Add bark (#24086)

* first raw version of the bark integration

* working code on small models with single run

* add converting script from suno weights 2 hf

* many changes

* correct past_kv output

* working implementation for inference

* update the converting script according to the architecture changes

* add a working end-to-end inference code

* remove some comments and make small changes

* remove unecessary comment

* add docstrings and ensure no unecessary intermediary output during audio generation

* remove done TODOs

* make style + add config docstrings

* modification for batch inference support on the whole model

* add details to .generation_audio method

* add copyright

* convert EncodecModel from original library to transformers implementation

* add two class in order to facilitate model and sub-models loading from the hub

* add support of loading the whole model

* add BarkProcessor

* correct modeling according to processor output

* Add proper __init__ and auto support

* Add up-to-date copyright/license message

* add relative import instead of absolute

* cleaner head_dim computation

* small comment removal or changes

* more verbose LayerNorm init method

* specify eps for clearer comprehension

* more verbose variable naming in the MLP module

* remove unecessary BarkBlock parameter

* clearer code in the forward pass of the BarkBlock

* remove _initialize_modules method for cleaner code

* Remove unnecessary methods from sub-models

* move code to remove unnecessary function

* rename a variable for clarity and change an assert

* move code and change variable name for clarity

* remove unnecessary asserts

* correct small bug

* correct a comment

* change variable names for clarity

* remove asserts

* change import from absolute to relative

* correct small error due to comma missing + correct import

* Add attribute Bark config

* add first version of tests

* update attention_map

* add tie_weights and resize_token_embeddings for fineModel

* correct getting attention_mask in generate_text_semantic

* remove Bark inference trick

* leave more choices in barkProcessor

* remove _no_split_modules

* fixe error in forward of block and introduce clearer notations

* correct converting script with last changes

* make style + add draft bark.mdx

* correct BarkModelTest::test_generate_text_semantic

* add Bark in main README

* add dummy_pt_objects for Bark

* add missing models in the main init

* correct test_decoder_model_past_with_large_inputs

* disable torchscript test

* change docstring of BarkProcessor

* Add test_processor_bark

* make style

* correct copyrights

* add bark.mdx + make style, quality and consistency

* Apply suggestions from code review
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Remove unnecessary test method

* simply logic of a test

* Only check first ids for slow audio generation

* split full end-to-end generation tests

* remove unneccessary comment

* change submodel names for clearer naming

* remove ModuleDict from modeling_bark

* combine two if statements

* ensure that an edge misued won't happen

* modify variable name

* move code snippet to the right place (coarse instead of semantic)

* change BarkSemanticModule -> BarkSemanticModel

* align BarkProcessor with transformers paradigm

* correct BarkProcessor tests with last commit changes

* change _validate_voice_preset to an instance method instead of a class method

* tie_weights already called with post_init

* add codec_model config to configuration

* update bark modeling tests with recent BarkProcessor changes

* remove SubModelPretrainedModel + change speakers embeddings prompt type in BarkModel

* change absolute imports to relative

* remove TODO

* change docstrings

* add examples to docs and docstrings

* make style

* uses BatchFeature in BarkProcessor insteads of dict

* continue improving docstrings and docs + make style

* correct docstrings examples

* more comprehensible speaker_embeddings load/Save

* rename speaker_embeddings_dict -> speaker_embeddings

* correct bark.mdx + add bark to documentation_tests

* correct docstrings configuration_bark

* integrate last nit suggestions

* integrate BarkGeneration configs

* make style

* remove bark tests from documentation_tests.txt because timeout - tested manually

* add proper generation config initialization

* small bark.mdx documentation changes

* rename bark.mdx -> bark.md

* add torch.no_grad behind BarkModel.generate_audio()

* replace assert by ValueError in convert_suno_to_hf.py

* integrate a series of short comments from reviewer

* move SemanticLogitsProcessors and remove .detach() from Bark docs and docstrings

* actually remove SemanticLogitsProcessor from modeling_bark.oy

* BarkProcessor returns a single output instead of tuple + correct docstrings

* make style + correct bug

* add initializer_range to BarkConfig + correct slow modeling tests

* add .clone() to history_prompt.coarse_prompt to avoid modifying input array

* Making sure no extra "`" are present

* remove extra characters in modeling_bark.py

* Correct output if history_prompt is None

* remove TODOs

* remove ravel comment

* completing generation_configuration_bark.py docstrings

* change docstrings - number of audio codebooks instead of Encodec codebooks

* change 'bias' docstrings in configuration_bark.py

* format code

* rename BarkModel.generate_audio -> BarkModel.generate_speech

* modify AutoConfig instead of EncodecConfig in BarkConfig

* correct AutoConfig wrong init

* refactor BarkModel and sub-models generate_coarse, generate_fine, generate_text_semantic

* remove SemanticLogitsProcessor and replace it with SuppressTokensLogitsProcessor

* move nb_codebook related config arguments to BarkFineConfig

* rename bark.mdx -> bark.md

* correcting BarkModelConfig from_pretrained + remove keys_to_ignore

* correct bark.md with correct hub path

* correct code bug in bark.md

* correct list tokens_to_suppress

* modify Processor to load nested speaker embeddings in a safer way

* correct batch sampling in BarkFineModel.generate_fine

* Apply suggestions from code review

Small docstrings correction and code improvements
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* give more details about num_layers in docstrings

* correct indentation mistake

* correct submodelconfig order of docstring variables

* put audio models in alphabetical order in utils/check_repo.my

* remove useless line from test_modeling_bark.py

* makes BarkCoarseModelTest inherits from (ModelTesterMixin, GenerationTesterMixin, unittest.TestCase) instead of BarkSemanticModelTest

* make a Tester class for each sub-model instead of inheriting

* add test_resize_embeddings=True for Bark sub-models

* add Copied from transformers.models.gpt_neo.modeling_gpt_neo.GPTNeoSelfAttention._split_heads

* remove 'Copied fom Bark' comment

* remove unneccessary comment

* change np.min -> min in modeling_bark.py

* refactored all custom layers to have Bark prefix

* add attention_mask as an argument of generate_text_semantic

* refactor sub-models start docstrings to have more precise config class definition

* move _tied_weights_keys overriding

* add docstrings to generate_xxx in modeling_bark.py

* add loading whole BarkModel to convert_suno_to_hf

* refactor attribute and variable names

* make style convert_suno

* update bark checkpoints

* remove never entered if statement

* move bark_modeling docstrings after BarkPretrainedModel class definition

* refactor modeling_bark.py: kv -> key_values

* small nits - code refactoring and removing unecessary lines from _init_weights

* nits - replace inplace method by variable assigning

* remove *optional* when necessary

* remove some lines in generate_speech

* add default value for optional parameter

* Refactor preprocess_histories_before_coarse -> preprocess_histories
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* correct usage after refactoring

* refactor Bark's generate_xxx -> generate and modify docstrings and tests accordingly

* update docstrings python in configuration_bark.py

* add bark files in utils/documentation_test.txt

* correct docstrings python snippet

* add the ability to use parameters in the form of e.g coarse_temperature

* add semantic_max_new_tokens in python snippet in docstrings for quicker generation

* Reformate sub-models kwargs in BakModel.generate
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* correct kwargs in BarkModel.generate

* correct attention_mask kwarg in BarkModel.generate

* add tests for sub-models args in BarkModel.generate and correct BarkFineModel.test_generate_fp16

* enrich BarkModel.generate docstrings with a description of how to use the kwargs

---------
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

README.md

@@ -291,6 +291,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.

