하면 Hugging Face 팀에 도움을 요청하는 것이 좋습니다.
+
+🤗 Transformers 구현이 올바르게 작동하는지 확인하는 마지막 단계는 출력이 `1e-3`의 정밀도로 동일한지 확인하는 것입니다. 먼저, 출력 모양이 동일하도록 보장해야 합니다. 즉, 🤗 Transformers 구현 스크립트와 원본 구현 사이에서 `outputs.shape`는 동일한 값을 반환해야 합니다. 그 다음으로, 출력 값이 동일하도록 해야 합니다. 이는 새로운 모델을 추가할 때 가장 어려운 부분 중 하나입니다. 출력이 동일하지 않은 일반적인 실수 사례는 다음과 같습니다:
+
+- 일부 레이어가 추가되지 않았습니다. 즉, *활성화* 레이어가 추가되지 않았거나 잔차 연결이 빠졌습니다.
+- 단어 임베딩 행렬이 연결되지 않았습니다.
+- 잘못된 위치 임베딩이 사용되었습니다. 원본 구현에서는 오프셋을 사용합니다.
+- 순방향 패스 중에 Dropout이 적용되었습니다. 이를 수정하려면 *model.training이 False*인지 확인하고 순방향 패스 중에 Dropout 레이어가 잘못 활성화되지 않도록 하세요. 즉, [PyTorch의 기능적 Dropout](https://pytorch.org/docs/stable/nn.functional.html?highlight=dropout#torch.nn.functional.dropout)에 *self.training*을 전달하세요.
+
+문제를 해결하는 가장 좋은 방법은 일반적으로 원본 구현과 🤗 Transformers 구현의 순방향 패스를 나란히 놓고 차이점이 있는지 확인하는 것입니다. 이상적으로는 순방향 패스의 중간 출력을 디버그/출력하여 원본 구현과 🤗 Transformers 구현의 정확한 위치를 찾을 수 있어야 합니다. 먼저, 두 스크립트의 하드코딩된 `input_ids`가 동일한지 확인하세요. 다음으로, `input_ids`의 첫 번째 변환의 출력(일반적으로 단어 임베딩)이 동일한지 확인하세요. 그런 다음 네트워크의 가장 마지막 레이어까지 진행해보세요. 어느 시점에서 두 구현 사이에 차이가 있는 것을 알게 되는데, 이는 🤗 Transformers 구현의 버그 위치를 가리킬 것입니다. 저희 경험상으로는 원본 구현과 🤗 Transformers 구현 모두에서 동일한 위치에 많은 출력 문을 추가하고 이들의 중간 표현에 대해 동일한 값을 보이는 출력 문을 연속적으로 제거하는 것이 간단하고 효과적인 방법입니다.
+
+`torch.allclose(original_output, output, atol=1e-3)`로 출력을 확인하여 두 구현이 동일한 출력을 하는 것을 확신한다면, 가장 어려운 부분은 끝났습니다! 축하드립니다. 남은 작업은 쉬운 일이 될 것입니다 😊.
+
+**8. 필요한 모든 모델 테스트 추가하기**
+
+이 시점에서 새로운 모델을 성공적으로 추가했습니다. 그러나 해당 모델이 요구되는 디자인에 완전히 부합하지 않을 수도 있습니다. 🤗 Transformers와 완벽하게 호환되는 구현인지 확인하기 위해 모든 일반 테스트를 통과해야 합니다. Cookiecutter는 아마도 모델을 위한 테스트 파일을 자동으로 추가했을 것입니다. 아마도 `tests/models/brand_new_bert/test_modeling_brand_new_bert.py`와 같은 경로에 위치할 것입니다. 이 테스트 파일을 실행하여 일반 테스트가 모두 통과하는지 확인하세요.
+
+```bash
+pytest tests/models/brand_new_bert/test_modeling_brand_new_bert.py
+```
+
+모든 일반 테스트를 수정한 후, 이제 수행한 작업을 충분히 테스트하여 다음 사항을 보장해야 합니다.
+
+- a) 커뮤니티가 *brand_new_bert*의 특정 테스트를 살펴봄으로써 작업을 쉽게 이해할 수 있도록 함
+- b) 모델에 대한 향후 변경 사항이 모델의 중요한 기능을 손상시키지 않도록 함
+
+먼저 통합 테스트를 추가해야 합니다. 이러한 통합 테스트는 이전에 모델을 🤗 Transformers로 구현하기 위해 사용한 디버깅 스크립트와 동일한 작업을 수행합니다. Cookiecutter에 이미 이러한 모델 테스트의 템플릿인 `BrandNewBertModelIntegrationTests`가 추가되어 있으며, 여러분이 작성해야 할 내용으로만 채워 넣으면 됩니다. 이러한 테스트가 통과하는지 확인하려면 다음을 실행하세요.
+
+```bash
+RUN_SLOW=1 pytest -sv tests/models/brand_new_bert/test_modeling_brand_new_bert.py::BrandNewBertModelIntegrationTests
+```
+
+<Tip>
+
+Windows를 사용하는 경우 `RUN_SLOW=1`을 `SET RUN_SLOW=1`로 바꿔야 합니다.
+
+</Tip>
+
+둘째로, *brand_new_bert*에 특화된 모든 기능도 별도의 테스트에서 추가로 테스트해야 합니다. 이 부분은 종종 잊히는데, 두 가지 측면에서 굉장히 유용합니다.
+
+- *brand_new_bert*의 특수 기능이 어떻게 작동해야 하는지 보여줌으로써 커뮤니티에게 모델 추가 과정에서 습득한 지식을 전달하는 데 도움이 됩니다.
+- 향후 기여자는 이러한 특수 테스트를 실행하여 모델에 대한 변경 사항을 빠르게 테스트할 수 있습니다.
+
+
+**9. 토크나이저 구현하기**
+
+다음으로, *brand_new_bert*의 토크나이저를 추가해야 합니다. 보통 토크나이저는 🤗 Transformers의 기존 토크나이저와 동일하거나 매우 유사합니다.
+
+토크나이저가 올바르게 작동하는지 확인하기 위해 먼저 원본 리포지토리에서 문자열을 입력하고 `input_ids`를 반환하는 스크립트를 생성하는 것이 좋습니다. 다음과 같은 유사한 스크립트일 수 있습니다 (의사 코드로 작성):
+
+```python
+input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
+model = BrandNewBertModel.load_pretrained_checkpoint("/path/to/checkpoint/")
+input_ids = model.tokenize(input_str)
+```
+
+원본 리포지토리를 자세히 살펴보고 올바른 토크나이저 함수를 찾거나, 복제본에서 변경 사항을 적용하여 `input_ids`만 출력하도록 해야 합니다. 원본 리포지토리를 사용하는 기능적인 토큰화 스크립트를 작성한 후, 🤗 Transformers의 유사한 스크립트를 생성해야 합니다. 다음과 같이 작성되어야 합니다:
+
+```python
+from transformers import BrandNewBertTokenizer
+
+input_str = "This is a long example input string containing special characters .$?-, numbers 2872 234 12 and words."
+
+tokenizer = BrandNewBertTokenizer.from_pretrained("/path/to/tokenizer/folder/")
+
+input_ids = tokenizer(input_str).input_ids
+```
+
+두 개의 `input_ids`가 동일한 값을 반환할 때, 마지막 단계로 토크나이저 테스트 파일도 추가해야 합니다.
+
+*brand_new_bert*의 모델링 테스트 파일과 유사하게, *brand_new_bert*의 토크나이제이션 테스트 파일에는 몇 가지 하드코딩된 통합 테스트가 포함되어야 합니다.
+
+**10. 종단 간 통합 테스트 실행**
+
+토크나이저를 추가한 후에는 모델과 토크나이저를 사용하여 몇 가지 종단 간 통합 테스트를 추가해야 합니다. `tests/models/brand_new_bert/test_modeling_brand_new_bert.py`에 추가해주세요. 이러한 테스트는 🤗 Transformers 구현이 예상대로 작동하는지를 의미 있는 text-to-text 예시로 보여줘야 합니다. 그 예시로는 *예를 들어* source-to-target 번역 쌍, article-to-summary 쌍, question-to-answer 쌍 등이 포함될 수 있습니다. 불러온 체크포인트 중 어느 것도 다운스트림 작업에서 미세 조정되지 않았다면, 모델 테스트만으로 충분합니다. 모델이 완전히 기능을 갖추었는지 확인하기 위해 마지막 단계로 GPU에서 모든 테스트를 실행하는 것이 좋습니다. 모델의 내부 텐서의 일부에 `.to(self.device)` 문을 추가하는 것을 잊었을 수 있으며, 이 경우 테스트에서 오류로 표시됩니다. GPU에 액세스할 수 없는 경우, Hugging Face 팀이 테스트를 대신 실행할 수 있습니다.
+
+**11. 기술문서 추가**
+
+이제 *brand_new_bert*에 필요한 모든 기능이 추가되었습니다. 거의 끝났습니다! 추가해야 할 것은 멋진 기술문서과 기술문서 페이지입니다. Cookiecutter가 `docs/source/model_doc/brand_new_bert.md`라는 템플릿 파일을 추가해줬을 것입니다. 이 페이지를 사용하기 전에 모델을 사용하는 사용자들은 일반적으로 이 페이지를 먼저 확인합니다. 따라서 문서는 이해하기 쉽고 간결해야 합니다. 모델을 사용하는 방법을 보여주기 위해 *팁*을 추가하는 것이 커뮤니티에 매우 유용합니다. 독스트링에 관련하여 Hugging Face 팀에 문의하는 것을 주저하지 마세요.
+
+다음으로, `src/transformers/models/brand_new_bert/modeling_brand_new_bert.py`에 추가된 독스트링이 올바르며 필요한 모든 입력 및 출력을 포함하도록 확인하세요. [여기](writing-documentation)에서 우리의 문서 작성 가이드와 독스트링 형식에 대한 상세 가이드가 있습니다. 문서는 일반적으로 커뮤니티와 모델의 첫 번째 접점이기 때문에, 문서는 적어도 코드만큼의 주의를 기울여야 합니다.
+
+**코드 리팩토링**
+
+좋아요, 이제 *brand_new_bert*를 위한 모든 필요한 코드를 추가했습니다. 이 시점에서 다음을 실행하여 잠재적으로 잘못된 코드 스타일을 수정해야 합니다:
+
+그리고 코딩 스타일이 품질 점검을 통과하는지 확인하기 위해 다음을 실행하고 확인해야 합니다:
+
+```bash
+make style
+```
+
+🤗 Transformers에는 여전히 실패할 수 있는 몇 가지 매우 엄격한 디자인 테스트가 있습니다. 이는 독스트링에 누락된 정보나 잘못된 명명 때문에 종종 발생합니다. 여기서 막히면 Hugging Face 팀이 도움을 줄 것입니다.
+
+```bash
+make quality
+```
+
+마지막으로, 코드가 정확히 작동하는 것을 확인한 후에는 항상 코드를 리팩토링하는 것이 좋은 생각입니다. 모든 테스트가 통과된 지금은 추가한 코드를 다시 검토하고 리팩토링하는 좋은 시기입니다.
+
+이제 코딩 부분을 완료했습니다. 축하합니다! 🎉 멋져요! 😎
+
+**12. 모델을 모델 허브에 업로드하세요**
+
+이 마지막 파트에서는 모든 체크포인트를 변환하여 모델 허브에 업로드하고 각 업로드된 모델 체크포인트에 대한 모델 카드를 추가해야 합니다. [Model sharing and uploading Page](model_sharing)를 읽고 허브 기능에 익숙해지세요. *brand_new_bert*의 저자 조직 아래에 모델을 업로드할 수 있는 필요한 액세스 권한을 얻기 위해 Hugging Face 팀과 협업해야 합니다. `transformers`의 모든 모델에 있는 `push_to_hub` 메서드는 체크포인트를 허브에 빠르고 효율적으로 업로드하는 방법입니다. 아래에 작은 코드 조각이 붙여져 있습니다:
+
+각 체크포인트에 적합한 모델 카드를 만드는 데 시간을 할애하는 것은 가치가 있습니다. 모델 카드는 체크포인트의 특성을 강조해야 합니다. *예를 들어* 이 체크포인트는 어떤 데이터셋에서 사전 훈련/세부 훈련되었는지? 이 모델은 어떤 하위 작업에서 사용해야 하는지? 그리고 모델을 올바르게 사용하는 방법에 대한 몇 가지 코드도 포함해야 합니다.
