dnrops
/
transformers
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Add jamba (#29943) 

* Add jamba arch

* apply "make fix-copies" changes

* fix link to model in JambaConfig docstring

* Add n_ctx in modeling file because repo-consistency wants that

* Add jamba to flash attention and sdpa documentation

* mamba dt_proj quant fix now works for LoRA as well

* override test_left_padding_compatibility and use a more permissive tolerance. left padding numerical difference are accentuated by mamba layers

* add jamba to tokenization auto

* fix comments of shape (PR #24 in the model page: https://huggingface.co/ai21labs/Jamba-v0.1/discussions/24)

* simple PR fixes

* remove unnecessary kwargs from JambaAttentionDecoderLayer and JambaMambaDecoderLayer

* remove the LoRA hack for the mamba dt_proj bias. It was solved in huggingface/peft#1530 (https://github.com/huggingface/peft/pull/1530)

* Add copied comment on JambaMLP (it's the same as MixtralMLP)

* remove padding_mask warnings. It's not supported anymore

* fix docstring. Float instead of int

* A few more minor PR fixes

* (1) lowercase names for mamba layernorms (2) remove _apply_inner_layernorms and do it directly in the forward pass

* Return None attention weights from mamba layers. Append to all attentions only if not None.

* remove some leftover jamba archive lists

* Better separation between expert vs non-expert layers. non-expert layers return None as router_logits, and it is not concatenated to all_router_logits returned from JambaModel

* no need to take router_logits at config.expert_layer_offset anymore. result.router_logits now holds results only for expert layers

* Add Jamba paper on READMEs

* (1) rename n_ctx -> max_position_embeddings (2) don't use it in the modeling file since it's not needed (set it as an exception to check_config_attributes)

* Add copied from comment

* remove the code path for apply_inner_layernorms=False. Jamba always has the inner mamba layernorms

* clearer docstring for _convert_to_standard_cache

* style fixes

* Change calc_logits_for_entire_prompt (bool) to num_logits_to_keep (int). Adapt assisted decoding code tp use it. Also small change in low memory beam search decoding path to support this new int value in model_inputs

* rename test so it still overrides what its meant to override

* draft

* oups

* nit

* remove more complexe logic

* fix names used in config

* fix fix fix

* style

* fix some more failing tests

* generate did not init the cache 🙃

* more small nits

* typo

* config.mamba_expand * config.hidden_size for the intermediate size of the mamba shapes

* fix init of pkv with torch.tensor()

* empty tensor

* fix some init issues

* stupid changes required by generate because it does not even support it's own DynamicCache class

* more fixes

* fix general assisted gen cache_position bug

* tests passing

* Add offsets and periods as SPECIAL_CASES_TO_ALLOW in check_config_attributes.py

* fix reorder_cache to reorder mamba states and override some more functions in HybridMambaAttentionDynamicCache

* no need to override test_past_key_values_format() and _check_past_key_values_for_generate() in tests anymore

* fix docstrings and typehints for past_key_values

* style fixes

* fix docs

* change typehint due to copy from Mixtral

* forgot import

* import order

* Add configuration_jamba and modeling_jamba to not_doctested because the model is too big to download (in docstring of JambaForCausalLM.forward)

* Add integration test with tiny tandom Jamba model on hub

* fix flash attention cache shapes

* bring back forgotten hidden states

* rename HybridMambaAttentionDynamicCache.seqlen_offset to has_previous_state (and make bool) and bugfix - it should be set to True after a finished forward pass of the entire model

* align integration test after modeling fixes

* bugfix - mamba can use precomputed states only of forward pass is on a single token

* bugfix - mamba can use precomputed states only if they match the batch size

* typo

* remove making _prepare_4d_causal_attention_mask a leaf function

* stop using past_seq_len.get_seq_length(). Use cache positions instead. Adjust test (test_decoder_model_past_with_large_inputs) accordingly

---------

Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
Co-authored-by: Joao Gante <joao@huggingface.co>
This commit is contained in:
tomeras91 2024-04-18 12:04:02 +03:00 committed by GitHub
parent 28a22834bf
commit 3f20877da9
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
35 changed files with 3161 additions and 26 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
README.md
View File

@ -399,6 +399,7 @@ Current number of checkpoints: ![](https://img.shields.io/endpoint?url=https://h
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_de.md
View File

@ -395,6 +395,7 @@ Aktuelle Anzahl der Checkpoints: ![](https://img.shields.io/endpoint?url=https:/
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_es.md
View File

@ -372,6 +372,7 @@ Número actual de puntos de control: ![](https://img.shields.io/endpoint?url=htt
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_fr.md
View File

@ -393,6 +393,7 @@ Nombre actuel de points de contrôle : ![](https://img.shields.io/endpoint?url=h
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (d'OpenAI) a été publié dans l'article [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) par Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (de l'Université de Beihang, UC Berkeley, Rutgers University, SEDD Company) a été publié dans l'article [Informer : Au-delà du Transformer efficace pour la prévision de séries temporel
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (de Salesforce) a été publié dans l'article [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) de Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (d'OpenAI) a été publié dans l'article [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) de Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (de Microsoft Research Asia) a été publié dans l'article [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) de Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (de Microsoft Research Asia) a été publié dans l'article [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) de Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_hd.md
View File

@ -346,6 +346,7 @@ conda install conda-forge::transformers
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce से) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. द्वाराअनुसंधान पत्र [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) के साथ जारी किया गया
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_ja.md
View File

@ -406,6 +406,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI から) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever から公開された研究論文: [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/)
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce から) Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi. から公開された研究論文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI から) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever から公開された研究論文: [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf)
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia から) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou から公開された研究論文: [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318)

1
README_ko.md
View File

@ -321,6 +321,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (OpenAI 에서) Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 의 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 논문과 함께 발표했습니다.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (Salesforce 에서 제공)은 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.의 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500)논문과 함께 발표했습니다.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (OpenAI 에서) Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever 의 [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) 논문과 함께 발표했습니다.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (Microsoft Research Asia 에서) Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 의 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 논문과 함께 발표했습니다.