README_es.md

@@ -268,6 +268,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.

README_hd.md

@@ -240,6 +240,7 @@ conda install -c huggingface transformers
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (फेसबुक) साथ थीसिस [बार्ट: प्राकृतिक भाषा निर्माण, अनुवाद के लिए अनुक्रम-से-अनुक्रम पूर्व प्रशिक्षण , और समझ] (https://arxiv.org/pdf/1910.13461.pdf) पर निर्भर माइक लुईस, यिनहान लियू, नमन गोयल, मार्जन ग़ज़विनिनेजाद, अब्देलरहमान मोहम्मद, ओमर लेवी, वेस स्टोयानोव और ल्यूक ज़ेटलमॉयर
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (से École polytechnique) साथ थीसिस [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) पर निर्भर Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis रिहाई।
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (VinAI Research से) साथ में पेपर [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701)गुयेन लुओंग ट्रान, डुओंग मिन्ह ले और डाट क्वोक गुयेन द्वारा पोस्ट किया गया।

README_ja.md

@@ -302,6 +302,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (BAAI から) Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell から公開された研究論文: [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679)
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (MIT から) Yuan Gong, Yu-An Chung, James Glass から公開された研究論文: [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778)
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (Facebook から) Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer から公開された研究論文: [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461)
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (École polytechnique から) Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis から公開された研究論文: [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321)
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (VinAI Research から) Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen から公開された研究論文: [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701)

README_ko.md

@@ -217,6 +217,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.

README_zh-hans.md

@@ -241,6 +241,7 @@ conda install -c huggingface transformers
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (来自 BAAI) 伴随论文 [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) 由 Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell 发布。
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (来自 MIT) 伴随论文 [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) 由 Yuan Gong, Yu-An Chung, James Glass 发布。
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (来自 Facebook) 伴随论文 [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) 由 Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer 发布。
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (来自 École polytechnique) 伴随论文 [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) 由 Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis 发布。
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (来自 VinAI Research) 伴随论文 [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) 由 Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen 发布。

README_zh-hant.md

@@ -253,6 +253,7 @@ conda install -c huggingface transformers
 1. **[AltCLIP](https://huggingface.co/docs/transformers/model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](https://huggingface.co/docs/transformers/model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](https://huggingface.co/docs/transformers/model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](https://huggingface.co/docs/transformers/main/model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](https://huggingface.co/docs/transformers/model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/pdf/1910.13461.pdf) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](https://huggingface.co/docs/transformers/model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](https://huggingface.co/docs/transformers/model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.

docs/source/en/_toctree.yml

@@ -545,6 +545,8 @@
       sections:
       - local: model_doc/audio-spectrogram-transformer
         title: Audio Spectrogram Transformer
+      - local: model_doc/bark
+        title: Bark
       - local: model_doc/clap
         title: CLAP
       - local: model_doc/encodec

docs/source/en/index.md

@@ -57,6 +57,7 @@ The documentation is organized into five sections:
 1. **[AltCLIP](model_doc/altclip)** (from BAAI) released with the paper [AltCLIP: Altering the Language Encoder in CLIP for Extended Language Capabilities](https://arxiv.org/abs/2211.06679) by Chen, Zhongzhi and Liu, Guang and Zhang, Bo-Wen and Ye, Fulong and Yang, Qinghong and Wu, Ledell.
 1. **[Audio Spectrogram Transformer](model_doc/audio-spectrogram-transformer)** (from MIT) released with the paper [AST: Audio Spectrogram Transformer](https://arxiv.org/abs/2104.01778) by Yuan Gong, Yu-An Chung, James Glass.
 1. **[Autoformer](model_doc/autoformer)** (from Tsinghua University) released with the paper [Autoformer: Decomposition Transformers with Auto-Correlation for Long-Term Series Forecasting](https://arxiv.org/abs/2106.13008) by Haixu Wu, Jiehui Xu, Jianmin Wang, Mingsheng Long.
+1. **[Bark](model_doc/bark)** (from Suno) released in the repository [suno-ai/bark](https://github.com/suno-ai/bark) by Suno AI team.
 1. **[BART](model_doc/bart)** (from Facebook) released with the paper [BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension](https://arxiv.org/abs/1910.13461) by Mike Lewis, Yinhan Liu, Naman Goyal, Marjan Ghazvininejad, Abdelrahman Mohamed, Omer Levy, Ves Stoyanov and Luke Zettlemoyer.
 1. **[BARThez](model_doc/barthez)** (from École polytechnique) released with the paper [BARThez: a Skilled Pretrained French Sequence-to-Sequence Model](https://arxiv.org/abs/2010.12321) by Moussa Kamal Eddine, Antoine J.-P. Tixier, Michalis Vazirgiannis.
 1. **[BARTpho](model_doc/bartpho)** (from VinAI Research) released with the paper [BARTpho: Pre-trained Sequence-to-Sequence Models for Vietnamese](https://arxiv.org/abs/2109.09701) by Nguyen Luong Tran, Duong Minh Le and Dat Quoc Nguyen.
@@ -282,6 +283,7 @@ Flax), PyTorch, and/or TensorFlow.
 |            AltCLIP            |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 | Audio Spectrogram Transformer |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |          Autoformer           |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
+|             Bark              |       ❌       |       ❌       |       ✅        |         ❌         |      ❌      |
 |             BART              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |
 |             BEiT              |       ❌       |       ❌       |       ✅        |         ❌         |      ✅      |
 |             BERT              |       ✅       |       ✅       |       ✅        |         ✅         |      ✅      |

