Add TF whisper (#19378)

* simplify loop

* add featur extractor

* add model

* start conversion

* add dropout

* initial commit of test files

* copnversion for all models

* update processor for correct padding

* update feature extraction

* update integration test logits match

* fmnt: off for the logits

* on the fly mel bank

* small nit

* update test

* update tokenizer

* nit feature extraction

* update

* update tokenizer test

* adds logit processor and update tokenizer to get supress tokens

* style

* clean convert

* revert to original modeling tf utils

* Update

* update

* nit

* clean convert file

* update tests and nits

* quality

* slow generation test

* ffn_dim to allow customization

* update readme

* add to toctreee

* start fixing integration tests

* update tests and code

* fix feature extractor

* fix config tests common

* update code to fix tests

* fix feature exctractor

* nit feature extraction

* update test for new feature extractor

* style

* add absrtact

* large logits wioth custom decoder input ids

* wraap around is otrch available

* fix feature extractor

* correct logits for whisper small.en

* nit

* fix encoder_attentino_mask

* some fixes

* remove unnecessary inputs

* nits

* add normalizer file

* update etst tokenization

* fix attention mask not defined

* fix generate

* remove uncoder attention mask useless

* update test modeling whisper

* update condfig to add second non supress tokens

* nits on feature exrtactor

* nit for test tokenizers

* update etsts

* update tests

* update tokenization test

* fixup

* invalidated hf token. Clean convert openai to whisper

* fix logit tests

* fixup

* Add model to README

* Fix doc tests

* clean merge

* revert toc_tree changes

* remove useless LogitProcessor

* Update whisper .mdx

* update config file doc

* update configuration docstring

* update test tokenization

* update test tokenization

* update tokenization whisper
Added copied from where needed

* update feature extraction

* nit test name

* style

* quality

* remove get suppress tokens and update non_speech tokens global variables

* Update src/transformers/models/whisper/feature_extraction_whisper.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* clean modeling whisper and test
Removed the attention mask arguments that are deprecated

* fix large test

* Add multilingual audio test, and translate test

* style

* fix larg multilingual test

* nits

* add copied from for attention layer

* remove attention masks in doc

* add english normalizer

* Update docs/source/en/model_doc/whisper.mdx
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* update tokenization test

* remove copied from in whisper attention : no bias in k_proj only

* wrap around dependencies in english normalizer

* style

* correct import generation logits

* for now, wrap feature extractor with torch

* remove torch depencies for feature extraction and style

* Update src/transformers/models/whisper/convert_openai_whisper_to_tfms.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/models/whisper/configuration_whisper.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update docs/source/en/model_doc/whisper.mdx
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* fixup

* nit

* update logitds

* style

* nit

* nits and fix final tests

* add `is_more_itertools_available` to utils

* quality

* add begin supress tokens, supress tokens to generate args and config

* clean supressTokensLogitProcessor in generation logits

* Nit naming

* add supressTokensAtBegin

* udpate tests, supress tokens to None or correct values

* nit and style

* update RAG to fit test and generate_logit

* add copy pasted statment on english normalizer

* add arguments to config_common_kwargs

* Update src/transformers/generation_utils.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Update src/transformers/generation_logits_process.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* revert changes based on reviews

* update doc and nits

* Update src/transformers/models/whisper/configuration_whisper.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* more nits

* last nits

* update test configuration common

* add BART name in decoder attention mask documentation

* Update src/transformers/models/whisper/modeling_whisper.py
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>

* style

* nit

* nit

* add english.json file to git

* nits on documentation

* nit

* nits

* last styling

* add main toctree file

* remove sentence piece dependency

* clean init file

* fix tokenizer that has no dependencies on sentencepiece

* update whisper init file, nit

* remove english.json file

* add get decoder prompt id

* All weights loading

* Remove hanging pdb

* Fixup and tidy up

* Use same copied from as PT model

* Remove whitespace changes

* Remove torch references

* Tie embeddings

* Remove logits processor input to generate

* Update logit values

* revert changes and add forced logit processor

* nit

* clean normalizer

* remove protected

* Add logit processors and update generation code & tests

* Some tidy up

* Update docstring

* update

* update based on review

* Update src/transformers/models/whisper/configuration_whisper.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update src/transformers/models/whisper/configuration_whisper.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Update to reflect changes on the PT model branch

* Tidy up

* Remove extra whitespace

* Fix test - make input ids small enough we can append

* Include upstream changes on main

* PR comments - add batch tests, remove comments & defaults

* Fix model output imports

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/generation_tf_logits_process.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update tests/models/whisper/test_modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update docstring example

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* Remove changes to adjust_logits_during_generation function

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Tidy up imports that don't require TF

* Update tests - skip and no more skip

* Update tests/generation/test_generation_tf_logits_process.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Update src/transformers/models/whisper/modeling_tf_whisper.py

* Update src/transformers/models/whisper/modeling_tf_whisper.py
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* Add training flags

* Add (skipped) XLA generation tests

* Add embedding correctness test

* Add constant ids for generation tests

* Make logits finding a bit tidier

* Remove unused args

* xla generation enabled

* Don't skip XLA tests anymore

* Fix tests - add position ids to expected signature and update rag generation

* Undo method reorder

* Remove added whitespace

* Remove copy-paste gradient checkopint ref

* Remove

* Trigger CI - (issue with refs when pulling)
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: NielsRogge <niels.rogge1@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: NielsRogge <48327001+NielsRogge@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>
Co-authored-by: Joao Gante <joao@huggingface.co>

README.md

@@ -55,13 +55,13 @@ limitations under the License.
     <a href="https://hf.co/course"><img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/course_banner.png"></a>
 </h3>
 
-🤗 Transformers provides thousands of pretrained models to perform tasks on different modalities such as text, vision, and audio. 
+🤗 Transformers provides thousands of pretrained models to perform tasks on different modalities such as text, vision, and audio.
 
 These models can be applied on:
 
-* 📝 Text, for tasks like text classification, information extraction, question answering, summarization, translation, text generation, in over 100 languages. 
-* 🖼️ Images, for tasks like image classification, object detection, and segmentation. 
-* 🗣️ Audio, for tasks like speech recognition and audio classification. 
+* 📝 Text, for tasks like text classification, information extraction, question answering, summarization, translation, text generation, in over 100 languages.
+* 🖼️ Images, for tasks like image classification, object detection, and segmentation.
+* 🗣️ Audio, for tasks like speech recognition and audio classification.
 
 Transformer models can also perform tasks on **several modalities combined**, such as table question answering, optical character recognition, information extraction from scanned documents, video classification, and visual question answering.
 

README_ko.md

@@ -59,7 +59,7 @@ limitations under the License.
 
 🤗 Transformers는 이러한 사전학습 모델을 빠르게 다운로드해 특정 텍스트에 사용하고, 원하는 데이터로 fine-tuning해 커뮤니티나 우리의 [모델 허브](https://huggingface.co/models)에 공유할 수 있도록 API를 제공합니다. 또한, 모델 구조를 정의하는 각 파이썬 모듈은 완전히 독립적이여서 연구 실험을 위해 손쉽게 수정할 수 있습니다.
 
-🤗 Transformers는 가장 유명한 3개의 딥러닝 라이브러리를 지원합니다. 이들은 서로 완벽히 연동됩니다 — [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/), [TensorFlow](https://www.tensorflow.org/). 간단하게 이 라이브러리 중 하나로 모델을 학습하고, 또 다른 라이브러리로 추론을 위해 모델을 불러올 수 있습니다. 
+🤗 Transformers는 가장 유명한 3개의 딥러닝 라이브러리를 지원합니다. 이들은 서로 완벽히 연동됩니다 — [Jax](https://jax.readthedocs.io/en/latest/), [PyTorch](https://pytorch.org/), [TensorFlow](https://www.tensorflow.org/). 간단하게 이 라이브러리 중 하나로 모델을 학습하고, 또 다른 라이브러리로 추론을 위해 모델을 불러올 수 있습니다.
 
 ## 온라인 데모
 
@@ -74,7 +74,7 @@ limitations under the License.
 - [DistilBERT를 이용한 질문 답변](https://huggingface.co/distilbert-base-uncased-distilled-squad?text=Which+name+is+also+used+to+describe+the+Amazon+rainforest+in+English%3F&context=The+Amazon+rainforest+%28Portuguese%3A+Floresta+Amaz%C3%B4nica+or+Amaz%C3%B4nia%3B+Spanish%3A+Selva+Amaz%C3%B3nica%2C+Amazon%C3%ADa+or+usually+Amazonia%3B+French%3A+For%C3%AAt+amazonienne%3B+Dutch%3A+Amazoneregenwoud%29%2C+also+known+in+English+as+Amazonia+or+the+Amazon+Jungle%2C+is+a+moist+broadleaf+forest+that+covers+most+of+the+Amazon+basin+of+South+America.+This+basin+encompasses+7%2C000%2C000+square+kilometres+%282%2C700%2C000+sq+mi%29%2C+of+which+5%2C500%2C000+square+kilometres+%282%2C100%2C000+sq+mi%29+are+covered+by+the+rainforest.+This+region+includes+territory+belonging+to+nine+nations.+The+majority+of+the+forest+is+contained+within+Brazil%2C+with+60%25+of+the+rainforest%2C+followed+by+Peru+with+13%25%2C+Colombia+with+10%25%2C+and+with+minor+amounts+in+Venezuela%2C+Ecuador%2C+Bolivia%2C+Guyana%2C+Suriname+and+French+Guiana.+States+or+departments+in+four+nations+contain+%22Amazonas%22+in+their+names.+The+Amazon+represents+over+half+of+the+planet%27s+remaining+rainforests%2C+and+comprises+the+largest+and+most+biodiverse+tract+of+tropical+rainforest+in+the+world%2C+with+an+estimated+390+billion+individual+trees+divided+into+16%2C000+species)
 - [T5로 번역하기](https://huggingface.co/t5-base?text=My+name+is+Wolfgang+and+I+live+in+Berlin)
 
-**[Transformer와 글쓰기](https://transformer.huggingface.co)** 는 이 저장소의 텍스트 생성 능력에 관한 Hugging Face 팀의 공식 데모입니다. 
+**[Transformer와 글쓰기](https://transformer.huggingface.co)** 는 이 저장소의 텍스트 생성 능력에 관한 Hugging Face 팀의 공식 데모입니다.
 
 ## Hugging Face 팀의 커스텀 지원을 원한다면
 
@@ -258,7 +258,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
 1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
 1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
-1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 
+1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori.
 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
 1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
@@ -297,7 +297,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
 1. **[OWL-ViT](https://huggingface.co/docs/transformers/model_doc/owlvit)** (from Google AI) released with the paper [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) by Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby.
 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. **[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. **[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. **[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.
@@ -307,9 +307,9 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
 1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
-1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár. 
+1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
 1. **[RemBERT](https://huggingface.co/docs/transformers/model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/pdf/2010.12821.pdf) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
-1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. 
+1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
 1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
@@ -325,11 +325,11 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released in the repository [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
-1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace).
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft), released together with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
-1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler 
+1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
@@ -345,7 +345,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
 1. **[Whisper](https://huggingface.co/docs/transformers/main/model_doc/whisper)** (from OpenAI) released with the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
 1. **[X-CLIP](https://huggingface.co/docs/transformers/model_doc/xclip)** (from Microsoft Research) released with the paper [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) by Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling.
-1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
+1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-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. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.
@@ -355,7 +355,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 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. 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. 
+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

@@ -245,7 +245,7 @@ conda install -c huggingface transformers
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (来自 Google Research) 伴随论文 [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) 由 Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed 发布。
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (来自 Facebook) 伴随论文 [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) 由 Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston 发布。
-1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/). 
+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)** (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. **[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。
@@ -282,7 +282,7 @@ conda install -c huggingface transformers
 1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (来自 KAIST) 伴随论文 [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) 由 Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim 发布。
 1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (来自 OpenAI) 伴随论文 [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) 由 Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever 发布。
 1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (来自 EleutherAI) 随仓库 [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) 发布。作者为 Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy 发布。
-1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 
+1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (来自 ABEJA) 由 Shinya Otani, Takayoshi Makabe, Anuj Arora, Kyo Hattori。
 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (来自 OpenAI) 伴随论文 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 由 Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 发布。
 1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (来自 EleutherAI) 伴随论文 [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) 由 Ben Wang and Aran Komatsuzaki 发布。
@@ -314,14 +314,14 @@ conda install -c huggingface transformers
 1. **[MobileViT](https://huggingface.co/docs/transformers/model_doc/mobilevit)** (来自 Apple) 伴随论文 [MobileViT: Light-weight, General-purpose, and Mobile-friendly Vision Transformer](https://arxiv.org/abs/2110.02178) 由 Sachin Mehta and Mohammad Rastegari 发布。
 1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (来自 Microsoft Research) 伴随论文 [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) 由 Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu 发布。
 1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (来自 Google AI) 伴随论文 [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) 由 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel 发布。
-1. **[MVP](https://huggingface.co/docs/transformers/model_doc/mvp)** (来自 中国人民大学 AI Box) 伴随论文 [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) 由 Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen 发布。 
+1. **[MVP](https://huggingface.co/docs/transformers/model_doc/mvp)** (来自 中国人民大学 AI Box) 伴随论文 [MVP: Multi-task Supervised Pre-training for Natural Language Generation](https://arxiv.org/abs/2206.12131) 由 Tianyi Tang, Junyi Li, Wayne Xin Zhao and Ji-Rong Wen 发布。
 1. **[Nezha](https://huggingface.co/docs/transformers/model_doc/nezha)** (来自华为诺亚方舟实验室) 伴随论文 [NEZHA: Neural Contextualized Representation for Chinese Language Understanding](https://arxiv.org/abs/1909.00204) 由 Junqiu Wei, Xiaozhe Ren, Xiaoguang Li, Wenyong Huang, Yi Liao, Yasheng Wang, Jiashu Lin, Xin Jiang, Xiao Chen and Qun Liu 发布。
 1. **[NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)** (来自 Meta) 伴随论文 [No Language Left Behind: Scaling Human-Centered Machine Translation](https://arxiv.org/abs/2207.04672) 由 the NLLB team 发布。
 1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (来自 the University of Wisconsin - Madison) 伴随论文 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 由 Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 发布。
 1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (来自 Meta AI) 伴随论文 [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) 由 Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al 发布。
 1. **[OWL-ViT](https://huggingface.co/docs/transformers/model_doc/owlvit)** (来自 Google AI) 伴随论文 [Simple Open-Vocabulary Object Detection with Vision Transformers](https://arxiv.org/abs/2205.06230) 由 Matthias Minderer, Alexey Gritsenko, Austin Stone, Maxim Neumann, Dirk Weissenborn, Alexey Dosovitskiy, Aravindh Mahendran, Anurag Arnab, Mostafa Dehghani, Zhuoran Shen, Xiao Wang, Xiaohua Zhai, Thomas Kipf, and Neil Houlsby 发布。
 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. **[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. **[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. **[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。
@@ -331,9 +331,9 @@ conda install -c huggingface transformers
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (来自 Facebook) 伴随论文 [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) 由 Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela 发布。
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (来自 Google Research) 伴随论文 [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) 由 Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang 发布。
 1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (来自 Google Research) 伴随论文 [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) 由 Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya 发布。
-1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár. 
+1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
 1. **[RemBERT](https://huggingface.co/docs/transformers/model_doc/rembert)** (来自 Google Research) 伴随论文 [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/pdf/2010.12821.pdf) 由 Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder 发布。
-1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. 
+1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
 1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (来自 Facebook), 伴随论文 [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) 由 Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov 发布。
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (来自 ZhuiyiTechnology), 伴随论文 [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) 由 Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu 发布。
 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (来自 NVIDIA) 伴随论文 [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) 由 Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo 发布。
@@ -349,11 +349,11 @@ conda install -c huggingface transformers
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (来自 Google AI) 伴随论文 [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) 由 Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu 发布。
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (来自 Google AI) 伴随论文 [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) 由 Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos 发布。
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (来自 Microsoft Research) 伴随论文 [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) 由 Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou 发布。
-1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace).
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (来自 Google/CMU) 伴随论文 [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) 由 Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov 发布。
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (来自 Microsoft) 伴随论文 [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) 由 Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei 发布。
-1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler 
+1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (来自 Microsoft Research) 伴随论文 [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) 由 Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang 发布。
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (来自 Microsoft Research) 伴随论文 [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) 由 Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu 发布。
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (来自 Tsinghua University and Nankai University) 伴随论文 [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) 由 Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu 发布。
@@ -369,7 +369,7 @@ conda install -c huggingface transformers
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
 1. **[Whisper](https://huggingface.co/docs/transformers/main/model_doc/whisper)** (来自 OpenAI) 伴随论文 [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) 由 Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever 发布。
 1. **[X-CLIP](https://huggingface.co/docs/transformers/model_doc/xclip)** (来自 Microsoft Research) 伴随论文 [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) 由 Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling 发布。
-1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
+1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (来自 Facebook) 伴随论文 [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) 由 Guillaume Lample and Alexis Conneau 发布。
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-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. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (来自 Facebook AI), 伴随论文 [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) 由 Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov 发布。