+
+```python
+brand_new_bert.push_to_hub("brand_new_bert")
+# Uncomment the following line to push to an organization.
+# brand_new_bert.push_to_hub("<organization>/brand_new_bert")
+```
+
+**13. (선택 사항) 노트북 추가**
+
+*brand_new_bert*를 다운스트림 작업에서 추론 또는 미세 조정에 사용하는 방법을 자세히 보여주는 노트북을 추가하는 것이 매우 유용합니다. 이것은 PR을 병합하는 데 필수적이지는 않지만 커뮤니티에 매우 유용합니다.
+
+**14. 완료된 PR 제출**
+
+이제 프로그래밍을 마쳤으며, 마지막 단계로 PR을 메인 브랜치에 병합해야 합니다. 보통 Hugging Face 팀은 이미 여기까지 도움을 주었을 것입니다. 그러나 PR에 멋진 설명을 추가하고 리뷰어에게 특정 디자인 선택 사항을 강조하려면 완료된 PR에 약간의 설명을 추가하는 시간을 할애하는 것이 가치가 있습니다.
+
+### 작업물을 공유하세요!! [[share-your-work]]
+
+이제 커뮤니티에서 작업물을 인정받을 시간입니다! 모델 추가 작업을 완료하는 것은 Transformers와 전체 NLP 커뮤니티에 큰 기여입니다. 당신의 코드와 이식된 사전 훈련된 모델은 수백, 심지어 수천 명의 개발자와 연구원에 의해 확실히 사용될 것입니다. 당신의 작업에 자랑스러워해야 하며 이를 커뮤니티와 공유해야 합니다.
+
+**당신은 커뮤니티 내 모든 사람들에게 매우 쉽게 접근 가능한 또 다른 모델을 만들었습니다! 🤯**

docs/source/ko/add_new_pipeline.md

@@ -0,0 +1,248 @@
+<!--Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 어떻게 사용자 정의 파이프라인을 생성하나요? [[how-to-create-a-custom-pipeline]]
+
+이 가이드에서는 사용자 정의 파이프라인을 어떻게 생성하고 [허브](hf.co/models)에 공유하거나 🤗 Transformers 라이브러리에 추가하는 방법을 살펴보겠습니다.
+
+먼저 파이프라인이 수용할 수 있는 원시 입력을 결정해야 합니다.
+문자열, 원시 바이트, 딕셔너리 또는 가장 원하는 입력일 가능성이 높은 것이면 무엇이든 가능합니다.
+이 입력을 가능한 한 순수한 Python 형식으로 유지해야 (JSON을 통해 다른 언어와도) 호환성이 좋아집니다.
+이것이 전처리(`preprocess`) 파이프라인의 입력(`inputs`)이 될 것입니다.
+
+그런 다음 `outputs`를 정의하세요.
+`inputs`와 같은 정책을 따르고, 간단할수록 좋습니다.
+이것이 후처리(`postprocess`) 메소드의 출력이 될 것입니다.
+
+먼저 4개의 메소드(`preprocess`, `_forward`, `postprocess` 및 `_sanitize_parameters`)를 구현하기 위해 기본 클래스 `Pipeline`을 상속하여 시작합니다.
+
+
+```python
+from transformers import Pipeline
+
+
+class MyPipeline(Pipeline):
+    def _sanitize_parameters(self, **kwargs):
+        preprocess_kwargs = {}
+        if "maybe_arg" in kwargs:
+            preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
+        return preprocess_kwargs, {}, {}
+
+    def preprocess(self, inputs, maybe_arg=2):
+        model_input = Tensor(inputs["input_ids"])
+        return {"model_input": model_input}
+
+    def _forward(self, model_inputs):
+        # model_inputs == {"model_input": model_input}
+        outputs = self.model(**model_inputs)
+        # Maybe {"logits": Tensor(...)}
+        return outputs
+
+    def postprocess(self, model_outputs):
+        best_class = model_outputs["logits"].softmax(-1)
+        return best_class
+```
+
+이 분할 구조는 CPU/GPU에 대한 비교적 원활한 지원을 제공하는 동시에, 다른 스레드에서 CPU에 대한 사전/사후 처리를 수행할 수 있게 지원하는 것입니다.
+
+`preprocess`는 원래 정의된 입력을 가져와 모델에 공급할 수 있는 형식으로 변환합니다.
+더 많은 정보를 포함할 수 있으며 일반적으로 `Dict` 형태입니다.
+
+`_forward`는 구현 세부 사항이며 직접 호출할 수 없습니다. 
+`forward`는 예상 장치에서 모든 것이 작동하는지 확인하기 위한 안전장치가 포함되어 있어 선호되는 호출 메소드입니다.
+실제 모델과 관련된 것은 `_forward` 메소드에 속하며, 나머지는 전처리/후처리 과정에 있습니다.
+
+`postprocess` 메소드는 `_forward`의 출력을 가져와 이전에 결정한 최종 출력 형식으로 변환합니다.
+
+`_sanitize_parameters`는 초기화 시간에 `pipeline(...., maybe_arg=4)`이나 호출 시간에 `pipe = pipeline(...); output = pipe(...., maybe_arg=4)`과 같이, 사용자가 원하는 경우 언제든지 매개변수를 전달할 수 있도록 허용합니다.
+
+`_sanitize_parameters`의 반환 값은 `preprocess`, `_forward`, `postprocess`에 직접 전달되는 3개의 kwargs 딕셔너리입니다.
+호출자가 추가 매개변수로 호출하지 않았다면 아무것도 채우지 마십시오.
+이렇게 하면 항상 더 "자연스러운" 함수 정의의 기본 인수를 유지할 수 있습니다.
+
+분류 작업에서 `top_k` 매개변수가 대표적인 예입니다.
+
+```python
+>>> pipe = pipeline("my-new-task")
+>>> pipe("This is a test")
+[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}, {"label": "3-star", "score": 0.05}
+{"label": "4-star", "score": 0.025}, {"label": "5-star", "score": 0.025}]
+
+>>> pipe("This is a test", top_k=2)
+[{"label": "1-star", "score": 0.8}, {"label": "2-star", "score": 0.1}]
+```
+
+이를 달성하기 위해 우리는 `postprocess` 메소드를 기본 매개변수인 `5`로 업데이트하고 `_sanitize_parameters`를 수정하여 이 새 매개변수를 허용합니다.
+
+
+```python
+def postprocess(self, model_outputs, top_k=5):
+    best_class = model_outputs["logits"].softmax(-1)
+    # top_k를 처리하는 로직 추가
+    return best_class
+
+
+def _sanitize_parameters(self, **kwargs):
+    preprocess_kwargs = {}
+    if "maybe_arg" in kwargs:
+        preprocess_kwargs["maybe_arg"] = kwargs["maybe_arg"]
+
+    postprocess_kwargs = {}
+    if "top_k" in kwargs:
+        postprocess_kwargs["top_k"] = kwargs["top_k"]
+    return preprocess_kwargs, {}, postprocess_kwargs
+```
+
+입/출력을 가능한 한 간단하고 완전히 JSON 직렬화 가능한 형식으로 유지하려고 노력하십시오.
+이렇게 하면 사용자가 새로운 종류의 개체를 이해하지 않고도 파이프라인을 쉽게 사용할 수 있습니다.
+또한 사용 용이성을 위해 여러 가지 유형의 인수(오디오 파일은 파일 이름, URL 또는 순수한 바이트일 수 있음)를 지원하는 것이 비교적 일반적입니다.
+
+
+
+## 지원되는 작업 목록에 추가하기 [[adding-it-to-the-list-of-supported-tasks]]
+
+`new-task`를 지원되는 작업 목록에 등록하려면 `PIPELINE_REGISTRY`에 추가해야 합니다:
+
+```python
+from transformers.pipelines import PIPELINE_REGISTRY
+
+PIPELINE_REGISTRY.register_pipeline(
+    "new-task",
+    pipeline_class=MyPipeline,
+    pt_model=AutoModelForSequenceClassification,
+)
+```
+
+원하는 경우 기본 모델을 지정할 수 있으며, 이 경우 특정 개정(분기 이름 또는 커밋 해시일 수 있음, 여기서는 "abcdef")과 타입을 함께 가져와야 합니다:
+
+```python
+PIPELINE_REGISTRY.register_pipeline(
+    "new-task",
+    pipeline_class=MyPipeline,
+    pt_model=AutoModelForSequenceClassification,
+    default={"pt": ("user/awesome_model", "abcdef")},
+    type="text",  # 현재 지원 유형: text, audio, image, multimodal
+)
+```
+
+## Hub에 파이프라인 공유하기 [[share-your-pipeline-on-the-hub]]
+
+Hub에 사용자 정의 파이프라인을 공유하려면 `Pipeline` 하위 클래스의 사용자 정의 코드를 Python 파일에 저장하기만 하면 됩니다.
+예를 들어, 다음과 같이 문장 쌍 분류를 위한 사용자 정의 파이프라인을 사용한다고 가정해 보겠습니다:
+
+```py
+import numpy as np
+
+from transformers import Pipeline
+
+
+def softmax(outputs):
+    maxes = np.max(outputs, axis=-1, keepdims=True)
+    shifted_exp = np.exp(outputs - maxes)
+    return shifted_exp / shifted_exp.sum(axis=-1, keepdims=True)
+
+
+class PairClassificationPipeline(Pipeline):
+    def _sanitize_parameters(self, **kwargs):
+        preprocess_kwargs = {}
+        if "second_text" in kwargs:
+            preprocess_kwargs["second_text"] = kwargs["second_text"]
+        return preprocess_kwargs, {}, {}
+
+    def preprocess(self, text, second_text=None):
+        return self.tokenizer(text, text_pair=second_text, return_tensors=self.framework)
+
+    def _forward(self, model_inputs):
+        return self.model(**model_inputs)
+
+    def postprocess(self, model_outputs):
+        logits = model_outputs.logits[0].numpy()
+        probabilities = softmax(logits)
+
+        best_class = np.argmax(probabilities)
+        label = self.model.config.id2label[best_class]
+        score = probabilities[best_class].item()
+        logits = logits.tolist()
+        return {"label": label, "score": score, "logits": logits}
+```
+
+구현은 프레임워크에 구애받지 않으며, PyTorch와 TensorFlow 모델에 대해 작동합니다.
+이를 `pair_classification.py`라는 파일에 저장한 경우, 다음과 같이 가져오고 등록할 수 있습니다:
+
+```py
+from pair_classification import PairClassificationPipeline
+from transformers.pipelines import PIPELINE_REGISTRY
+from transformers import AutoModelForSequenceClassification, TFAutoModelForSequenceClassification
+
+PIPELINE_REGISTRY.register_pipeline(
+    "pair-classification",
+    pipeline_class=PairClassificationPipeline,
+    pt_model=AutoModelForSequenceClassification,
+    tf_model=TFAutoModelForSequenceClassification,
+)
+```
+
+이 작업이 완료되면 사전훈련된 모델과 함께 사용할 수 있습니다.
+예를 들어, `sgugger/finetuned-bert-mrpc`은 MRPC 데이터 세트에서 미세 조정되어 문장 쌍을 패러프레이즈인지 아닌지를 분류합니다.
+
+```py
+from transformers import pipeline
+
+classifier = pipeline("pair-classification", model="sgugger/finetuned-bert-mrpc")
+```
+
+그런 다음 `Repository`의 `save_pretrained` 메소드를 사용하여 허브에 공유할 수 있습니다:
+
+```py
+from huggingface_hub import Repository
+
+repo = Repository("test-dynamic-pipeline", clone_from="{your_username}/test-dynamic-pipeline")
+classifier.save_pretrained("test-dynamic-pipeline")
+repo.push_to_hub()
+```
+
+이렇게 하면 "test-dynamic-pipeline" 폴더 내에 `PairClassificationPipeline`을 정의한 파일이 복사되며, 파이프라인의 모델과 토크나이저도 저장한 후, `{your_username}/test-dynamic-pipeline` 저장소에 있는 모든 것을 푸시합니다.
+이후에는 `trust_remote_code=True` 옵션만 제공하면 누구나 사용할 수 있습니다.