docs/source/en/model_doc/bark.md

+<!--Copyright 2023 The HuggingFace Team. All rights reserved.
+
+Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
+the License. You may obtain a copy of the License at
+
+http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
+an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
+specific language governing permissions and limitations under the License.
+-->
+
+# Bark
+
+## Overview
+
+Bark is a transformer-based text-to-speech model proposed by Suno AI in [suno-ai/bark](https://github.com/suno-ai/bark). 
+
+
+Bark is made of 4 main models:
+
+- [`BarkSemanticModel`] (also referred to as the 'text' model): a causal auto-regressive transformer model that takes as input tokenized text, and predicts semantic text tokens that capture the meaning of the text.
+- [`BarkCoarseModel`] (also referred to as the 'coarse acoustics' model): a causal autoregressive transformer, that takes as input the results of the [`BarkSemanticModel`] model. It aims at predicting the first two audio codebooks necessary for EnCodec.
+- [`BarkFineModel`] (the 'fine acoustics' model), this time a non-causal autoencoder transformer, which iteratively predicts the last codebooks based on the sum of the previous codebooks embeddings.
+- having predicted all the codebook channels from the [`EncodecModel`], Bark uses it to decode the output audio array.
+
+It should be noted that each of the first three modules can support conditional speaker embeddings to condition the output sound according to specific predefined voice.
+
+
+### Tips:
+
+Suno offers a library of voice presets in a number of languages [here](https://suno-ai.notion.site/8b8e8749ed514b0cbf3f699013548683?v=bc67cff786b04b50b3ceb756fd05f68c).
+These presets are also uploaded in the hub [here](https://huggingface.co/suno/bark-small/tree/main/speaker_embeddings) or [here](https://huggingface.co/suno/bark/tree/main/speaker_embeddings).
+
+```python
+>>> from transformers import AutoProcessor, BarkModel
+
+>>> processor = AutoProcessor.from_pretrained("suno/bark")
+>>> model = BarkModel.from_pretrained("suno/bark")
+
+>>> voice_preset = "v2/en_speaker_6"
+
+>>> inputs = processor("Hello, my dog is cute", voice_preset=voice_preset)
+
+>>> audio_array = model.generate(**inputs)
+>>> audio_array = audio_array.cpu().numpy().squeeze()
+```
+
+Bark can generate highly realistic, **multilingual** speech as well as other audio - including music, background noise and simple sound effects. 
+
+```python
+>>> # Multilingual speech - simplified Chinese
+>>> inputs = processor("惊人的！我会说中文")
+
+>>> # Multilingual speech - French - let's use a voice_preset as well
+>>> inputs = processor("Incroyable! Je peux générer du son.", voice_preset="fr_speaker_5")
+
+>>> # Bark can also generate music. You can help it out by adding music notes around your lyrics.
+>>> inputs = processor("♪ Hello, my dog is cute ♪")
+
+>>> audio_array = model.generate(**inputs)
+>>> audio_array = audio_array.cpu().numpy().squeeze()
+```
+
+The model can also produce **nonverbal communications** like laughing, sighing and crying.
+
+
+```python
+>>> # Adding non-speech cues to the input text
+>>> inputs = processor("Hello uh ... [clears throat], my dog is cute [laughter]")
+
+>>> audio_array = model.generate(**inputs)
+>>> audio_array = audio_array.cpu().numpy().squeeze()
+```
+
+To save the audio, simply take the sample rate from the model config and some scipy utility:
+
+```python
+>>> from scipy.io.wavfile import write as write_wav
+
+>>> # save audio to disk, but first take the sample rate from the model config
+>>> sample_rate = model.generation_config.sample_rate
+>>> write_wav("bark_generation.wav", sample_rate, audio_array)
+```
+
+
+This model was contributed by [Yoach Lacombe (ylacombe)](https://huggingface.co/ylacombe) and [Sanchit Gandhi (sanchit-gandhi)](https://github.com/sanchit-gandhi).
+The original code can be found [here](https://github.com/suno-ai/bark).
+
+
+## BarkConfig
+
+[[autodoc]] BarkConfig
+    - all
+
+## BarkProcessor
+
+[[autodoc]] BarkProcessor
+    - all
+    - __call__
+
+## BarkModel
+
+[[autodoc]] BarkModel
+    - generate
+
+## BarkSemanticModel
+
+[[autodoc]] BarkSemanticModel
+    - forward
+
+## BarkCoarseModel
+
+[[autodoc]] BarkCoarseModel
+    - forward
+
+## BarkFineModel
+
+[[autodoc]] BarkFineModel
+    - forward
+
+## BarkCausalModel
+
+[[autodoc]] BarkCausalModel
+    - forward
+
+## BarkCoarseConfig
+
+[[autodoc]] BarkCoarseConfig
+    - all
+
+## BarkFineConfig
+
+[[autodoc]] BarkFineConfig
+    - all
+
+## BarkSemanticConfig
+
+[[autodoc]] BarkSemanticConfig
+    - all
+

src/transformers/__init__.py

@@ -160,6 +160,13 @@ _import_structure = {
         "AUTOFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP",
         "AutoformerConfig",
     ],
+    "models.bark": [
+        "BarkCoarseConfig",
+        "BarkConfig",
+        "BarkFineConfig",
+        "BarkProcessor",
+        "BarkSemanticConfig",
+    ],
     "models.bart": ["BartConfig", "BartTokenizer"],
     "models.barthez": [],
     "models.bartpho": [],
@@ -1136,6 +1143,17 @@ else:
             "AutoformerPreTrainedModel",
         ]
     )
+    _import_structure["models.bark"].extend(
+        [
+            "BARK_PRETRAINED_MODEL_ARCHIVE_LIST",
+            "BarkCausalModel",
+            "BarkCoarseModel",
+            "BarkFineModel",
+            "BarkModel",
+            "BarkPreTrainedModel",
+            "BarkSemanticModel",
+        ]
+    )
     _import_structure["models.bart"].extend(
         [
             "BART_PRETRAINED_MODEL_ARCHIVE_LIST",
@@ -4098,6 +4116,13 @@ if TYPE_CHECKING:
         AUTOFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP,
         AutoformerConfig,
     )
+    from .models.bark import (
+        BarkCoarseConfig,
+        BarkConfig,
+        BarkFineConfig,
+        BarkProcessor,
+        BarkSemanticConfig,
+    )
     from .models.bart import BartConfig, BartTokenizer
     from .models.beit import BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP, BeitConfig
     from .models.bert import (
@@ -4978,6 +5003,15 @@ if TYPE_CHECKING:
             AutoformerModel,
             AutoformerPreTrainedModel,
         )
+        from .models.bark import (
+            BARK_PRETRAINED_MODEL_ARCHIVE_LIST,
+            BarkCausalModel,
+            BarkCoarseModel,
+            BarkFineModel,
+            BarkModel,
+            BarkPreTrainedModel,
+            BarkSemanticModel,
+        )
         from .models.bart import (
             BART_PRETRAINED_MODEL_ARCHIVE_LIST,
             BartForCausalLM,

src/transformers/generation/logits_process.py

@@ -1148,3 +1148,39 @@ class ClassifierFreeGuidanceLogitsProcessor(LogitsProcessor):
         cond_logits, uncond_logits = scores.split(unguided_bsz, dim=0)
         scores = uncond_logits + (cond_logits - uncond_logits) * self.guidance_scale
         return scores
+
+
+class AlternatingCodebooksLogitsProcessor(LogitsProcessor):
+    r"""
+    [`LogitsProcessor`] enforcing alternated generation between the two codebooks of [`Bark`]'s fine submodel.
+
+    Args:
+        input_start_len (`int`):
+            The length of the initial input sequence.
+        semantic_vocab_size (`int`):
+            Vocabulary size of the semantic part, i.e number of tokens associated to the semantic vocabulary.
+        codebook_size (`int`):
+            Number of tokens associated to the codebook.
+    """
+
+    def __init__(self, input_start_len: int, semantic_vocab_size: int, codebook_size: int):
+        if not isinstance(input_start_len, int) or input_start_len < 0:
+            raise ValueError(f"`input_starting_length` has to be a non-negative integer, but is {input_start_len}")
+
+        self.input_start_len = input_start_len
+        self.semantic_vocab_size = semantic_vocab_size
+        self.codebook_size = codebook_size
+
+    def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
+        curr_len = input_ids.shape[-1]
+
+        # even -> first codebook, odd -> second codebook
+        is_first_codebook = ((curr_len - self.input_start_len) % 2) == 0
+
+        if is_first_codebook:
+            scores[:, : self.semantic_vocab_size] = -float("inf")
+            scores[:, self.semantic_vocab_size + self.codebook_size :] = -float("inf")
+        else:
+            scores[:, : self.semantic_vocab_size + self.codebook_size] = -float("inf")
+
+        return scores

src/transformers/models/__init__.py

@@ -19,6 +19,7 @@ from . import (
     audio_spectrogram_transformer,
     auto,
     autoformer,
+    bark,
     bart,
     barthez,
     bartpho,