README_zh-hant.md

@@ -257,7 +257,7 @@ conda install -c huggingface transformers
 1. **[BigBird-RoBERTa](https://huggingface.co/docs/transformers/model_doc/big_bird)** (from Google Research) released with the paper [Big Bird: Transformers for Longer Sequences](https://arxiv.org/abs/2007.14062) by Manzil Zaheer, Guru Guruganesh, Avinava Dubey, Joshua Ainslie, Chris Alberti, Santiago Ontanon, Philip Pham, Anirudh Ravula, Qifan Wang, Li Yang, Amr Ahmed.
 1. **[Blenderbot](https://huggingface.co/docs/transformers/model_doc/blenderbot)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
 1. **[BlenderbotSmall](https://huggingface.co/docs/transformers/model_doc/blenderbot-small)** (from Facebook) released with the paper [Recipes for building an open-domain chatbot](https://arxiv.org/abs/2004.13637) by Stephen Roller, Emily Dinan, Naman Goyal, Da Ju, Mary Williamson, Yinhan Liu, Jing Xu, Myle Ott, Kurt Shuster, Eric M. Smith, Y-Lan Boureau, Jason Weston.
-1. **[BLOOM](https://huggingface.co/docs/transformers/model_doc/bloom)** (from BigScience workshop) released by the [BigSicence Workshop](https://bigscience.huggingface.co/). 
+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)** (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. **[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.
@@ -294,7 +294,7 @@ conda install -c huggingface transformers
 1. **[GLPN](https://huggingface.co/docs/transformers/model_doc/glpn)** (from KAIST) released with the paper [Global-Local Path Networks for Monocular Depth Estimation with Vertical CutDepth](https://arxiv.org/abs/2201.07436) by Doyeon Kim, Woonghyun Ga, Pyungwhan Ahn, Donggyu Joo, Sehwan Chun, Junmo Kim.
 1. **[GPT](https://huggingface.co/docs/transformers/model_doc/openai-gpt)** (from OpenAI) released with the paper [Improving Language Understanding by Generative Pre-Training](https://blog.openai.com/language-unsupervised/) by Alec Radford, Karthik Narasimhan, Tim Salimans and Ilya Sutskever.
 1. **[GPT Neo](https://huggingface.co/docs/transformers/model_doc/gpt_neo)** (from EleutherAI) released in the repository [EleutherAI/gpt-neo](https://github.com/EleutherAI/gpt-neo) by Sid Black, Stella Biderman, Leo Gao, Phil Wang and Connor Leahy.
-1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach 
+1. **[GPT NeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox)** (from EleutherAI) released with the paper [GPT-NeoX-20B: An Open-Source Autoregressive Language Model](https://arxiv.org/abs/2204.06745) by Sid Black, Stella Biderman, Eric Hallahan, Quentin Anthony, Leo Gao, Laurence Golding, Horace He, Connor Leahy, Kyle McDonell, Jason Phang, Michael Pieler, USVSN Sai Prashanth, Shivanshu Purohit, Laria Reynolds, Jonathan Tow, Ben Wang, Samuel Weinbach
 1. **[GPT NeoX Japanese](https://huggingface.co/docs/transformers/model_doc/gpt_neox_japanese)** (from ABEJA) released by Shinya Otani, Takayoshi Makabe, Anuj Arora, and Kyo Hattori.
 1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
 1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released with the paper [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
@@ -316,7 +316,7 @@ conda install -c huggingface transformers
 1. **[M2M100](https://huggingface.co/docs/transformers/model_doc/m2m_100)** (from Facebook) released with the paper [Beyond English-Centric Multilingual Machine Translation](https://arxiv.org/abs/2010.11125) by Angela Fan, Shruti Bhosale, Holger Schwenk, Zhiyi Ma, Ahmed El-Kishky, Siddharth Goyal, Mandeep Baines, Onur Celebi, Guillaume Wenzek, Vishrav Chaudhary, Naman Goyal, Tom Birch, Vitaliy Liptchinsky, Sergey Edunov, Edouard Grave, Michael Auli, Armand Joulin.
 1. **[MarianMT](https://huggingface.co/docs/transformers/model_doc/marian)** Machine translation models trained using [OPUS](http://opus.nlpl.eu/) data by Jörg Tiedemann. The [Marian Framework](https://marian-nmt.github.io/) is being developed by the Microsoft Translator Team.
 1. **[MarkupLM](https://huggingface.co/docs/transformers/main/model_doc/markuplm)** (from Microsoft Research Asia) released with the paper [MarkupLM: Pre-training of Text and Markup Language for Visually-rich Document Understanding](https://arxiv.org/abs/2110.08518) by Junlong Li, Yiheng Xu, Lei Cui, Furu Wei.
-1. **[MaskFormer](https://huggingface.co/docs/transformers/model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov 
+1. **[MaskFormer](https://huggingface.co/docs/transformers/model_doc/maskformer)** (from Meta and UIUC) released with the paper [Per-Pixel Classification is Not All You Need for Semantic Segmentation](https://arxiv.org/abs/2107.06278) by Bowen Cheng, Alexander G. Schwing, Alexander Kirillov
 1. **[mBART](https://huggingface.co/docs/transformers/model_doc/mbart)** (from Facebook) released with the paper [Multilingual Denoising Pre-training for Neural Machine Translation](https://arxiv.org/abs/2001.08210) by Yinhan Liu, Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov, Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
 1. **[mBART-50](https://huggingface.co/docs/transformers/model_doc/mbart)** (from Facebook) released with the paper [Multilingual Translation with Extensible Multilingual Pretraining and Finetuning](https://arxiv.org/abs/2008.00401) by Yuqing Tang, Chau Tran, Xian Li, Peng-Jen Chen, Naman Goyal, Vishrav Chaudhary, Jiatao Gu, Angela Fan.
 1. **[Megatron-BERT](https://huggingface.co/docs/transformers/model_doc/megatron-bert)** (from NVIDIA) released with the paper [Megatron-LM: Training Multi-Billion Parameter Language Models Using Model Parallelism](https://arxiv.org/abs/1909.08053) by Mohammad Shoeybi, Mostofa Patwary, Raul Puri, Patrick LeGresley, Jared Casper and Bryan Catanzaro.
@@ -343,9 +343,9 @@ conda install -c huggingface transformers
 1. **[RAG](https://huggingface.co/docs/transformers/model_doc/rag)** (from Facebook) released with the paper [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401) by Patrick Lewis, Ethan Perez, Aleksandara Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, Douwe Kiela.
 1. **[REALM](https://huggingface.co/docs/transformers/model_doc/realm.html)** (from Google Research) released with the paper [REALM: Retrieval-Augmented Language Model Pre-Training](https://arxiv.org/abs/2002.08909) by Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat and Ming-Wei Chang.
 1. **[Reformer](https://huggingface.co/docs/transformers/model_doc/reformer)** (from Google Research) released with the paper [Reformer: The Efficient Transformer](https://arxiv.org/abs/2001.04451) by Nikita Kitaev, Łukasz Kaiser, Anselm Levskaya.
-1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár. 
+1. **[RegNet](https://huggingface.co/docs/transformers/model_doc/regnet)** (from META Research) released with the paper [Designing Network Design Space](https://arxiv.org/abs/2003.13678) by Ilija Radosavovic, Raj Prateek Kosaraju, Ross Girshick, Kaiming He, Piotr Dollár.
 1. **[RemBERT](https://huggingface.co/docs/transformers/model_doc/rembert)** (from Google Research) released with the paper [Rethinking embedding coupling in pre-trained language models](https://arxiv.org/pdf/2010.12821.pdf) by Hyung Won Chung, Thibault Févry, Henry Tsai, M. Johnson, Sebastian Ruder.
-1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun. 
+1. **[ResNet](https://huggingface.co/docs/transformers/model_doc/resnet)** (from Microsoft Research) released with the paper [Deep Residual Learning for Image Recognition](https://arxiv.org/abs/1512.03385) by Kaiming He, Xiangyu Zhang, Shaoqing Ren, Jian Sun.
 1. **[RoBERTa](https://huggingface.co/docs/transformers/model_doc/roberta)** (from Facebook), released together with the paper a [Robustly Optimized BERT Pretraining Approach](https://arxiv.org/abs/1907.11692) by Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, Veselin Stoyanov.
 1. **[RoFormer](https://huggingface.co/docs/transformers/model_doc/roformer)** (from ZhuiyiTechnology), released together with the paper a [RoFormer: Enhanced Transformer with Rotary Position Embedding](https://arxiv.org/pdf/2104.09864v1.pdf) by Jianlin Su and Yu Lu and Shengfeng Pan and Bo Wen and Yunfeng Liu.
 1. **[SegFormer](https://huggingface.co/docs/transformers/model_doc/segformer)** (from NVIDIA) released with the paper [SegFormer: Simple and Efficient Design for Semantic Segmentation with Transformers](https://arxiv.org/abs/2105.15203) by Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, Ping Luo.
@@ -361,17 +361,17 @@ conda install -c huggingface transformers
 1. **[T5v1.1](https://huggingface.co/docs/transformers/model_doc/t5v1.1)** (from Google AI) released with the paper [google-research/text-to-text-transfer-transformer](https://github.com/google-research/text-to-text-transfer-transformer/blob/main/released_checkpoints.md#t511) by Colin Raffel and Noam Shazeer and Adam Roberts and Katherine Lee and Sharan Narang and Michael Matena and Yanqi Zhou and Wei Li and Peter J. Liu.
 1. **[TAPAS](https://huggingface.co/docs/transformers/model_doc/tapas)** (from Google AI) released with the paper [TAPAS: Weakly Supervised Table Parsing via Pre-training](https://arxiv.org/abs/2004.02349) by Jonathan Herzig, Paweł Krzysztof Nowak, Thomas Müller, Francesco Piccinno and Julian Martin Eisenschlos.
 1. **[TAPEX](https://huggingface.co/docs/transformers/model_doc/tapex)** (from Microsoft Research) released with the paper [TAPEX: Table Pre-training via Learning a Neural SQL Executor](https://arxiv.org/abs/2107.07653) by Qian Liu, Bei Chen, Jiaqi Guo, Morteza Ziyadi, Zeqi Lin, Weizhu Chen, Jian-Guang Lou.
-1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace). 
-1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine 
+1. **[Time Series Transformer](https://huggingface.co/docs/transformers/main/model_doc/time_series_transformer)** (from HuggingFace).
+1. **[Trajectory Transformer](https://huggingface.co/docs/transformers/model_doc/trajectory_transformers)** (from the University of California at Berkeley) released with the paper [Offline Reinforcement Learning as One Big Sequence Modeling Problem](https://arxiv.org/abs/2106.02039) by Michael Janner, Qiyang Li, Sergey Levine
 1. **[Transformer-XL](https://huggingface.co/docs/transformers/model_doc/transfo-xl)** (from Google/CMU) released with the paper [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context](https://arxiv.org/abs/1901.02860) by Zihang Dai*, Zhilin Yang*, Yiming Yang, Jaime Carbonell, Quoc V. Le, Ruslan Salakhutdinov.
 1. **[TrOCR](https://huggingface.co/docs/transformers/model_doc/trocr)** (from Microsoft) released with the paper [TrOCR: Transformer-based Optical Character Recognition with Pre-trained Models](https://arxiv.org/abs/2109.10282) by Minghao Li, Tengchao Lv, Lei Cui, Yijuan Lu, Dinei Florencio, Cha Zhang, Zhoujun Li, Furu Wei.
-1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler 
+1. **[UL2](https://huggingface.co/docs/transformers/model_doc/ul2)** (from Google Research) released with the paper [Unifying Language Learning Paradigms](https://arxiv.org/abs/2205.05131v1) by Yi Tay, Mostafa Dehghani, Vinh Q. Tran, Xavier Garcia, Dara Bahri, Tal Schuster, Huaixiu Steven Zheng, Neil Houlsby, Donald Metzler
 1. **[UniSpeech](https://huggingface.co/docs/transformers/model_doc/unispeech)** (from Microsoft Research) released with the paper [UniSpeech: Unified Speech Representation Learning with Labeled and Unlabeled Data](https://arxiv.org/abs/2101.07597) by Chengyi Wang, Yu Wu, Yao Qian, Kenichi Kumatani, Shujie Liu, Furu Wei, Michael Zeng, Xuedong Huang.
 1. **[UniSpeechSat](https://huggingface.co/docs/transformers/model_doc/unispeech-sat)** (from Microsoft Research) released with the paper [UNISPEECH-SAT: UNIVERSAL SPEECH REPRESENTATION LEARNING WITH SPEAKER AWARE PRE-TRAINING](https://arxiv.org/abs/2110.05752) by Sanyuan Chen, Yu Wu, Chengyi Wang, Zhengyang Chen, Zhuo Chen, Shujie Liu, Jian Wu, Yao Qian, Furu Wei, Jinyu Li, Xiangzhan Yu.
 1. **[VAN](https://huggingface.co/docs/transformers/model_doc/van)** (from Tsinghua University and Nankai University) released with the paper [Visual Attention Network](https://arxiv.org/pdf/2202.09741.pdf) by Meng-Hao Guo, Cheng-Ze Lu, Zheng-Ning Liu, Ming-Ming Cheng, Shi-Min Hu.
 1. **[VideoMAE](https://huggingface.co/docs/transformers/model_doc/videomae)** (from Multimedia Computing Group, Nanjing University) released with the paper [VideoMAE: Masked Autoencoders are Data-Efficient Learners for Self-Supervised Video Pre-Training](https://arxiv.org/abs/2203.12602) by Zhan Tong, Yibing Song, Jue Wang, Limin Wang.
 1. **[ViLT](https://huggingface.co/docs/transformers/model_doc/vilt)** (from NAVER AI Lab/Kakao Enterprise/Kakao Brain) released with the paper [ViLT: Vision-and-Language Transformer Without Convolution or Region Supervision](https://arxiv.org/abs/2102.03334) by Wonjae Kim, Bokyung Son, Ildoo Kim.
-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. **[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. **[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/main/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.
@@ -381,7 +381,7 @@ conda install -c huggingface transformers
 1. **[WavLM](https://huggingface.co/docs/transformers/model_doc/wavlm)** (from Microsoft Research) released with the paper [WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing](https://arxiv.org/abs/2110.13900) by Sanyuan Chen, Chengyi Wang, Zhengyang Chen, Yu Wu, Shujie Liu, Zhuo Chen, Jinyu Li, Naoyuki Kanda, Takuya Yoshioka, Xiong Xiao, Jian Wu, Long Zhou, Shuo Ren, Yanmin Qian, Yao Qian, Jian Wu, Michael Zeng, Furu Wei.
 1. **[Whisper](https://huggingface.co/docs/transformers/main/model_doc/whisper)** (from OpenAI) released with the paper [Robust Speech Recognition via Large-Scale Weak Supervision](https://cdn.openai.com/papers/whisper.pdf) by Alec Radford, Jong Wook Kim, Tao Xu, Greg Brockman, Christine McLeavey, Ilya Sutskever.
 1. **[X-CLIP](https://huggingface.co/docs/transformers/model_doc/xclip)** (from Microsoft Research) released with the paper [Expanding Language-Image Pretrained Models for General Video Recognition](https://arxiv.org/abs/2208.02816) by Bolin Ni, Houwen Peng, Minghao Chen, Songyang Zhang, Gaofeng Meng, Jianlong Fu, Shiming Xiang, Haibin Ling.
-1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li. 
+1. **[XGLM](https://huggingface.co/docs/transformers/model_doc/xglm)** (From Facebook AI) released with the paper [Few-shot Learning with Multilingual Language Models](https://arxiv.org/abs/2112.10668) by Xi Victoria Lin, Todor Mihaylov, Mikel Artetxe, Tianlu Wang, Shuohui Chen, Daniel Simig, Myle Ott, Naman Goyal, Shruti Bhosale, Jingfei Du, Ramakanth Pasunuru, Sam Shleifer, Punit Singh Koura, Vishrav Chaudhary, Brian O'Horo, Jeff Wang, Luke Zettlemoyer, Zornitsa Kozareva, Mona Diab, Veselin Stoyanov, Xian Li.
 1. **[XLM](https://huggingface.co/docs/transformers/model_doc/xlm)** (from Facebook) released together with the paper [Cross-lingual Language Model Pretraining](https://arxiv.org/abs/1901.07291) by Guillaume Lample and Alexis Conneau.
 1. **[XLM-ProphetNet](https://huggingface.co/docs/transformers/model_doc/xlm-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. **[XLM-RoBERTa](https://huggingface.co/docs/transformers/model_doc/xlm-roberta)** (from Facebook AI), released together with the paper [Unsupervised Cross-lingual Representation Learning at Scale](https://arxiv.org/abs/1911.02116) by Alexis Conneau*, Kartikay Khandelwal*, Naman Goyal, Vishrav Chaudhary, Guillaume Wenzek, Francisco Guzmán, Edouard Grave, Myle Ott, Luke Zettlemoyer and Veselin Stoyanov.

docs/source/en/index.mdx

@@ -330,7 +330,7 @@ Flax), PyTorch, and/or TensorFlow.
 |          Wav2Vec2           |       ✅       |       ❌       |       ✅        |         ✅         |      ✅      |
 |     Wav2Vec2-Conformer      |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            WavLM            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
-|           Whisper           |       ✅       |       ❌       |       ✅        |         ❌         |      ❌      |
+|           Whisper           |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |
 |           X-CLIP            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |            XGLM             |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             XLM             |       ✅       |       ❌       |       ✅        |         ✅         |      ❌      |

docs/source/en/model_doc/whisper.mdx

@@ -23,11 +23,11 @@ The abstract from the paper is the following:
 
 Tips:
 
-- The model usually performs well without requiring any finetuning. 
+- The model usually performs well without requiring any finetuning.
 - The architecture follows a classic encoder-decoder architecture, which means that it relies on the [`~generation_utils.GenerationMixin.generate`] function for inference.
 - One can use [`WhisperProcessor`] to prepare audio for the model, and decode the predicted ID's back into text.
 
-This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ).
+This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ). The Tensorflow version of this model was contributed by [amyeroberts](https://huggingface.co/amyeroberts).
 The original code can be found [here](https://github.com/openai/whisper).
 