+
+```py
+from transformers import pipeline
+
+classifier = pipeline(model="{your_username}/test-dynamic-pipeline", trust_remote_code=True)
+```
+
+## 🤗 Transformers에 파이프라인 추가하기 [[add-the-pipeline-to-transformers]]
+
+🤗 Transformers에 사용자 정의 파이프라인을 기여하려면, `pipelines` 하위 모듈에 사용자 정의 파이프라인 코드와 함께 새 모듈을 추가한 다음, `pipelines/__init__.py`에서 정의된 작업 목록에 추가해야 합니다.
+
+그런 다음 테스트를 추가해야 합니다.
+`tests/test_pipelines_MY_PIPELINE.py`라는 새 파일을 만들고 다른 테스트와 예제를 함께 작성합니다.
+
+`run_pipeline_test` 함수는 매우 일반적이며, `model_mapping` 및 `tf_model_mapping`에서 정의된 가능한 모든 아키텍처의 작은 무작위 모델에서 실행됩니다.
+
+이는 향후 호환성을 테스트하는 데 매우 중요하며, 누군가 `XXXForQuestionAnswering`을 위한 새 모델을 추가하면 파이프라인 테스트가 해당 모델에서 실행을 시도한다는 의미입니다.
+모델이 무작위이기 때문에 실제 값을 확인하는 것은 불가능하므로, 단순히 파이프라인 출력 `TYPE`과 일치시키기 위한 도우미 `ANY`가 있습니다.
+
+또한 2개(이상적으로는 4개)의 테스트를 구현해야 합니다.
+
+- `test_small_model_pt`: 이 파이프라인에 대한 작은 모델 1개를 정의(결과가 의미 없어도 상관없음)하고 파이프라인 출력을 테스트합니다.
+결과는 `test_small_model_tf`와 동일해야 합니다.
+- `test_small_model_tf`: 이 파이프라인에 대한 작은 모델 1개를 정의(결과가 의미 없어도 상관없음)하고 파이프라인 출력을 테스트합니다.
+결과는 `test_small_model_pt`와 동일해야 합니다.
+- `test_large_model_pt`(`선택사항`): 결과가 의미 있을 것으로 예상되는 실제 파이프라인에서 파이프라인을 테스트합니다.
+이러한 테스트는 속도가 느리므로 이를 표시해야 합니다.
+여기서의 목표는 파이프라인을 보여주고 향후 릴리즈에서의 변화가 없는지 확인하는 것입니다.
+- `test_large_model_tf`(`선택사항`): 결과가 의미 있을 것으로 예상되는 실제 파이프라인에서 파이프라인을 테스트합니다.
+이러한 테스트는 속도가 느리므로 이를 표시해야 합니다.
+여기서의 목표는 파이프라인을 보여주고 향후 릴리즈에서의 변화가 없는지 확인하는 것입니다.

docs/source/ko/add_tensorflow_model.md

@@ -0,0 +1,263 @@
+<!--Copyright 2022 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+
+⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+
+-->
+
+# 어떻게 🤗 Transformers 모델을 TensorFlow로 변환하나요? [[how-to-convert-a-transformers-model-to-tensorflow]]
+
+🤗 Transformers에서처럼 사용할 수 있는 여러 가지 프레임워크가 있다는 것은 애플리케이션을 설계할 때 그들의 강점을 유연하게 이용할 수 있다는 장점이 있지만, 모델 별로 호환성을 추가해야 한다는 단점 또한 존재한다는 것을 의미합니다. 좋은 소식은 기존 모델에 TensorFlow 호환성을 추가하는 것이 [처음부터 새로운 모델을 추가하는 것](add_new_model)보다도 간단하다는 것입니다! 
+
+만약 대규모 TensorFlow 모델을 더 깊이 이해하려거나, 오픈 소스에 큰 기여를 하려거나, 선택한 모델에 Tensorflow를 활용하려한다면, 이 안내서는 여러분께 도움이 될 것입니다.
+
+이 가이드는 Hugging Face 팀의 최소한의 감독 아래에서 🤗 Transformers에서 사용되는 TensorFlow 모델 가중치와/또는 아키텍처를 기여할 수 있는 커뮤니티 구성원인 여러분을 대상으로 합니다. 
+새로운 모델을 작성하는 것은 쉬운 일이 아니지만, 이 가이드를 통해 조금 덜 힘들고 훨씬 쉬운 작업으로 만들 수 있습니다. 
+모두의 경험을 모으는 것은 이 작업을 점차적으로 더 쉽게 만드는 데 굉장히 중요하기 때문에, 이 가이드를 개선시킬만한 제안이 떠오르면 공유하시는걸 적극적으로 권장합니다!
+
+더 깊이 알아보기 전에, 🤗 Transformers를 처음 접하는 경우 다음 자료를 확인하는 것이 좋습니다:
+- [🤗 Transformers의 일반 개요](add_new_model#general-overview-of-transformers)
+- [Hugging Face의 TensorFlow 철학](https://huggingface.co/blog/tensorflow-philosophy)
+
+이 가이드의 나머지 부분에서는 새로운 TensorFlow 모델 아키텍처를 추가하는 데 필요한 단계, Pytorch를 TensorFlow 모델 가중치로 변환하는 절차 및 ML 프레임워크 간의 불일치를 효율적으로 디버깅하는 방법을 알게 될 것입니다. 시작해봅시다!
+
+<Tip>
+
+사용하려는 모델이 이미 해당하는 TensorFlow 아키텍처가 있는지 확실하지 않나요?
+
+선택한 모델([예](https://huggingface.co/bert-base-uncased/blob/main/config.json#L14))의 `config.json`의 `model_type` 필드를 확인해보세요. 🤗 Transformers의 해당 모델 폴더에는 "modeling_tf"로 시작하는 파일이 있는 경우, 해당 모델에는 해당 TensorFlow 아키텍처([예](https://github.com/huggingface/transformers/tree/main/src/transformers/models/bert))가 있다는 의미입니다.
+
+</Tip>
+
+## TensorFlow 모델 아키텍처 코드 추가하는 단계별 가이드 [[step-by-step-guide-to add-tensorFlow-model-architecture-code]]
+
+대규모 아키텍처를 가진 모델을 설계하는 방법에는 여러가지가 있으며, 해당 설계를 구현하는 방법도 여러 가지입니다. 
+그러나 우리는 [🤗 Transformers 일반 개요](add_new_model#general-overview-of-transformers)에서 언급한 대로 일관된 설계 선택에 따라야지만 🤗 Transformers를 사용하기 편할 것이라는 확고한 의견을 가지고 있습니다.
+우리의 경험을 통해 TensorFlow 모델을 추가하는 데 관련된 중요한 몇 가지 사항을 알려 드릴 수 있습니다:
+
+- 이미 있는걸 다시 개발하려 하지 마세요! 최소한 2개의 이미 구현된 모델을 대개 참조해야 합니다. 구현하려는 모델과 기능상 동일한 Pytorch 모델 하나와 같은 문제 유형을 풀고 있는 다른 TensorFlow 모델 하나를 살펴보세요.
+- 우수한 모델 구현은 시간이 지나도 남아있습니다. 이것은 코드가 아름답다는 이유가 아니라 코드가 명확하고 디버깅 및 개선이 쉽기 때문입니다. TensorFlow 구현에서 다른 모델들과 패턴을 똑같이 하고 Pytorch 구현과의 불일치를 최소화하여 메인테이너의 업무를 쉽게 한다면, 기여한 코드가 오래도록 유지될 수 있습니다.
+- 필요하다면 도움을 요청하세요! 🤗 Transformers 팀은 여러분을 돕기 위해 있으며, 여러분이 직면한 동일한 문제에 대한 해결책을 이미 찾은 경우도 있을 수 있습니다.
+
+TensorFlow 모델 아키텍처를 추가하는 데 필요한 단계를 개략적으로 써보면:
+1. 변환하려는 모델 선택
+2. transformers 개발 환경 준비
+3. (선택 사항) 이론적 측면 및 기존 구현 이해
+4. 모델 아키텍처 구현
+5. 모델 테스트 구현
+6. PR (pull request) 제출
+7. (선택 사항) 데모 빌드 및 공유
+
+### 1.-3. 모델 기여 준비 [[1.-3.-prepare-your-model-contribution]]
+
+**1. 변환하려는 모델 선택**
+
+우선 기본 사항부터 시작해 보겠습니다. 먼저 변환하려는 아키텍처를 알아야 합니다. 
+특정 아키텍처에 대한 관심 없는 경우, 🤗 Transformers 팀에게 제안을 요청하는 것은 여러분의 영향력을 극대화하는 좋은 방법입니다. 
+우리는 TensorFlow에서 빠져 있는 가장 유명한 아키텍처로 이끌어 드리겠습니다. 
+TensorFlow에서 사용할 모델이 이미 🤗 Transformers에 TensorFlow 아키텍처 구현이 있지만 가중치가 없는 경우, 
+이 페이지의 [가중치 추가 섹션](#adding-tensorflow-weights-to-hub)으로 바로 이동하셔도 됩니다.
+
+간단히 말해서, 이 안내서의 나머지 부분은 TensorFlow 버전의 *BrandNewBert*([가이드](add_new_model)와 동일한 예제)를 기여하려고 결정했다고 가정합니다.
+
+<Tip>
+
+TensorFlow 모델 아키텍처에 작업을 시작하기 전에 해당 작업이 진행 중인지 확인하세요. 
+`BrandNewBert`를 검색하여
+[pull request GitHub 페이지](https://github.com/huggingface/transformers/pulls?q=is%3Apr)에서 TensorFlow 관련 pull request가 없는지 확인할 수 있습니다.
+
+</Tip>
+
+**2. transformers 개발 환경 준비**
+
+
+모델 아키텍처를 선택한 후, 관련 작업을 수행할 의도를 미리 알리기 위해 Draft PR을 여세요. 아래 지침대로 하시면 환경을 설정하고 Draft PR을 열 수 있습니다.
+
+1. 'Fork' 버튼을 클릭하여 [리포지터리](https://github.com/huggingface/transformers)를 포크하세요. 이렇게 하면 GitHub 사용자 계정에 코드의 사본이 생성됩니다.
+
+
+2. `transformers` 포크를 로컬 디스크에 클론하고 원본 리포지터리를 원격 리포지터리로 추가하세요.
+
+```bash
+git clone https://github.com/[your Github handle]/transformers.git
+cd transformers
+git remote add upstream https://github.com/huggingface/transformers.git
+```
+
+3. 개발 환경을 설정하세요. 예를 들어, 다음 명령을 실행하여 개발 환경을 설정할 수 있습니다.
+
+```bash
+python -m venv .env
+source .env/bin/activate
+pip install -e ".[dev]"
+```
+
+운영 체제에 따라서 Transformers의 선택적 종속성이 증가하면서 위 명령이 실패할 수도 있습니다. 그런 경우 TensorFlow를 설치한 후 다음을 실행하세요.
+
+```bash
+pip install -e ".[quality]"
+```
+
+**참고:** CUDA를 설치할 필요는 없습니다. 새로운 모델이 CPU에서 작동하도록 만드는 것만으로 충분합니다.
+
+4. 메인 브랜치에서 만드려는 기능이 잘 표현되는 이름으로 브랜치를 만듭니다.
+
+```bash
+git checkout -b add_tf_brand_new_bert
+```
+
+5. 메인 브랜치의 현재 상태를 페치(fetch)하고 리베이스하세요.
+
+```bash
+git fetch upstream
+git rebase upstream/main
+```
+
+6. `transformers/src/models/brandnewbert/`에 `modeling_tf_brandnewbert.py`라는 빈 `.py` 파일을 추가하세요. 이 파일이 TensorFlow 모델 파일이 될 것입니다.
+
+7. 변경 사항을 계정에 푸시하세요.
+
+```bash
+git add .
+git commit -m "initial commit"
+git push -u origin add_tf_brand_new_bert
+```
+
+8. 만족스러운 경우 GitHub에서 포크된 웹 페이지로 이동합니다. "Pull request"를 클릭합니다. Hugging Face 팀의 GitHub ID를 리뷰어로 추가해서, 앞으로의 변경 사항에 대해 Hugging Face 팀이 알림을 받을 수 있도록 합니다.
+
+
+9. GitHub Pull Requests 페이지의 오른쪽에 있는 "Convert to draft"를 클릭하여 PR을 초안으로 변경하세요.
+
+이제 🤗 Transformers에서 *BrandNewBert*를 TensorFlow로 변환할 개발 환경을 설정했습니다.