1
README_pt-br.md
View File

@ -404,6 +404,7 @@ Número atual de pontos de verificação: ![](https://img.shields.io/endpoint?ur
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_ru.md
View File

@ -394,6 +394,7 @@ conda install conda-forge::transformers
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_te.md
View File

@ -396,6 +396,7 @@ Flax, PyTorch లేదా TensorFlow యొక్క ఇన్‌స్టా
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_vi.md
View File

@ -395,6 +395,7 @@ Số lượng điểm kiểm tra hiện tại: ![](https://img.shields.io/endpoi
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (từ OpenAI) được phát hành với bài báo [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (từ Beihang University, UC Berkeley, Rutgers University, SEDD Company) được phát hành với bài báo [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (từ Salesforce) được phát hành với bài báo [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (từ OpenAI) được phát hành với bài báo [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (từ Microsoft Research Asia) được phát hành với bài báo [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (từ Microsoft Research Asia) được phát hành với bài báo [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

1
README_zh-hans.md
View File

@ -345,6 +345,7 @@ conda install conda-forge::transformers
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (来自 OpenAI) 伴随论文 [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) 由 Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever 发布。
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (来自 Salesforce) 伴随论文 [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) 由 Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi 发布。
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (来自 Microsoft Research Asia) 伴随论文 [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) 由 Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou 发布。

1
README_zh-hant.md
View File

@ -357,6 +357,7 @@ conda install conda-forge::transformers
1. **[ImageGPT](https://huggingface.co/docs/transformers/model_doc/imagegpt)** (from OpenAI) released with the paper [Generative Pretraining from Pixels](https://openai.com/blog/image-gpt/) by Mark Chen, Alec Radford, Rewon Child, Jeffrey Wu, Heewoo Jun, David Luan, Ilya Sutskever.
1. **[Informer](https://huggingface.co/docs/transformers/model_doc/informer)** (from Beihang University, UC Berkeley, Rutgers University, SEDD Company) released with the paper [Informer: Beyond Efficient Transformer for Long Sequence Time-Series Forecasting](https://arxiv.org/abs/2012.07436) by Haoyi Zhou, Shanghang Zhang, Jieqi Peng, Shuai Zhang, Jianxin Li, Hui Xiong, and Wancai Zhang.
1. **[InstructBLIP](https://huggingface.co/docs/transformers/model_doc/instructblip)** (from Salesforce) released with the paper [InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuning](https://arxiv.org/abs/2305.06500) by Wenliang Dai, Junnan Li, Dongxu Li, Anthony Meng Huat Tiong, Junqi Zhao, Weisheng Wang, Boyang Li, Pascale Fung, Steven Hoi.
1. **[Jamba](https://huggingface.co/docs/transformers/main/model_doc/jamba)** (from AI21 Labs Ltd.) released with the paper [Jamba: A Hybrid Transformer-Mamba Language Model](https://arxiv.org/abs/2403.19887) by Opher Lieber, Barak Lenz, Hofit Bata, Gal Cohen, Jhonathan Osin, Itay Dalmedigos, Erez Safahi, Shaked Meirom, Yonatan Belinkov, Shai Shalev-Shwartz, Omri Abend, Raz Alon, Tomer Asida, Amir Bergman, Roman Glozman, Michael Gokhman, Avshalom Manevich, Nir Ratner, Noam Rozen, Erez Shwartz, Mor Zusman, Yoav Shoham.
1. **[Jukebox](https://huggingface.co/docs/transformers/model_doc/jukebox)** (from OpenAI) released with the paper [Jukebox: A Generative Model for Music](https://arxiv.org/pdf/2005.00341.pdf) by Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, Ilya Sutskever.
1. **[KOSMOS-2](https://huggingface.co/docs/transformers/model_doc/kosmos-2)** (from Microsoft Research Asia) released with the paper [Kosmos-2: Grounding Multimodal Large Language Models to the World](https://arxiv.org/abs/2306.14824) by Zhiliang Peng, Wenhui Wang, Li Dong, Yaru Hao, Shaohan Huang, Shuming Ma, Furu Wei.
1. **[LayoutLM](https://huggingface.co/docs/transformers/model_doc/layoutlm)** (from Microsoft Research Asia) released with the paper [LayoutLM: Pre-training of Text and Layout for Document Image Understanding](https://arxiv.org/abs/1912.13318) by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, Ming Zhou.

2
docs/source/en/_toctree.yml
View File

@ -382,6 +382,8 @@
title: HerBERT
- local: model_doc/ibert
title: I-BERT
- local: model_doc/jamba
title: Jamba
- local: model_doc/jukebox
title: Jukebox
- local: model_doc/led

1
docs/source/en/index.md
View File

@ -164,6 +164,7 @@ Flax), PyTorch, and/or TensorFlow.
| [ImageGPT](model_doc/imagegpt) | ✅ | ❌ | ❌ |
| [Informer](model_doc/informer) | ✅ | ❌ | ❌ |
| [InstructBLIP](model_doc/instructblip) | ✅ | ❌ | ❌ |
| [Jamba](model_doc/jamba) | ✅ | ❌ | ❌ |
| [Jukebox](model_doc/jukebox) | ✅ | ❌ | ❌ |
| [KOSMOS-2](model_doc/kosmos-2) | ✅ | ❌ | ❌ |
| [LayoutLM](model_doc/layoutlm) | ✅ | ✅ | ❌ |