 
@@ -66,3 +66,14 @@ The original code can be found [here](https://github.com/openai/whisper).
 
 [[autodoc]] WhisperForConditionalGeneration
     - forward
+
+
+## TFWhisperModel
+
+[[autodoc]] TFWhisperModel
+    - call
+
+## TFWhisperForConditionalGeneration
+
+[[autodoc]] TFWhisperForConditionalGeneration
+    - call

src/transformers/__init__.py

@@ -2754,6 +2754,14 @@ else:
             "TFWav2Vec2PreTrainedModel",
         ]
     )
+    _import_structure["models.whisper"].extend(
+        [
+            "TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "TFWhisperForConditionalGeneration",
+            "TFWhisperModel",
+            "TFWhisperPreTrainedModel",
+        ]
+    )
     _import_structure["models.xglm"].extend(
         [
             "TF_XGLM_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -5303,6 +5311,12 @@ if TYPE_CHECKING:
             TFWav2Vec2Model,
             TFWav2Vec2PreTrainedModel,
         )
+        from .models.whisper import (
+            TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFWhisperForConditionalGeneration,
+            TFWhisperModel,
+            TFWhisperPreTrainedModel,
+        )
         from .models.xglm import (
             TF_XGLM_PRETRAINED_MODEL_ARCHIVE_LIST,
             TFXGLMForCausalLM,

src/transformers/generation_tf_logits_process.py

@@ -504,3 +504,84 @@ class TFForcedEOSTokenLogitsProcessor(TFLogitsProcessor):
                     axis=-1,
                 )
         return scores
+
+
+class TFSuppressTokensAtBeginLogitsProcessor(TFLogitsProcessor):
+    r"""
+    [`TFSuppressTokensAtBeginLogitsProcessor`] suppresses a list of tokens as soon as the `generate` function starts
+    generating using `begin_index` tokens. This should ensure that the tokens defined by `begin_suppress_tokens` at not
+    sampled at the begining of the generation.
+    """
+
+    def __init__(self, begin_suppress_tokens, begin_index):
+        self.begin_suppress_tokens = list(begin_suppress_tokens)
+        self.begin_index = begin_index
+
+    def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
+        scores = tf.cond(
+            tf.equal(cur_len, self.begin_index),
+            lambda: tf.tensor_scatter_nd_update(
+                scores,
+                indices=[[i, token] for i in range(scores.shape[0]) for token in self.begin_suppress_tokens],
+                updates=[-float("inf") for _ in range(scores.shape[0] * len(self.begin_suppress_tokens))],
+            ),
+            lambda: scores,
+        )
+        return scores
+
+
+class TFSuppressTokensLogitsProcessor(TFLogitsProcessor):
+    r"""This processor can be used to suppress a list of tokens. The processor will set their log probs to `-inf` so that they
+    are not sampled."""
+
+    def __init__(self, suppress_tokens):
+        self.suppress_tokens = list(suppress_tokens)
+
+    def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
+        scores = tf.tensor_scatter_nd_update(
+            scores,
+            indices=[[i, token] for i in range(scores.shape[0]) for token in self.suppress_tokens],
+            updates=[-float("inf") for _ in range(scores.shape[0] * len(self.suppress_tokens))],
+        )
+        return scores
+
+
+class TFForceTokensLogitsProcessor(TFLogitsProcessor):
+    r"""This processor can be used to force a list of tokens. The processor will set their log probs to `0` and all
+    other tokens to `-inf` so that they are sampled at their corresponding index."""
+
+    def __init__(self, force_token_map):
+        force_token_map = dict(force_token_map)
+        # Converts the dictionary of format {index: token} containing the tokens to be forced to an array, where the
+        # index of the array corresponds to the index of the token to be forced, for XLA compatibility.
+        # Indexes without forced tokens will have an negative value.
+        force_token_array = np.ones((max(force_token_map.keys()) + 1), dtype=np.int32) * -1
+        for index, token in force_token_map.items():
+            force_token_array[index] = token
+        self.force_token_array = tf.convert_to_tensor(force_token_array, dtype=tf.int32)
+
+    def __call__(self, input_ids: tf.Tensor, scores: tf.Tensor, cur_len: int) -> tf.Tensor:
+        def _force_token(generation_idx):
+            batch_size = scores.shape[0]
+            current_token = self.force_token_array[generation_idx]
+
+            new_scores = tf.ones_like(scores, dtype=scores.dtype) * -float("inf")
+            indices = tf.stack((tf.range(batch_size), tf.tile([current_token], [batch_size])), axis=1)
+            updates = tf.zeros((batch_size,), dtype=scores.dtype)
+            new_scores = tf.tensor_scatter_nd_update(new_scores, indices, updates)
+            return new_scores
+
+        scores = tf.cond(
+            tf.greater_equal(cur_len, tf.shape(self.force_token_array)[0]),
+            # If the current length is geq than the length of force_token_array, the processor does nothing.
+            lambda: tf.identity(scores),
+            # Otherwise, it may force a certain token.
+            lambda: tf.cond(
+                tf.greater_equal(self.force_token_array[cur_len], 0),
+                # Only valid (positive) tokens are forced
+                lambda: _force_token(cur_len),
+                # Otherwise, the processor does nothing.
+                lambda: scores,
+            ),
+        )
+        return scores

src/transformers/generation_tf_utils.py

@@ -26,11 +26,14 @@ from tensorflow.compiler.tf2xla.python.xla import dynamic_update_slice
 from .generation_tf_logits_process import (
     TFForcedBOSTokenLogitsProcessor,
     TFForcedEOSTokenLogitsProcessor,
+    TFForceTokensLogitsProcessor,
     TFLogitsProcessorList,
     TFMinLengthLogitsProcessor,
     TFNoBadWordsLogitsProcessor,
     TFNoRepeatNGramLogitsProcessor,
     TFRepetitionPenaltyLogitsProcessor,
+    TFSuppressTokensAtBeginLogitsProcessor,
+    TFSuppressTokensLogitsProcessor,
     TFTemperatureLogitsWarper,
     TFTopKLogitsWarper,
     TFTopPLogitsWarper,
@@ -401,6 +404,9 @@ class TFGenerationMixin:
         return_dict_in_generate=None,
         forced_bos_token_id=None,
         forced_eos_token_id=None,
+        suppress_tokens: Optional[List[int]] = None,
+        begin_suppress_tokens: Optional[List[int]] = None,
+        forced_decoder_ids: Optional[List[int]] = None,
         **model_kwargs,
     ) -> Union[TFGreedySearchOutput, TFSampleOutput, TFBeamSearchOutput, TFBeamSampleOutput, tf.Tensor]:
         r"""
@@ -494,6 +500,14 @@ class TFGenerationMixin:
                 the target language token.
             forced_eos_token_id (`int`, *optional*):
                 The id of the token to force as the last generated token when `max_length` is reached.
+            suppress_tokens  (`List[int]`, *optional*, defaults to `model.config.suppress_tokens`):
+                A list of tokens that will be supressed at generation. The `SupressTokens` logit processor will set
+                their log probs to `-inf` so that they are not sampled.
+            begin_suppress_tokens  (`List[int]`, *optional*, defaults to `model.config.begin_suppress_tokens`):
+                A list of tokens that will be supressed at the begining of the generation. The `SupressBeginTokens`
+                logit processor will set their log probs to `-inf` so that they are not sampled.
+            forced_decoder_ids (`List[int]`, *optional*, defaults to `model.config.forced_decoder_ids`):
+                A list of tokens that will be forced as beginning tokens, before sampling.
             model_specific_kwargs:
                 Additional model specific kwargs will be forwarded to the `forward` function of the model.
 
@@ -609,6 +623,9 @@ class TFGenerationMixin:
                 return_dict_in_generate=return_dict_in_generate,
                 forced_bos_token_id=forced_bos_token_id,
                 forced_eos_token_id=forced_eos_token_id,
+                suppress_tokens=suppress_tokens,
+                begin_suppress_tokens=begin_suppress_tokens,
+                forced_decoder_ids=forced_decoder_ids,
                 **model_kwargs,
             )
 
@@ -648,6 +665,12 @@ class TFGenerationMixin:
         forced_eos_token_id = (
             forced_eos_token_id if forced_eos_token_id is not None else self.config.forced_eos_token_id
         )
+        suppress_tokens = suppress_tokens if suppress_tokens is not None else self.config.suppress_tokens
+        begin_suppress_tokens = (
+            begin_suppress_tokens if begin_suppress_tokens is not None else self.config.begin_suppress_tokens
+        )
+        if forced_decoder_ids is None and hasattr(self.config, "forced_decoder_ids"):
+            forced_decoder_ids = self.config.forced_decoder_ids
 
         output_scores = output_scores if output_scores is not None else self.config.output_scores
         output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
@@ -1368,6 +1391,9 @@ class TFGenerationMixin:
         return_dict_in_generate=None,
         forced_bos_token_id=None,
         forced_eos_token_id=None,
+        suppress_tokens=None,
+        begin_suppress_tokens=None,
+        forced_decoder_ids=None,
         **model_kwargs,
     ) -> Union[TFGreedySearchOutput, TFSampleOutput, TFBeamSearchOutput, TFBeamSampleOutput, tf.Tensor]:
         r"""
@@ -1461,6 +1487,15 @@ class TFGenerationMixin:
                 the target language token.
             forced_eos_token_id (`int`, *optional*):
                 The id of the token to force as the last generated token when `max_length` is reached.
+            suppress_tokens  (`List[int]`, *optional*, defaults to `model.config.suppress_tokens`):
+                A list of tokens that will be supressed at generation. The `SupressTokens` logit processor will set
+                their log probs to `-inf` so that they are not sampled.
+            begin_suppress_tokens  (`List[int]`, *optional*, defaults to `model.config.begin_suppress_tokens`):
+                A list of tokens that will be supressed at the begining of the generation. The `SupressBeginTokens`
+                logit processor will set their log probs to `-inf` so that they are not sampled.
+            forced_decoder_ids (`List[int]`, *optional*, defaults to `model.config.forced_decoder_ids`):
+                A list of tokens that will be forced as beginning tokens.
+
             model_kwargs:
                 Additional model specific kwargs will be forwarded to the `call` function of the model.
 
@@ -1695,12 +1730,16 @@ class TFGenerationMixin:
         logits_processor = self._get_logits_processor(
             repetition_penalty=repetition_penalty,
             no_repeat_ngram_size=no_repeat_ngram_size,
+            input_ids_seq_length=input_ids_seq_length,
             bad_words_ids=bad_words_ids,
             min_length=min_length,
             max_length=max_length,
             eos_token_id=eos_token_id,
             forced_bos_token_id=forced_bos_token_id,
             forced_eos_token_id=forced_eos_token_id,
+            suppress_tokens=suppress_tokens,
+            begin_suppress_tokens=begin_suppress_tokens,
+            forced_decoder_ids=forced_decoder_ids,
         )
 
         # 9. go into different generation modes
@@ -1994,7 +2033,7 @@ class TFGenerationMixin:
         def _initialize_past(past, num_padding_values, batch_axis):
             """initialize past with zeros -- the structure depends on `batch_axis`"""
             if batch_axis == 0:
-                padding_values = tf.scatter_nd(indices=[[2, 1]], updates=[num_padding_values], shape=(4, 2))
+                padding_values = tf.constant([[0, 0], [0, 0], [0, num_padding_values], [0, 0]], dtype=tf.int32)
                 new_past = ()
                 for past_layer in past:
                     new_past_layer = list(past_layer)
@@ -2099,12 +2138,16 @@ class TFGenerationMixin:
         self,
         repetition_penalty: float,
         no_repeat_ngram_size: int,
+        input_ids_seq_length: int,
         bad_words_ids: List[List[int]],
         min_length: int,
         max_length: int,
         eos_token_id: int,
         forced_bos_token_id: int,
         forced_eos_token_id: int,
+        suppress_tokens: Optional[List[int]] = None,
+        begin_suppress_tokens: Optional[List[int]] = None,
+        forced_decoder_ids: Optional[List[int]] = None,
     ) -> TFLogitsProcessorList:
         """
         This class returns a [`TFLogitsProcessorList`] list object that contains all relevant [`TFLogitsProcessor`]
@@ -2118,6 +2161,12 @@ class TFGenerationMixin:
         )
         bad_words_ids = bad_words_ids if bad_words_ids is not None else self.config.bad_words_ids
         eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
+        suppress_tokens = suppress_tokens if suppress_tokens is not None else self.config.suppress_tokens
+        begin_suppress_tokens = (
+            begin_suppress_tokens if begin_suppress_tokens is not None else self.config.begin_suppress_tokens
+        )
+        if forced_decoder_ids is None and hasattr(self.config, "forced_decoder_ids"):
+            forced_decoder_ids = self.config.forced_decoder_ids
 
         # instantiate processors list
         if repetition_penalty is not None and repetition_penalty != 1.0:
@@ -2132,7 +2181,16 @@ class TFGenerationMixin:
             processors.append(TFForcedBOSTokenLogitsProcessor(forced_bos_token_id))
         if forced_eos_token_id is not None:
             processors.append(TFForcedEOSTokenLogitsProcessor(max_length, forced_eos_token_id))
-
+        if suppress_tokens is not None:
+            processors.append(TFSuppressTokensLogitsProcessor(suppress_tokens))
+        if begin_suppress_tokens is not None:
+            begin_index = input_ids_seq_length
+            begin_index = begin_index if (input_ids_seq_length > 1 or forced_bos_token_id is None) else begin_index + 1
+            if forced_decoder_ids is not None:
+                begin_index += forced_decoder_ids[-1][0]  # generation starts after the last token that is forced
+            processors.append(TFSuppressTokensAtBeginLogitsProcessor(begin_suppress_tokens, begin_index))
+        if forced_decoder_ids is not None:
+            processors.append(TFForceTokensLogitsProcessor(forced_decoder_ids))
         return processors
 
     def greedy_search(

src/transformers/models/auto/modeling_tf_auto.py

@@ -80,6 +80,7 @@ TF_MODEL_MAPPING_NAMES = OrderedDict(
         ("vit", "TFViTModel"),
         ("vit_mae", "TFViTMAEModel"),
         ("wav2vec2", "TFWav2Vec2Model"),
+        ("whisper", "TFWhisperModel"),
         ("xglm", "TFXGLMModel"),
         ("xlm", "TFXLMModel"),
         ("xlm-roberta", "TFXLMRobertaModel"),
@@ -145,6 +146,7 @@ TF_MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
         ("t5", "TFT5ForConditionalGeneration"),
         ("tapas", "TFTapasForMaskedLM"),
         ("transfo-xl", "TFTransfoXLLMHeadModel"),
+        ("whisper", "TFWhisperForConditionalGeneration"),
         ("xlm", "TFXLMWithLMHeadModel"),
         ("xlm-roberta", "TFXLMRobertaForMaskedLM"),
         ("xlnet", "TFXLNetLMHeadModel"),
@@ -253,6 +255,7 @@ TF_MODEL_FOR_SEQ_TO_SEQ_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
 TF_MODEL_FOR_SPEECH_SEQ_2_SEQ_MAPPING_NAMES = OrderedDict(
     [
         ("speech_to_text", "TFSpeech2TextForConditionalGeneration"),
+        ("whisper", "TFWhisperForConditionalGeneration"),
     ]
 )
 

src/transformers/models/rag/modeling_tf_rag.py

@@ -1262,6 +1262,7 @@ class TFRagTokenForGeneration(TFRagPreTrainedModel, TFCausalLanguageModelingLoss
                 eos_token_id=eos_token_id,
                 forced_bos_token_id=None,
                 forced_eos_token_id=None,
+                input_ids_seq_length=tf.shape(decoder_input_ids)[-1],
             )
             model_kwargs["attention_mask"] = context_attention_mask
 

src/transformers/models/whisper/__init__.py

@@ -17,7 +17,7 @@
 # limitations under the License.
 from typing import TYPE_CHECKING
 
-from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_torch_available
+from ...utils import OptionalDependencyNotAvailable, _LazyModule, is_tf_available, is_torch_available
 
 
 _import_structure = {
@@ -41,6 +41,18 @@ else:
         "WhisperPreTrainedModel",
     ]
 
+try:
+    if not is_tf_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_tf_whisper"] = [
+        "TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "TFWhisperForConditionalGeneration",
+        "TFWhisperModel",
+        "TFWhisperPreTrainedModel",
+    ]
 
 if TYPE_CHECKING:
     from .configuration_whisper import WHISPER_PRETRAINED_CONFIG_ARCHIVE_MAP, WhisperConfig
@@ -61,6 +73,19 @@ if TYPE_CHECKING:
             WhisperPreTrainedModel,
         )
 
+    try:
+        if not is_tf_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_tf_whisper import (
+            TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST,
+            TFWhisperForConditionalGeneration,
+            TFWhisperModel,
+            TFWhisperPreTrainedModel,
+        )
+
 else:
     import sys
 

src/transformers/models/whisper/feature_extraction_whisper.py

@@ -218,7 +218,6 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
         return_attention_mask: Optional[bool] = None,
         padding: Optional[str] = "max_length",
         max_length: Optional[int] = None,
-        sampling_rate: Optional[int] = None,
         **kwargs
     ) -> BatchFeature:
         """
@@ -262,19 +261,6 @@ class WhisperFeatureExtractor(SequenceFeatureExtractor):
                 The value that is used to fill the padding values / vectors.
         """
 