src/transformers/models/auto/configuration_auto.py

@@ -35,6 +35,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
         ("altclip", "AltCLIPConfig"),
         ("audio-spectrogram-transformer", "ASTConfig"),
         ("autoformer", "AutoformerConfig"),
+        ("bark", "BarkConfig"),
         ("bart", "BartConfig"),
         ("beit", "BeitConfig"),
         ("bert", "BertConfig"),
@@ -237,6 +238,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
         ("altclip", "ALTCLIP_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("audio-spectrogram-transformer", "AUDIO_SPECTROGRAM_TRANSFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("autoformer", "AUTOFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP"),
+        ("bark", "BARK_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bart", "BART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("beit", "BEIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
         ("bert", "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
@@ -419,6 +421,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
         ("altclip", "AltCLIP"),
         ("audio-spectrogram-transformer", "Audio Spectrogram Transformer"),
         ("autoformer", "Autoformer"),
+        ("bark", "Bark"),
         ("bart", "BART"),
         ("barthez", "BARThez"),
         ("bartpho", "BARTpho"),

src/transformers/models/auto/modeling_auto.py

@@ -33,6 +33,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
         ("altclip", "AltCLIPModel"),
         ("audio-spectrogram-transformer", "ASTModel"),
         ("autoformer", "AutoformerModel"),
+        ("bark", "BarkModel"),
         ("bart", "BartModel"),
         ("beit", "BeitModel"),
         ("bert", "BertModel"),

src/transformers/models/auto/processing_auto.py

@@ -44,6 +44,7 @@ PROCESSOR_MAPPING_NAMES = OrderedDict(
     [
         ("align", "AlignProcessor"),
         ("altclip", "AltCLIPProcessor"),
+        ("bark", "BarkProcessor"),
         ("blip", "BlipProcessor"),
         ("blip-2", "Blip2Processor"),
         ("bridgetower", "BridgeTowerProcessor"),

src/transformers/models/bark/__init__.py

+# Copyright 2023 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import TYPE_CHECKING
+
+from ...utils import (
+    OptionalDependencyNotAvailable,
+    _LazyModule,
+    is_torch_available,
+)
+
+
+_import_structure = {
+    "configuration_bark": [
+        "BARK_PRETRAINED_CONFIG_ARCHIVE_MAP",
+        "BarkCoarseConfig",
+        "BarkConfig",
+        "BarkFineConfig",
+        "BarkSemanticConfig",
+    ],
+    "processing_bark": ["BarkProcessor"],
+}
+
+try:
+    if not is_torch_available():
+        raise OptionalDependencyNotAvailable()
+except OptionalDependencyNotAvailable:
+    pass
+else:
+    _import_structure["modeling_bark"] = [
+        "BARK_PRETRAINED_MODEL_ARCHIVE_LIST",
+        "BarkFineModel",
+        "BarkSemanticModel",
+        "BarkCoarseModel",
+        "BarkModel",
+        "BarkPreTrainedModel",
+        "BarkCausalModel",
+    ]
+
+if TYPE_CHECKING:
+    from .configuration_bark import (
+        BARK_PRETRAINED_CONFIG_ARCHIVE_MAP,
+        BarkCoarseConfig,
+        BarkConfig,
+        BarkFineConfig,
+        BarkSemanticConfig,
+    )
+    from .processing_bark import BarkProcessor
+
+    try:
+        if not is_torch_available():
+            raise OptionalDependencyNotAvailable()
+    except OptionalDependencyNotAvailable:
+        pass
+    else:
+        from .modeling_bark import (
+            BARK_PRETRAINED_MODEL_ARCHIVE_LIST,
+            BarkCausalModel,
+            BarkCoarseModel,
+            BarkFineModel,
+            BarkModel,
+            BarkPreTrainedModel,
+            BarkSemanticModel,
+        )
+
+else:
+    import sys
+
+    sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