122
docs/source/en/model_doc/jamba.md Normal file
View File

@ -0,0 +1,122 @@
<!--Copyright 2024 The HuggingFace Team. All rights reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
specific language governing permissions and limitations under the License.
⚠️ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.
-->
# Jamba
## Overview
Jamba is a state-of-the-art, hybrid SSM-Transformer LLM. It is the first production-scale Mamba implementation, which opens up interesting research and application opportunities. While this initial experimentation shows encouraging gains, we expect these to be further enhanced with future optimizations and explorations.
For full details of this model please read the [release blog post](https://www.ai21.com/blog/announcing-jamba).
### Model Details
Jamba is a pretrained, mixture-of-experts (MoE) generative text model, with 12B active parameters and an overall of 52B parameters across all experts. It supports a 256K context length, and can fit up to 140K tokens on a single 80GB GPU.
As depicted in the diagram below, Jamba's architecture features a blocks-and-layers approach that allows Jamba to successfully integrate Transformer and Mamba architectures altogether. Each Jamba block contains either an attention or a Mamba layer, followed by a multi-layer perceptron (MLP), producing an overall ratio of one Transformer layer out of every eight total layers.
<img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/model_doc/jamba_architecture.png"
alt="drawing" width="600"/>
## Usage
### Presequities
Jamba requires you use `transformers` version 4.39.0 or higher:
```bash
pip install transformers>=4.39.0
```
In order to run optimized Mamba implementations, you first need to install `mamba-ssm` and `causal-conv1d`:
```bash
pip install mamba-ssm causal-conv1d>=1.2.0
```
You also have to have the model on a CUDA device.
You can run the model not using the optimized Mamba kernels, but it is **not** recommended as it will result in significantly lower latencies. In order to do that, you'll need to specify `use_mamba_kernels=False` when loading the model.
### Run the model
```python
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1")
tokenizer = AutoTokenizer.from_pretrained("ai21labs/Jamba-v0.1")
input_ids = tokenizer("In the recent Super Bowl LVIII,", return_tensors='pt').to(model.device)["input_ids"]
outputs = model.generate(input_ids, max_new_tokens=216)
print(tokenizer.batch_decode(outputs))
# ["<|startoftext|>In the recent Super Bowl LVIII, the Kansas City Chiefs emerged victorious, defeating the San Francisco 49ers in a thrilling overtime showdown. The game was a nail-biter, with both teams showcasing their skills and determination.\n\nThe Chiefs, led by their star quarterback Patrick Mahomes, displayed their offensive prowess, while the 49ers, led by their strong defense, put up a tough fight. The game went into overtime, with the Chiefs ultimately securing the win with a touchdown.\n\nThe victory marked the Chiefs' second Super Bowl win in four years, solidifying their status as one of the top teams in the NFL. The game was a testament to the skill and talent of both teams, and a thrilling end to the NFL season.\n\nThe Super Bowl is not just about the game itself, but also about the halftime show and the commercials. This year's halftime show featured a star-studded lineup, including Usher, Alicia Keys, and Lil Jon. The show was a spectacle of music and dance, with the performers delivering an energetic and entertaining performance.\n"]
```
<details>
<summary><strong>Loading the model in half precision</strong></summary>
The published checkpoint is saved in BF16. In order to load it into RAM in BF16/FP16, you need to specify `torch_dtype`:
```python
from transformers import AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1", torch_dtype=torch.bfloat16)
# you can also use torch_dtype=torch.float16
```
When using half precision, you can enable the [FlashAttention2](https://github.com/Dao-AILab/flash-attention) implementation of the Attention blocks. In order to use it, you also need the model on a CUDA device. Since in this precision the model is to big to fit on a single 80GB GPU, you'll also need to parallelize it using [accelerate](https://huggingface.co/docs/accelerate/index):
```python
from transformers import AutoModelForCausalLM
import torch
model = AutoModelForCausalLM.from_pretrained("ai21labs/Jamba-v0.1",
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
device_map="auto")
```
</details>
<details><summary><strong>Load the model in 8-bit</strong></summary>
**Using 8-bit precision, it is possible to fit up to 140K sequence lengths on a single 80GB GPU.** You can easily quantize the model to 8-bit using [bitsandbytes](https://huggingface.co/docs/bitsandbytes/index). In order to not degrade model quality, we recommend to exclude the Mamba blocks from the quantization:
```python
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
quantization_config = BitsAndBytesConfig(load_in_8bit=True, llm_int8_skip_modules=["mamba"])
model = AutoModelForCausalLM.from_pretrained(
"ai21labs/Jamba-v0.1", torch_dtype=torch.bfloat16, attn_implementation="flash_attention_2", quantization_config=quantization_config
)
```
</details>
## JambaConfig
[[autodoc]] JambaConfig
## JambaModel
[[autodoc]] JambaModel
- forward
## JambaForCausalLM
[[autodoc]] JambaForCausalLM
- forward
## JambaForSequenceClassification
[[autodoc]] transformers.JambaForSequenceClassification
- forward

2
docs/source/en/perf_infer_gpu_one.md
View File

@ -49,6 +49,7 @@ FlashAttention-2 is currently supported for the following architectures:
* [GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj#transformers.GPTJModel)
* [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
* [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
* [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
* [Llava](https://huggingface.co/docs/transformers/model_doc/llava)
* [Llava-NeXT](https://huggingface.co/docs/transformers/model_doc/llava_next)
@ -186,6 +187,7 @@ For now, Transformers supports SDPA inference and training for the following arc
* [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
* [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
* [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
* [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
* [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
* [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)

2
docs/source/en/tasks/language_modeling.md
View File

@ -37,7 +37,7 @@ You can finetune other architectures for causal language modeling following the
Choose one of the following architectures:
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [Cohere](../model_doc/cohere), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [Fuyu](../model_doc/fuyu), [Gemma](../model_doc/gemma), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Mamba](../model_doc/mamba), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MusicGen Melody](../model_doc/musicgen_melody), [MVP](../model_doc/mvp), [OLMo](../model_doc/olmo), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Qwen2MoE](../model_doc/qwen2_moe), [RecurrentGemma](../model_doc/recurrent_gemma), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [StableLm](../model_doc/stablelm), [Starcoder2](../model_doc/starcoder2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [Whisper](../model_doc/whisper), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeLlama](../model_doc/code_llama), [CodeGen](../model_doc/codegen), [Cohere](../model_doc/cohere), [CPM-Ant](../model_doc/cpmant), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [Falcon](../model_doc/falcon), [Fuyu](../model_doc/fuyu), [Gemma](../model_doc/gemma), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [Jamba](../model_doc/jamba), [LLaMA](../model_doc/llama), [Mamba](../model_doc/mamba), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MPT](../model_doc/mpt), [MusicGen](../model_doc/musicgen), [MusicGen Melody](../model_doc/musicgen_melody), [MVP](../model_doc/mvp), [OLMo](../model_doc/olmo), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Qwen2MoE](../model_doc/qwen2_moe), [RecurrentGemma](../model_doc/recurrent_gemma), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [RWKV](../model_doc/rwkv), [Speech2Text2](../model_doc/speech_to_text_2), [StableLm](../model_doc/stablelm), [Starcoder2](../model_doc/starcoder2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [Whisper](../model_doc/whisper), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)

2
docs/source/en/tasks/sequence_classification.md
View File

@ -33,7 +33,7 @@ The task illustrated in this tutorial is supported by the following model archit
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [Gemma](../model_doc/gemma), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Qwen2MoE](../model_doc/qwen2_moe), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [StableLm](../model_doc/stablelm), [Starcoder2](../model_doc/starcoder2), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [CodeLlama](../model_doc/code_llama), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [Falcon](../model_doc/falcon), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [Gemma](../model_doc/gemma), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [Jamba](../model_doc/jamba), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [Mistral](../model_doc/mistral), [Mixtral](../model_doc/mixtral), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MPT](../model_doc/mpt), [MRA](../model_doc/mra), [MT5](../model_doc/mt5), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenLlama](../model_doc/open-llama), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [Persimmon](../model_doc/persimmon), [Phi](../model_doc/phi), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Qwen2](../model_doc/qwen2), [Qwen2MoE](../model_doc/qwen2_moe), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [StableLm](../model_doc/stablelm), [Starcoder2](../model_doc/starcoder2), [T5](../model_doc/t5), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [UMT5](../model_doc/umt5), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)