-        if sampling_rate is not None:
-            if sampling_rate != self.sampling_rate:
-                raise ValueError(
-                    f"The model corresponding to this feature extractor: {self} was trained using a sampling rate of"
-                    f" {self.sampling_rate}. Please make sure that the provided `raw_speech` input was sampled with"
-                    f" {self.sampling_rate} and not {sampling_rate}."
-                )
-        else:
-            logger.warning(
-                "It is strongly recommended to pass the `sampling_rate` argument to this function. "
-                "Failing to do so can result in silent errors that might be hard to debug."
-            )
-
         is_batched = bool(
             isinstance(raw_speech, (list, tuple))
             and (isinstance(raw_speech[0], np.ndarray) or isinstance(raw_speech[0], (tuple, list)))

src/transformers/models/whisper/modeling_tf_whisper.py

+# coding=utf-8
+# Copyright 2022 The OpenAI Authors and The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" TensorFlow Whisper model."""
+
+
+import math
+import random
+from typing import Dict, Optional, Tuple
+
+import tensorflow as tf
+
+from ...activations_tf import get_tf_activation
+from ...modeling_tf_outputs import (
+    TFBaseModelOutput,
+    TFBaseModelOutputWithPastAndCrossAttentions,
+    TFSeq2SeqLMOutput,
+    TFSeq2SeqModelOutput,
+)
+from ...modeling_tf_utils import TFCausalLanguageModelingLoss, TFPreTrainedModel, keras_serializable, unpack_inputs
+from ...tf_utils import shape_list, stable_softmax
+from ...utils import add_start_docstrings, add_start_docstrings_to_model_forward, logging, replace_return_docstrings
+from .configuration_whisper import WhisperConfig
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "WhisperConfig"
+
+
+TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST = [
+    "openai/whisper-base",
+    # See all Whisper models at https://huggingface.co/models?filter=whisper
+]
+
+LARGE_NEGATIVE = -1e8
+
+
+# Copied from transformers.models.bart.modeling_tf_bart.shift_tokens_right
+def shift_tokens_right(input_ids: tf.Tensor, pad_token_id: int, decoder_start_token_id: int):
+    pad_token_id = tf.cast(pad_token_id, input_ids.dtype)
+    decoder_start_token_id = tf.cast(decoder_start_token_id, input_ids.dtype)
+    start_tokens = tf.fill(
+        (shape_list(input_ids)[0], 1), tf.convert_to_tensor(decoder_start_token_id, input_ids.dtype)
+    )
+    shifted_input_ids = tf.concat([start_tokens, input_ids[:, :-1]], -1)
+    # replace possible -100 values in labels by `pad_token_id`
+    shifted_input_ids = tf.where(
+        shifted_input_ids == -100,
+        tf.fill(shape_list(shifted_input_ids), tf.convert_to_tensor(pad_token_id, input_ids.dtype)),
+        shifted_input_ids,
+    )
+
+    # "Verify that `labels` has only positive values and -100"
+    assert_gte0 = tf.debugging.assert_greater_equal(shifted_input_ids, tf.constant(0, dtype=input_ids.dtype))
+
+    # Make sure the assertion op is called by wrapping the result in an identity no-op
+    with tf.control_dependencies([assert_gte0]):
+        shifted_input_ids = tf.identity(shifted_input_ids)
+
+    return shifted_input_ids
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._make_causal_mask
+def _make_causal_mask(input_ids_shape: tf.TensorShape, past_key_values_length: int = 0):
+    """
+    Make causal mask used for bi-directional self-attention.
+    """
+    bsz = input_ids_shape[0]
+    tgt_len = input_ids_shape[1]
+    mask = tf.ones((tgt_len, tgt_len)) * LARGE_NEGATIVE
+    mask_cond = tf.range(shape_list(mask)[-1])
+
+    mask = tf.where(mask_cond < tf.reshape(mask_cond + 1, (shape_list(mask)[-1], 1)), 0.0, mask)
+
+    if past_key_values_length > 0:
+        mask = tf.concat([tf.zeros((tgt_len, past_key_values_length)), mask], axis=-1)
+
+    return tf.tile(mask[None, None, :, :], (bsz, 1, 1, 1))
+
+
+# Copied from transformers.models.bart.modeling_tf_bart._expand_mask
+def _expand_mask(mask: tf.Tensor, tgt_len: Optional[int] = None):
+    """
+    Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`.
+    """
+    src_len = shape_list(mask)[1]
+    tgt_len = tgt_len if tgt_len is not None else src_len
+    one_cst = tf.constant(1.0)
+    mask = tf.cast(mask, dtype=one_cst.dtype)
+    expanded_mask = tf.tile(mask[:, None, None, :], (1, 1, tgt_len, 1))
+
+    return (one_cst - expanded_mask) * LARGE_NEGATIVE
+
+
+class TFWhisperPositionalEmbedding(tf.keras.layers.Layer):
+    def __init__(self, num_positions: int, embedding_dim: int, padding_idx: Optional[int] = None, **kwargs):
+        super().__init__(**kwargs)
+        self.num_positions = num_positions
+        self.embedding_dim = embedding_dim
+        self.padding_idx = padding_idx
+
+    def build(self, input_shape):
+        self.weight = self.add_weight(
+            name="weight",
+            shape=[self.num_positions, self.embedding_dim],
+            trainable=True,
+        )
+        super().build(input_shape)
+
+    def call(self, input_ids, past_key_values_length=0):
+        past_key_values_length = tf.cast(past_key_values_length, tf.int32)
+        gather_indices = tf.range(tf.shape(input_ids)[-1], delta=1) + past_key_values_length
+        return tf.gather(self.weight, gather_indices)
+
+
+class TFWhisperAttention(tf.keras.layers.Layer):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    def __init__(
+        self,
+        embed_dim: int,
+        num_heads: int,
+        dropout: float = 0.0,
+        is_decoder: bool = False,
+        bias: bool = True,
+        **kwargs
+    ):
+        super().__init__(**kwargs)
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+        self.dropout = tf.keras.layers.Dropout(dropout)
+        self.head_dim = embed_dim // num_heads
+
+        if (self.head_dim * num_heads) != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim}"
+                f" and `num_heads`: {num_heads})."
+            )
+        self.scaling = self.head_dim**-0.5
+        self.is_decoder = is_decoder
+
+        self.k_proj = tf.keras.layers.Dense(embed_dim, use_bias=False, name="k_proj")
+        self.v_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="v_proj")
+        self.q_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="q_proj")
+        self.out_proj = tf.keras.layers.Dense(embed_dim, use_bias=bias, name="out_proj")
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention._shape with BART->whisper
+    def _shape(self, tensor: tf.Tensor, seq_len: int, bsz: int):
+        return tf.transpose(tf.reshape(tensor, (bsz, seq_len, self.num_heads, self.head_dim)), (0, 2, 1, 3))
+
+    # Copied from transformers.models.bart.modeling_tf_bart.TFBartAttention.call with BART->whisper
+    def call(
+        self,
+        hidden_states: tf.Tensor,
+        key_value_states: Optional[tf.Tensor] = None,
+        past_key_value: Optional[Tuple[Tuple[tf.Tensor]]] = None,
+        attention_mask: Optional[tf.Tensor] = None,
+        layer_head_mask: Optional[tf.Tensor] = None,
+        training: Optional[bool] = False,
+    ) -> Tuple[tf.Tensor, Optional[tf.Tensor]]:
+        """Input shape: Batch x Time x Channel"""
+
+        # if key_value_states are provided this layer is used as a cross-attention layer
+        # for the decoder
+        is_cross_attention = key_value_states is not None
+        bsz, tgt_len, embed_dim = shape_list(hidden_states)
+
+        # get query proj
+        query_states = self.q_proj(hidden_states) * self.scaling
+        # get key, value proj
+        if is_cross_attention and past_key_value is not None:
+            # reuse k,v, cross_attentions
+            key_states = past_key_value[0]
+            value_states = past_key_value[1]
+        elif is_cross_attention:
+            # cross_attentions
+            key_states = self._shape(self.k_proj(key_value_states), -1, bsz)
+            value_states = self._shape(self.v_proj(key_value_states), -1, bsz)
+        elif past_key_value is not None:
+            # reuse k, v, self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+            key_states = tf.concat([past_key_value[0], key_states], axis=2)
+            value_states = tf.concat([past_key_value[1], value_states], axis=2)
+        else:
+            # self_attention
+            key_states = self._shape(self.k_proj(hidden_states), -1, bsz)
+            value_states = self._shape(self.v_proj(hidden_states), -1, bsz)
+
+        if self.is_decoder:
+            # if cross_attention save Tuple(tf.Tensor, tf.Tensor) of all cross attention key/value_states.
+            # Further calls to cross_attention layer can then reuse all cross-attention
+            # key/value_states (first "if" case)
+            # if uni-directional self-attention (decoder) save Tuple(tf.Tensor, tf.Tensor) of
+            # all previous decoder key/value_states. Further calls to uni-directional self-attention
+            # can concat previous decoder key/value_states to current projected key/value_states (third "elif" case)
+            # if encoder bi-directional self-attention `past_key_value` is always `None`
+            past_key_value = (key_states, value_states)
+
+        proj_shape = (bsz * self.num_heads, -1, self.head_dim)
+        query_states = tf.reshape(self._shape(query_states, tgt_len, bsz), proj_shape)
+        key_states = tf.reshape(key_states, proj_shape)
+        value_states = tf.reshape(value_states, proj_shape)
+
+        src_len = shape_list(key_states)[1]
+        attn_weights = tf.matmul(query_states, key_states, transpose_b=True)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_weights),
+            [bsz * self.num_heads, tgt_len, src_len],
+            message=(
+                f"Attention weights should be of size {(bsz * self.num_heads, tgt_len, src_len)}, but is"
+                f" {shape_list(attn_weights)}"
+            ),
+        )
+
+        if attention_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(attention_mask),
+                [bsz, 1, tgt_len, src_len],
+                message=(
+                    f"Attention mask should be of size {(bsz, 1, tgt_len, src_len)}, but is"
+                    f" {shape_list(attention_mask)}"
+                ),
+            )
+
+            attention_mask = tf.cast(attention_mask, dtype=attn_weights.dtype)
+            attn_weights = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len)) + attention_mask
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_weights = stable_softmax(attn_weights, axis=-1)
+
+        if layer_head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(layer_head_mask),
+                [self.num_heads],
+                message=(
+                    f"Head mask for a single layer should be of size {(self.num_heads)}, but is"
+                    f" {shape_list(layer_head_mask)}"
+                ),
+            )
+
+            attn_weights = tf.reshape(layer_head_mask, (1, -1, 1, 1)) * tf.reshape(
+                attn_weights, (bsz, self.num_heads, tgt_len, src_len)
+            )
+            attn_weights = tf.reshape(attn_weights, (bsz * self.num_heads, tgt_len, src_len))
+
+        attn_probs = self.dropout(attn_weights, training=training)
+        attn_output = tf.matmul(attn_probs, value_states)
+
+        tf.debugging.assert_equal(
+            shape_list(attn_output),
+            [bsz * self.num_heads, tgt_len, self.head_dim],
+            message=(
+                f"`attn_output` should be of size {(bsz, self.num_heads, tgt_len, self.head_dim)}, but is"
+                f" {shape_list(attn_output)}"
+            ),
+        )
+
+        attn_output = tf.transpose(
+            tf.reshape(attn_output, (bsz, self.num_heads, tgt_len, self.head_dim)), (0, 2, 1, 3)
+        )
+        attn_output = tf.reshape(attn_output, (bsz, tgt_len, embed_dim))
+
+        attn_output = self.out_proj(attn_output)
+        attn_weights: tf.Tensor = tf.reshape(attn_weights, (bsz, self.num_heads, tgt_len, src_len))
+
+        return attn_output, attn_weights, past_key_value
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextEncoderLayer with Speech2Text->Whisper
+class TFWhisperEncoderLayer(tf.keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+        self.self_attn = TFWhisperAttention(
+            self.embed_dim, config.encoder_attention_heads, dropout=config.attention_dropout, name="self_attn"
+        )
+        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)
+        self.fc1 = tf.keras.layers.Dense(config.encoder_ffn_dim, name="fc1")
+        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+
+    def call(
+        self, hidden_states: tf.Tensor, attention_mask: tf.Tensor, layer_head_mask: tf.Tensor, training: bool = False
+    ):
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(encoder_attention_heads,)`
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+        hidden_states, self_attn_weights, _ = self.self_attn(
+            hidden_states=hidden_states,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+
+        tf.debugging.assert_equal(
+            shape_list(hidden_states),
+            shape_list(residual),
+            message=f"Self attn modified the shape of query {shape_list(residual)} to {shape_list(hidden_states)}",
+        )
+
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return hidden_states, self_attn_weights
+
+
+# Copied from transformers.models.speech_to_text.modeling_tf_speech_to_text.TFSpeech2TextDecoderLayer with Speech2Text->Whisper
+class TFWhisperDecoderLayer(tf.keras.layers.Layer):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.embed_dim = config.d_model
+
+        self.self_attn = TFWhisperAttention(
+            embed_dim=self.embed_dim,
+            num_heads=config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="self_attn",
+            is_decoder=True,
+        )
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.activation_fn = get_tf_activation(config.activation_function)
+        self.activation_dropout = tf.keras.layers.Dropout(config.activation_dropout)
+
+        self.self_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="self_attn_layer_norm")
+        self.encoder_attn = TFWhisperAttention(
+            self.embed_dim,
+            config.decoder_attention_heads,
+            dropout=config.attention_dropout,
+            name="encoder_attn",
+            is_decoder=True,
+        )
+        self.encoder_attn_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="encoder_attn_layer_norm")
+        self.fc1 = tf.keras.layers.Dense(config.decoder_ffn_dim, name="fc1")
+        self.fc2 = tf.keras.layers.Dense(self.embed_dim, name="fc2")
+        self.final_layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="final_layer_norm")
+
+    def call(
+        self,
+        hidden_states,
+        attention_mask: Optional[tf.Tensor] = None,
+        encoder_hidden_states: Optional[tf.Tensor] = None,
+        encoder_attention_mask: Optional[tf.Tensor] = None,
+        layer_head_mask: Optional[tf.Tensor] = None,
+        cross_attn_layer_head_mask: Optional[tf.Tensor] = None,
+        past_key_value: Optional[Tuple[tf.Tensor]] = None,
+        training=False,
+    ) -> Tuple[tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]:
+        """
+        Args:
+            hidden_states (`tf.Tensor`): input to the layer of shape `(seq_len, batch, embed_dim)`
+            attention_mask (`tf.Tensor`): attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            encoder_hidden_states (`tf.Tensor`):
+                cross attention input to the layer of shape `(seq_len, batch, embed_dim)`
+            encoder_attention_mask (`tf.Tensor`): encoder attention mask of size
+                `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values.
+            layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size
+                `(decoder_attention_heads,)`
+            cross_attn_layer_head_mask (`tf.Tensor`): mask for heads of the cross-attention module.
+                `(decoder_attention_heads,)`
+            past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states
+        """
+        residual = hidden_states
+        hidden_states = self.self_attn_layer_norm(hidden_states)
+
+        # Self Attention
+        # decoder uni-directional self-attention cached key/values tuple is at positions 1,2
+        self_attn_past_key_value = past_key_value[:2] if past_key_value is not None else None
+        # add present self-attn cache to positions 1,2 of present_key_value tuple
+        hidden_states, self_attn_weights, present_key_value = self.self_attn(
+            hidden_states=hidden_states,
+            past_key_value=self_attn_past_key_value,
+            attention_mask=attention_mask,
+            layer_head_mask=layer_head_mask,
+            training=training,
+        )
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        # Cross-Attention Block
+        cross_attn_present_key_value = None
+        cross_attn_weights = None
+        if encoder_hidden_states is not None:
+            residual = hidden_states
+            hidden_states = self.encoder_attn_layer_norm(hidden_states)
+
+            # cross_attn cached key/values tuple is at positions 3,4 of present_key_value tuple
+            cross_attn_past_key_value = past_key_value[-2:] if past_key_value is not None else None
+            hidden_states, cross_attn_weights, cross_attn_present_key_value = self.encoder_attn(
+                hidden_states=hidden_states,
+                key_value_states=encoder_hidden_states,
+                attention_mask=encoder_attention_mask,
+                layer_head_mask=cross_attn_layer_head_mask,
+                past_key_value=cross_attn_past_key_value,
+                training=training,
+            )
+            hidden_states = self.dropout(hidden_states, training=training)
+            hidden_states = residual + hidden_states
+
+            # add cross-attn to positions 3,4 of present_key_value tuple
+            present_key_value = present_key_value + cross_attn_present_key_value
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.final_layer_norm(hidden_states)
+        hidden_states = self.activation_fn(self.fc1(hidden_states))
+        hidden_states = self.activation_dropout(hidden_states, training=training)
+        hidden_states = self.fc2(hidden_states)
+        hidden_states = self.dropout(hidden_states, training=training)
+        hidden_states = residual + hidden_states
+
+        return (
+            hidden_states,
+            self_attn_weights,
+            cross_attn_weights,
+            present_key_value,
+        )
+
+
+class TFWhisperPreTrainedModel(TFPreTrainedModel):
+    config_class = WhisperConfig
+    base_model_prefix = "model"
+    main_input_name = "input_features"
+
+    def _get_feat_extract_output_lengths(self, input_lengths: tf.Tensor) -> int:
+        """
+        Computes the output length of the convolutional layers
+        """
+        input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+    @property
+    def dummy_inputs(self) -> Dict[str, tf.Tensor]:
+        """
+        Dummy inputs to build the network.
+
+        Returns:
+            `Dict[str, tf.Tensor]`: The dummy inputs.
+        """
+        return {
+            self.main_input_name: tf.random.uniform(
+                [2, self.config.num_mel_bins, self.config.max_source_positions * 2 - 1], dtype=tf.float32
+            ),
+            "decoder_input_ids": tf.constant([[2, 3]], dtype=tf.int64),
+        }
+
+    @tf.function(
+        input_signature=[
+            {
+                "input_features": tf.TensorSpec((None, None, None), tf.float32, name="input_features"),
+                "decoder_input_ids": tf.TensorSpec((None, None), tf.int64, name="decoder_input_ids"),
+                "decoder_attention_mask": tf.TensorSpec((None, None), tf.int64, name="decoder_attention_mask"),
+            }
+        ]
+    )
+    def serving(self, inputs):
+        output = self.call(inputs)
+        return self.serving_output(output)
+
+
+WHISPER_START_DOCSTRING = r"""
+    This model inherits from [`TFPreTrainedModel`]. Check the superclass documentation for the generic methods the
+    library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+    etc.)
+
+    This model is also a [tf.keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it
+    as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and
+    behavior.
+
+    Parameters:
+        config ([`WhisperConfig`]):
+            Model configuration class with all the parameters of the model. Initializing with a config file does not
+            load the weights associated with the model, only the configuration. Check out the
+            [`~TFPreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+WHISPER_INPUTS_DOCSTRING = r"""
+    Args:
+        input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+            Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be obtained
+            by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a `numpy.ndarray`, *e.g.*
+            via the soundfile library (`pip install soundfile`). To prepare the array into `input_features`, the
+            [`WhisperFeatureExtractor`] should be used for extracting the fbank features, padding and conversion into a
+            tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+        decoder_input_ids (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Indices of decoder input sequence tokens in the vocabulary.
+
+            Indices can be obtained using [`SpeechToTextTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+            [`PreTrainedTokenizer.__call__`] for details.
+
+            [What are decoder input IDs?](../glossary#decoder-input-ids)
+
+            SpeechToText uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If
+            `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+            `past_key_values`).
+        decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*):
+            Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also
+            be used by default.
+
+            If you want to change padding behavior, you should read
+            [`modeling_whisper._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the
+            paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy.
+        head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the encoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        decoder_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the attention modules in the decoder. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+            Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`:
+
+            - 1 indicates the head is **not masked**,
+            - 0 indicates the head is **masked**.
+
+        encoder_outputs (`tuple(tuple(tf.Tensor)`, *optional*):
+            Tuple consists of (`last_hidden_state`, *optional*: `hidden_states`, *optional*: `attentions`)
+            `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)`, *optional*) is a sequence of
+            hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder.
+        past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+            Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+            `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+            `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+            Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+            blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+            If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+            don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+            `decoder_input_ids` of shape `(batch_size, sequence_length)`. decoder_inputs_embeds (`tf.Tensor` of shape
+            `(batch_size, target_sequence_length, hidden_size)`, *optional*): Optionally, instead of passing
+            `decoder_input_ids` you can choose to directly pass an embedded representation. If `past_key_values` is
+            used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is
+            useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors
+            than the model's internal embedding lookup matrix.
+
+            If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value
+            of `inputs_embeds`.
+        use_cache (`bool`, *optional*):
+            If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+            `past_key_values`).
+        output_attentions (`bool`, *optional*):
+            Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+            tensors for more detail.
+        output_hidden_states (`bool`, *optional*):
+            Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+            more detail.
+        return_dict (`bool`, *optional*):
+            Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+"""
+
+
+@keras_serializable
+class TFWhisperEncoder(tf.keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a
+    [`TFWhisperEncoderLayer`].
+
+    Args:
+        config: WhisperConfig
+        embed_tokens (TFWhisperEmbedding): output embedding
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.layerdrop = config.encoder_layerdrop
+
+        self.embed_dim = config.d_model
+        self.num_mel_bins = config.num_mel_bins
+        self.padding_idx = config.pad_token_id
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(self.embed_dim) if config.scale_embedding else 1.0
+
+        # Padding is added in call() to match the PyTorch implementation
+        self.conv1 = tf.keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=1, padding="valid", name="conv1")
+        self.conv2 = tf.keras.layers.Conv1D(self.embed_dim, kernel_size=3, strides=2, padding="valid", name="conv2")
+
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            self.max_source_positions, self.embed_dim, name="embed_positions"
+        )
+
+        self.encoder_layers = [TFWhisperEncoderLayer(config, name=f"layers.{i}") for i in range(config.encoder_layers)]
+        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        head_mask=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_features (`tf.Tensor` of shape `(batch_size, feature_size, sequence_length)`):