+
+
+**3. (선택 사항) 이론적 측면 및 기존 구현 이해**
+
+
+*BrandNewBert*처럼 자세한 글이 있다면 시간을 내어 논문을 읽는걸 추천드립니다. 이해하기 어려운 부분이 많을 수 있습니다. 그렇다고 해서 걱정하지 마세요! 목표는 논문의 심도있는 이론적 이해가 아니라 TensorFlow를 사용하여 🤗 Transformers에 모델을 효과적으로 다시 구현하는 데 필요한 필수 정보를 추출하는 것입니다. 많은 시간을 이론적 이해에 투자할 필요는 없지만 실용적인 측면에서 현재 존재하는 모델 문서 페이지(e.g. [model docs for BERT](model_doc/bert))에 집중하는 것이 좋습니다.
+
+
+모델의 기본 사항을 이해한 후, 기존 구현을 이해하는 것이 중요합니다. 이는 작업 중인 모델에 대한 실제 구현이 여러분의 기대와 일치함을 확인하고, TensorFlow 측면에서의 기술적 문제를 예상할 수 있습니다.
+
+막대한 양의 정보를 처음으로 학습할 때 압도당하는 것은 자연스러운 일입니다. 이 단계에서 모델의 모든 측면을 이해해야 하는 필요는 전혀 없습니다. 그러나 우리는 Hugging Face의 [포럼](https://discuss.huggingface.co/)을 통해 질문이 있는 경우 대답을 구할 것을 권장합니다.
+
+### 4. 모델 구현 [[4-model-implementation]]
+
+
+이제 드디어 코딩을 시작할 시간입니다. 우리의 제안된 시작점은 PyTorch 파일 자체입니다: `modeling_brand_new_bert.py`의 내용을 
+`src/transformers/models/brand_new_bert/` 내부의
+`modeling_tf_brand_new_bert.py`에 복사합니다. 이 섹션의 목표는 파일을 수정하고 🤗 Transformers의 import 구조를 업데이트하여 `TFBrandNewBert` 및 `TFBrandNewBert.from_pretrained(model_repo, from_pt=True)`가 성공적으로 작동하는 TensorFlow *BrandNewBert* 모델을 가져올 수 있도록 하는 것입니다.
+
+유감스럽게도, PyTorch 모델을 TensorFlow로 변환하는 규칙은 없습니다. 그러나 프로세스를 가능한한 원활하게 만들기 위해 다음 팁을 따를 수 있습니다.
+
+- 모든 클래스 이름 앞에 `TF`를 붙입니다(예: `BrandNewBert`는 `TFBrandNewBert`가 됩니다).
+- 대부분의 PyTorch 작업에는 직접적인 TensorFlow 대체가 있습니다. 예를 들어, `torch.nn.Linear`는 `tf.keras.layers.Dense`에 해당하고, `torch.nn.Dropout`은 `tf.keras.layers.Dropout`에 해당합니다. 특정 작업에 대해 확신이 없는 경우 [TensorFlow 문서](https://www.tensorflow.org/api_docs/python/tf)나 [PyTorch 문서](https://pytorch.org/docs/stable/)를 참조할 수 있습니다.
+- 🤗 Transformers 코드베이스에서 패턴을 찾으세요. 직접적인 대체가 없는 특정 작업을 만나면 다른 사람이 이미 동일한 문제를 해결한 경우가 많습니다.
+- 기본적으로 PyTorch와 동일한 변수 이름과 구조를 유지하세요. 이렇게 하면 디버깅과 문제 추적, 그리고 문제 해결 추가가 더 쉬워집니다.
+- 일부 레이어는 각 프레임워크마다 다른 기본값을 가지고 있습니다. 대표적인 예로 배치 정규화 레이어의 epsilon은 [PyTorch](https://pytorch.org/docs/stable/generated/torch.nn.BatchNorm2d.html#torch.nn.BatchNorm2d)에서 `1e-5`이고 [TensorFlow](https://www.tensorflow.org/api_docs/python/tf/keras/layers/BatchNormalization)에서 `1e-3`입니다. 문서를 모두 확인하세요!
+- PyTorch의 `nn.Parameter` 변수는 일반적으로 TF 레이어의 `build()` 내에서 초기화해야 합니다. 다음 예를 참조하세요: [PyTorch](https://github.com/huggingface/transformers/blob/655f72a6896c0533b1bdee519ed65a059c2425ac/src/transformers/models/vit_mae/modeling_vit_mae.py#L212) /
+   [TensorFlow](https://github.com/huggingface/transformers/blob/655f72a6896c0533b1bdee519ed65a059c2425ac/src/transformers/models/vit_mae/modeling_tf_vit_mae.py#L220)
+- PyTorch 모델의 함수 상단에 `#copied from ...`가 있는 경우, TensorFlow 모델에 TensorFlow 아키텍처가 있다면 TensorFlow 모델이 해당 함수를 복사한 아키텍처에서 사용할 수 있습니다.
+- TensorFlow 함수에서 `name` 속성을 올바르게 할당하는 것은 `from_pt=True` 가중치 교차 로딩을 수행하는 데 중요합니다. `name`은 대부분 PyTorch 코드의 해당 변수의 이름입니다. `name`이 제대로 설정되지 않으면 모델 가중치를 로드할 때 오류 메시지에서 확인할 수 있습니다.
+- 기본 모델 클래스인 `BrandNewBertModel`의 로직은 실제로 Keras 레이어 서브클래스([예시](https://github.com/huggingface/transformers/blob/4fd32a1f499e45f009c2c0dea4d81c321cba7e02/src/transformers/models/bert/modeling_tf_bert.py#L719))인 `TFBrandNewBertMainLayer`에 있습니다. `TFBrandNewBertModel`은 이 레이어를 감싸기만 하는 래퍼 역할을 합니다.
+- Keras 모델은 사전 훈련된 가중치를 로드하기 위해 빌드되어야 합니다. 따라서 `TFBrandNewBertPreTrainedModel`은 모델의 입력 예제인 `dummy_inputs`([예시](https://github.com/huggingface/transformers/blob/4fd32a1f499e45f009c2c0dea4d81c321cba7e02/src/transformers/models/bert/modeling_tf_bert.py#L916)) 유지해야 합니다.
+- 도움이 필요한 경우 도움을 요청하세요. 우리는 여기 있어서 도움을 드리기 위해 있는 것입니다! 🤗
+
+모델 파일 자체 외에도 모델 클래스 및 관련 문서 페이지에 대한 포인터를 추가해야 합니다. 이 부분은 다른 PR([예시](https://github.com/huggingface/transformers/pull/18020/files))의 패턴을 따라 완전히 완료할 수 있습니다. 다음은 필요한 수동 변경 목록입니다.
+
+- `src/transformers/__init__.py`에 *BrandNewBert*의 모든 공개 클래스를 포함합니다.
+- `src/transformers/models/auto/modeling_tf_auto.py`에서 *BrandNewBert* 클래스를 해당 Auto 클래스에 추가합니다.
+- `utils/documentation_tests.txt`에 모델 파일을 문서화하는 테스트 파일 목록을 추가합니다.
+- `src/transformers/utils/dummy_tf_objects.py`에 *BrandNewBert*와 관련된 레이지 로딩 클래스를 추가합니다.
+- `src/transformers/models/brand_new_bert/__init__.py`에서 공개 클래스에 대한 import 구조를 업데이트합니다.
+- `docs/source/en/model_doc/brand_new_bert.md`에서 *BrandNewBert*의 공개 메서드에 대한 문서 포인터를 추가합니다.
+- `docs/source/en/model_doc/brand_new_bert.md`의 *BrandNewBert* 기여자 목록에 자신을 추가합니다.
+- 마지막으로 ✅ 녹색 체크박스를 TensorFlow 열 docs/source/en/index.md 안 BrandNewBert에 추가합니다.
+
+구현이 만족하면 다음 체크리스트를 실행하여 모델 아키텍처가 준비되었는지 확인하세요.  
+
+1. 훈련 시간에 다르게 동작하는 `training` 인수로 불리는 모든 레이어(예: Dropout)는 최상위 클래스에서 전파됩니다.
+2. #copied from ...가능할 때마다 사용했습니다.
+3. `TFBrandNewBertMainLayer`와 그것을 사용하는 모든 클래스는 `call`함수로 `@unpack_inputs`와 함께 데코레이터 됩니다.
+4. `TFBrandNewBertMainLayer`는 `@keras_serializable`로 데코레이터 됩니다.
+5. TensorFlow 모델은 `TFBrandNewBert.from_pretrained(model_repo, from_pt=True)`를 사용하여 PyTorch 가중치에서 로드할 수 있습니다.
+6. 예상 입력 형식을 사용하여 TensorFlow 모델을 호출할 수 있습니다.
+
+### 5. 모델 테스트 구현 [[5-add-model-tests]]
+
+TensorFlow 모델 아키텍처를 구현하는 데 성공했습니다! 이제 TensorFlow 모델을 테스트하는 구현을 작성할 차례입니다. 이를 통해 모델이 예상대로 작동하는지 확인할 수 있습니다. 이전에 우리는 `test_modeling_brand_new_bert.py` 파일을 `tests/models/brand_new_bert/ into test_modeling_tf_brand_new_bert.py`에 복사한 뒤, TensorFlow로 교체하는 것이 좋습니다. 지금은, 모든 `.from_pretrained()`을 `from_pt=True`를 사용하여 존재하는 Pytorch 가중치를 가져오도록 해야합니다.  
+
+완료하셨으면, 이제 진실의 순간이 찾아왔습니다: 테스트를 실행해 보세요! 😬
+
+```bash
+NVIDIA_TF32_OVERRIDE=0 RUN_SLOW=1 RUN_PT_TF_CROSS_TESTS=1 \
+py.test -vv tests/models/brand_new_bert/test_modeling_tf_brand_new_bert.py
+```
+
+오류가 많이 나타날 것이지만 괜찮습니다! 기계 학습 모델을 디버깅하는 것은 악명높게 어려우며 성공의 핵심 요소는 인내심입니다 (`breakpoint()`도 필요합니다). 우리의 경험상으로는 ML 프레임워크 사이의 미묘한 불일치로 인해 가장 어려운 문제가 발생합니다. 이에 대한 몇 가지 지침이 이 가이드의 끝 부분에 있습니다. 다른 경우에는 일반 테스트가 직접 모델에 적용되지 않을 수 있으며, 이 경우 모델 테스트 클래스 레벨에서 재정의를 제안합니다. 문제가 무엇이든지 상관없이 문제가 있으면 당신이 고립되었다면 draft pull request에서 도움을 요청하는 것이 좋습니다.
+
+모든 테스트가 통과되면 축하합니다. 이제 모델을 🤗 Transformers 라이브러리에 추가할 준비가 거의 완료된 것입니다! 🎉
+
+
+테스트를 추가하는 방법에 대한 자세한 내용은 [🤗 Transformers의 테스트 가이드](https://huggingface.co/transformers/contributing.html#running-tests)를 참조하세요.
+
+### 6.-7. 모든 사용자가 당신의 모델을 사용할 수 있게 하기 [[6.-7.-ensure-everyone -can-use-your-model]]
+
+**6. 풀 요청 제출하기**
+
+구현과 테스트가 완료되면 풀 요청을 제출할 시간입니다. 코드를 푸시하기 전에 코드 서식 맞추기 유틸리티인 `make fixup` 🪄 를 실행하세요. 이렇게 하면 자동으로 서식 오류를 수정하며 자동 검사가 실패하는 것을 방지할 수 있습니다.
+
+이제 드래프트 풀 요청을 실제 풀 요청으로 변환하는 시간입니다. "리뷰 준비됨" 버튼을 클릭하고 Joao (`@gante`)와 Matt (`@Rocketknight1`)를 리뷰어로 추가하세요. 모델 풀 요청에는 적어도 3명의 리뷰어가 필요하지만, 그들이 당신의 모델에 적절한 추가 리뷰어를 찾을 것입니다.
+
+모든 리뷰어들이 PR 상태에 만족하면 마지막으로 `.from_pretrained()` 호출에서 `from_pt=True` 플래그를 제거하는 것입니다. TensorFlow 가중치가 없기 때문에 이를 추가해야 합니다! 이를 수행하는 방법은 아래 섹션의 지침을 확인하세요.
+
+마침내 TensorFlow 가중치가 병합되고, 적어도 3명의 리뷰어 승인을 받았으며 모든 CI 검사가 통과되었다면, 로컬로 테스트를 한 번 더 확인하세요.