20
src/transformers/__init__.py
View File

@ -20,6 +20,7 @@
__version__ = "4.40.0.dev0"
from typing import TYPE_CHECKING
# Check the dependencies satisfy the minimal versions required.
@ -517,6 +518,7 @@ _import_structure = {
"InstructBlipQFormerConfig",
"InstructBlipVisionConfig",
],
"models.jamba": ["JambaConfig"],
"models.jukebox": [
"JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP",
"JukeboxConfig",
@ -1473,6 +1475,7 @@ else:
"AlignVisionModel",
]
)
_import_structure["models.altclip"].extend(
[
"ALTCLIP_PRETRAINED_MODEL_ARCHIVE_LIST",
@ -2480,6 +2483,14 @@ else:
"InstructBlipVisionModel",
]
)
_import_structure["models.jamba"].extend(
[
"JambaForCausalLM",
"JambaForSequenceClassification",
"JambaModel",
"JambaPreTrainedModel",
]
)
_import_structure["models.jukebox"].extend(
[
"JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST",
@ -5439,6 +5450,7 @@ if TYPE_CHECKING:
InstructBlipQFormerConfig,
InstructBlipVisionConfig,
)
from .models.jamba import JambaConfig
from .models.jukebox import (
JUKEBOX_PRETRAINED_CONFIG_ARCHIVE_MAP,
JukeboxConfig,
@ -7213,6 +7225,12 @@ if TYPE_CHECKING:
InstructBlipQFormerModel,
InstructBlipVisionModel,
)
from .models.jamba import (
JambaForCausalLM,
JambaForSequenceClassification,
JambaModel,
JambaPreTrainedModel,
)
from .models.jukebox import (
JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST,
JukeboxModel,
@ -7852,8 +7870,6 @@ if TYPE_CHECKING:
SamModel,
SamPreTrainedModel,
)
# PyTorch model imports
from .models.seamless_m4t import (
SEAMLESS_M4T_PRETRAINED_MODEL_ARCHIVE_LIST,
SeamlessM4TCodeHifiGan,

10
src/transformers/generation/candidate_generator.py
View File

@ -18,6 +18,8 @@ from typing import TYPE_CHECKING, Any, Dict, Optional, Tuple
import torch
from ..cache_utils import DynamicCache
if TYPE_CHECKING:
from ..modeling_utils import PreTrainedModel
@ -371,7 +373,13 @@ def _crop_past_key_values(model, past_key_values, maximum_length):
else:
for idx in range(len(past_key_values)):
past_key_values[idx] = past_key_values[idx][:, :, :maximum_length, :]
else:
elif isinstance(past_key_values, DynamicCache):
for idx in range(len(past_key_values.key_cache)):
if past_key_values.value_cache[idx].shape[-1] != 0:
past_key_values.key_cache[idx] = past_key_values.key_cache[idx][:, :, :maximum_length, :]
past_key_values.value_cache[idx] = past_key_values.value_cache[idx][:, :, :maximum_length, :]
elif past_key_values is not None:
for idx in range(len(past_key_values)):
new_past.append(
(

58
src/transformers/generation/utils.py
View File

@ -598,7 +598,11 @@ class GenerationMixin:
def _expand_dict_for_generation(dict_to_expand):
for key in dict_to_expand:
if dict_to_expand[key] is not None and isinstance(dict_to_expand[key], torch.Tensor):
if (
key != "cache_position"
and dict_to_expand[key] is not None
and isinstance(dict_to_expand[key], torch.Tensor)
):
dict_to_expand[key] = dict_to_expand[key].repeat_interleave(expand_size, dim=0)
return dict_to_expand
@ -2094,7 +2098,8 @@ class GenerationMixin:
# Replicates the new past_key_values to match the `top_k` candidates
new_key_values = []
for layer in model_kwargs["past_key_values"]:
past = model_kwargs["past_key_values"]
for layer in past:
items = []
# item is either the key or the value matrix
for item in layer:
@ -2103,7 +2108,13 @@ class GenerationMixin:
else:
items.append(item.repeat_interleave(top_k, dim=0))
new_key_values.append(tuple(items))
model_kwargs["past_key_values"] = tuple(new_key_values)
if not isinstance(past, DynamicCache):
past = tuple(new_key_values)
else:
for layer_idx in range(len(new_key_values)):
past.key_cache[layer_idx] = new_key_values[layer_idx][0]
past.value_cache[layer_idx] = new_key_values[layer_idx][1]
model_kwargs["past_key_values"] = past
if sequential:
all_outputs = []
@ -2178,16 +2189,22 @@ class GenerationMixin:
else:
next_past_key_values = self._extract_past_from_model_output(outputs, standardize_cache_format=True)
new_key_values = ()
new_key_values = []
for layer in next_past_key_values:
items = ()
items = []
# item is either the key or the value matrix
for item in layer:
item = torch.stack(torch.split(item, top_k, dim=0)) # [B, K, num_head, seq_len, esz]
item = item[range(batch_size), selected_idx, ...] # [B, num_head, seq_len, esz]
items += (item,)
new_key_values += (items,)
next_past_key_values = new_key_values
items += [item]
new_key_values += [items]
if not isinstance(next_past_key_values, DynamicCache):
next_past_key_values = tuple(new_key_values)
else:
for layer_idx in range(len(new_key_values)):
next_past_key_values.key_cache[layer_idx] = new_key_values[layer_idx][0]
next_past_key_values.value_cache[layer_idx] = new_key_values[layer_idx][1]
logit_for_next_step = torch.stack(torch.split(logits, top_k))[range(batch_size), selected_idx, :]
@ -3127,6 +3144,7 @@ class GenerationMixin:
"transo_xl",
"xlnet",
"cpm",
"jamba",
]
):
raise RuntimeError(
@ -4645,21 +4663,22 @@ class GenerationMixin:
# we use this forward pass to also pick the subsequent logits in the original model.
# 2.1. Prepare the model inputs
candidate_kwargs = copy.copy(model_kwargs)
candidate_kwargs = _prepare_attention_mask(
candidate_kwargs, candidate_input_ids.shape[1], self.config.is_encoder_decoder
model_kwargs = _prepare_attention_mask(
model_kwargs, candidate_input_ids.shape[1], self.config.is_encoder_decoder
)
candidate_kwargs = _prepare_token_type_ids(candidate_kwargs, candidate_input_ids.shape[1])
if "cache_position" in candidate_kwargs:
candidate_kwargs["cache_position"] = torch.cat(
model_kwargs = _prepare_token_type_ids(model_kwargs, candidate_input_ids.shape[1])
if "cache_position" in model_kwargs:
model_kwargs["cache_position"] = torch.cat(
(
candidate_kwargs["cache_position"],
model_kwargs["cache_position"],
torch.arange(cur_len, cur_len + candidate_length, device=input_ids.device, dtype=torch.long),
),
dim=0,
)
model_inputs = self.prepare_inputs_for_generation(candidate_input_ids, **candidate_kwargs)
model_inputs = self.prepare_inputs_for_generation(candidate_input_ids, **model_kwargs)
if "num_logits_to_keep" in model_inputs:
model_inputs["num_logits_to_keep"] = candidate_length + 1
# 2.2. Run a forward pass on the candidate sequence
outputs = self(
@ -4985,7 +5004,7 @@ def _split_model_inputs(
# ModelOutput object.
# bool should not be split but replicated for each split
bool_keys = [k for k in keys if isinstance(model_input[k], bool) or k == "cache_position"]
keys_to_ignore = ["cache_position", "encoder_outputs"]
keys_to_ignore = ["cache_position", "encoder_outputs", "num_logits_to_keep"]
non_bool_keys = [k for k in keys if not isinstance(model_input[k], bool) and k not in keys_to_ignore]
# we split the tensors and tuples of tensors
@ -5001,6 +5020,11 @@ def _split_model_inputs(
data_split_list = [
{**data_split, "encoder_outputs": encoder_outputs_split[i]} for i, data_split in enumerate(data_split_list)
]
# num_logits_to_keep should be replicated for each split, similar to bool values
if "num_logits_to_keep" in model_input:
data_split_list = [
{**data_split, "num_logits_to_keep": model_input["num_logits_to_keep"]} for data_split in data_split_list
]
# Convert each dictionary in the list to an object of the inferred class
split_model_inputs: List[Union[ModelOutput, Dict]] = [