+                Float values of fbank features extracted from the raw speech waveform. Raw speech waveform can be
+                obtained by loading a `.flac` or `.wav` audio file into an array of type `List[float]` or a
+                `numpy.ndarray`, *e.g.* via the soundfile library (`pip install soundfile`). To prepare the array into
+                `input_features`, the [`WhisperFeatureExtractor`] should be used for extracting the fbank features,
+                padding and conversion into a tensor of type `tf.Tensor`. See [`~WhisperFeatureExtractor.__call__`]
+            head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # TF 2.0 layers can't use channels first format when running on CPU.
+        input_features = tf.transpose(input_features, perm=(0, 2, 1))
+        input_features = tf.pad(input_features, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = tf.keras.activations.gelu(self.conv1(input_features))
+        inputs_embeds = tf.pad(inputs_embeds, [[0, 0], [1, 1], [0, 0]])
+        inputs_embeds = tf.keras.activations.gelu(self.conv2(inputs_embeds))
+        inputs_embeds = tf.transpose(inputs_embeds, perm=(0, 1, 2))
+
+        embed_pos = self.embed_positions(input_ids=tf.zeros((1, self.max_source_positions), dtype=tf.int32))
+
+        hidden_states = inputs_embeds + embed_pos
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        encoder_states = () if output_hidden_states else None
+        all_attentions = () if output_attentions else None
+
+        # check if head_mask has a correct number of layers specified if desired
+        if head_mask is not None:
+            tf.debugging.assert_equal(
+                shape_list(head_mask)[0],
+                len(self.encoder_layers),
+                message=(
+                    f"The head_mask should be specified for {len(self.encoder_layers)} layers, but it is for"
+                    f" {shape_list(head_mask)[0]}."
+                ),
+            )
+
+        for idx, encoder_layer in enumerate(self.encoder_layers):
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):  # skip the layer
+                continue
+
+            hidden_states, attn = encoder_layer(
+                hidden_states,
+                None,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                training=training,
+            )
+
+            if output_attentions:
+                all_attentions += (attn,)
+
+        hidden_states = self.layer_norm(hidden_states)
+        if output_hidden_states:
+            encoder_states = encoder_states + (hidden_states,)
+
+        if not return_dict:
+            return tuple(v for v in [hidden_states, encoder_states, all_attentions] if v is not None)
+        return TFBaseModelOutput(
+            last_hidden_state=hidden_states, hidden_states=encoder_states, attentions=all_attentions
+        )
+
+
+@keras_serializable
+class TFWhisperDecoder(tf.keras.layers.Layer):
+    config_class = WhisperConfig
+    """
+    Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`TFWhisperDecoderLayer`]
+
+    Args:
+        config: WhisperConfig
+    """
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.dropout = tf.keras.layers.Dropout(config.dropout)
+        self.layerdrop = config.decoder_layerdrop
+        self.padding_idx = config.pad_token_id
+        self.max_target_positions = config.max_target_positions
+        self.max_source_positions = config.max_source_positions
+        self.embed_scale = math.sqrt(config.d_model) if config.scale_embedding else 1.0
+
+        self.embed_tokens = tf.keras.layers.Embedding(
+            input_dim=config.vocab_size,
+            output_dim=config.d_model,
+            embeddings_initializer=tf.keras.initializers.TruncatedNormal(stddev=self.config.init_std),
+            name="embed_tokens",
+        )
+        self.embed_positions = TFWhisperPositionalEmbedding(
+            self.max_target_positions, config.d_model, name="embed_positions"
+        )
+
+        self.decoder_layers = [TFWhisperDecoderLayer(config, name=f"layers.{i}") for i in range(config.decoder_layers)]
+
+        self.layer_norm = tf.keras.layers.LayerNormalization(epsilon=1e-5, name="layer_norm")
+
+    def get_input_embeddings(self):
+        return self.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.embed_tokens = value
+
+    def _prepare_decoder_attention_mask(self, attention_mask, input_shape, past_key_values_length):
+        # create causal mask
+        # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+        batch_size, seq_len = input_shape[0], input_shape[1]
+
+        combined_attention_mask = tf.cond(
+            tf.math.greater(seq_len, 1),
+            lambda: _make_causal_mask(input_shape, past_key_values_length=past_key_values_length),
+            lambda: _expand_mask(tf.ones((batch_size, seq_len + past_key_values_length)), tgt_len=seq_len),
+        )
+
+        if attention_mask is not None:
+            attention_mask = tf.cond(
+                tf.greater(tf.shape(attention_mask)[-1], seq_len) & tf.greater(seq_len, 0),
+                lambda: attention_mask[:, : seq_len + past_key_values_length],
+                lambda: attention_mask,
+            )
+            # [bsz, seq_len] -> [bsz, 1, tgt_seq_len, src_seq_len]
+            expanded_attn_mask = _expand_mask(attention_mask, tgt_len=input_shape[-1])
+            combined_attention_mask = (
+                expanded_attn_mask if combined_attention_mask is None else expanded_attn_mask + combined_attention_mask
+            )
+        return combined_attention_mask
+
+    @unpack_inputs
+    def call(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        position_ids=None,
+        encoder_hidden_states=None,
+        head_mask=None,
+        cross_attn_head_mask=None,
+        past_key_values=None,
+        inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Args:
+            input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`):
+                Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you
+                provide it.
+
+                Indices can be obtained using [`WhisperTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+                [`PreTrainedTokenizer.__call__`] for details.
+
+                [What are input IDs?](../glossary#input-ids)
+            attention_mask (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+                - 1 for tokens that are **not masked**,
+                - 0 for tokens that are **masked**.
+
+                [What are attention masks?](../glossary#attention-mask)
+            position_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+                Indices of positions of each decoder input sequence tokens in the position embeddings. Selected in the
+                range `[0, config.max_position_embeddings - 1]`.
+            encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*):
+                Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention
+                of the decoder.
+            head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            cross_attn_head_mask (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*):
+                Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention
+                on hidden heads. Mask values selected in `[0, 1]`:
+
+                - 1 indicates the head is **not masked**,
+                - 0 indicates the head is **masked**.
+
+            past_key_values (`tuple(tuple(tf.Tensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+                Tuple of `tuple(tf.Tensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+                `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+                `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+                Contains pre-computed hidden-states (key and values in the self-attention blocks and in the
+                cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+                If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those
+                that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of
+                all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`tf.Tensor` of shape
+                `(batch_size, sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `input_ids`
+                you can choose to directly pass an embedded representation. This is useful if you want more control
+                over how to convert `input_ids` indices into associated vectors than the model's internal embedding
+                lookup matrix.
+            output_attentions (`bool`, *optional*):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more detail.
+            output_hidden_states (`bool`, *optional*):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more detail.
+            return_dict (`bool`, *optional*):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+        """
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # retrieve input_ids and inputs_embeds
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both decoder_input_ids and decoder_inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = tf.shape(input_ids)
+            input_ids = tf.reshape(input_ids, (-1, input_shape[-1]))
+        elif inputs_embeds is not None:
+            input_shape = tf.shape(inputs_embeds)[:-1]
+        else:
+            raise ValueError("You have to specify either decoder_input_ids or decoder_inputs_embeds")
+
+        # past_key_values_length
+        past_key_values_length = past_key_values[0][0].shape[2] if past_key_values is not None else 0
+
+        if inputs_embeds is None:
+            # Note: tf.gather, on which the embedding layer is based, won't check positive out of bound
+            # indices on GPU, returning zeros instead. This is a dangerous silent behavior.
+            tf.debugging.assert_less(
+                input_ids,
+                tf.cast(self.embed_tokens.input_dim, dtype=input_ids.dtype),
+                message=(
+                    "input_ids must be smaller than the embedding layer's input dimension (got"
+                    f" {tf.math.reduce_max(input_ids)} >= {self.embed_tokens.input_dim})"
+                ),
+            )
+            inputs_embeds = self.embed_tokens(input_ids)
+
+        attention_mask = self._prepare_decoder_attention_mask(attention_mask, input_shape, past_key_values_length)
+
+        # embed positions
+        filled_past_positions = past_key_values_length if position_ids is None else position_ids[0, -1]
+        positions = self.embed_positions(input_ids, past_key_values_length=filled_past_positions)
+
+        hidden_states = inputs_embeds + positions
+        hidden_states = self.dropout(hidden_states, training=training)
+
+        # decoder layers
+        all_hidden_states = () if output_hidden_states else None
+        all_self_attns = () if output_attentions else None
+        all_cross_attentions = () if (output_attentions and encoder_hidden_states is not None) else None
+        next_decoder_cache = () if use_cache else None
+
+        # check if head_mask/cross_attn_head_mask has a correct number of layers specified if desired
+        for attn_mask_name, attn_mask in [("head_mask", head_mask), ("cross_attn_head_mask", cross_attn_head_mask)]:
+            if attn_mask is not None:
+                tf.debugging.assert_equal(
+                    shape_list(attn_mask)[0],
+                    len(self.decoder_layers),
+                    message=(
+                        f"The {attn_mask_name} should be specified for {len(self.decoder_layers)} layers, but it is"
+                        f" for {shape_list(attn_mask)[0]}."
+                    ),
+                )
+
+        for idx, decoder_layer in enumerate(self.decoder_layers):
+            # add LayerDrop (see https://arxiv.org/abs/1909.11556 for description)
+            if output_hidden_states:
+                all_hidden_states += (hidden_states,)
+            dropout_probability = random.uniform(0, 1)
+            if training and (dropout_probability < self.layerdrop):
+                continue
+
+            past_key_value = past_key_values[idx] if past_key_values is not None else None
+
+            layer_outputs = decoder_layer(
+                hidden_states,
+                attention_mask=attention_mask,
+                encoder_hidden_states=encoder_hidden_states,
+                layer_head_mask=(head_mask[idx] if head_mask is not None else None),
+                cross_attn_layer_head_mask=(cross_attn_head_mask[idx] if cross_attn_head_mask is not None else None),
+                past_key_value=past_key_value,
+                training=training,
+            )
+            hidden_states = layer_outputs[0]
+
+            if use_cache:
+                next_decoder_cache += (layer_outputs[3],)
+
+            if output_attentions:
+                all_self_attns += (layer_outputs[1],)
+
+                if encoder_hidden_states is not None:
+                    all_cross_attentions += (layer_outputs[2],)
+
+        hidden_states = self.layer_norm(hidden_states)
+        # add hidden states from the last decoder layer
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        next_cache = next_decoder_cache if use_cache else None
+        if not return_dict:
+            return tuple(
+                v
+                for v in [hidden_states, next_cache, all_hidden_states, all_self_attns, all_cross_attentions]
+                if v is not None
+            )
+        return TFBaseModelOutputWithPastAndCrossAttentions(
+            last_hidden_state=hidden_states,
+            past_key_values=next_cache,
+            hidden_states=all_hidden_states,
+            attentions=all_self_attns,
+            cross_attentions=all_cross_attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+@keras_serializable
+class TFWhisperMainLayer(tf.keras.layers.Layer):
+    config_class = WhisperConfig
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(**kwargs)
+        self.config = config
+        self.encoder = TFWhisperEncoder(config, name="encoder")
+        self.decoder = TFWhisperDecoder(config, name="decoder")
+
+    def get_input_embeddings(self):
+        return self.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.encoder
+
+    def get_decoder(self):
+        return self.decoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, WhisperFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(
+         ...     ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["sampling_rate"], return_tensors="tf"
+         ... )
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        use_cache = use_cache if use_cache is not None else self.config.use_cache
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if encoder_outputs is None:
+            encoder_outputs = self.encoder(
+                input_features,
+                head_mask=head_mask,
+                output_attentions=output_attentions,
+                output_hidden_states=output_hidden_states,
+                return_dict=return_dict,
+                training=training,
+            )
+        # If the user passed a tuple for encoder_outputs, we wrap it in a TFBaseModelOutput when return_dict=True
+        elif return_dict and not isinstance(encoder_outputs, TFBaseModelOutput):
+            encoder_outputs = TFBaseModelOutput(
+                last_hidden_state=encoder_outputs[0],
+                hidden_states=encoder_outputs[1] if len(encoder_outputs) > 1 else None,
+                attentions=encoder_outputs[2] if len(encoder_outputs) > 2 else None,
+            )
+
+        # decoder outputs consists of (dec_features, past_key_value, dec_hidden, dec_attn)
+        decoder_outputs = self.decoder(
+            input_ids=decoder_input_ids,
+            attention_mask=decoder_attention_mask,
+            position_ids=decoder_position_ids,
+            encoder_hidden_states=encoder_outputs[0],
+            head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+
+        if not return_dict:
+            return decoder_outputs + encoder_outputs
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=decoder_outputs.last_hidden_state,
+            past_key_values=decoder_outputs.past_key_values,
+            decoder_hidden_states=decoder_outputs.hidden_states,
+            decoder_attentions=decoder_outputs.attentions,
+            cross_attentions=decoder_outputs.cross_attentions,
+            encoder_last_hidden_state=encoder_outputs.last_hidden_state,
+            encoder_hidden_states=encoder_outputs.hidden_states,
+            encoder_attentions=encoder_outputs.attentions,
+        )
+
+
+@add_start_docstrings(
+    "The bare Whisper Model outputting raw hidden-states without any specific head on top.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperModel(TFWhisperPreTrainedModel):
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_input_embeddings(self):
+        return self.model.decoder.embed_tokens
+
+    def set_input_embeddings(self, value):
+        self.model.decoder.embed_tokens = value
+
+    def get_encoder(self):
+        return self.model.encoder
+
+    def get_decoder(self):
+        return self.model.decoder
+
+    def decoder(self):
+        return self.model.decoder
+
+    def encoder(self):
+        return self.model.encoder
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        Returns:
+
+        Example:
+
+         ```python
+         >>> import tensorflow as tf
+         >>> from transformers import TFWhisperModel, WhisperFeatureExtractor
+         >>> from datasets import load_dataset
+
+         >>> model = TFWhisperModel.from_pretrained("openai/whisper-base")
+         >>> feature_extractor = WhisperFeatureExtractor.from_pretrained("openai/whisper-base")
+         >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+         >>> inputs = feature_extractor(
+         ...     ds[0]["audio"]["array"], sampling_rate=ds[0]["audio"]["sampling_rate"], return_tensors="tf"
+         ... )
+         >>> input_features = inputs.input_features
+         >>> decoder_input_ids = tf.convert_to_tensor([[1, 1]]) * model.config.decoder_start_token_id
+         >>> last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+         >>> list(last_hidden_state.shape)
+         [1, 2, 512]
+         ```"""
+        outputs = self.model(
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            encoder_outputs=encoder_outputs,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        return outputs
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqModelOutput(
+            last_hidden_state=output.last_hidden_state,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+
+@add_start_docstrings(
+    "The Whisper Model with a language modeling head. Can be used for automatic speech recognition.",
+    WHISPER_START_DOCSTRING,
+)
+class TFWhisperForConditionalGeneration(TFWhisperPreTrainedModel, TFCausalLanguageModelingLoss):
+    base_model_prefix = "model"
+    _keys_to_ignore_on_load_missing = [
+        r"encoder.version",
+        r"decoder.version",
+        r"proj_out.weight",
+    ]
+    _keys_to_ignore_on_save = [
+        r"proj_out.weight",
+    ]
+
+    def __init__(self, config: WhisperConfig, **kwargs):
+        super().__init__(config, **kwargs)
+        self.model = TFWhisperMainLayer(config, name="model")
+
+    def get_encoder(self):
+        return self.model.get_encoder()
+
+    def get_decoder(self):
+        return self.model.get_decoder()
+
+    def get_output_embeddings(self):
+        return self.get_input_embeddings()
+
+    def set_output_embeddings(self, value):
+        self.set_input_embeddings(value)
+
+    def resize_token_embeddings(self, new_num_tokens: int) -> tf.keras.layers.Embedding:
+        new_embeddings = super().resize_token_embeddings(new_num_tokens)
+        return new_embeddings
+
+    @add_start_docstrings_to_model_forward(WHISPER_INPUTS_DOCSTRING)
+    @replace_return_docstrings(output_type=TFSeq2SeqLMOutput, config_class=_CONFIG_FOR_DOC)
+    @unpack_inputs
+    def call(
+        self,
+        input_features=None,
+        decoder_input_ids=None,
+        decoder_attention_mask=None,
+        decoder_position_ids=None,
+        head_mask=None,
+        decoder_head_mask=None,
+        cross_attn_head_mask=None,
+        encoder_outputs=None,
+        past_key_values=None,
+        decoder_inputs_embeds=None,
+        labels=None,
+        use_cache=None,
+        output_attentions=None,
+        output_hidden_states=None,
+        return_dict=None,
+        training=False,
+    ):
+        r"""
+        labels (`tf.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+            Labels for computing the language modeling loss. Indices should either be in `[0, ..., config.vocab_size]`
+            or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is
+            only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+        Returns:
+
+        Example:
+
+        ```python
+        >>> import tensorflow as tf
+        >>> from transformers import WhisperProcessor, TFWhisperForConditionalGeneration
+        >>> from datasets import load_dataset
+
+        >>> processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        >>> model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        >>> ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+
+        >>> inputs = processor(ds[0]["audio"]["array"], return_tensors="tf")
+        >>> input_features = inputs.input_features
+
+        >>> generated_ids = model.generate(inputs=input_features)
+
+        >>> transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+        >>> transcription
+        ' Mr. Quilter is the apostle of the middle classes, and we are glad to'
+        ```"""
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        if labels is not None:
+            if decoder_input_ids is None:
+                decoder_input_ids = shift_tokens_right(
+                    labels, self.config.pad_token_id, self.config.decoder_start_token_id
+                )
+
+        outputs = self.model(
+            input_features,
+            decoder_input_ids=decoder_input_ids,
+            encoder_outputs=encoder_outputs,
+            decoder_attention_mask=decoder_attention_mask,
+            decoder_position_ids=decoder_position_ids,
+            head_mask=head_mask,
+            decoder_head_mask=decoder_head_mask,
+            cross_attn_head_mask=cross_attn_head_mask,
+            past_key_values=past_key_values,
+            decoder_inputs_embeds=decoder_inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            training=training,
+        )
+        decoder_last_hidden_state = outputs[0]
+        # Decoder and encoder embeddings are tied
+        lm_logits = tf.matmul(decoder_last_hidden_state, self.get_output_embeddings().weights, transpose_b=True)
+
+        loss = None if labels is None else self.hf_compute_loss(labels, lm_logits)
+
+        if not return_dict:
+            output = (lm_logits,) + outputs[1:]
+            return ((loss,) + output) if loss is not None else output
+
+        return TFSeq2SeqLMOutput(
+            loss=loss,
+            logits=lm_logits,
+            past_key_values=outputs.past_key_values,
+            decoder_hidden_states=outputs.decoder_hidden_states,
+            decoder_attentions=outputs.decoder_attentions,
+            cross_attentions=outputs.cross_attentions,
+            encoder_last_hidden_state=outputs.encoder_last_hidden_state,
+            encoder_hidden_states=outputs.encoder_hidden_states,
+            encoder_attentions=outputs.encoder_attentions,
+        )
+
+    def serving_output(self, output):
+        pkv = tf.tuple(output.past_key_values)[1] if self.config.use_cache else None
+        dec_hs = tf.convert_to_tensor(output.decoder_hidden_states) if self.config.output_hidden_states else None
+        dec_attns = tf.convert_to_tensor(output.decoder_attentions) if self.config.output_attentions else None
+        cross_attns = tf.convert_to_tensor(output.cross_attentions) if self.config.output_attentions else None
+        enc_hs = tf.convert_to_tensor(output.encoder_hidden_states) if self.config.output_hidden_states else None
+        enc_attns = tf.convert_to_tensor(output.encoder_attentions) if self.config.output_attentions else None
+
+        return TFSeq2SeqLMOutput(
+            logits=output.logits,
+            past_key_values=pkv,
+            decoder_hidden_states=dec_hs,
+            decoder_attentions=dec_attns,
+            cross_attentions=cross_attns,
+            encoder_last_hidden_state=output.encoder_last_hidden_state,
+            encoder_hidden_states=enc_hs,
+            encoder_attentions=enc_attns,
+        )
+
+    def prepare_inputs_for_generation(
+        self,
+        decoder_input_ids,
+        past=None,
+        use_cache=None,
+        encoder_outputs=None,
+        attention_mask=None,
+        decoder_attention_mask=None,
+        **kwargs
+    ):
+        # cut decoder_input_ids if past is used
+        if past is not None:
+            decoder_input_ids = decoder_input_ids[:, -1:]
+
+        if decoder_attention_mask is not None:  # xla
+            decoder_position_ids = tf.math.cumsum(decoder_attention_mask, axis=-1, exclusive=True)[:, -1:]
+        elif past is not None:  # no xla + past
+            decoder_position_ids = past[0][0].shape[2]
+        else:  # no xla + no past
+            decoder_position_ids = tf.range(decoder_input_ids.shape[1])
+        decoder_position_ids = tf.broadcast_to(decoder_position_ids, decoder_input_ids.shape)
+
+        return {
+            "input_features": None,  # Needs to be passed to make Keras.layer.__call__ happy
+            "encoder_outputs": encoder_outputs,
+            "past_key_values": past,
+            "decoder_input_ids": decoder_input_ids,
+            "use_cache": use_cache,
+            "decoder_attention_mask": decoder_attention_mask,
+            "decoder_position_ids": decoder_position_ids,
+        }
+
+    #
+    @staticmethod
+    def _reorder_cache(past, beam_idx):
+        reordered_past = ()
+        for layer_past in past:
+            reordered_past += (tuple(tf.gather(past_state, beam_idx) for past_state in layer_past),)
+        return reordered_past