src/transformers/models/bark/configuration_bark.py

+# coding=utf-8
+# Copyright 2023 The Suno AI 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.
+""" BARK model configuration"""
+
+import copy
+import os
+from typing import Dict, Optional, Union
+
+from ...configuration_utils import PretrainedConfig
+from ...utils import add_start_docstrings, logging
+from ..auto import AutoConfig
+
+
+logger = logging.get_logger(__name__)
+
+
+BARK_PRETRAINED_CONFIG_ARCHIVE_MAP = {
+    "suno/bark-small": "https://huggingface.co/suno/bark-small/resolve/main/config.json",
+    "suno/bark": "https://huggingface.co/suno/bark/resolve/main/config.json",
+}
+
+BARK_SUBMODELCONFIG_START_DOCSTRING = """
+    This is the configuration class to store the configuration of a [`{model}`]. It is used to instantiate the model
+    according to the specified arguments, defining the model architecture. Instantiating a configuration with the
+    defaults will yield a similar configuration to that of the Bark [suno/bark](https://huggingface.co/suno/bark)
+    architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+        block_size (`int`, *optional*, defaults to 1024):
+            The maximum sequence length that this model might ever be used with. Typically set this to something large
+            just in case (e.g., 512 or 1024 or 2048).
+        input_vocab_size (`int`, *optional*, defaults to 10_048):
+            Vocabulary size of a Bark sub-model. Defines the number of different tokens that can be represented by the
+            `inputs_ids` passed when calling [`{model}`]. Defaults to 10_048 but should be carefully thought with
+            regards to the chosen sub-model.
+        output_vocab_size (`int`, *optional*, defaults to 10_048):
+            Output vocabulary size of a Bark sub-model. Defines the number of different tokens that can be represented
+            by the: `output_ids` when passing forward a [`{model}`]. Defaults to 10_048 but should be carefully thought
+            with regards to the chosen sub-model.
+        num_layers (`int`, *optional*, defaults to 12):
+            Number of hidden layers in the given sub-model.
+        num_heads (`int`, *optional*, defaults to 12):
+            Number of attention heads for each attention layer in the Transformer architecture.
+        hidden_size (`int`, *optional*, defaults to 768):
+            Dimensionality of the "intermediate" (often named feed-forward) layer in the architecture.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
+        bias (`bool`, *optional*, defaults to `True`):
+            Whether or not to use bias in the linear layers and layer norm layers.
+        initializer_range (`float`, *optional*, defaults to 0.02):
+            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+        use_cache (`bool`, *optional*, defaults to `True`):
+            Whether or not the model should return the last key/values attentions (not used by all models).
+"""
+
+
+class BarkSubModelConfig(PretrainedConfig):
+    model_type = "bark_module"
+    keys_to_ignore_at_inference = ["past_key_values"]
+
+    attribute_map = {
+        "num_attention_heads": "num_heads",
+        "num_hidden_layers": "num_layers",
+        "vocab_size": "input_vocab_size",
+        "window_size": "block_size",
+    }
+
+    def __init__(
+        self,
+        block_size=1024,
+        input_vocab_size=10_048,
+        output_vocab_size=10_048,
+        num_layers=12,
+        num_heads=12,
+        hidden_size=768,
+        dropout=0.0,
+        bias=True,  # True: bias in Linears and LayerNorms, like GPT-2. False: a bit better and faster
+        initializer_range=0.02,
+        use_cache=True,
+        **kwargs,
+    ):
+        self.block_size = block_size
+        self.input_vocab_size = input_vocab_size
+        self.output_vocab_size = output_vocab_size
+        self.num_layers = num_layers
+        self.num_heads = num_heads
+        self.hidden_size = hidden_size
+        self.dropout = dropout
+        self.bias = bias
+        self.use_cache = use_cache
+        self.initializer_range = initializer_range
+
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_pretrained(
+        cls,
+        pretrained_model_name_or_path: Union[str, os.PathLike],
+        cache_dir: Optional[Union[str, os.PathLike]] = None,
+        force_download: bool = False,
+        local_files_only: bool = False,
+        token: Optional[Union[str, bool]] = None,
+        revision: str = "main",
+        **kwargs,
+    ) -> "PretrainedConfig":
+        kwargs["cache_dir"] = cache_dir
+        kwargs["force_download"] = force_download
+        kwargs["local_files_only"] = local_files_only
+        kwargs["revision"] = revision
+
+        cls._set_token_in_kwargs(kwargs, token)
+
+        config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+        # get the config dict if we are loading from Bark
+        if config_dict.get("model_type") == "bark":
+            config_dict = config_dict[f"{cls.model_type}_config"]
+
+        if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+            logger.warning(
+                f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+                f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+            )
+
+        return cls.from_dict(config_dict, **kwargs)
+
+
+@add_start_docstrings(
+    BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkSemanticConfig", model="BarkSemanticModel"),
+    """
+    Example:
+
+    ```python
+    >>> from transformers import BarkSemanticConfig, BarkSemanticModel
+
+    >>> # Initializing a Bark sub-module style configuration
+    >>> configuration = BarkSemanticConfig()
+
+    >>> # Initializing a model (with random weights) from the suno/bark style configuration
+    >>> model = BarkSemanticModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```""",
+)
+class BarkSemanticConfig(BarkSubModelConfig):
+    model_type = "semantic"
+
+
+@add_start_docstrings(
+    BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkCoarseConfig", model="BarkCoarseModel"),
+    """
+    Example:
+
+    ```python
+    >>> from transformers import BarkCoarseConfig, BarkCoarseModel
+
+    >>> # Initializing a Bark sub-module style configuration
+    >>> configuration = BarkCoarseConfig()
+
+    >>> # Initializing a model (with random weights) from the suno/bark style configuration
+    >>> model = BarkCoarseModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```""",
+)
+class BarkCoarseConfig(BarkSubModelConfig):
+    model_type = "coarse_acoustics"
+
+
+@add_start_docstrings(
+    BARK_SUBMODELCONFIG_START_DOCSTRING.format(config="BarkFineConfig", model="BarkFineModel"),
+    """
+        n_codes_total (`int`, *optional*, defaults to 8):
+            The total number of audio codebooks predicted. Used in the fine acoustics sub-model.
+        n_codes_given (`int`, *optional*, defaults to 1):
+            The number of audio codebooks predicted in the coarse acoustics sub-model. Used in the acoustics
+            sub-models.
+    Example:
+
+    ```python
+    >>> from transformers import BarkFineConfig, BarkFineModel
+
+    >>> # Initializing a Bark sub-module style configuration
+    >>> configuration = BarkFineConfig()
+
+    >>> # Initializing a model (with random weights) from the suno/bark style configuration
+    >>> model = BarkFineModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```""",
+)
+class BarkFineConfig(BarkSubModelConfig):
+    model_type = "fine_acoustics"
+
+    def __init__(self, tie_word_embeddings=True, n_codes_total=8, n_codes_given=1, **kwargs):
+        self.n_codes_total = n_codes_total
+        self.n_codes_given = n_codes_given
+
+        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
+
+
+class BarkConfig(PretrainedConfig):
+    """
+    This is the configuration class to store the configuration of a [`BarkModel`]. It is used to instantiate a Bark
+    model according to the specified sub-models configurations, defining the model architecture.
+
+    Instantiating a configuration with the defaults will yield a similar configuration to that of the Bark
+    [suno/bark](https://huggingface.co/suno/bark) architecture.
+
+    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+    documentation from [`PretrainedConfig`] for more information.
+
+    Args:
+    semantic_config ([`BarkSemanticConfig`], *optional*):
+        Configuration of the underlying semantic sub-model.
+    coarse_acoustics_config ([`BarkCoarseConfig`], *optional*):
+        Configuration of the underlying coarse acoustics sub-model.
+    fine_acoustics_config ([`BarkFineConfig`], *optional*):
+        Configuration of the underlying fine acoustics sub-model.
+    codec_config ([`AutoConfig`], *optional*):
+        Configuration of the underlying codec sub-model.
+
+    Example:
+
+    ```python
+    >>> from transformers import (
+    ...     BarkSemanticConfig,
+    ...     BarkCoarseConfig,
+    ...     BarkFineConfig,
+    ...     BarkModel,
+    ...     BarkConfig,
+    ...     AutoConfig,
+    ... )
+
+    >>> # Initializing Bark sub-modules configurations.
+    >>> semantic_config = BarkSemanticConfig()
+    >>> coarse_acoustics_config = BarkCoarseConfig()
+    >>> fine_acoustics_config = BarkFineConfig()
+    >>> codec_config = AutoConfig.from_pretrained("facebook/encodec_24khz")
+
+
+    >>> # Initializing a Bark module style configuration
+    >>> configuration = BarkConfig.from_sub_model_configs(
+    ...     semantic_config, coarse_acoustics_config, fine_acoustics_config, codec_config
+    ... )
+
+    >>> # Initializing a model (with random weights)
+    >>> model = BarkModel(configuration)
+
+    >>> # Accessing the model configuration
+    >>> configuration = model.config
+    ```
+    """
+
+    model_type = "bark"
+    is_composition = True
+
+    def __init__(
+        self,
+        semantic_config: Dict = None,
+        coarse_acoustics_config: Dict = None,
+        fine_acoustics_config: Dict = None,
+        codec_config: Dict = None,
+        initializer_range=0.02,
+        **kwargs,
+    ):
+        if semantic_config is None:
+            semantic_config = {}
+            logger.info("semantic_config is None. initializing the semantic model with default values.")
+
+        if coarse_acoustics_config is None:
+            coarse_acoustics_config = {}
+            logger.info("coarse_acoustics_config is None. initializing the coarse model with default values.")
+
+        if fine_acoustics_config is None:
+            fine_acoustics_config = {}
+            logger.info("fine_acoustics_config is None. initializing the fine model with default values.")
+
+        if codec_config is None:
+            codec_config = {}
+            logger.info("codec_config is None. initializing the codec model with default values.")
+
+        self.semantic_config = BarkSemanticConfig(**semantic_config)
+        self.coarse_acoustics_config = BarkCoarseConfig(**coarse_acoustics_config)
+        self.fine_acoustics_config = BarkFineConfig(**fine_acoustics_config)
+        self.codec_config = AutoConfig.for_model(**codec_config)
+
+        self.initializer_range = initializer_range
+
+        super().__init__(**kwargs)
+
+    @classmethod
+    def from_sub_model_configs(
+        cls,
+        semantic_config: BarkSemanticConfig,
+        coarse_acoustics_config: BarkCoarseConfig,
+        fine_acoustics_config: BarkFineConfig,
+        codec_config: AutoConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`BarkConfig`] (or a derived class) from bark sub-models configuration.
+
+        Returns:
+            [`BarkConfig`]: An instance of a configuration object
+        """
+        return cls(
+            semantic_config=semantic_config.to_dict(),
+            coarse_acoustics_config=coarse_acoustics_config.to_dict(),
+            fine_acoustics_config=fine_acoustics_config.to_dict(),
+            codec_config=codec_config.to_dict(),
+            **kwargs,
+        )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+
+        output["semantic_config"] = self.semantic_config.to_dict()
+        output["coarse_acoustics_config"] = self.coarse_acoustics_config.to_dict()
+        output["fine_acoustics_config"] = self.fine_acoustics_config.to_dict()
+        output["codec_config"] = self.codec_config.to_dict()
+
+        output["model_type"] = self.__class__.model_type
+        return output

src/transformers/models/bark/convert_suno_to_hf.py

+"""Convert Bark checkpoint."""
+import argparse
+import os
+from pathlib import Path
+
+import torch
+from bark.generation import _load_model as _bark_load_model
+from huggingface_hub import hf_hub_download
+
+from transformers import EncodecConfig, EncodecModel, set_seed
+from transformers.models.bark.configuration_bark import (
+    BarkCoarseConfig,
+    BarkConfig,
+    BarkFineConfig,
+    BarkSemanticConfig,
+)
+from transformers.models.bark.generation_configuration_bark import (
+    BarkCoarseGenerationConfig,
+    BarkFineGenerationConfig,
+    BarkGenerationConfig,
+    BarkSemanticGenerationConfig,
+)
+from transformers.models.bark.modeling_bark import BarkCoarseModel, BarkFineModel, BarkModel, BarkSemanticModel
+from transformers.utils import logging
+
+
+logging.set_verbosity_info()
+logger = logging.get_logger(__name__)
+
+set_seed(770)
+
+
+new_layer_name_dict = {
+    "c_attn": "att_proj",
+    "c_proj": "out_proj",
+    "c_fc": "in_proj",
+    "transformer.": "",
+    "h.": "layers.",
+    "ln_1": "layernorm_1",
+    "ln_2": "layernorm_2",
+    "ln_f": "layernorm_final",
+    "wpe": "position_embeds_layer",
+    "wte": "input_embeds_layer",
+}
+
+
+REMOTE_MODEL_PATHS = {
+    "text_small": {
+        "repo_id": "suno/bark",
+        "file_name": "text.pt",
+    },
+    "coarse_small": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse.pt",
+    },
+    "fine_small": {
+        "repo_id": "suno/bark",
+        "file_name": "fine.pt",
+    },
+    "text": {
+        "repo_id": "suno/bark",
+        "file_name": "text_2.pt",
+    },
+    "coarse": {
+        "repo_id": "suno/bark",
+        "file_name": "coarse_2.pt",
+    },
+    "fine": {
+        "repo_id": "suno/bark",
+        "file_name": "fine_2.pt",
+    },
+}
+
+CUR_PATH = os.path.dirname(os.path.abspath(__file__))
+default_cache_dir = os.path.join(os.path.expanduser("~"), ".cache")
+CACHE_DIR = os.path.join(os.getenv("XDG_CACHE_HOME", default_cache_dir), "suno", "bark_v0")
+
+
+def _get_ckpt_path(model_type, use_small=False):
+    key = model_type
+    if use_small:
+        key += "_small"
+    return os.path.join(CACHE_DIR, REMOTE_MODEL_PATHS[key]["file_name"])
+
+
+def _download(from_hf_path, file_name):
+    os.makedirs(CACHE_DIR, exist_ok=True)
+    hf_hub_download(repo_id=from_hf_path, filename=file_name, local_dir=CACHE_DIR)
+
+
+def _load_model(ckpt_path, device, use_small=False, model_type="text"):
+    if model_type == "text":
+        ModelClass = BarkSemanticModel
+        ConfigClass = BarkSemanticConfig
+        GenerationConfigClass = BarkSemanticGenerationConfig
+    elif model_type == "coarse":
+        ModelClass = BarkCoarseModel
+        ConfigClass = BarkCoarseConfig
+        GenerationConfigClass = BarkCoarseGenerationConfig
+    elif model_type == "fine":
+        ModelClass = BarkFineModel
+        ConfigClass = BarkFineConfig
+        GenerationConfigClass = BarkFineGenerationConfig
+    else:
+        raise NotImplementedError()
+    model_key = f"{model_type}_small" if use_small else model_type
+    model_info = REMOTE_MODEL_PATHS[model_key]
+    if not os.path.exists(ckpt_path):
+        logger.info(f"{model_type} model not found, downloading into `{CACHE_DIR}`.")
+        _download(model_info["repo_id"], model_info["file_name"])
+    checkpoint = torch.load(ckpt_path, map_location=device)
+    # this is a hack
+    model_args = checkpoint["model_args"]
+    if "input_vocab_size" not in model_args:
+        model_args["input_vocab_size"] = model_args["vocab_size"]
+        model_args["output_vocab_size"] = model_args["vocab_size"]
+        del model_args["vocab_size"]
+
+    # convert Bark model arguments to HF Bark model arguments
+    model_args["num_heads"] = model_args.pop("n_head")
+    model_args["hidden_size"] = model_args.pop("n_embd")
+    model_args["num_layers"] = model_args.pop("n_layer")
+
+    model_config = ConfigClass(**checkpoint["model_args"])
+    model = ModelClass(config=model_config)
+    model_generation_config = GenerationConfigClass()
+
+    model.generation_config = model_generation_config
+    state_dict = checkpoint["model"]
+    # fixup checkpoint
+    unwanted_prefix = "_orig_mod."
+    for k, v in list(state_dict.items()):
+        if k.startswith(unwanted_prefix):
+            # replace part of the key with corresponding layer name in HF implementation
+            new_k = k[len(unwanted_prefix) :]
+            for old_layer_name in new_layer_name_dict:
+                new_k = new_k.replace(old_layer_name, new_layer_name_dict[old_layer_name])
+
+            state_dict[new_k] = state_dict.pop(k)
+
+    extra_keys = set(state_dict.keys()) - set(model.state_dict().keys())
+    extra_keys = {k for k in extra_keys if not k.endswith(".attn.bias")}
+    missing_keys = set(model.state_dict().keys()) - set(state_dict.keys())
+    missing_keys = {k for k in missing_keys if not k.endswith(".attn.bias")}
+    if len(extra_keys) != 0:
+        raise ValueError(f"extra keys found: {extra_keys}")
+    if len(missing_keys) != 0:
+        raise ValueError(f"missing keys: {missing_keys}")
+    model.load_state_dict(state_dict, strict=False)
+    n_params = model.num_parameters(exclude_embeddings=True)