1
src/transformers/models/__init__.py
View File

@ -115,6 +115,7 @@ from . import (
imagegpt,
informer,
instructblip,
jamba,
jukebox,
kosmos2,
layoutlm,

2
src/transformers/models/auto/configuration_auto.py
View File

@ -129,6 +129,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
("imagegpt", "ImageGPTConfig"),
("informer", "InformerConfig"),
("instructblip", "InstructBlipConfig"),
("jamba", "JambaConfig"),
("jukebox", "JukeboxConfig"),
("kosmos-2", "Kosmos2Config"),
("layoutlm", "LayoutLMConfig"),
@ -397,6 +398,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
("imagegpt", "ImageGPT"),
("informer", "Informer"),
("instructblip", "InstructBLIP"),
("jamba", "Jamba"),
("jukebox", "Jukebox"),
("kosmos-2", "KOSMOS-2"),
("layoutlm", "LayoutLM"),

3
src/transformers/models/auto/modeling_auto.py
View File

@ -123,6 +123,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
("idefics2", "Idefics2Model"),
("imagegpt", "ImageGPTModel"),
("informer", "InformerModel"),
("jamba", "JambaModel"),
("jukebox", "JukeboxModel"),
("kosmos-2", "Kosmos2Model"),
("layoutlm", "LayoutLMModel"),
@ -451,6 +452,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
("gpt_neox", "GPTNeoXForCausalLM"),
("gpt_neox_japanese", "GPTNeoXJapaneseForCausalLM"),
("gptj", "GPTJForCausalLM"),
("jamba", "JambaForCausalLM"),
("llama", "LlamaForCausalLM"),
("mamba", "MambaForCausalLM"),
("marian", "MarianForCausalLM"),
@ -851,6 +853,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
("gpt_neox", "GPTNeoXForSequenceClassification"),
("gptj", "GPTJForSequenceClassification"),
("ibert", "IBertForSequenceClassification"),
("jamba", "JambaForSequenceClassification"),
("layoutlm", "LayoutLMForSequenceClassification"),
("layoutlmv2", "LayoutLMv2ForSequenceClassification"),
("layoutlmv3", "LayoutLMv3ForSequenceClassification"),

7
src/transformers/models/auto/tokenization_auto.py
View File

@ -203,6 +203,13 @@ else:
("idefics", (None, "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("idefics2", ("LlamaTokenizer", "LlamaTokenizerFast" if is_tokenizers_available() else None)),
("instructblip", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
(
"jamba",
(
"LlamaTokenizer" if is_sentencepiece_available() else None,
"LlamaTokenizerFast" if is_tokenizers_available() else None,
),
),
("jukebox", ("JukeboxTokenizer", None)),
(
"kosmos-2",

58
src/transformers/models/jamba/__init__.py Normal file
View File

@ -0,0 +1,58 @@
# Copyright 2024 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_jamba": ["JambaConfig"],
}
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
_import_structure["modeling_jamba"] = [
"JambaForCausalLM",
"JambaForSequenceClassification",
"JambaModel",
"JambaPreTrainedModel",
]
if TYPE_CHECKING:
from .configuration_jamba import JambaConfig
try:
if not is_torch_available():
raise OptionalDependencyNotAvailable()
except OptionalDependencyNotAvailable:
pass
else:
from .modeling_jamba import (
JambaForCausalLM,
JambaForSequenceClassification,
JambaModel,
JambaPreTrainedModel,
)
else:
import sys
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)