src/transformers/utils/dummy_tf_objects.py

@@ -2394,6 +2394,30 @@ class TFWav2Vec2PreTrainedModel(metaclass=DummyObject):
         requires_backends(self, ["tf"])
 
 
+TF_WHISPER_PRETRAINED_MODEL_ARCHIVE_LIST = None
+
+
+class TFWhisperForConditionalGeneration(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFWhisperModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
+class TFWhisperPreTrainedModel(metaclass=DummyObject):
+    _backends = ["tf"]
+
+    def __init__(self, *args, **kwargs):
+        requires_backends(self, ["tf"])
+
+
 TF_XGLM_PRETRAINED_MODEL_ARCHIVE_LIST = None
 
 

tests/generation/test_generation_tf_logits_process.py

@@ -29,11 +29,14 @@ if is_tf_available():
     from transformers.generation_tf_logits_process import (
         TFForcedBOSTokenLogitsProcessor,
         TFForcedEOSTokenLogitsProcessor,
+        TFForceTokensLogitsProcessor,
         TFLogitsProcessorList,
         TFMinLengthLogitsProcessor,
         TFNoBadWordsLogitsProcessor,
         TFNoRepeatNGramLogitsProcessor,
         TFRepetitionPenaltyLogitsProcessor,
+        TFSuppressTokensAtBeginLogitsProcessor,
+        TFSuppressTokensLogitsProcessor,
         TFTemperatureLogitsWarper,
         TFTopKLogitsWarper,
         TFTopPLogitsWarper,
@@ -331,6 +334,86 @@ class TFLogitsProcessorTest(unittest.TestCase):
         scores = logits_processor(input_ids, scores, cur_len)
         self.assertFalse(tf.math.reduce_any(tf.math.is_inf((scores))))
 
+    @parameterized.expand([(False,), (True,)])
+    def test_suppress_tokens_at_begin_logits_processor(self, use_xla):
+        vocab_size = 20
+        batch_size = 4
+
+        begin_suppress_tokens = [1, 2, 3]
+        begin_index = 5
+
+        logits_processor = TFSuppressTokensAtBeginLogitsProcessor(
+            begin_suppress_tokens=begin_suppress_tokens, begin_index=begin_index
+        )
+        if use_xla:
+            logits_processor = tf.function(logits_processor, jit_compile=True)
+
+        # Check that no scores are suppressed if begin_index is not reached
+        cur_len = 4
+        input_ids = tf.convert_to_tensor([[11, 17, 15, 8], [14, 0, 19, 5], [13, 11, 18, 19], [11, 12, 16, 15]])
+        scores = self._get_uniform_logits(batch_size, vocab_size)
+        scores = logits_processor(input_ids, scores, cur_len)
+        self.assertFalse(tf.math.reduce_any(tf.math.is_inf((scores))))
+
+        # Check that scores are suppressed if begin_index is reached
+        cur_len = 5
+        input_ids = tf.convert_to_tensor([[5, 5, 5, 0, 17], [18, 1, 9, 14, 17], [18, 6, 8, 15, 19], [8, 12, 17, 1, 2]])
+        scores = self._get_uniform_logits(batch_size, vocab_size)
+        scores = logits_processor(input_ids, scores, cur_len)
+        self.assertTrue(tf.math.reduce_all(tf.math.is_inf(tf.gather(scores, begin_suppress_tokens, axis=1))))
+
+    @parameterized.expand([(False,), (True,)])
+    def test_suppress_tokens_logits_processor(self, use_xla):
+        vocab_size = 20
+        batch_size = 4
+
+        suppress_tokens = [1, 3, 5]
+        keep_tokens = [i for i in range(vocab_size) if i not in suppress_tokens]
+
+        logits_processor = TFSuppressTokensLogitsProcessor(suppress_tokens=suppress_tokens)
+        if use_xla:
+            logits_processor = tf.function(logits_processor, jit_compile=True)
+
+        # Check that suppress_tokens are suppressed and others are not
+        cur_len = 5
+        input_ids = tf.convert_to_tensor([[0, 10, 19, 6, 3], [17, 4, 8, 17, 2], [7, 1, 11, 6, 15], [5, 8, 13, 16, 0]])
+        scores = self._get_uniform_logits(batch_size, vocab_size)
+        scores = logits_processor(input_ids, scores, cur_len)
+        self.assertTrue(tf.math.reduce_all(tf.math.is_inf(tf.gather(scores, suppress_tokens, axis=1))))
+        self.assertFalse(tf.math.reduce_any(tf.math.is_inf(tf.gather(scores, keep_tokens, axis=1))))
+
+    @parameterized.expand([(False,), (True,)])
+    def test_force_tokens_logits_processor(self, use_xla):
+        vocab_size = 20
+        batch_size = 4
+
+        force_token_map = {1: 2, 3: 2}
+
+        logits_processor = TFForceTokensLogitsProcessor(force_token_map=force_token_map)
+        if use_xla:
+            logits_processor = tf.function(logits_processor, jit_compile=True)
+
+        # check that if the cur_len is contained in the force_token_map, the logits are the same
+        # for all tokens except the one the force_token_map points to
+        cur_len = 1
+        input_ids = tf.convert_to_tensor([[11], [7], [5], [15]])
+        ids_tensor((batch_size, cur_len), vocab_size=20)
+        scores = self._get_uniform_logits(batch_size, vocab_size)
+        scores = logits_processor(input_ids, scores, cur_len)
+        tf.debugging.assert_near(tf.gather(scores, [force_token_map[cur_len]], axis=1), 0.0)
+
+        non_forced_inds = [i for i in range(vocab_size) if i != force_token_map[cur_len]]
+        self.assertTrue(
+            tf.math.reduce_all(tf.math.is_inf(tf.gather(scores, [non_forced_inds], axis=1))),
+        )
+
+        # check that if the cur_len is not contained in the force_token_map, the logits are not modified
+        cur_len = 2
+        input_ids = tf.convert_to_tensor([[2, 19], [19, 15], [4, 9], [7, 6]])
+        scores = self._get_uniform_logits(batch_size, vocab_size)
+        scores = logits_processor(input_ids, scores, cur_len)
+        self.assertFalse(tf.math.reduce_any(tf.math.is_inf((scores))))
+
     @parameterized.expand([(False,), (True,)])
     def test_processor_list(self, use_xla):
         # TODO (Joao): reintroduce TFNoRepeatNGramLogitsProcessor when it gets compatible with XLA