+
+```bash
+NVIDIA_TF32_OVERRIDE=0 RUN_SLOW=1 RUN_PT_TF_CROSS_TESTS=1 \
+py.test -vv tests/models/brand_new_bert/test_modeling_tf_brand_new_bert.py
+```
+
+그리고 우리는 당신의 PR을 병합할 것입니다! 마일스톤 달성을 축하드립니다! 🎉
+
+**7. (선택 사항) 데모를 만들고 세상과 공유하기**
+
+오픈 소스의 가장 어려운 부분 중 하나는 발견입니다. 다른 사용자들이 당신의 멋진 TensorFlow 기여를 어떻게 알 수 있을까요? 물론 적절한 커뮤니케이션으로 가능합니다! 📣
+
+커뮤니티와 모델을 공유하는 두 가지 주요 방법이 있습니다:
+- 데모 만들기. Gradio 데모, 노트북 및 모델을 자랑하는 다른 재미있는 방법을 포함합니다. [커뮤니티 기반 데모](https://huggingface.co/docs/transformers/community)에 노트북을 추가하는 것을 적극 권장합니다.
+- Twitter와 LinkedIn과 같은 소셜 미디어에 이야기 공유하기. 당신의 작업에 자랑스러워하고 커뮤니티와 당신의 업적을 공유해야 합니다. 이제 당신의 모델은 전 세계의 수천 명의 엔지니어와 연구원들에 의해 사용될 수 있습니다 🌍! 우리는 당신의 게시물을 리트윗하고 커뮤니티와 함께 당신의 작업을 공유하는 데 도움이 될 것입니다.
+
+
+## 🤗 허브에 TensorFlow 가중치 추가하기 [[adding-tensorFlow-weights-to-🤗-hub]]
+
+TensorFlow 모델 아키텍처가 🤗 Transformers에서 사용 가능하다고 가정하고, PyTorch 가중치를 TensorFlow 가중치로 변환하는 것은 쉽습니다!
+
+다음은 그 방법입니다:
+1. 터미널에서 Hugging Face 계정으로 로그인되어 있는지 확인하십시오. `huggingface-cli login` 명령어를 사용하여 로그인할 수 있습니다. (액세스 토큰은 [여기](https://huggingface.co/settings/tokens)에서 찾을 수 있습니다.)
+2. `transformers-cli pt-to-tf --model-name foo/bar`를 실행하십시오. 여기서 `foo/bar`는 변환하려는 PyTorch 가중치가 있는 모델 저장소의 이름입니다.
+3. 방금 만든 🤗 허브 PR에서 `@joaogante`와 `@Rocketknight1`을 태그합니다.
+
+그게 다입니다! 🎉
+
+
+## ML 프레임워크 간 디버깅 🐛[[debugging-mismatches-across-ml-frameworks]]
+
+새로운 아키텍처를 추가하거나 기존 아키텍처에 대한 TensorFlow 가중치를 생성할 때, PyTorch와 TensorFlow 간의 불일치로 인한 오류가 발생할 수 있습니다. 심지어 두 프레임워크의 모델 아키텍처 코드가 동일해 보일 수도 있습니다. 무슨 일이 벌어지고 있는 걸까요? 🤔
+
+먼저, 이러한 불일치를 이해하는 이유에 대해 이야기해 보겠습니다. 많은 커뮤니티 멤버들은 🤗 Transformers 모델을 그대로 사용하고, 우리의 모델이 예상대로 작동할 것이라고 믿습니다. 두 프레임워크 간에 큰 불일치가 있으면 모델이 적어도 하나의 프레임워크에 대한 참조 구현을 따르지 않음을 의미합니다. 이는 모델이 의도한 대로 작동하지 않을 수 있음을 나타냅니다. 이는 아예 실행되지 않는 모델보다 나쁠 수 있습니다! 따라서 우리는 모든 모델의 프레임워크 불일치를 `1e-5`보다 작게 유지하는 것을 목표로 합니다.
+
+기타 숫자 문제와 마찬가지로, 세세한 문제가 있습니다. 그리고 세세함에 집중하는 공정에서 필수 요소는 인내심입니다. 이러한 종류의 문제가 발생할 때 권장되는 작업 흐름은 다음과 같습니다:
+1. 불일치의 원인을 찾아보십시오. 변환 중인 모델은 아마도 특정 지점까지 거의 동일한 내부 변수를 가지고 있을 것입니다. 두 프레임워크의 아키텍처에 `breakpoint()` 문을 넣고, 위에서 아래로 숫자 변수의 값을 비교하여 문제의 근원을 찾아냅니다.
+2. 이제 문제의 근원을 찾았으므로 🤗 Transformers 팀에 연락하세요. 우리는 비슷한 문제를 이전에 겪었을 수 있으며 빠르게 해결책을 제공할 수 있습니다. 예외적인 경우에는 StackOverflow와 GitHub 이슈와 같은 인기있는 페이지를 확인하십시오.
+3. 더 이상 해결책이 없는 경우, 더 깊이 들어가야 합니다. 좋은 소식은 문제의 원인을 찾았으므로 나머지 모델을 추상화하고 문제가 있는 명령어에 초점을 맞출 수 있습니다! 나쁜 소식은 해당 명령어의 소스 구현에 대해 알아봐야 한다는 것입니다. 일부 경우에는 참조 구현에 문제가 있을 수도 있으니 업스트림 저장소에서 이슈를 열기를 꺼리지 마십시오.
+
+어떤 경우에는 🤗 Transformers 팀과의 토론을 통해 불일치를 수정할 수 없을 수도 있습니다. 모델의 출력 레이어에서 불일치가 매우 작지만 숨겨진 상태에서 크게 나타날 수 있기 때문입니다. 이 경우 모델을 배포하는 것을 우선시하기 위해 불일치를 무시하기로 결정할 수도 있습니다. 위에서 언급한 `pt-to-tf` CLI에는 가중치 변환 시 오류 메시지를 무시하는 `--max-error` 플래그가 있습니다.

docs/source/ko/community.md

@@ -0,0 +1,69 @@
+<!--⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
+rendered properly in your Markdown viewer.
+-->
+
+# 커뮤니티 [[community]]
+
+이 페이지는 커뮤니티에서 개발한 🤗 Transformers 리소스를 재구성한 페이지입니다.
+
+## 커뮤니티 리소스: [[community-resources]]
+
+| 리소스     |      설명      |      만든이     |
+|:----------|:-------------|------:|
+| [Hugging Face Transformers 용어집 플래시카드](https://www.darigovresearch.com/huggingface-transformers-glossary-flashcards) | [Transformers 문서 용어집](glossary)을 기반으로 한 플래시카드 세트로, 지식을 장기적으로 유지하기 위해 특별히 설계된 오픈소스 크로스 플랫폼 앱인 [Anki](https://apps.ankiweb.net/)를 사용하여 쉽게 학습/수정할 수 있는 형태로 제작되었습니다. [플래시카드 사용법에 대한 소개 동영상](https://www.youtube.com/watch?v=Dji_h7PILrw)을 참조하세요. | [Darigov 리서치](https://www.darigovresearch.com/) |
+
+## 커뮤니티 노트북: [[community-notebooks]]
+
+| 노트북     |      설명      |      만든이      |      |
+|:----------|:-------------|:-------------|------:|
+| [가사를 생성하기 위해 사전훈련된 트랜스포머를 미세 조정하기](https://github.com/AlekseyKorshuk/huggingartists) | GPT-2 모델을 미세 조정하여 좋아하는 아티스트의 스타일로 가사를 생성하는 방법 | [Aleksey Korshuk](https://github.com/AlekseyKorshuk) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/AlekseyKorshuk/huggingartists/blob/master/huggingartists-demo.ipynb) |
+| [Tensorflow 2로 T5 훈련하기](https://github.com/snapthat/TF-T5-text-to-text) |  Tensorflow 2를 사용하여 T5를 훈련시키는 방법. 이 노트북은 Tensorflow 2로 SQUAD를 사용하여 구현한 질의응답 작업을 보여줍니다. | [Muhammad Harris](https://github.com/HarrisDePerceptron) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/snapthat/TF-T5-text-to-text/blob/master/snapthatT5/notebooks/TF-T5-Datasets%20Training.ipynb) |
+| [TPU에서 T5 훈련하기](https://github.com/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb) | Transformers와 Nlp를 사용하여 SQUAD로 T5를 훈련하는 방법 | [Suraj Patil](https://github.com/patil-suraj) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/T5_on_TPU.ipynb#scrollTo=QLGiFCDqvuil) |
+| [분류 및 객관식 문제를 위해 T5 미세 조정하기](https://github.com/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) | 분류 및 객관식 문제에 맞게 텍스트-텍스트 형식을 사용하여 PyTorch Lightning으로 T5를 미세 조정하는 방법 | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/exploring-T5/blob/master/t5_fine_tuning.ipynb) |
+| [새로운 데이터 세트와 언어로 DialoGPT 미세 조정하기](https://github.com/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) | 자유 대화형 챗봇을 만들기 위해 새로운 데이터 세트로 DialoGPT 모델을 미세 조정하는 방법 | [Nathan Cooper](https://github.com/ncoop57) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ncoop57/i-am-a-nerd/blob/master/_notebooks/2020-05-12-chatbot-part-1.ipynb) |
+| [Reformer로 긴 시퀀스 모델링하기](https://github.com/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb)  | Reformer로 최대 50만 토큰의 시퀀스를 훈련하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/PyTorch_Reformer.ipynb) |
+| [요약을 위해 BART 미세 조정하기](https://github.com/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) | blurr를 사용하여 fastai로 요약하기 위해 BART를 미세 조정하는 방법 | [Wayde Gilliam](https://ohmeow.com/) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/ohmeow/ohmeow_website/blob/master/posts/2021-05-25-mbart-sequence-classification-with-blurr.ipynb) |
+| [다른 사람의 트윗으로 사전훈련된 트랜스포머 미세 조정하기](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) | GPT-2 모델을 미세 조정하여 좋아하는 트위터 계정 스타일로 트윗을 생성하는 방법 | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/borisdayma/huggingtweets/blob/master/huggingtweets-demo.ipynb) |
+| [Weights & Biases로 🤗 Hugging Face 모델 최적화하기](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) | W&B와 Hugging Face의 통합을 보여주는 전체 튜토리얼 | [Boris Dayma](https://github.com/borisdayma) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/wandb/examples/blob/master/colabs/huggingface/Optimize_Hugging_Face_models_with_Weights_%26_Biases.ipynb) |
+| [Longformer 사전훈련하기](https://github.com/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb)  | 기존 사전훈련된 모델의 "긴" 버전을 빌드하는 방법 |  [Iz Beltagy](https://beltagy.net) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/allenai/longformer/blob/master/scripts/convert_model_to_long.ipynb) |
+| [QA를 위해 Longformer 미세 조정하기](https://github.com/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) | QA 작업을 위해 Longformer를 미세 조정하는 방법 | [Suraj Patil](https://github.com/patil-suraj) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patil-suraj/Notebooks/blob/master/longformer_qa_training.ipynb) |
+| [🤗 Nlp로 모델 평가하기](https://github.com/patrickvonplaten/notebooks/blob/master/How_to_evaluate_Longformer_on_TriviaQA_using_NLP.ipynb) | `Nlp`로 TriviaQA에서 Longformer를 평가하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1m7eTGlPmLRgoPkkA7rkhQdZ9ydpmsdLE?usp=sharing) |
+| [감정 범위 추출을 위해 T5 미세 조정하기](https://github.com/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb)  | 감정 범위 추출을 위해 텍스트-텍스트 형식을 사용하여 PyTorch Lightning으로 T5를 미세 조정하는 방법 |  [Lorenzo Ampil](https://github.com/enzoampil) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/enzoampil/t5-intro/blob/master/t5_qa_training_pytorch_span_extraction.ipynb) |
+| [다중 클래스 분류를 위해 DistilBert 미세 조정하기](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb) | 다중 클래스 분류를 위해 PyTorch를 사용하여 DistilBert를 미세 조정하는 방법 | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multiclass_classification.ipynb)|
+| [다중 레이블 분류를 위해 BERT 미세 조정하기](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb) | 다중 레이블 분류를 위해 PyTorch를 사용하여 BERT를 미세 조정하는 방법 | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_multi_label_classification.ipynb)|
+| [요약을 위해 T5 미세 조정하기](https://github.com/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb) | 요약을 위해 PyTorch로 T5를 미세 조정하고 WandB로 실험을 추적하는 방법 | [Abhishek Kumar Mishra](https://github.com/abhimishra91) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/abhimishra91/transformers-tutorials/blob/master/transformers_summarization_wandb.ipynb)|
+| [동적 패딩/버켓팅으로 Transformers 미세 조정 속도 높이기](https://github.com/ELS-RD/transformers-notebook/blob/master/Divide_Hugging_Face_Transformers_training_time_by_2_or_more.ipynb)| 동적 패딩/버켓팅을 사용하여 미세 조정 속도를 2배로 높이는 방법 |[Michael Benesty](https://github.com/pommedeterresautee) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1CBfRU1zbfu7-ijiOqAAQUA-RJaxfcJoO?usp=sharing)|
+|[마스킹된 언어 모델링을 위해 Reformer 사전훈련하기](https://github.com/patrickvonplaten/notebooks/blob/master/Reformer_For_Masked_LM.ipynb)| 양방향 셀프 어텐션 레이어를 이용해서 Reformer 모델을 훈련하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1tzzh0i8PgDQGV3SMFUGxM7_gGae3K-uW?usp=sharing)|
+| [Sci-BERT 확장 및 미세 조정하기](https://github.com/lordtt13/word-embeddings/blob/master/COVID-19%20Research%20Data/COVID-SciBERT.ipynb)| CORD 데이터 세트로 AllenAI에서 사전훈련된 SciBERT 모델의 어휘를 늘리고 파이프라인을 구축하는 방법 | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1rqAR40goxbAfez1xvF3hBJphSCsvXmh8)|
+| [요약을 위해 Trainer API로 BlenderBotSmall 미세 조정하기](https://github.com/lordtt13/transformers-experiments/blob/master/Custom%20Tasks/fine-tune-blenderbot_small-for-summarization.ipynb)| 요약을 위해 Trainer API를 사용하여 사용자 지정 데이터 세트로 BlenderBotSmall 미세 조정하기 | [Tanmay Thakur](https://github.com/lordtt13) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/19Wmupuls7mykSGyRN_Qo6lPQhgp56ymq?usp=sharing)|