223
src/transformers/models/jamba/configuration_jamba.py Normal file
View File

@ -0,0 +1,223 @@
# coding=utf-8
# Copyright 2024 AI21 Labs Ltd. 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.
""" Jamba model configuration"""
import math
from ...configuration_utils import PretrainedConfig
from ...utils import logging
logger = logging.get_logger(__name__)
class JambaConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`JambaModel`]. It is used to instantiate a
Jamba 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 Jamba-v0.1 model.
[ai21labs/Jamba-v0.1](https://huggingface.co/ai21labs/Jamba-v0.1)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
vocab_size (`int`, *optional*, defaults to 65536):
Vocabulary size of the Jamba model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`JambaModel`]
tie_word_embeddings (`bool`, *optional*, defaults to `False`):
Whether the model's input and output word embeddings should be tied. Note that this is only relevant if the
model has a output word embedding layer.
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 14336):
Dimension of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer encoder.
num_key_value_heads (`int`, *optional*, defaults to 8):
This is the number of key_value heads that should be used to implement Grouped Query Attention. If
`num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
`num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When
converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
by meanpooling all the original heads within that group. For more details checkout [this
paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `8`.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
rms_norm_eps (`float`, *optional*, defaults to 1e-06):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
num_logits_to_keep (`int` or `None`, *optional*, defaults to 1):
Number of prompt logits to calculate during generation. If `None`, all logits will be calculated. If an
integer value, only last `num_logits_to_keep` logits will be calculated. Default is 1 because only the
logits of the last prompt token are needed for generation. For long sequences, the logits for the entire
sequence may use a lot of memory so, setting `num_logits_to_keep=1` will reduce memory footprint
significantly.
output_router_logits (`bool`, *optional*, defaults to `False`):
Whether or not the router logits should be returned by the model. Enabling this will also
allow the model to output the auxiliary loss. See [here]() for more details
router_aux_loss_coef (`float`, *optional*, defaults to 0.001):
The aux loss factor for the total loss.
pad_token_id (`int`, *optional*, defaults to 0):
The id of the padding token.
bos_token_id (`int`, *optional*, defaults to 1):
The id of the "beginning-of-sequence" token.
eos_token_id (`int`, *optional*, defaults to 2):
The id of the "end-of-sequence" token.
sliding_window (`int`, *optional*):
Sliding window attention window size. If not specified, will default to `None`.
max_position_embeddings (`int`, *optional*, defaults to 262144):
This value doesn't have any real effect. The maximum sequence length that this model is intended to be
used with. It can be used with longer sequences, but performance may degrade.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
num_experts_per_tok (`int`, *optional*, defaults to 2):
The number of experts to root per-token, can be also interpreted as the `top-p` routing
parameter
num_experts (`int`, *optional*, defaults to 16):
Number of experts per Sparse MLP layer.
expert_layer_period (`int`, *optional*, defaults to 2):
Once in this many layers, we will have an expert layer
expert_layer_offset (`int`, *optional*, defaults to 1):
The first layer index that contains an expert mlp layer
attn_layer_period (`int`, *optional*, defaults to 8):
Once in this many layers, we will have a vanilla attention layer
attn_layer_offset (`int`, *optional*, defaults to 4):
The first layer index that contains a vanilla attention mlp layer
use_mamba_kernels (`bool`, *optional*, defaults to `True`):
Flag indicating whether or not to use the fast mamba kernels. These are available only if `mamba-ssm` and
`causal-conv1d` are installed, and the mamba modules are running on a CUDA device. Raises ValueError if
`True` and kernels are not available
mamba_d_state (`int`, *optional*, defaults to 16):
The dimension the mamba state space latents
mamba_d_conv (`int`, *optional*, defaults to 4):
The size of the mamba convolution kernel
mamba_expand (`int`, *optional*, defaults to 2):
Expanding factor (relative to hidden_size) used to determine the mamba intermediate size
mamba_dt_rank (`Union[int,str]`, *optional*, defaults to `"auto"`):
Rank of the the mamba discretization projection matrix. `"auto"` means that it will default to `math.ceil(self.hidden_size / 16)`
mamba_conv_bias (`bool`, *optional*, defaults to `True`):
Flag indicating whether or not to use bias in the convolution layer of the mamba mixer block.
mamba_proj_bias (`bool`, *optional*, defaults to `False`):
Flag indicating whether or not to use bias in the input and output projections (["in_proj", "out_proj"]) of the mamba mixer block
"""
model_type = "jamba"
keys_to_ignore_at_inference = ["past_key_values"]
def __init__(
self,
vocab_size=65536,
tie_word_embeddings=False,
hidden_size=4096,
intermediate_size=14336,
num_hidden_layers=32,
num_attention_heads=32,
num_key_value_heads=8,
hidden_act="silu",
initializer_range=0.02,
rms_norm_eps=1e-6,
use_cache=True,
num_logits_to_keep=1,
output_router_logits=False,
router_aux_loss_coef=0.001,
pad_token_id=0,
bos_token_id=1,
eos_token_id=2,
sliding_window=None,
max_position_embeddings=262144,
attention_dropout=0.0,
num_experts_per_tok=2,
num_experts=16,
expert_layer_period=2,
expert_layer_offset=1,
attn_layer_period=8,
attn_layer_offset=4,
use_mamba_kernels=True,
mamba_d_state=16,
mamba_d_conv=4,
mamba_expand=2,
mamba_dt_rank="auto",
mamba_conv_bias=True,
mamba_proj_bias=False,
**kwargs,
):
self.vocab_size = vocab_size
self.tie_word_embeddings = tie_word_embeddings
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.sliding_window = sliding_window
self.max_position_embeddings = max_position_embeddings
self.attention_dropout = attention_dropout
# for backward compatibility
if num_key_value_heads is None:
num_key_value_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.num_logits_to_keep = num_logits_to_keep
self.output_router_logits = output_router_logits
self.router_aux_loss_coef = router_aux_loss_coef
self.num_experts_per_tok = num_experts_per_tok
self.num_experts = num_experts
self.expert_layer_period = expert_layer_period
self.expert_layer_offset = expert_layer_offset
self.attn_layer_period = attn_layer_period
self.attn_layer_offset = attn_layer_offset
self.use_mamba_kernels = use_mamba_kernels
self.mamba_d_state = mamba_d_state
self.mamba_d_conv = mamba_d_conv
self.mamba_expand = mamba_expand
self.mamba_dt_rank = math.ceil(self.hidden_size / 16) if mamba_dt_rank == "auto" else mamba_dt_rank
self.mamba_conv_bias = mamba_conv_bias
self.mamba_proj_bias = mamba_proj_bias
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
@property
def layers_block_type(self):
return [
"attention" if i % self.attn_layer_period == self.attn_layer_offset else "mamba"
for i in range(self.num_hidden_layers)
]
@property
def layers_num_experts(self):
return [
self.num_experts if i % self.expert_layer_period == self.expert_layer_offset else 1
for i in range(self.num_hidden_layers)
]

1882
src/transformers/models/jamba/modeling_jamba.py Executable file
View File

File diff suppressed because it is too large Load Diff

28
src/transformers/utils/dummy_pt_objects.py
View File

@ -4526,6 +4526,34 @@ class InstructBlipVisionModel(metaclass=DummyObject):
requires_backends(self, ["torch"])
class JambaForCausalLM(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class JambaForSequenceClassification(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class JambaModel(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
class JambaPreTrainedModel(metaclass=DummyObject):
_backends = ["torch"]
def __init__(self, *args, **kwargs):
requires_backends(self, ["torch"])
JUKEBOX_PRETRAINED_MODEL_ARCHIVE_LIST = None