tests/models/whisper/test_modeling_tf_whisper.py

+# coding=utf-8
+# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+""" Testing suite for the TensorFlow Whisper model. """
+
+import inspect
+import tempfile
+import unittest
+
+import numpy as np
+
+from transformers import WhisperConfig, WhisperFeatureExtractor, WhisperProcessor
+from transformers.testing_utils import is_tf_available, require_tf, require_tokenizers, slow
+from transformers.utils import cached_property
+from transformers.utils.import_utils import is_datasets_available
+
+from ...test_configuration_common import ConfigTester
+from ...test_modeling_tf_common import TFModelTesterMixin, floats_tensor, ids_tensor
+
+
+if is_datasets_available():
+    import datasets
+    from datasets import load_dataset
+
+
+if is_tf_available():
+    import tensorflow as tf
+
+    from transformers import TFWhisperForConditionalGeneration, TFWhisperModel, set_seed
+    from transformers.models.whisper.modeling_tf_whisper import TFWhisperDecoder, TFWhisperEncoder
+
+
+def prepare_whisper_inputs_dict(
+    config,
+    input_features,
+    decoder_input_ids,
+    attention_mask=None,
+    decoder_attention_mask=None,
+    head_mask=None,
+    decoder_head_mask=None,
+    cross_attn_head_mask=None,
+):
+    if decoder_attention_mask is None:
+        decoder_attention_mask = tf.where(decoder_input_ids != config.pad_token_id, 1, 0)
+    if head_mask is None:
+        head_mask = tf.ones((config.encoder_layers, config.encoder_attention_heads))
+    if decoder_head_mask is None:
+        decoder_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
+    if cross_attn_head_mask is None:
+        cross_attn_head_mask = tf.ones((config.decoder_layers, config.decoder_attention_heads))
+    return {
+        "input_features": input_features,
+        "decoder_input_ids": decoder_input_ids,
+        "decoder_attention_mask": decoder_attention_mask,
+        "head_mask": head_mask,
+        "decoder_head_mask": decoder_head_mask,
+        "cross_attn_head_mask": cross_attn_head_mask,
+    }
+
+
+@require_tf
+class TFWhisperModelTester:
+    def __init__(
+        self,
+        parent,
+        batch_size=13,
+        seq_length=60,
+        is_training=True,
+        use_labels=False,
+        vocab_size=99,
+        hidden_size=16,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        input_channels=1,
+        hidden_act="gelu",
+        hidden_dropout_prob=0.1,
+        attention_probs_dropout_prob=0.1,
+        max_position_embeddings=20,
+        max_source_positions=30,
+        max_target_positions=60,
+        bos_token_id=98,
+        eos_token_id=98,
+        pad_token_id=0,
+        num_mel_bins=80,
+        decoder_start_token_id=85,
+        num_conv_layers=1,
+        suppress_tokens=None,
+        begin_suppress_tokens=None,
+    ):
+        self.parent = parent
+        self.batch_size = batch_size
+        self.seq_length = seq_length
+        self.is_training = is_training
+        self.use_labels = use_labels
+        self.vocab_size = vocab_size
+        self.hidden_size = hidden_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.input_channels = input_channels
+        self.hidden_act = hidden_act
+        self.hidden_dropout_prob = hidden_dropout_prob
+        self.attention_probs_dropout_prob = attention_probs_dropout_prob
+        self.num_mel_bins = num_mel_bins
+        self.max_position_embeddings = max_position_embeddings
+        self.max_source_positions = max_source_positions
+        self.max_target_positions = max_target_positions
+        self.eos_token_id = eos_token_id
+        self.pad_token_id = pad_token_id
+        self.bos_token_id = bos_token_id
+        self.decoder_start_token_id = decoder_start_token_id
+        self.num_conv_layers = num_conv_layers
+        self.suppress_tokens = suppress_tokens
+        self.begin_suppress_tokens = begin_suppress_tokens
+
+    def prepare_config_and_inputs(self):
+        input_features = floats_tensor([self.batch_size, self.num_mel_bins, self.seq_length], self.vocab_size)
+
+        decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
+
+        config = self.get_config()
+        inputs_dict = prepare_whisper_inputs_dict(
+            config,
+            attention_mask=None,
+            input_features=input_features,
+            decoder_input_ids=decoder_input_ids,
+        )
+        return config, inputs_dict
+
+    def get_config(self):
+        return WhisperConfig(
+            vocab_size=self.vocab_size,
+            d_model=self.hidden_size,
+            encoder_layers=self.num_hidden_layers,
+            decoder_layers=self.num_hidden_layers,
+            encoder_attention_heads=self.num_attention_heads,
+            decoder_attention_heads=self.num_attention_heads,
+            input_channels=self.input_channels,
+            dropout=self.hidden_dropout_prob,
+            attention_dropout=self.attention_probs_dropout_prob,
+            max_position_embeddings=self.max_position_embeddings,
+            max_source_positions=self.max_source_positions,
+            max_target_positions=self.max_target_positions,
+            eos_token_id=self.eos_token_id,
+            bos_token_id=self.bos_token_id,
+            pad_token_id=self.pad_token_id,
+            decoder_ffn_dim=self.hidden_size,
+            encoder_ffn_dim=self.hidden_size,
+            decoder_start_token_id=self.decoder_start_token_id,
+            suppress_tokens=self.suppress_tokens,
+            begin_suppress_tokens=self.begin_suppress_tokens,
+        )
+
+    def prepare_config_and_inputs_for_common(self):
+        config, inputs_dict = self.prepare_config_and_inputs()
+        return config, inputs_dict
+
+    def get_subsampled_output_lengths(self, input_lengths):
+        """
+        Computes the output length of the convolutional layers
+        """
+
+        for i in range(self.num_conv_layers):
+            input_lengths = (input_lengths - 1) // 2 + 1
+
+        return input_lengths
+
+    def create_and_check_model_forward(self, config, inputs_dict):
+        model = TFWhisperModel(config=config)
+
+        input_features = inputs_dict["input_features"]
+        decoder_input_ids = inputs_dict["decoder_input_ids"]
+
+        # first forward pass
+        last_hidden_state = model(input_features, decoder_input_ids=decoder_input_ids).last_hidden_state
+
+        self.parent.assertTrue(last_hidden_state.shape, (13, 7, 16))
+
+    def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
+        model = TFWhisperModel(config=config).get_decoder()
+        # take a slice so we're shorter than the seqeuence length and can append later
+        input_ids = inputs_dict["decoder_input_ids"][:, :-10]
+        attention_mask = inputs_dict["decoder_attention_mask"][:, :-10]
+
+        # first forward pass
+        outputs = model(input_ids, attention_mask=attention_mask, use_cache=True)
+
+        output, past_key_values = outputs.to_tuple()
+
+        # create hypothetical multiple next token and extent to next_input_ids
+        next_token = ids_tensor((self.batch_size, 3), config.vocab_size)
+        next_tokens = tf.where(next_token <= 2, 2, next_token)
+        next_attn_mask = ids_tensor((self.batch_size, 3), 2)
+
+        # append to next input_ids and
+        next_input_ids = tf.concat([input_ids, next_tokens], axis=-1)
+        next_attention_mask = tf.concat([attention_mask, next_attn_mask], axis=-1)
+
+        output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
+        output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
+            "last_hidden_state"
+        ]
+
+        # select random slice
+        random_slice_idx = np.random.randint(0, output_from_past.shape[-1])
+        output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx]
+        output_from_past_slice = output_from_past[:, :, random_slice_idx]
+
+        self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
+
+        # test that outputs are equal for slice
+        self.parent.assertTrue(np.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-2))
+
+    def check_encoder_decoder_model_standalone(self, config, inputs_dict):
+        model = TFWhisperModel(config=config)
+        outputs = model(**inputs_dict)
+
+        encoder_last_hidden_state = outputs.encoder_last_hidden_state
+        last_hidden_state = outputs.last_hidden_state
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            encoder = model.get_encoder()
+            encoder.save_pretrained(tmpdirname)
+            encoder = TFWhisperEncoder.from_pretrained(tmpdirname)
+
+        encoder_last_hidden_state_2 = encoder(inputs_dict["input_features"])[0]
+
+        self.parent.assertTrue((encoder_last_hidden_state_2 - encoder_last_hidden_state).abs().max() < 1e-3)
+
+        with tempfile.TemporaryDirectory() as tmpdirname:
+            decoder = model.get_decoder()
+            decoder.save_pretrained(tmpdirname)
+            decoder = TFWhisperDecoder.from_pretrained(tmpdirname)
+
+        last_hidden_state_2 = decoder(
+            input_ids=inputs_dict["decoder_input_ids"],
+            attention_mask=inputs_dict["decoder_attention_mask"],
+            encoder_hidden_states=encoder_last_hidden_state,
+        )[0]
+
+        self.parent.assertTrue((last_hidden_state_2 - last_hidden_state).abs().max() < 1e-3)
+
+
+@require_tf
+class TFWhisperModelTest(TFModelTesterMixin, unittest.TestCase):
+    all_model_classes = (TFWhisperModel, TFWhisperForConditionalGeneration) if is_tf_available() else ()
+    all_generative_model_classes = (TFWhisperForConditionalGeneration,) if is_tf_available() else ()
+    is_encoder_decoder = True
+    fx_compatible = False
+    test_pruning = False
+    test_missing_keys = False
+    test_onnx = False
+
+    input_name = "input_features"
+
+    def setUp(self):
+        self.model_tester = TFWhisperModelTester(self)
+        self.config_tester = ConfigTester(self, config_class=WhisperConfig)
+        self.maxDiff = 3000
+
+    def test_config(self):
+        self.config_tester.run_common_tests()
+
+    def test_save_load_strict(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs()
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+
+            model(model.dummy_inputs)
+
+            with tempfile.TemporaryDirectory() as tmpdirname:
+                model.save_pretrained(tmpdirname, saved_model=False)
+                model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
+            self.assertEqual(info["missing_keys"], [])
+
+    def test_model_forward(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_model_forward(*config_and_inputs)
+
+    def test_decoder_model_past_with_large_inputs(self):
+        config_and_inputs = self.model_tester.prepare_config_and_inputs()
+        self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
+
+    def _get_input_ids_and_config(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        input_ids = inputs_dict[self.input_name]
+
+        # cut to half length & take max batch_size 3
+        max_batch_size = 3
+        input_ids = input_ids[:max_batch_size, :, :]
+
+        # generate max 3 tokens
+        max_length = input_ids.shape[-1] + 3
+        if config.eos_token_id is not None and config.pad_token_id is None:
+            # hack to allow generate for models such as GPT2 as is done in `generate()`
+            config.pad_token_id = config.eos_token_id
+
+        return config, input_ids, None, max_length
+
+    # not implemented currently
+    def test_inputs_embeds(self):
+        pass
+
+    @unittest.skip("Training is not yet supported")
+    def test_training(self):
+        pass
+
+    def test_generate_with_head_masking(self):
+        pass
+
+    @unittest.skip("fp16 is not yet supported for TF models")
+    def test_generate_fp16(self):
+        config, input_dict = self.model_tester.prepare_config_and_inputs()
+        config.max_target_positions = 400
+        input_features = input_dict["input_features"]
+        model = TFWhisperForConditionalGeneration(config)
+        model.generate(input_features)
+        model.generate(input_features, num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
+
+    def test_forward_signature(self):
+        config, _ = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            model = model_class(config)
+            signature = inspect.signature(model.call)
+            # signature.parameters is an OrderedDict => so arg_names order is deterministic
+            arg_names = [*signature.parameters.keys()]
+
+            expected_arg_names = [
+                "input_features",
+                "decoder_input_ids",
+                "decoder_attention_mask",
+            ]
+            expected_arg_names.extend(
+                ["decoder_position_ids", "head_mask", "decoder_head_mask", "cross_attn_head_mask", "encoder_outputs"]
+                if "head_mask" and "decoder_head_mask" and "cross_attn_head_mask" in arg_names
+                else ["encoder_outputs"]
+            )
+            self.assertListEqual(arg_names[: len(expected_arg_names)], expected_arg_names)
+
+    def test_hidden_states_output(self):
+        def check_hidden_states_output(inputs_dict, config, model_class):
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            hidden_states = outputs.encoder_hidden_states if config.is_encoder_decoder else outputs.hidden_states
+
+            expected_num_layers = getattr(
+                self.model_tester, "expected_num_hidden_layers", self.model_tester.num_hidden_layers + 1
+            )
+            self.assertEqual(len(hidden_states), expected_num_layers)
+
+            if hasattr(self.model_tester, "encoder_seq_length"):
+                seq_length = self.model_tester.encoder_seq_length
+            else:
+                seq_length = self.model_tester.seq_length
+
+            subsampled_seq_length = model._get_feat_extract_output_lengths(seq_length)
+
+            self.assertListEqual(
+                list(hidden_states[0].shape[-2:]),
+                [subsampled_seq_length, self.model_tester.hidden_size],
+            )
+
+            if config.is_encoder_decoder:
+                hidden_states = outputs.decoder_hidden_states
+
+                self.assertIsInstance(hidden_states, (list, tuple))
+                self.assertEqual(len(hidden_states), expected_num_layers)
+
+                decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_length)
+
+                self.assertListEqual(
+                    list(hidden_states[0].shape[-2:]),
+                    [decoder_seq_length, self.model_tester.hidden_size],
+                )
+
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_hidden_states"] = True
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+            # check that output_hidden_states also work using config
+            del inputs_dict["output_hidden_states"]
+            config.output_hidden_states = True
+
+            check_hidden_states_output(inputs_dict, config, model_class)
+
+    def test_attention_outputs(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        config.return_dict = True
+
+        seq_len = getattr(self.model_tester, "seq_length", None)
+        decoder_seq_length = getattr(self.model_tester, "decoder_seq_length", seq_len)
+        encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
+        encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
+        decoder_key_length = getattr(self.model_tester, "decoder_key_length", encoder_key_length)
+
+        for model_class in self.all_model_classes:
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = False
+            config.return_dict = True
+            model = model_class(config)
+
+            subsampled_encoder_seq_length = model._get_feat_extract_output_lengths(encoder_seq_length)
+            subsampled_encoder_key_length = model._get_feat_extract_output_lengths(encoder_key_length)
+
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            # check that output_attentions also work using config
+            del inputs_dict["output_attentions"]
+            config.output_attentions = True
+            model = model_class(config)
+
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+            attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+            self.assertEqual(len(attentions), self.model_tester.num_hidden_layers)
+
+            self.assertListEqual(
+                list(attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length],
+            )
+            out_len = len(outputs)
+
+            correct_outlen = 5
+
+            # loss is at first position
+            if "labels" in inputs_dict:
+                correct_outlen += 1  # loss is added to beginning
+            if "past_key_values" in outputs:
+                correct_outlen += 1  # past_key_values have been returned
+
+            self.assertEqual(out_len, correct_outlen)
+
+            # decoder attentions
+            decoder_attentions = outputs.decoder_attentions
+            self.assertIsInstance(decoder_attentions, (list, tuple))
+            self.assertEqual(len(decoder_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(decoder_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, decoder_seq_length, decoder_key_length],
+            )
+
+            # cross attentions
+            cross_attentions = outputs.cross_attentions
+            self.assertIsInstance(cross_attentions, (list, tuple))
+            self.assertEqual(len(cross_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(cross_attentions[0].shape[-3:]),
+                [
+                    self.model_tester.num_attention_heads,
+                    decoder_seq_length,
+                    subsampled_encoder_key_length,
+                ],
+            )
+
+            # Check attention is always last and order is fine
+            inputs_dict["output_attentions"] = True
+            inputs_dict["output_hidden_states"] = True
+            model = model_class(config)
+            outputs = model(**self._prepare_for_class(inputs_dict, model_class))
+
+            added_hidden_states = 2
+            self.assertEqual(out_len + added_hidden_states, len(outputs))
+
+            self_attentions = outputs.encoder_attentions if config.is_encoder_decoder else outputs.attentions
+
+            self.assertEqual(len(self_attentions), self.model_tester.num_hidden_layers)
+            self.assertListEqual(
+                list(self_attentions[0].shape[-3:]),
+                [self.model_tester.num_attention_heads, subsampled_encoder_seq_length, subsampled_encoder_key_length],
+            )
+
+    def test_generate_without_input_ids(self):
+        pass
+
+    @staticmethod
+    def _get_encoder_outputs(
+        model, input_ids, attention_mask, output_attentions=None, output_hidden_states=None, num_interleave=1
+    ):
+        encoder = model.get_encoder()
+        encoder_outputs = encoder(
+            input_ids,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+        )
+        encoder_outputs["last_hidden_state"] = encoder_outputs.last_hidden_state.repeat_interleave(
+            num_interleave, dim=0
+        )
+        input_ids = input_ids[:, :, 0]
+        input_ids = tf.zeros_like(input_ids[:, :1], dtype=tf.int64) + tf.convert_to_tensor(
+            [model._get_decoder_start_token_id()]
+        )
+        attention_mask = None
+        return encoder_outputs, input_ids, attention_mask
+
+    def _check_outputs(self, output, input_ids, config, use_cache=False, num_return_sequences=1):
+        batch_size, mel, seq_length = input_ids.shape
+        subsampled_seq_length = self.model_tester.get_subsampled_output_lengths(seq_length)
+        num_sequences_in_output = batch_size * num_return_sequences
+        gen_len = (
+            output.sequences.shape[-1] - 1 if config.is_encoder_decoder else output.sequences.shape[-1] - seq_length
+        )
+
+        # scores
+        self._check_scores(num_sequences_in_output, output.scores, length=gen_len, config=config)
+
+        # Attentions
+        # encoder
+        self._check_encoder_attention_for_generate(
+            output.encoder_attentions, batch_size, config, subsampled_seq_length
+        )
+        # decoder
+        self._check_attentions_for_generate(
+            num_sequences_in_output,
+            output.decoder_attentions,
+            min_length=1,
+            max_length=output.sequences.shape[-1],
+            config=config,
+            use_cache=use_cache,
+        )
+
+        # Hidden States
+        # encoder
+        self._check_encoder_hidden_states_for_generate(
+            output.encoder_hidden_states, batch_size, config, subsampled_seq_length
+        )
+
+        # decoder
+        self._check_hidden_states_for_generate(
+            num_sequences_in_output,
+            output.decoder_hidden_states,
+            min_length=1,
+            max_length=output.sequences.shape[-1],
+            config=config,
+            use_cache=use_cache,
+        )
+
+    # overwritten from parent due to the inability to work when non-text inputs are not passed AND because the input is
+    # `input_features`
+    def test_lm_head_model_random_no_beam_search_generate(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        input_features = inputs_dict.get("input_features", None)
+
+        # iterate over all generative models
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            if config.bos_token_id is None:
+                # if bos token id is not defined model needs input_features
+                with self.assertRaises(AssertionError):
+                    model.generate(do_sample=True, max_length=5)
+                # num_return_sequences = 1
+                self._check_generated_ids(model.generate(input_features, do_sample=True))
+
+            with self.assertRaises(ValueError):
+                # generating multiple sequences when no beam search generation
+                # is not allowed as it would always generate the same sequences
+                model.generate(input_features, do_sample=False, num_return_sequences=2)
+
+            # num_return_sequences > 1, sample
+            self._check_generated_ids(model.generate(input_features, do_sample=True, num_return_sequences=2))
+
+            # check bad words tokens language generation
+            # create list of 1-seq bad token and list of 2-seq of bad tokens
+            bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
+            output_tokens = model.generate(
+                input_features, do_sample=True, bad_words_ids=bad_words_ids, num_return_sequences=2
+            )
+            # only count generated tokens
+            generated_ids = output_tokens[:, input_features.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
+
+    # overwritten from parent due to the inability to work when non-text inputs are not passed AND because the input is
+    # `input_features`
+    def test_lm_head_model_random_beam_search_generate(self):
+        config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+        input_features = inputs_dict.get("input_features", None)
+
+        for model_class in self.all_generative_model_classes:
+            model = model_class(config)
+
+            if config.bos_token_id is None:
+                # if bos token id is not defined model needs input_ids, num_return_sequences = 1
+                self._check_generated_ids(model.generate(input_features, do_sample=True, num_beams=2))
+
+            with self.assertRaises(ValueError):
+                # generating more sequences than having beams leads is not possible
+                model.generate(input_features, do_sample=False, num_return_sequences=3, num_beams=2)
+
+            # num_return_sequences > 1, sample
+            self._check_generated_ids(
+                model.generate(
+                    input_features,
+                    do_sample=True,
+                    num_beams=2,
+                    num_return_sequences=2,
+                )
+            )
+            # num_return_sequences > 1, greedy
+            self._check_generated_ids(
+                model.generate(input_features, do_sample=False, num_beams=2, num_return_sequences=2)
+            )
+
+            # check bad words tokens language generation
+            # create list of 1-seq bad token and list of 2-seq of bad tokens
+            bad_words_ids = [self._generate_random_bad_tokens(1, model), self._generate_random_bad_tokens(2, model)]
+            output_tokens = model.generate(
+                input_features, do_sample=False, bad_words_ids=bad_words_ids, num_beams=2, num_return_sequences=2
+            )
+            # only count generated tokens
+            generated_ids = output_tokens[:, input_features.shape[-1] :]
+            self.assertFalse(self._check_match_tokens(generated_ids.numpy().tolist(), bad_words_ids))
+
+
+@require_tf
+@require_tokenizers
+class TFWhisperModelIntegrationTests(unittest.TestCase):
+    @cached_property
+    def default_processor(self):
+        return WhisperProcessor.from_pretrained("openai/whisper-base")
+
+    def _load_datasamples(self, num_samples):
+
+        ds = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation")
+        # automatic decoding with librispeech
+        speech_samples = ds.sort("id").select(range(num_samples))[:num_samples]["audio"]
+
+        return [x["array"] for x in speech_samples]
+
+    @slow
+    def test_tiny_logits_librispeech(self):
+        set_seed(0)
+        model = TFWhisperModel.from_pretrained("openai/whisper-tiny")
+        input_speech = self._load_datasamples(1)
+        feature_extractor = WhisperFeatureExtractor()
+        input_features = feature_extractor(input_speech, return_tensors="tf").input_features
+
+        logits = model(
+            input_features,
+            decoder_input_ids=tf.convert_to_tensor([[50258, 50259, 50359]]),
+            output_hidden_states=False,
+            output_attentions=False,
+            return_dict=False,
+            use_cache=False,
+        )
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                2.9892, -6.7607, 5.7348, 3.6096, 0.2152, -5.7321, 4.8855, -1.6407,
+                0.2823, -1.5718, 10.4269, 3.4427, 0.0219, -8.0612, 3.4784, 8.4246,
+                4.0575, -2.2864, 11.1084, 0.9963, 0.9884, -8.5154, -3.5469, -9.3713,
+                0.9786, 3.5435, 7.4850, -5.2579, -1.4366, 10.4841
+            ]
+        )
+        # fmt: on
+        self.assertTrue(np.allclose(logits[0][0, 0, :30], EXPECTED_LOGITS, atol=1e-4))
+
+        # fmt: off
+        EXPECTED_GENERATION = tf.convert_to_tensor(
+            [
+                -1.4651, -2.6944, 2.7821, 2.3793, 4.0738, 0.0188, -3.3203, 1.9836,
+                0.0520, 0.7095, 1.1063, 0.2952, -3.6786, -0.5249, 0.3105, 4.7691,
+                1.1562, 1.3046, 0.5810, -0.3624, 1.7006, 1.3424, 0.9817, 2.1958,
+                1.8775, -5.7046, -0.7679, 4.0113, 2.6848, 2.8609
+            ]
+        )
+        # fmt: on
+
+        head_logits = logits[0] @ tf.transpose(model.model.decoder.embed_tokens.weights[0])
+        self.assertTrue(np.allclose(head_logits[0, 0, :30], EXPECTED_GENERATION, atol=1e-4))
+
+    @slow
+    def test_small_en_logits_librispeech(self):
+        set_seed(0)
+        model = TFWhisperModel.from_pretrained("openai/whisper-small.en")
+
+        input_speech = self._load_datasamples(1)
+
+        feaure_extractor = WhisperFeatureExtractor()
+        input_features = feaure_extractor(input_speech, return_tensors="tf").input_features
+
+        logits = model(
+            input_features,
+            decoder_input_ids=tf.convert_to_tensor([[model.config.decoder_start_token_id]]),
+            output_hidden_states=False,
+            output_attentions=False,
+            use_cache=False,
+        )
+
+        logits = logits.last_hidden_state @ tf.transpose(model.model.decoder.embed_tokens.weights[0])
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                -3.6784, -7.7211, -9.5070, -11.9286, -7.6489, -9.7026, -5.6188,
+                -8.0104, -4.6238, -5.1833, -9.0485, -3.4079, -5.4874, -2.6935,
+                -6.3479, -7.3398, -6.9558, -7.6867, -7.4748, -8.3463, -9.9781,
+                -10.8389, -10.3105, -11.7201, -9.7261, -7.1590, -5.9272, -12.4509,
+                -11.1146, -8.1918
+            ]
+        )
+        # fmt: on
+        self.assertTrue(np.allclose(logits[0, 0, :30], EXPECTED_LOGITS, atol=1e-4))
+
+    @slow
+    def test_large_logits_librispeech(self):
+        set_seed(0)
+
+        model = TFWhisperModel.from_pretrained("openai/whisper-large")
+
+        input_speech = self._load_datasamples(1)
+
+        processor = WhisperProcessor.from_pretrained("openai/whisper-large")
+        processed_inputs = processor(audio=input_speech, text="This part of the speech", return_tensors="tf")
+        input_features = processed_inputs.input_features
+        labels = processed_inputs.labels
+
+        logits = model(
+            input_features,
+            decoder_input_ids=labels,
+            output_hidden_states=False,
+            output_attentions=False,
+            use_cache=False,
+        )
+
+        logits = logits.last_hidden_state @ tf.transpose(model.model.decoder.embed_tokens.weights[0])
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                2.1382, 0.9381, 4.4671, 3.5589, 2.4022, 3.8576, -0.6521, 2.5472,
+                1.8301, 1.9957, 2.3432, 1.4678, 0.5459, 2.2597, 1.5179, 2.5357,
+                1.1624, 0.6194, 1.0757, 1.8259, 2.4076, 1.6601, 2.3503, 1.3376,
+                1.9891, 1.8635, 3.8931, 5.3699, 4.4772, 3.9184
+            ]
+        )
+        # fmt: on
+
+        self.assertTrue(np.allclose(logits[0, 0, :30], EXPECTED_LOGITS, atol=1e-4))
+
+    @slow
+    def test_tiny_en_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+        model.config.decoder_start_token_id = 50257
+
+        input_speech = self._load_datasamples(1)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        generated_ids = model.generate(input_features, num_beams=5)
+        transcript = processor.tokenizer.batch_decode(generated_ids)[0]
+
+        EXPECTED_TRANSCRIPT = (
+            "<|startoftranscript|><|notimestamps|> Mr. Quilter is the apostle of the middle"
+            " classes, and we are glad to"
+        )
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_tiny_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
+
+        input_speech = self._load_datasamples(1)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        generated_ids = model.generate(input_features, num_beams=5)
+        transcript = processor.tokenizer.decode(generated_ids[0])
+
+        EXPECTED_TRANSCRIPT = (
+            "<|startoftranscript|><|en|><|transcribe|><|notimestamps|> Mr. Quilter is the apostle of the middle"
+            " classes and we are glad"
+        )
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_tiny_xla_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny")
+
+        input_speech = self._load_datasamples(1)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        xla_generate = tf.function(model.generate, jit_compile=True)
+
+        generated_ids = model.generate(input_features, num_beams=5)
+        generated_ids_xla = xla_generate(input_features, num_beams=5)
+
+        transcript = processor.tokenizer.decode(generated_ids[0])
+        transcript_xla = processor.tokenizer.decode(generated_ids_xla[0])
+
+        EXPECTED_TRANSCRIPT = (
+            "<|startoftranscript|><|en|><|transcribe|><|notimestamps|> Mr. Quilter is the apostle of the middle"
+            " classes and we are glad"
+        )
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+        self.assertEqual(transcript_xla, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_large_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-large")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-large")
+
+        input_speech = self._load_datasamples(1)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="en", task="transcribe")
+        generated_ids = model.generate(
+            input_features,
+            do_sample=False,
+        )
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+
+        EXPECTED_TRANSCRIPT = " Mr. Quilter is the apostle of the middle classes and we are glad"
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_large_generation_multilingual(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-large")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-large")
+
+        ds = load_dataset("common_voice", "ja", split="test", streaming=True)
+        ds = ds.cast_column("audio", datasets.Audio(sampling_rate=16_000))
+        input_speech = next(iter(ds))["audio"]["array"]
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="ja", task="transcribe")
+        generated_ids = model.generate(input_features, do_sample=False)
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+
+        EXPECTED_TRANSCRIPT = "木村さんに電話を貸してもらいました"
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+        model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="en", task="transcribe")
+        generated_ids = model.generate(
+            input_features,
+            do_sample=False,
+        )
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+
+        EXPECTED_TRANSCRIPT = " Kimura san ni denwa wo kaite moraimashita"
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+        model.config.forced_decoder_ids = processor.get_decoder_prompt_ids(language="ja", task="translate")
+        generated_ids = model.generate(input_features, do_sample=False)
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
+
+        EXPECTED_TRANSCRIPT = " I borrowed a phone from Kimura san"
+        self.assertEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_large_batched_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-large")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-large")
+
+        input_speech = self._load_datasamples(4)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+        generated_ids = model.generate(input_features)
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                [50258, 50358, 50363, 2221, 13, 2326, 388, 391, 307, 264, 50244, 295, 264, 2808, 5359, 293, 321, 366, 5404, 281],
+                [50258, 50358, 50363, 6966, 307, 2221, 13, 2326, 388, 391, 311, 9060, 1570, 1880, 813, 702, 1871, 13, 50257, 50257],
+                [50258, 50358, 50363, 634, 5112, 505, 300, 412, 341, 42729, 3196, 295, 264, 1064, 11, 365, 5272, 293, 12904, 9256],
+                [50258, 50358, 50363, 634, 575, 12525, 22618, 1968, 6144, 35617, 20084, 1756, 311, 589, 307, 534, 10281, 934, 439, 11]
+            ]
+        )
+        # fmt: on
+
+        self.assertTrue(np.allclose(generated_ids, EXPECTED_LOGITS))
+
+        # fmt: off
+        EXPECTED_TRANSCRIPT = [
+            ' Mr. Quilter is the apostle of the middle classes, and we are glad to',
+            " Nor is Mr. Quilter's manner less interesting than his matter.",
+            " He tells us that at this festive season of the year, with Christmas and roast beef",
+            " He has grave doubts whether Sir Frederick Layton's work is really Greek after all,"
+        ]
+        # fmt: on
+
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        self.assertListEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_tiny_en_batched_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        input_speech = self._load_datasamples(4)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+        generated_ids = model.generate(input_features)
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                [50257, 50362, 1770, 13, 2264, 346, 353, 318, 262, 46329, 286, 262, 3504, 6097, 11, 290, 356, 389, 9675, 284],
+                [50257, 50362, 5414, 318, 1770, 13, 2264, 346, 353, 338, 5642, 1342, 3499, 621, 465, 2300, 13, 50256, 50256, 50256],
+                [50257, 50362, 679, 4952, 514, 326, 379, 428, 43856, 1622, 286, 262, 614, 11, 351, 6786, 290, 32595, 12023, 28236],
+                [50257, 50362, 679, 468, 12296, 17188, 1771, 7361, 26113, 18881, 1122, 338, 670, 318, 1107, 8312, 706, 477, 290, 460]
+            ]
+
+        )
+        # fmt: on
+
+        self.assertTrue(np.allclose(generated_ids, EXPECTED_LOGITS))
+
+        # fmt: off
+        EXPECTED_TRANSCRIPT = [
+            " Mr. Quilter is the apostle of the middle classes, and we are glad to",
+            " Nor is Mr. Quilter's manner less interesting than his matter.",
+            " He tells us that at this festive season of the year, with Christmas and roast beef looming",
+            " He has grave doubts whether Sir Frederick Layton's work is really Greek after all and can",
+        ]
+        # fmt: on
+
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        self.assertListEqual(transcript, EXPECTED_TRANSCRIPT)
+
+    @slow
+    def test_tiny_en_batched_xla_generation(self):
+        set_seed(0)
+        processor = WhisperProcessor.from_pretrained("openai/whisper-tiny.en")
+        model = TFWhisperForConditionalGeneration.from_pretrained("openai/whisper-tiny.en")
+
+        input_speech = self._load_datasamples(4)
+        input_features = processor.feature_extractor(raw_speech=input_speech, return_tensors="tf").input_features
+
+        xla_generate = tf.function(model.generate, jit_compile=True)
+
+        generated_ids = model.generate(input_features)
+        generated_ids_xla = xla_generate(input_features)
+
+        # fmt: off
+        EXPECTED_LOGITS = tf.convert_to_tensor(
+            [
+                [50257, 50362, 1770, 13, 2264, 346, 353, 318, 262, 46329, 286, 262, 3504, 6097, 11, 290, 356, 389, 9675, 284],
+                [50257, 50362, 5414, 318, 1770, 13, 2264, 346, 353, 338, 5642, 1342, 3499, 621, 465, 2300, 13, 50256, 50256, 50256],
+                [50257, 50362, 679, 4952, 514, 326, 379, 428, 43856, 1622, 286, 262, 614, 11, 351, 6786, 290, 32595, 12023, 28236],
+                [50257, 50362, 679, 468, 12296, 17188, 1771, 7361, 26113, 18881, 1122, 338, 670, 318, 1107, 8312, 706, 477, 290, 460]
+            ]
+
+        )
+        # fmt: on
+
+        self.assertTrue(np.allclose(generated_ids, EXPECTED_LOGITS))
+        self.assertTrue(np.allclose(generated_ids_xla, EXPECTED_LOGITS))
+
+        # fmt: off
+        EXPECTED_TRANSCRIPT = [
+            " Mr. Quilter is the apostle of the middle classes, and we are glad to",
+            " Nor is Mr. Quilter's manner less interesting than his matter.",
+            " He tells us that at this festive season of the year, with Christmas and roast beef looming",
+            " He has grave doubts whether Sir Frederick Layton's work is really Greek after all and can",
+        ]
+        # fmt: on
+
+        transcript = processor.batch_decode(generated_ids, skip_special_tokens=True)
+        transcript_xla = processor.batch_decode(generated_ids_xla, skip_special_tokens=True)
+        self.assertListEqual(transcript, EXPECTED_TRANSCRIPT)
+        self.assertListEqual(transcript_xla, EXPECTED_TRANSCRIPT)