+| [통합 기울기(Integrated Gradient)를 이용하여 Electra 미세 조정하고 해석하기](https://github.com/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb) | 감정 분석을 위해 Electra를 미세 조정하고 Captum 통합 기울기로 예측을 해석하는 방법 | [Eliza Szczechla](https://elsanns.github.io) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/electra_fine_tune_interpret_captum_ig.ipynb)|
+| [Trainer 클래스로 비영어권 GPT-2 모델 미세 조정하기](https://github.com/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb) | Trainer 클래스로 비영어권 GPT-2 모델을 미세 조정하는 방법 | [Philipp Schmid](https://www.philschmid.de) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/philschmid/fine-tune-GPT-2/blob/master/Fine_tune_a_non_English_GPT_2_Model_with_Huggingface.ipynb)|
+|[다중 라벨 분류 작업을 위해 DistilBERT 모델 미세 조정하기](https://github.com/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb) | 다중 라벨 분류 작업을 위해 DistilBERT 모델을 미세 조정하는 방법 | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/Transformers_scripts/blob/master/Transformers_multilabel_distilbert.ipynb)|
+|[문장쌍 분류를 위해 ALBERT 미세 조정하기](https://github.com/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb) | 문장쌍 분류 작업을 위해 ALBERT 모델 또는 다른 BERT 기반 모델을 미세 조정하는 방법 | [Nadir El Manouzi](https://github.com/NadirEM) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NadirEM/nlp-notebooks/blob/master/Fine_tune_ALBERT_sentence_pair_classification.ipynb)|
+|[감정 분석을 위해 Roberta 미세 조정하기](https://github.com/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb) | 감정 분석을 위해 Roberta 모델을 미세 조정하는 방법 | [Dhaval Taunk](https://github.com/DhavalTaunk08) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/DhavalTaunk08/NLP_scripts/blob/master/sentiment_analysis_using_roberta.ipynb)|
+|[질문 생성 모델 평가하기](https://github.com/flexudy-pipe/qugeev) | seq2seq 트랜스포머 모델이 생성한 질문과 이에 대한 답변이 얼마나 정확한가요? | [Pascal Zoleko](https://github.com/zolekode) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1bpsSqCQU-iw_5nNoRm_crPq6FRuJthq_?usp=sharing)|
+|[DistilBERT와 Tensorflow로 텍스트 분류하기](https://github.com/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb) | 텍스트 분류를 위해 TensorFlow로  DistilBERT를 미세 조정하는 방법 | [Peter Bayerle](https://github.com/peterbayerle) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/peterbayerle/huggingface_notebook/blob/main/distilbert_tf.ipynb)|
+|[CNN/Dailail 요약을 위해 인코더-디코더 모델에 BERT 활용하기](https://github.com/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb) | CNN/Dailail 요약을 위해 *bert-base-uncased* 체크포인트를 활용하여 *EncoderDecoderModel*을 워밍업하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/BERT2BERT_for_CNN_Dailymail.ipynb)|
+|[BBC XSum 요약을 위해 인코더-디코더 모델에 RoBERTa 활용하기](https://github.com/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb) | BBC/XSum 요약을 위해 *roberta-base* 체크포인트를 활용하여 공유 *EncoderDecoderModel*을 워밍업하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/RoBERTaShared_for_BBC_XSum.ipynb)|
+|[순차적 질문 답변(SQA)을 위해 TAPAS 미세 조정하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb) | *tapas-base* 체크포인트를 활용하여 순차적 질문 답변(SQA) 데이터 세트로 *TapasForQuestionAnswering*을 미세 조정하는 방법 | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Fine_tuning_TapasForQuestionAnswering_on_SQA.ipynb)|
+|[표 사실 검사(TabFact)로 TAPAS 평가하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb) | 🤗 Datasets와 🤗 Transformer 라이브러리를 함께 사용하여 *tapas-base-finetuned-tabfact* 체크포인트로 미세 조정된 *TapasForSequenceClassification*을 평가하는 방법 | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/TAPAS/Evaluating_TAPAS_on_the_Tabfact_test_set.ipynb)|
+|[번역을 위해 mBART 미세 조정하기](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb) | 힌디어에서 영어로 번역하기 위해 Seq2SeqTrainer를 사용하여 mBART를 미세 조정하는 방법 | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/huggingface-tutorials/blob/main/translation_training.ipynb)|
+|[FUNSD(양식 이해 데이터 세트)로 LayoutLM 미세 조정하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb) | 스캔한 문서에서 정보 추출을 위해 FUNSD 데이터 세트로 *LayoutLMForTokenClassification*을 미세 조정하는 방법 | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForTokenClassification_on_FUNSD.ipynb)|
+|[DistilGPT2 미세 조정하고 및 텍스트 생성하기](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb) | DistilGPT2를 미세 조정하고 텍스트를 생성하는 방법 | [Aakash Tripathi](https://github.com/tripathiaakash) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/tripathiaakash/DistilGPT2-Tutorial/blob/main/distilgpt2_fine_tuning.ipynb)|
+|[최대 8K 토큰에서 LED 미세 조정하기](https://github.com/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb) | 긴 범위를 요약하기 위해 PubMed로 LED를 미세 조정하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Fine_tune_Longformer_Encoder_Decoder_(LED)_for_Summarization_on_pubmed.ipynb)|
+|[Arxiv로 LED 평가하기](https://github.com/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb) | 긴 범위 요약에 대해 LED를 효과적으로 평가하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/LED_on_Arxiv.ipynb)|
+|[RVL-CDIP(문서 이미지 분류 데이터 세트)로 LayoutLM 미세 조정하기)](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb) | 스캔 문서 분류를 위해 RVL-CDIP 데이터 세트로 *LayoutLMForSequenceClassification*을 미세 조정하는 방법 | [Niels Rogge](https://github.com/nielsrogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/LayoutLM/Fine_tuning_LayoutLMForSequenceClassification_on_RVL_CDIP.ipynb)|
+|[GPT2 조정을 통한 Wav2Vec2 CTC 디코딩](https://github.com/voidful/huggingface_notebook/blob/main/xlsr_gpt.ipynb) | 언어 모델 조정을 통해 CTC 시퀀스를 디코딩하는 방법 | [Eric Lam](https://github.com/voidful) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1e_z5jQHYbO2YKEaUgzb1ww1WwiAyydAj?usp=sharing)|
+|[Trainer 클래스로 두 개 언어로 요약하기 위해 BART 미세 조정하기](https://github.com/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb) | Trainer 클래스로 두 개 언어로 요약하기 위해 BART 미세 조정하는 방법 | [Eliza Szczechla](https://github.com/elsanns) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/elsanns/xai-nlp-notebooks/blob/master/fine_tune_bart_summarization_two_langs.ipynb)|
+|[Trivia QA로 Big Bird 평가하기](https://github.com/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb) | Trivia QA로 긴 문서 질문에 대한 답변에 대해 BigBird를 평가하는 방법 | [Patrick von Platen](https://github.com/patrickvonplaten) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/patrickvonplaten/notebooks/blob/master/Evaluating_Big_Bird_on_TriviaQA.ipynb)|
+| [Wav2Vec2를 사용하여 동영상 캡션 만들기](https://github.com/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) | Wav2Vec으로 오디오를 텍스트로 변환하여 모든 동영상에서 YouTube 캡션 만드는 방법 | [Niklas Muennighoff](https://github.com/Muennighoff) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/Muennighoff/ytclipcc/blob/main/wav2vec_youtube_captions.ipynb) |
+| [PyTorch Lightning을 사용하여 CIFAR-10으로 비전 트랜스포머 미세 조정하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) | HuggingFace Transformers, Datasets, PyTorch Lightning을 사용하여 CIFAR-10으로 비전 트랜스포머(ViT)를 미세 조정하는 방법  | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_PyTorch_Lightning.ipynb) |
+| [🤗 Trainer를 사용하여 CIFAR-10에서 비전 트랜스포머 미세 조정하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) | Datasets, 🤗 Trainer를 사용하여 CIFAR-10에서 비전 트랜스포머(ViT)를 미세 조정하는 방법 | [Niels Rogge](https://github.com/nielsrogge) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/VisionTransformer/Fine_tuning_the_Vision_Transformer_on_CIFAR_10_with_the_%F0%9F%A4%97_Trainer.ipynb) |
+| [개체 입력 데이터 세트인 Open Entity로 LUKE 평가하기](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) | Open Entity 데이터 세트로 *LukeForEntityClassification*을 평가하는 방법 | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_open_entity.ipynb) |
+| [관계 추출 데이터 세트인 TACRED로 LUKE 평가하기](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) | TACRED 데이터 세트로 *LukeForEntityPairClassification*을 평가하는 방법 | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_tacred.ipynb) |
+| [중요 NER 벤치마크인 CoNLL-2003으로 LUKE 평가하기](https://github.com/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) | CoNLL-2003 데이터 세트로 *LukeForEntitySpanClassification*를 평가하는 방법 | [Ikuya Yamada](https://github.com/ikuyamada) |[![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/studio-ousia/luke/blob/master/notebooks/huggingface_conll_2003.ipynb) |
+| [PubMed 데이터 세트로 BigBird-Pegasus 평가하기](https://github.com/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) | PubMed 데이터 세트로 *BigBirdPegasusForConditionalGeneration*를 평가하는 방법 | [Vasudev Gupta](https://github.com/vasudevgupta7) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/vasudevgupta7/bigbird/blob/main/notebooks/bigbird_pegasus_evaluation.ipynb) |
+| [Wav2Vec2를 사용해서 음성 감정 분류하기](https://github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) | 감정 분류를 위해 사전훈련된 Wav2Vec2 모델을 MEGA 데이터 세트에 활용하는 방법 | [Mehrdad Farahani](https://github.com/m3hrdadfi) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/m3hrdadfi/soxan/blob/main/notebooks/Emotion_recognition_in_Greek_speech_using_Wav2Vec2.ipynb) |
+| [DETR로 이미지에서 객체 탐지하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) | 훈련된 *DetrForObjectDetection* 모델을 사용하여 이미지에서 객체를 탐지하고 어텐션을 시각화하는 방법 | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/DETR_minimal_example_(with_DetrFeatureExtractor).ipynb) |
+| [사용자 지정 객체 탐지 데이터 세트로 DETR 미세 조정하기](https://github.com/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) | 사용자 지정 객체 탐지 데이터 세트로 *DetrForObjectDetection*을 미세 조정하는 방법 | [Niels Rogge](https://github.com/NielsRogge) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/github/NielsRogge/Transformers-Tutorials/blob/master/DETR/Fine_tuning_DetrForObjectDetection_on_custom_dataset_(balloon).ipynb) |
+| [개체명 인식을 위해 T5 미세 조정하기](https://github.com/ToluClassics/Notebooks/blob/main/T5_Ner_Finetuning.ipynb) | 개체명 인식 작업을 위해 *T5*를 미세 조정하는 방법 | [Ogundepo Odunayo](https://github.com/ToluClassics) | [![Open In Colab](https://colab.research.google.com/assets/colab-badge.svg)](https://colab.research.google.com/drive/1obr78FY_cBmWY5ODViCmzdY6O1KB65Vc?usp=sharing) |

docs/source/ko/contributing.md

@@ -0,0 +1,332 @@
+<!---
+Copyright 2020 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+-->
+
+# 🤗 Transformers에 기여하기 [[contribute-to-transformers]]
+
+누구나 🤗 Transformers에 기여할 수 있으며, 우리는 모든 사람의 기여를 소중히 생각합니다. 코드 기여는 커뮤니티를 돕는 유일한 방법이 아닙니다. 질문에 답하거나 다른 사람을 도와 문서를 개선하는 것도 매우 가치가 있습니다.