10
tests/generation/test_utils.py
View File

@ -1050,7 +1050,7 @@ class GenerationTesterMixin:
for model_class in self.all_generative_model_classes:
if any(model_name in model_class.__name__.lower() for model_name in ["fsmt", "reformer", "speech2text"]):
self.skipTest("Won't fix: old model with different cache format")
if any(model_name in model_class.__name__.lower() for model_name in ["gptbigcode"]):
if any(model_name in model_class.__name__.lower() for model_name in ["gptbigcode", "jamba"]):
self.skipTest("TODO: fix me")
config, input_ids, attention_mask, max_length = self._get_input_ids_and_config(batch_size=1)
@ -1098,6 +1098,7 @@ class GenerationTesterMixin:
"transo_xl",
"xlnet",
"cpm",
"jamba",
]
):
self.skipTest("May fix in the future: need model-specific fixes")
@ -1735,11 +1736,12 @@ class GenerationTesterMixin:
use_cache=use_cache,
)
# Past Key Value States -- two notes here:
# Past Key Value States -- a few notes here:
# 1. Its inner sequence length is with respect to the inputs of the latest forward pass, hence the "-1"
# 2. Some old models still return `output.past_key_values` even without `use_cache=True`
# 3. TODO (joao): A few models have different formats, skipping those until the cache refactor is complete
models_without_standard_cache = ("bloom", "ctrl", "fsmt", "gptbigcode", "mega", "reformer")
# 3. TODO (joao): A few models have different formats/types, skipping those until the cache refactor is
# complete
models_without_standard_cache = ("bloom", "ctrl", "fsmt", "gptbigcode", "mega", "reformer", "jamba")
has_standard_cache = not any(
model_name in config.__class__.__name__.lower() for model_name in models_without_standard_cache
)