tests/test_modeling_tf_common.py

@@ -736,6 +736,23 @@ class TFModelTesterMixin:
                         dtype="float32",
                     ),
                 }
+            elif model_class.__name__ in ["TFWhisperModel", "TFWhisperForConditionalGeneration"]:
+                inputs = {
+                    "decoder_input_ids": tf.keras.Input(
+                        batch_shape=(2, max_input),
+                        name="decoder_input_ids",
+                        dtype="int32",
+                    ),
+                    "input_features": tf.keras.Input(
+                        batch_shape=(
+                            2,
+                            self.model_tester.num_mel_bins,
+                            self.model_tester.seq_length,
+                        ),
+                        name="input_features",
+                        dtype="float32",
+                    ),
+                }
             elif self.is_encoder_decoder:
                 inputs = {
                     "decoder_input_ids": tf.keras.Input(
@@ -1223,8 +1240,17 @@ class TFModelTesterMixin:
 
             # fetch the output for an input exclusively made of new members of the vocabulary
             inputs_dict = copy.deepcopy(original_inputs_dict)
-            new_vocab_input_ids = ids_tensor(inputs_dict["input_ids"].shape, new_tokens_size)
+            ids_feat_name = None
+            if "input_ids" in inputs_dict:
+                ids_feat_name = "input_ids"
+            elif "decoder_input_ids" in inputs_dict:
+                ids_feat_name = "decoder_input_ids"
+            else:
+                assert False, "No input ids feature found in the inputs dict"
+
+            new_vocab_input_ids = ids_tensor(inputs_dict[ids_feat_name].shape, new_tokens_size)
             new_vocab_input_ids += old_total_size
+            inputs_dict[ids_feat_name] = new_vocab_input_ids
             if "input_ids" in inputs_dict:
                 inputs_dict["input_ids"] = new_vocab_input_ids
             if "decoder_input_ids" in inputs_dict:

utils/check_repo.py

@@ -105,6 +105,8 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
     "TFElectraMainLayer",  # Building part of bigger (tested) model (should it be a TFPreTrainedModel ?)
     "TFRobertaForMultipleChoice",  # TODO: fix
     "TrOCRDecoderWrapper",  # Building part of bigger (tested) model.
+    "TFWhisperEncoder",  # Building part of bigger (tested) model.
+    "TFWhisperDecoder",  # Building part of bigger (tested) model.
     "SeparableConv1D",  # Building part of bigger (tested) model.
     "FlaxBartForCausalLM",  # Building part of bigger (tested) model.
     "FlaxBertForCausalLM",  # Building part of bigger (tested) model. Tested implicitly through FlaxRobertaForCausalLM.

utils/documentation_tests.txt

@@ -97,4 +97,5 @@ src/transformers/models/wav2vec2_conformer/modeling_wav2vec2_conformer.py
 src/transformers/models/wav2vec2_with_lm/processing_wav2vec2_with_lm.py
 src/transformers/models/wavlm/modeling_wavlm.py
 src/transformers/models/whisper/modeling_whisper.py
+src/transformers/models/whisper/modeling_tf_whisper.py
 src/transformers/models/yolos/modeling_yolos.py