+
+🤗 Transformers를 널리 알리는 것도 큰 도움이 됩니다! 멋진 프로젝트들을 가능하게 한 🤗 Transformers 라이브러리에 대해 블로그 게시글에 언급하거나, 도움이 되었을 때마다 Twitter에 알리거나, 저장소에 ⭐️ 를 표시하여 감사 인사를 전해주세요.
+
+어떤 방식으로 기여하든 [행동 규칙](https://github.com/huggingface/transformers/blob/main/CODE_OF_CONDUCT.md)을 숙지하고 존중해주세요.
+
+**이 안내서는 멋진 [scikit-learn 기여 안내서](https://github.com/scikit-learn/scikit-learn/blob/main/CONTRIBUTING.md)에서 큰 영감을 받았습니다.**
+
+## 기여하는 방법 [[ways-to-contribute]]
+
+여러 가지 방법으로 🤗 Transformers에 기여할 수 있습니다:
+
+* 기존 코드의 미해결된 문제를 수정합니다.
+* 버그 또는 새로 추가되길 원하는 기능과 관련된 이슈를 제출합니다.
+* 새로운 모델을 구현합니다.
+* 예제나 문서에 기여합니다.
+
+어디서부터 시작할지 모르겠다면, [Good First Issue](https://github.com/huggingface/transformers/contribute) 목록을 확인해보세요. 이 목록은 초보자도 참여하기 쉬운 오픈 이슈 목록을 제공하며, 당신이 오픈소스에 처음으로 기여하는 데 큰 도움이 될 것입니다. 그저 작업하고 싶은 이슈에 댓글만 달아주면 됩니다.
+
+조금 더 도전적인 작업을 원한다면, [Good Second Issue](https://github.com/huggingface/transformers/labels/Good%20Second%20Issue) 목록도 확인해보세요. 이미 당신이 잘 하고 있다고 생각되더라도, 한 번 시도해보세요! 우리도 여러분을 도울 것입니다. 🚀
+
+> 커뮤니티에 이루어지는 모든 기여는 똑같이 소중합니다. 🥰
+
+## 미해결된 문제 수정하기 [[fixing-outstanding-issues]]
+
+기존 코드에서 발견한 문제점에 대한 해결책이 떠오른 경우, 언제든지 [기여를 시작](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md/#create-a-pull-request)하고 Pull Request를 생성해주세요!
+
+## 버그 관련 이슈를 제기하거나 새로운 기능 요청하기 [[submitting-a-bugrelated-issue-or-feature-request]]
+
+버그 관련 이슈를 제기하거나 새로운 기능을 요청할 때는 다음 가이드라인을 최대한 준수해주세요. 이렇게 하면 좋은 피드백과 함께 빠르게 답변해 드릴 수 있습니다.
+
+### 버그를 발견하셨나요? [[did-you-find-a-bug]]
+
+🤗 Transformers 라이브러리는 사용 중에 겪는 문제를 보고해주는 사용자들 덕분에 더욱 견고해지고 신뢰할 수 있게 되었습니다.
+
+이슈를 보고하기 전에, 버그가 이미 **보고되지 않았는지** 확인해주세요. (GitHub의 이슈 탭 아래의 검색 바를 사용하세요). 이슈는 라이브러리 자체에서 발생한 버그어야 하며, 코드의 다른 부분과 관련된 것이 아니어야 합니다. 버그가 라이브러리의 문제로 발생하였는지 확실하지 않은 경우 먼저 [포럼](https://discuss.huggingface.co/)에서 질문해 주세요. 이렇게 하면 일반적인 질문보다 라이브러리와 관련된 문제를 더 빠르게 해결할 수 있습니다.
+
+버그가 이미 보고되지 않았다는 것을 확인했다면, 다음 정보를 포함하여 이슈를 제출해 주세요. 그러면 우리가 빠르게 해결할 수 있습니다:
+
+* 사용 중인 **운영체제 종류와 버전**, 그리고 **Python**, **PyTorch** 또는 **TensorFlow** 버전.
+* 버그를 30초 이내로 재현할 수 있는 간단하고 독립적인 코드 스니펫.
+* 예외가 발생한 경우 *전체* 트레이스백.
+* 스크린샷과 같이 도움이 될 것으로 생각되는 추가 정보를 첨부해 주세요.
+
+운영체제와 소프트웨어 버전을 자동으로 가져오려면 다음 명령을 실행하세요:
+
+```bash
+transformers-cli env
+```
+
+저장소의 루트 디렉터리에서도 같은 명령을 실행할 수 있습니다:
+
+```bash
+python src/transformers/commands/transformers_cli.py env
+```
+
+
+### 새로운 기능을 원하시나요? [[do-you-want-a-new-feature]]
+
+🤗 Transformers에서 사용하고 싶은 새로운 기능이 있다면, 다음 내용을 포함하여 이슈를 제출해 주세요:
+
+1. 이 기능이 필요한 *이유*는 무엇인가요? 라이브러리에 대한 문제나 불만과 관련이 있나요? 프로젝트에 필요한 기능인가요? 커뮤니티에 도움이 될 만한 기능인가요?
+
+   어떤 내용이든 여러분의 이야기를 듣고 싶습니다!
+
+2. 요청하는 기능을 최대한 자세히 설명해 주세요. 더 많은 정보를 제공할수록 더 나은 도움을 드릴 수 있습니다.
+3. 해당 기능의 사용법을 보여주는 *코드 스니펫*을 제공해 주세요.
+4. 기능과 관련된 논문이 있는 경우 링크를 포함해 주세요.
+
+이슈가 잘 작성되었다면 이슈가 생성된 순간, 이미 80% 정도의 작업이 완료된 것입니다. 
+
+이슈를 제기하는 데 도움이 될 만한 [템플릿](https://github.com/huggingface/transformers/tree/main/templates)도 준비되어 있습니다.
+
+## 새로운 모델을 구현하고 싶으신가요? [[do-you-want-to-implement-a-new-model]]
+
+새로운 모델은 계속해서 출시됩니다. 만약 여러분이 새로운 모델을 구현하고 싶다면 다음 정보를 제공해 주세요.
+
+* 모델에 대한 간단한 설명과 논문 링크.
+* 구현이 공개되어 있다면 구현 링크.
+* 모델 가중치가 사용 가능하다면 가중치 링크.
+
+만약 모델을 직접 기여하고 싶으시다면, 알려주세요. 🤗 Transformers에 추가할 수 있도록 도와드리겠습니다!
+
+새로운 모델을 추가하는 방법에 대한 [상세 안내서와 템플릿](https://github.com/huggingface/transformers/tree/main/templates)을 제공하고 있으며, [🤗 Transformers에 새로운 모델을 추가하는 방법](https://huggingface.co/docs/transformers/add_new_model)에 대한 기술적인 안내서도 있습니다.
+
+## 문서를 추가하고 싶으신가요? [[do-you-want-to-add-documentation]]
+
+우리는 언제나 더 명확하고 정확한 문서를 제공하기 위하여 개선점을 찾고 있습니다. 오탈자나 부족한 내용, 분명하지 않거나 부정확한 내용 등을 알려주시면 개선하는 데 도움이 됩니다. 관심이 있으시다면 변경하거나 기여하실 수 있도록 도와드리겠습니다!
+
+문서를 생성, 빌드 및 작성하는 방법에 대한 자세한 내용은 [README](https://github.com/huggingface/transformers/tree/main/docs) 문서를 확인해 주세요.
+
+## 풀 리퀘스트(Pull Request) 생성하기 [[create-a-pull-request]]
+
+코드를 작성하기 전에 기존의 Pull Request나 이슈를 검색하여 누군가 이미 동일한 작업을 하고 있는지 확인하는 것이 좋습니다. 확실하지 않다면 피드백을 받기 위해 이슈를 열어보는 것이 좋습니다.
+
+🤗 Transformers에 기여하기 위해서는 기본적인 `git` 사용 능력이 필요합니다. `git`은 사용하기 쉬운 도구는 아니지만, 매우 훌륭한 매뉴얼을 제공합니다. 쉘(shell)에서 `git --help`을 입력하여 확인해보세요! 만약 책을 선호한다면, [Pro Git](https://git-scm.com/book/en/v2)은 매우 좋은 참고 자료가 될 것입니다.
+
+🤗 Transformers에 기여하려면 **[Python 3.8]((https://github.com/huggingface/transformers/blob/main/setup.py#L426))** 이상의 버전이 필요합니다. 기여를 시작하려면 다음 단계를 따르세요:
+
+1. 저장소 페이지에서 **[Fork](https://github.com/huggingface/transformers/fork)** 버튼을 클릭하여 저장소를 포크하세요. 이렇게 하면 코드의 복사본이 여러분의 GitHub 사용자 계정 아래에 생성됩니다.
+
+2. 포크한 저장소를 로컬 디스크로 클론하고, 기본 저장소를 원격(remote)으로 추가하세요:
+
+   ```bash
+   git clone git@github.com:<your Github handle>/transformers.git
+   cd transformers
+   git remote add upstream https://github.com/huggingface/transformers.git
+   ```
+
+3. 개발 변경 사항을 저장할 새 브랜치를 생성하세요:
+
+   ```bash
+   git checkout -b a-descriptive-name-for-my-changes
+   ```
+
+   🚨 절대 `main` 브랜치에서 작업하지 **마세요!**
+
+4. 가상 환경에서 다음 명령을 실행하여 개발 환경을 설정하세요:
+
+   ```bash
+   pip install -e ".[dev]"
+   ```
+
+   만약 이미 가상 환경에 🤗 Transformers가 설치되어 있다면, `-e` 플래그를 사용하여 설치하기 전에 `pip uninstall transformers`로 제거해주세요.
+   
+   여러분의 운영체제에 따라서, 그리고 🤗 Transformers의 선택적 의존성의 수가 증가하면서, 이 명령이 실패할 수도 있습니다. 그럴 경우 사용하려는 딥러닝 프레임워크(PyTorch, TensorFlow, 그리고/또는 Flax)를 설치한 후 아래 명령을 실행해주세요:
+
+   ```bash
+   pip install -e ".[quality]"
+   ```
+
+   대부분의 경우 이것으로 충분할 것입니다.
+
+5. 브랜치에서 기능을 개발하세요.
+
+   코드를 작업하는 동안 테스트 스위트(test suite)가 통과하는지 확인하세요. 다음과 같이 변경 사항에 영향을 받는 테스트를 실행하세요:
+
+   ```bash
+   pytest tests/<TEST_TO_RUN>.py
+   ```
+
+   테스트에 대한 더 많은 정보는 [테스트](https://huggingface.co/docs/transformers/testing) 가이드를 확인하세요.
+
+   🤗 Transformers는 `black`과 `ruff`를 사용하여 소스 코드의 형식을 일관되게 유지합니다. 변경 사항을 적용한 후에는 다음 명령으로 자동으로 스타일 교정 및 코드 검증을 수행하세요:
+
+   ```bash
+   make fixup
+   ```
+
+   이것은 또한 작업 중인 PR에서 수정한 파일에서만 작동하도록 최적화되어 있습니다.