+    val_loss = checkpoint["best_val_loss"].item()
+    logger.info(f"model loaded: {round(n_params/1e6,1)}M params, {round(val_loss,3)} loss")
+    model.eval()
+    model.to(device)
+    del checkpoint, state_dict
+
+    return model
+
+
+def load_model(pytorch_dump_folder_path, use_small=False, model_type="text"):
+    if model_type not in ("text", "coarse", "fine"):
+        raise NotImplementedError()
+
+    device = "cpu"  # do conversion on cpu
+
+    ckpt_path = _get_ckpt_path(model_type, use_small=use_small)
+    model = _load_model(ckpt_path, device, model_type=model_type, use_small=use_small)
+
+    # load bark initial model
+    bark_model = _bark_load_model(ckpt_path, "cpu", model_type=model_type, use_small=use_small)
+
+    if model_type == "text":
+        bark_model = bark_model["model"]
+
+    if model.num_parameters(exclude_embeddings=True) != bark_model.get_num_params():
+        raise ValueError("initial and new models don't have the same number of parameters")
+
+    # check if same output as the bark model
+    batch_size = 5
+    sequence_length = 10
+
+    if model_type in ["text", "coarse"]:
+        vec = torch.randint(256, (batch_size, sequence_length), dtype=torch.int)
+        output_old_model = bark_model(vec)[0]
+
+        output_new_model_total = model(vec)
+
+        # take last logits
+        output_new_model = output_new_model_total.logits[:, [-1], :]
+
+    else:
+        prediction_codeboook_channel = 3
+        n_codes_total = 8
+        vec = torch.randint(256, (batch_size, sequence_length, n_codes_total), dtype=torch.int)
+
+        output_new_model_total = model(prediction_codeboook_channel, vec)
+        output_old_model = bark_model(prediction_codeboook_channel, vec)
+
+        output_new_model = output_new_model_total.logits
+
+    # output difference should come from the difference of self-attention implementation design
+    if output_new_model.shape != output_old_model.shape:
+        raise ValueError("initial and new outputs don't have the same shape")
+    if (output_new_model - output_old_model).abs().max().item() > 1e-3:
+        raise ValueError("initial and new outputs are not equal")
+
+    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+    model.save_pretrained(pytorch_dump_folder_path)
+
+
+def load_whole_bark_model(
+    semantic_path,
+    coarse_path,
+    fine_path,
+    append_text,
+    hub_path,
+    folder_path,
+):
+    pytorch_dump_folder_path = os.path.join(folder_path, append_text)
+
+    semanticConfig = BarkSemanticConfig.from_pretrained(os.path.join(semantic_path, "config.json"))
+    coarseAcousticConfig = BarkCoarseConfig.from_pretrained(os.path.join(coarse_path, "config.json"))
+    fineAcousticConfig = BarkFineConfig.from_pretrained(os.path.join(fine_path, "config.json"))
+    codecConfig = EncodecConfig.from_pretrained("facebook/encodec_24khz")
+
+    semantic = BarkSemanticModel.from_pretrained(semantic_path)
+    coarseAcoustic = BarkCoarseModel.from_pretrained(coarse_path)
+    fineAcoustic = BarkFineModel.from_pretrained(fine_path)
+    codec = EncodecModel.from_pretrained("facebook/encodec_24khz")
+
+    bark_config = BarkConfig.from_sub_model_configs(
+        semanticConfig, coarseAcousticConfig, fineAcousticConfig, codecConfig
+    )
+
+    bark_generation_config = BarkGenerationConfig.from_sub_model_configs(
+        semantic.generation_config, coarseAcoustic.generation_config, fineAcoustic.generation_config
+    )
+
+    bark = BarkModel(bark_config)
+
+    bark.semantic = semantic
+    bark.coarse_acoustics = coarseAcoustic
+    bark.fine_acoustics = fineAcoustic
+    bark.codec_model = codec
+
+    bark.generation_config = bark_generation_config
+
+    Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
+    bark.save_pretrained(pytorch_dump_folder_path, repo_id=hub_path, push_to_hub=True)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    # Required parameters
+
+    parser.add_argument("model_type", type=str, help="text, coarse or fine.")
+    parser.add_argument("pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
+    parser.add_argument("--is_small", action="store_true", help="convert the small version instead of the large.")
+
+    args = parser.parse_args()
+
+    load_model(args.pytorch_dump_folder_path, model_type=args.model_type, use_small=args.is_small)

src/transformers/models/bark/generation_configuration_bark.py

+# coding=utf-8
+# Copyright 2023 The Suno AI 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.
+""" BARK model generation configuration"""
+
+import copy
+from typing import Dict
+
+from ...generation.configuration_utils import GenerationConfig
+from ...utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+
+class BarkSemanticGenerationConfig(GenerationConfig):
+    model_type = "semantic"
+
+    def __init__(
+        self,
+        eos_token_id=10_000,
+        renormalize_logits=True,
+        max_new_tokens=768,
+        output_scores=False,
+        return_dict_in_generate=False,
+        output_hidden_states=False,
+        output_attentions=False,
+        temperature=0.7,
+        do_sample=True,
+        text_encoding_offset=10_048,
+        text_pad_token=129_595,
+        semantic_infer_token=129_599,
+        semantic_vocab_size=10_000,
+        max_input_semantic_length=256,
+        semantic_rate_hz=49.9,
+        **kwargs,
+    ):
+        """Class that holds a generation configuration for [`BarkSemanticModel`].
+
+        This configuration inherit from [`GenerationConfig`] and can be used to control the model generation. Read the
+        documentation from [`GenerationConfig`] for more information.
+
+        Args:
+            eos_token_id (`int`, *optional*, defaults to 10_000):
+                The id of the *end-of-sequence* token.
+            renormalize_logits (`bool`, *optional*, defaults to `True`):
+                Whether to renormalize the logits after applying all the logits processors or warpers (including the
+                custom ones). It's highly recommended to set this flag to `True` as the search algorithms suppose the
+                score logits are normalized but some logit processors or warpers break the normalization.
+            max_new_tokens (`int`, *optional*, defaults to 768):
+                The maximum numbers of tokens to generate, ignoring the number of tokens in the prompt.
+            output_scores (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
+            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+            output_hidden_states (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more details.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more details.
+            temperature (`float`, *optional*, defaults to 0.7):
+                The value used to modulate the next token probabilities.
+            do_sample (`bool`, *optional*, defaults to `True`):
+                Whether or not to use sampling ; use greedy decoding otherwise.
+            text_encoding_offset (`int`, *optional*, defaults to 10_048):
+                Text encoding offset.
+            text_pad_token (`int`, *optional*, defaults to 129_595):
+                Text pad token.
+            semantic_infer_token (`int`, *optional*, defaults to 129_599):
+                Semantic infer token.
+            semantic_vocab_size (`int`, *optional*, defaults to 10_000):
+                Semantic vocab size.
+            max_input_semantic_length (`int`, *optional*, defaults to 256):
+                Max length of semantic input vector.
+            semantic_rate_hz (`float`, *optional*, defaults to 49.9):
+                Semantic rate in Hertz.
+        """
+        super().__init__(
+            temperature=temperature,
+            do_sample=do_sample,
+            eos_token_id=eos_token_id,
+            renormalize_logits=renormalize_logits,
+            max_new_tokens=max_new_tokens,
+            output_scores=output_scores,
+            return_dict_in_generate=return_dict_in_generate,
+            output_hidden_states=output_hidden_states,
+            output_attentions=output_attentions,
+            **kwargs,
+        )
+
+        self.text_encoding_offset = text_encoding_offset
+        self.text_pad_token = text_pad_token
+        self.semantic_pad_token = eos_token_id
+        self.semantic_infer_token = semantic_infer_token
+        self.semantic_vocab_size = semantic_vocab_size
+        self.max_input_semantic_length = max_input_semantic_length
+        self.semantic_rate_hz = semantic_rate_hz
+
+
+class BarkCoarseGenerationConfig(GenerationConfig):
+    model_type = "coarse_acoustics"
+
+    def __init__(
+        self,
+        renormalize_logits=True,
+        output_scores=False,
+        return_dict_in_generate=False,
+        output_hidden_states=False,
+        output_attentions=False,
+        temperature=0.7,
+        do_sample=True,
+        coarse_semantic_pad_token=12_048,
+        coarse_rate_hz=75,
+        n_coarse_codebooks=2,
+        coarse_infer_token=12_050,
+        max_coarse_input_length=256,
+        max_coarse_history: int = 630,
+        sliding_window_len: int = 60,
+        **kwargs,
+    ):
+        """Class that holds a generation configuration for [`BarkCoarseModel`].
+
+        This configuration inherit from [`GenerationConfig`] and can be used to control the model generation. Read the
+        documentation from [`GenerationConfig`] for more information.
+
+        Args:
+            renormalize_logits (`bool`, *optional*, defaults to `True`):
+                Whether to renormalize the logits after applying all the logits processors or warpers (including the
+                custom ones). It's highly recommended to set this flag to `True` as the search algorithms suppose the
+                score logits are normalized but some logit processors or warpers break the normalization.
+            output_scores (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the prediction scores. See `scores` under returned tensors for more details.
+            return_dict_in_generate (`bool`, *optional*, defaults to `False`):
+                Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple.
+            output_hidden_states (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors
+                for more details.
+            output_attentions (`bool`, *optional*, defaults to `False`):
+                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+                returned tensors for more details.
+            temperature (`float`, *optional*, defaults to 0.7):
+                The value used to modulate the next token probabilities.
+            do_sample (`bool`, *optional*, defaults to `True`):
+                Whether or not to use sampling ; use greedy decoding otherwise.
+            coarse_semantic_pad_token (`int`, *optional*, defaults to 12_048):
+                Coarse semantic pad token.
+            coarse_rate_hz (`int`, *optional*, defaults to 75):
+                Coarse rate in Hertz.
+            n_coarse_codebooks (`int`, *optional*, defaults to 2):
+                Number of coarse codebooks.
+            coarse_infer_token (`int`, *optional*, defaults to 12_050):
+                Coarse infer token.
+            max_coarse_input_length (`int`, *optional*, defaults to 256):
+                Max length of input coarse vector.
+            max_coarse_history (`int`, *optional*, defaults to 630):
+                Max length of the output of the coarse acoustics model used in the fine generation step.
+            sliding_window_len (`int`, *optional*, defaults to 60):
+                The coarse generation step uses a sliding window to generate raw audio.
+        """
+        super().__init__(
+            temperature=temperature,
+            do_sample=do_sample,
+            renormalize_logits=renormalize_logits,
+            output_scores=output_scores,
+            return_dict_in_generate=return_dict_in_generate,
+            output_hidden_states=output_hidden_states,
+            output_attentions=output_attentions,
+            **kwargs,
+        )
+
+        self.coarse_semantic_pad_token = coarse_semantic_pad_token
+        self.coarse_rate_hz = coarse_rate_hz
+        self.n_coarse_codebooks = n_coarse_codebooks
+        self.coarse_infer_token = coarse_infer_token
+        self.max_coarse_input_length = max_coarse_input_length
+        self.max_coarse_history = max_coarse_history
+        self.sliding_window_len = sliding_window_len
+
+
+class BarkFineGenerationConfig(GenerationConfig):
+    model_type = "fine_acoustics"
+
+    def __init__(
+        self,
+        temperature=0.5,
+        max_fine_history_length=512,
+        max_fine_input_length=1024,
+        n_fine_codebooks=8,
+        **kwargs,
+    ):
+        """Class that holds a generation configuration for [`BarkFineModel`].
+
+        [`BarkFineModel`] is an autoencoder model, so should not usually be used for generation. However, under the
+        hood, it uses `temperature` when used by [`BarkModel`]
+
+        This configuration inherit from [`GenerationConfig`] and can be used to control the model generation. Read the
+        documentation from [`GenerationConfig`] for more information.
+
+        Args:
+            temperature (`float`, *optional*, defaults to 0.5):
+                The value used to modulate the next token probabilities.
+            max_fine_history_length (`int`, *optional*, defaults to 512):
+                Max length of the fine history vector.
+            max_fine_input_length (`int`, *optional*, defaults to 1024):
+                Max length of fine input vector.
+            n_fine_codebooks (`int`, *optional*, defaults to 8):
+                Number of codebooks used.
+        """
+        super().__init__(temperature=temperature)
+
+        self.max_fine_history_length = max_fine_history_length
+        self.max_fine_input_length = max_fine_input_length
+        self.n_fine_codebooks = n_fine_codebooks
+
+
+class BarkGenerationConfig(GenerationConfig):
+    model_type = "bark"
+    is_composition = True
+
+    # TODO (joao): nested from_dict
+
+    def __init__(
+        self,
+        semantic_config: Dict = None,
+        coarse_acoustics_config: Dict = None,
+        fine_acoustics_config: Dict = None,
+        sample_rate=24_000,
+        codebook_size=1024,
+        **kwargs,
+    ):
+        """Class that holds a generation configuration for [`BarkModel`].
+
+        The [`BarkModel`] does not have a `generate` method, but uses this class to generate speeches with a nested
+        [`BarkGenerationConfig`] which uses [`BarkSemanticGenerationConfig`], [`BarkCoarseGenerationConfig`],
+        [`BarkFineGenerationConfig`].
+
+        This configuration inherit from [`GenerationConfig`] and can be used to control the model generation. Read the
+        documentation from [`GenerationConfig`] for more information.
+
+        Args:
+            semantic_config (`Dict`, *optional*):
+                Semantic generation configuration.
+            coarse_acoustics_config (`Dict`, *optional*):
+                Coarse generation configuration.
+            fine_acoustics_config (`Dict`, *optional*):
+                Fine generation configuration.
+            sample_rate (`int`, *optional*, defaults to 24_000):
+                Sample rate.
+            codebook_size (`int`, *optional*, defaults to 1024):
+                Vector length for each codebook.
+        """
+        if semantic_config is None:
+            semantic_config = {}
+            logger.info("semantic_config is None. initializing the semantic model with default values.")
+
+        if coarse_acoustics_config is None:
+            coarse_acoustics_config = {}
+            logger.info("coarse_acoustics_config is None. initializing the coarse model with default values.")
+
+        if fine_acoustics_config is None:
+            fine_acoustics_config = {}
+            logger.info("fine_acoustics_config is None. initializing the fine model with default values.")
+
+        self.semantic_config = BarkSemanticGenerationConfig(**semantic_config)
+        self.coarse_acoustics_config = BarkCoarseGenerationConfig(**coarse_acoustics_config)
+        self.fine_acoustics_config = BarkFineGenerationConfig(**fine_acoustics_config)
+
+        self.sample_rate = sample_rate
+        self.codebook_size = codebook_size
+
+    @classmethod
+    def from_sub_model_configs(
+        cls,
+        semantic_config: BarkSemanticGenerationConfig,
+        coarse_acoustics_config: BarkCoarseGenerationConfig,
+        fine_acoustics_config: BarkFineGenerationConfig,
+        **kwargs,
+    ):
+        r"""
+        Instantiate a [`BarkGenerationConfig`] (or a derived class) from bark sub-models generation configuration.
+
+        Returns:
+            [`BarkGenerationConfig`]: An instance of a configuration object
+        """
+        return cls(
+            semantic_config=semantic_config.to_dict(),
+            coarse_acoustics_config=coarse_acoustics_config.to_dict(),
+            fine_acoustics_config=fine_acoustics_config.to_dict(),
+            **kwargs,
+        )
+
+    def to_dict(self):
+        """
+        Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
+
+        Returns:
+            `Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
+        """
+        output = copy.deepcopy(self.__dict__)
+
+        output["semantic_config"] = self.semantic_config.to_dict()
+        output["coarse_acoustics_config"] = self.coarse_acoustics_config.to_dict()
+        output["fine_acoustics_config"] = self.fine_acoustics_config.to_dict()
+
+        output["model_type"] = self.__class__.model_type
+        return output