+
+   검사를 하나씩 실행하려는 경우, 다음 명령으로 스타일 교정을 적용할 수 있습니다:
+
+   ```bash
+   make style
+   ```
+
+   🤗 Transformers는 또한 `ruff`와 몇 가지 사용자 정의 스크립트를 사용하여 코딩 실수를 확인합니다. CI를 통해 품질 관리가 수행되지만, 다음 명령으로 동일한 검사를 실행할 수 있습니다:
+
+   ```bash
+   make quality
+   ```
+
+   마지막으로, 새 모델을 추가할 때 일부 파일을 업데이트하는 것을 잊지 않도록 하기 위한 많은 스크립트가 있습니다. 다음 명령으로 이러한 스크립트를 실행할 수 있습니다:
+
+   ```bash
+   make repo-consistency
+   ```
+
+   이러한 검사에 대해 자세히 알아보고 관련 문제를 해결하는 방법은 [Pull Request에 대한 검사](https://huggingface.co/docs/transformers/pr_checks) 가이드를 확인하세요.
+
+   만약 `docs/source` 디렉터리 아래의 문서를 수정하는 경우, 문서가 빌드될 수 있는지 확인하세요. 이 검사는 Pull Request를 열 때도 CI에서 실행됩니다. 로컬 검사를 실행하려면 문서 빌더를 설치해야 합니다:
+   
+   ```bash
+   pip install ".[docs]"
+   ```
+
+   저장소의 루트 디렉터리에서 다음 명령을 실행하세요:
+
+   ```bash
+   doc-builder build transformers docs/source/en --build_dir ~/tmp/test-build
+   ```
+
+   이 명령은 `~/tmp/test-build` 폴더에 문서를 빌드하며, 생성된 Markdown 파일을 선호하는 편집기로 확인할 수 있습니다. Pull Request를 열 때 GitHub에서 문서를 미리 볼 수도 있습니다.
+
+   변경 사항에 만족하면 `git add`로 변경된 파일을 추가하고, `git commit`으로 변경 사항을 로컬에 기록하세요:
+
+   ```bash
+   git add modified_file.py
+   git commit
+   ```
+
+   [좋은 커밋 메시지](https://chris.beams.io/posts/git-commit/)를 작성하여 변경 사항을 명확하게 전달하세요!
+
+   변경 사항을 프로젝트 원본 저장소와 동기화하려면, PR을 *열기 전에* 브랜치를 `upstream/branch`로 리베이스(rebase)하세요. 또는 관리자의 요청에 이 작업이 필요할 수 있습니다:
+
+   ```bash
+   git fetch upstream
+   git rebase upstream/main
+   ```
+   
+   변경 사항을 브랜치에 푸시하세요:
+
+   ```bash
+   git push -u origin a-descriptive-name-for-my-changes
+   ```
+
+   이미 PR을 열었다면, `--force` 플래그와 함께 강제 푸시해야 합니다. 아직 PR이 열리지 않았다면 정상적으로 변경 사항을 푸시하면 됩니다.
+
+6. 이제 GitHub에서 포크한 저장소로 이동하고 **Pull request(풀 리퀘스트)**를 클릭하여 Pull Request를 열 수 있습니다. 아래의 [체크리스트](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md/#pull-request-checklist)에서 모든 항목에 체크 표시를 하세요. 준비가 완료되면 프로젝트 관리자에게 변경 사항을 보내 검토를 요청할 수 있습니다.
+
+7. 관리자가 변경 사항을 요청해도 괜찮습니다. 핵심 기여자들도 동일한 상황을 겪습니다! 모두가 변경 사항을 Pull Request에서 볼 수 있도록, 로컬 브랜치에서 작업하고 변경 사항을 포크한 저장소로 푸시하세요. 그러면 변경 사항이 자동으로 Pull Request에 나타납니다.
+
+### Pull Request 체크리스트 [[pull-request-checklist]]
+
+☐ Pull Request 제목은 기여 내용을 요약해야 합니다.<br>
+☐ Pull Request가 이슈를 해결하는 경우, Pull Request 설명에 이슈 번호를 언급하여 연관되어 있음을 알려주세요. (이슈를 확인하는 사람들이 해당 이슈에 대한 작업이 진행 중임을 알 수 있게 합니다).<br>
+☐ 작업이 진행중이라면 제목 앞에 `[WIP]`를 붙여주세요. 중복 작업을 피하고 병합할 준비가 된 PR과 구분하기에 유용합니다.<br>
+☐ 기존 테스트를 통과하는지 확인하세요.<br>
+☐ 새로운 기능을 추가하는 경우, 해당 기능에 대한 테스트도 추가하세요.<br>
+   - 새 모델을 추가하는 경우, `ModelTester.all_model_classes = (MyModel, MyModelWithLMHead,...)`을 사용하여 일반적인 테스트를 활성화하세요.
+   - 새 `@slow` 테스트를 추가하는 경우, 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/my_new_model/test_my_new_model.py`.
+   - 새 토크나이저를 추가하는 경우, 테스트를 작성하고 다음 명령으로 테스트를 통과하는지 확인하세요: `RUN_SLOW=1 python -m pytest tests/models/{your_model_name}/test_tokenization_{your_model_name}.py`. 
+   - CircleCI에서는 느린 테스트를 실행하지 않지만, GitHub Actions에서는 매일 밤 실행됩니다!<br>
+
+☐ 모든 공개 메소드는 유용한 기술문서를 가져야 합니다 (예를 들어 [`modeling_bert.py`](https://github.com/huggingface/transformers/blob/main/src/transformers/models/bert/modeling_bert.py) 참조).<br>
+☐ 저장소가 빠르게 성장하고 있으므로 저장소에 상당한 부담을 주는 이미지, 동영상 및 기타 텍스트가 아닌 파일은 추가하지 마세요. 대신 [`hf-internal-testing`](https://huggingface.co/hf-internal-testing)과 같은 Hub 저장소를 사용하여 이러한 파일을 호스팅하고 URL로 참조하세요. 문서와 관련된 이미지는 다음 저장소에 배치하는 것을 권장합니다: [huggingface/documentation-images](https://huggingface.co/datasets/huggingface/documentation-images). 이 데이터셋 저장소에서 PR을 열어서 Hugging Face 멤버에게 병합을 요청할 수 있습니다.
+
+Pull Request에서 실행되는 검사에 대한 자세한 정보는 [Pull Request에 대한 검사](https://huggingface.co/docs/transformers/pr_checks) 가이드를 확인하세요.
+
+### 테스트 [[tests]]
+
+라이브러리 동작과 여러 예제를 테스트할 수 있는 광범위한 테스트 스위트가 포함되어 있습니다. 라이브러리 테스트는 [tests](https://github.com/huggingface/transformers/tree/main/tests) 폴더에, 예제 테스트는 [examples](https://github.com/huggingface/transformers/tree/main/examples) 폴더에 있습니다.
+
+속도가 빠른 `pytest`와 `pytest-xdist`를 선호합니다. 저장소의 루트 디렉터리에서 테스트를 실행할 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요.
+
+```bash
+python -m pytest -n auto --dist=loadfile -s -v ./tests/models/my_new_model
+```
+
+마찬가지로 `examples` 디렉터리에서도 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요. 예를 들어, 다음 명령은 PyTorch `examples` 디렉터리의 텍스트 분류 하위 폴더를 테스트합니다:
+
+```bash
+pip install -r examples/xxx/requirements.txt  # only needed the first time
+python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/text-classification
+```
+
+이것이 실제로 `make test` 및 `make test-examples` 명령이 구현되는 방식입니다 (`pip install`은 제외합니다)!
+
+또한 특정 기능만 테스트하기 위한 더 작은 테스트를 지정할 수 있습니다.
+
+기본적으로 느린 테스트는 건너뛰지만 `RUN_SLOW` 환경 변수를 `yes`로 설정하여 실행할 수 있습니다. 이렇게 하면 많은 기가바이트 단위의 모델이 다운로드되므로 충분한 디스크 공간, 좋은 인터넷 연결과 많은 인내가 필요합니다!
+
+<Tip warning={true}>
+
+테스트를 실행하려면 *하위 폴더 경로 또는 테스트 파일 경로*를 지정하세요. 그렇지 않으면 `tests` 또는 `examples` 폴더의 모든 테스트를 실행하게 되어 매우 긴 시간이 걸립니다!
+
+</Tip>
+
+```bash
+RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./tests/models/my_new_model
+RUN_SLOW=yes python -m pytest -n auto --dist=loadfile -s -v ./examples/pytorch/text-classification
+```
+
+느린 테스트와 마찬가지로, 다음과 같이 테스트 중에 기본적으로 활성화되지 않는 다른 환경 변수도 있습니다:
+- `RUN_CUSTOM_TOKENIZERS`: 사용자 정의 토크나이저 테스트를 활성화합니다.
+- `RUN_PT_FLAX_CROSS_TESTS`: PyTorch + Flax 통합 테스트를 활성화합니다.
+- `RUN_PT_TF_CROSS_TESTS`: TensorFlow + PyTorch 통합 테스트를 활성화합니다.
+
+더 많은 환경 변수와 추가 정보는 [testing_utils.py](src/transformers/testing_utils.py)에서 찾을 수 있습니다.
+
+🤗 Transformers는 테스트 실행기로 `pytest`를 사용합니다. 그러나 테스트 스위트 자체에서는 `pytest` 관련 기능을 사용하지 않습니다.
+
+이것은 `unittest`가 완전히 지원된다는 것을 의미합니다. 다음은 `unittest`로 테스트를 실행하는 방법입니다:
+
+```bash
+python -m unittest discover -s tests -t . -v
+python -m unittest discover -s examples -t examples -v
+```
+
+### 스타일 가이드 [[style-guide]]
+
+문서는 [Google Python 스타일 가이드](https://google.github.io/styleguide/pyguide.html)를 따릅니다. 자세한 정보는 [문서 작성 가이드](https://github.com/huggingface/transformers/tree/main/docs#writing-documentation---specification)를 확인하세요.
+
+### Windows에서 개발 [[develop-on-windows]]
+
+Windows에서 개발할 경우([Windows Subsystem for Linux](https://learn.microsoft.com/en-us/windows/wsl/) 또는 WSL에서 작업하지 않는 한) Windows `CRLF` 줄 바꿈을 Linux `LF` 줄 바꿈으로 변환하도록 git을 구성해야 합니다:
+
+```bash
+git config core.autocrlf input
+```
+
+Windows에서 `make` 명령을 실행하는 한 가지 방법은 MSYS2를 사용하는 것입니다:
+
+1. [MSYS2](https://www.msys2.org/)를 다운로드합니다. `C:\msys64`에 설치되었다고 가정합니다.
+2. CLI에서 `C:\msys64\msys2.exe`를 엽니다 (시작 메뉴에서 사용 가능해야 함).
+3. 쉘에서 다음을 실행하여: `pacman -Syu` 및 `pacman -S make`로 `make`를 설치합니다.
+4. 환경 변수 PATH에 `C:\msys64\usr\bin`을 추가하세요.
+
+이제 모든 터미널 (Powershell, cmd.exe 등)에서 `make`를 사용할 수 있습니다! 🎉
+
+### 포크한 저장소를 상위 원본 브랜치(main)과 동기화하기 (Hugging Face 저장소) [[sync-a-forked-repository-with-upstream-main-the-hugging-face-repository]]
+
+포크한 저장소의 main 브랜치를 업데이트할 때, 다음 단계를 따라 수행해주세요. 이렇게 하면 각 upstream PR에 참조 노트가 추가되는 것을 피하고 이러한 PR에 관여하는 개발자들에게 불필요한 알림이 전송되는 것을 방지할 수 있습니다.
+
+1. 가능하면 포크된 저장소의 브랜치 및 PR을 사용하여 upstream과 동기화하지 마세요. 대신 포크된 main 저장소에 직접 병합하세요.
+2. PR이 반드시 필요한 경우, 브랜치를 확인한 후 다음 단계를 사용하세요:
+
+```bash
+git checkout -b your-branch-for-syncing
+git pull --squash --no-commit upstream main
+git commit -m '<your message without GitHub references>'
+git push --set-upstream origin your-branch-for-syncing
+```