0
tests/models/jamba/__init__.py Normal file
View File

730
tests/models/jamba/test_modeling_jamba.py Normal file
View File

@ -0,0 +1,730 @@
# coding=utf-8
# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
""" Testing suite for the PyTorch Jamba model. """
import math
import tempfile
import unittest
import pytest
from parameterized import parameterized
from transformers import AutoTokenizer, JambaConfig, is_torch_available
from transformers.testing_utils import (
require_bitsandbytes,
require_flash_attn,
require_torch,
require_torch_gpu,
slow,
torch_device,
)
from ...generation.test_utils import GenerationTesterMixin
from ...test_configuration_common import ConfigTester
from ...test_modeling_common import ModelTesterMixin, _config_zero_init, ids_tensor, random_attention_mask
from ...test_pipeline_mixin import PipelineTesterMixin
if is_torch_available():
import torch
from transformers import (
JambaForCausalLM,
JambaForSequenceClassification,
JambaModel,
)
from transformers.models.jamba.modeling_jamba import (
HybridMambaAttentionDynamicCache,
)
class JambaModelTester:
def __init__(
self,
parent,
batch_size=13,
seq_length=7,
is_training=True,
use_input_mask=True,
use_labels=True,
vocab_size=99,
hidden_size=32,
num_hidden_layers=5,
attn_layer_offset=1,
attn_layer_period=8,
num_attention_heads=4,
num_key_value_heads=2,
intermediate_size=37,
hidden_act="gelu",
hidden_dropout_prob=0.1,
attention_probs_dropout_prob=0.1,
max_position_embeddings=512,
type_vocab_size=16,
type_sequence_label_size=2,
initializer_range=0.02,
num_labels=3,
num_choices=4,
scope=None,
):
self.parent = parent
self.batch_size = batch_size
self.seq_length = seq_length
self.is_training = is_training
self.use_input_mask = use_input_mask
self.use_labels = use_labels
self.vocab_size = vocab_size
self.hidden_size = hidden_size
self.num_hidden_layers = num_hidden_layers
self.attn_layer_offset = attn_layer_offset
self.attn_layer_period = attn_layer_period
self.num_attention_heads = num_attention_heads
self.num_key_value_heads = num_key_value_heads
self.intermediate_size = intermediate_size
self.hidden_act = hidden_act
self.hidden_dropout_prob = hidden_dropout_prob
self.attention_probs_dropout_prob = attention_probs_dropout_prob
self.max_position_embeddings = max_position_embeddings
self.type_vocab_size = type_vocab_size
self.type_sequence_label_size = type_sequence_label_size
self.initializer_range = initializer_range
self.num_labels = num_labels
self.num_choices = num_choices
self.scope = scope
def prepare_config_and_inputs(self):
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
input_mask = None
if self.use_input_mask:
input_mask = random_attention_mask([self.batch_size, self.seq_length])
sequence_labels = None
token_labels = None
choice_labels = None
if self.use_labels:
sequence_labels = ids_tensor([self.batch_size], self.type_sequence_label_size)
token_labels = ids_tensor([self.batch_size, self.seq_length], self.num_labels)
choice_labels = ids_tensor([self.batch_size], self.num_choices)
config = self.get_config()
return config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
def get_config(self):
return JambaConfig(
vocab_size=self.vocab_size,
hidden_size=self.hidden_size,
num_hidden_layers=self.num_hidden_layers,
attn_layer_offset=self.attn_layer_offset,
attn_layer_period=self.attn_layer_period,
num_attention_heads=self.num_attention_heads,
num_key_value_heads=self.num_key_value_heads,
intermediate_size=self.intermediate_size,
hidden_act=self.hidden_act,
hidden_dropout_prob=self.hidden_dropout_prob,
attention_probs_dropout_prob=self.attention_probs_dropout_prob,
max_position_embeddings=self.max_position_embeddings,
type_vocab_size=self.type_vocab_size,
is_decoder=True,
initializer_range=self.initializer_range,
use_mamba_kernels=False,
num_experts=2,
)
def prepare_config_and_inputs_for_decoder(self):
(
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = self.prepare_config_and_inputs()
config.is_decoder = True
return (
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
)
def create_and_check_model(self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels):
model = JambaModel(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids)
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
def create_and_check_for_causal_lm(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
):
model = JambaForCausalLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=token_labels)
result = model(input_ids, attention_mask=input_mask)
result = model(input_ids, labels=token_labels)
result = model(input_ids)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_decoder_model_past_large_inputs(
self,
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
):
config.is_decoder = True
config.add_cross_attention = True
model = JambaForCausalLM(config=config)
model.to(torch_device)
model.eval()
# first forward pass
# Attention: Jamba needs the cache to be initialized to return a cache!
past_key_values = HybridMambaAttentionDynamicCache(
config, input_ids.shape[0], model.dtype, device=model.device
)
outputs = model(
input_ids,
attention_mask=input_mask,
past_key_values=past_key_values,
use_cache=True,
)
past_key_values = outputs.past_key_values
# create hypothetical multiple next token and extent to next_input_ids
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
# append to next input_ids and
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
output_from_no_past = model(
next_input_ids,
attention_mask=next_attention_mask,
output_hidden_states=True,
)["hidden_states"][0]
output_from_past = model(
next_tokens,
attention_mask=next_attention_mask,
past_key_values=past_key_values,
output_hidden_states=True,
cache_position=torch.arange(
input_ids.shape[1], input_ids.shape[1] + next_tokens.shape[1], device=model.device
),
)["hidden_states"][0]
# select random slice
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
output_from_no_past_slice = output_from_no_past[:, -3:, random_slice_idx].detach()
output_from_past_slice = output_from_past[:, :, random_slice_idx].detach()
self.parent.assertTrue(output_from_past_slice.shape[1] == next_tokens.shape[1])
# test that outputs are equal for slice
self.parent.assertTrue(torch.allclose(output_from_past_slice, output_from_no_past_slice, atol=1e-3))
def create_and_check_for_sequence_classification(
self, config, input_ids, input_mask, sequence_labels, token_labels, choice_labels
):
config.num_labels = self.num_labels
model = JambaForSequenceClassification(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=sequence_labels)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
def prepare_config_and_inputs_for_common(self):
config_and_inputs = self.prepare_config_and_inputs()
(
config,
input_ids,
input_mask,
sequence_labels,
token_labels,
choice_labels,
) = config_and_inputs
inputs_dict = {"input_ids": input_ids, "attention_mask": input_mask}
return config, inputs_dict
@require_torch
class JambaModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
all_model_classes = (
(
JambaModel,
JambaForCausalLM,
JambaForSequenceClassification,
)
if is_torch_available()
else ()
)
all_generative_model_classes = (JambaForCausalLM,) if is_torch_available() else ()
pipeline_model_mapping = (
{
"feature-extraction": JambaModel,
"text-classification": JambaForSequenceClassification,
"text-generation": JambaForCausalLM,
"zero-shot": JambaForSequenceClassification,
}
if is_torch_available()
else {}
)
test_headmasking = False
test_pruning = False
def setUp(self):
self.model_tester = JambaModelTester(self)
self.config_tester = ConfigTester(self, config_class=JambaConfig, hidden_size=37)
def test_config(self):
self.config_tester.run_common_tests()
def test_model(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_model(*config_and_inputs)
def test_for_casual_lm(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_causal_lm(*config_and_inputs)
def test_for_sequence_classification(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs()
self.model_tester.create_and_check_for_sequence_classification(*config_and_inputs)
def test_decoder_model_past_with_large_inputs(self):
config_and_inputs = self.model_tester.prepare_config_and_inputs_for_decoder()
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
def test_load_balancing_loss(self):
r"""
Let's make sure we can actually compute the loss and do a backward on it.
"""
config, input_dict = self.model_tester.prepare_config_and_inputs_for_common()
config.num_labels = 3
config.num_experts = 16
config.output_router_logits = True
input_ids = input_dict["input_ids"]
attention_mask = input_ids.ne(config.pad_token_id).to(torch_device)
model = JambaForCausalLM(config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=attention_mask)
bs, seqlen = input_ids.shape
self.assertEqual(result.router_logits[0].shape, (bs * seqlen, config.num_experts))
torch.testing.assert_close(result.aux_loss.cpu(), torch.tensor(2, dtype=torch.float32), rtol=1e-2, atol=1e-2)
# First, we make sure that adding padding tokens doesn't change the loss
# loss(input_ids, attention_mask=None) == loss(input_ids + padding, attention_mask=attention_mask_with_padding)
pad_length = 1000
# Add padding tokens to input_ids
padding_block = config.pad_token_id * torch.ones(input_ids.shape[0], pad_length, dtype=torch.int32).to(
torch_device
)
padded_input_ids = torch.cat((padding_block, input_ids), dim=1) # this is to simulate padding to the left
padded_attention_mask = padded_input_ids.ne(config.pad_token_id).to(torch_device)
padded_result = model(padded_input_ids, attention_mask=padded_attention_mask)
torch.testing.assert_close(result.aux_loss.cpu(), padded_result.aux_loss.cpu(), rtol=1e-4, atol=1e-4)
# We make sure that the loss of including padding tokens != the loss without padding tokens
# if attention_mask=None --> we don't exclude padding tokens
include_padding_result = model(padded_input_ids, attention_mask=None)
# This is to mimic torch.testing.assert_not_close
self.assertNotAlmostEqual(include_padding_result.aux_loss.item(), result.aux_loss.item())
def test_initialization(self):
r"""
Overriding the test_initialization test as the A_log and D params of the Mamba block are initialized differently
"""
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
configs_no_init = _config_zero_init(config)
for model_class in self.all_model_classes:
model = model_class(config=configs_no_init)
for name, param in model.named_parameters():
if param.requires_grad:
if "A_log" in name:
A = torch.arange(1, config.mamba_d_state + 1, dtype=torch.float32)[None, :]
self.assertTrue(torch.allclose(param.data, torch.log(A), atol=1e-5, rtol=1e-5))
elif "D" in name:
# check if it's a ones like
self.assertTrue(torch.allclose(param.data, torch.ones_like(param.data), atol=1e-5, rtol=1e-5))
else:
self.assertIn(
((param.data.mean() * 1e9).round() / 1e9).item(),
[0.0, 1.0],
msg=f"Parameter {name} of model {model_class} seems not properly initialized",
)
def test_mismatched_shapes_have_properly_initialized_weights(self):
r"""
Overriding the test_mismatched_shapes_have_properly_initialized_weights test because A_log and D params of the
Mamba block are initialized differently and we tested that in test_initialization
"""
self.skipTest("Cumbersome and redundant for Jamba")
def test_attention_outputs(self):
r"""
Overriding the test_attention_outputs test as the Jamba model outputs attention only for its attention layers
"""
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)
encoder_seq_length = getattr(self.model_tester, "encoder_seq_length", seq_len)
encoder_key_length = getattr(self.model_tester, "key_length", encoder_seq_length)
expected_num_attentions = math.ceil(
(self.model_tester.num_hidden_layers - self.model_tester.attn_layer_offset)
/ self.model_tester.attn_layer_period
)
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)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), expected_num_attentions)
# check that output_attentions also work using config
del inputs_dict["output_attentions"]
config.output_attentions = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
attentions = outputs.attentions
self.assertEqual(len(attentions), expected_num_attentions)
self.assertListEqual(
list(attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
out_len = len(outputs)
# Check attention is always last and order is fine
inputs_dict["output_attentions"] = True
inputs_dict["output_hidden_states"] = True
model = model_class(config)
model.to(torch_device)
model.eval()
with torch.no_grad():
outputs = model(**self._prepare_for_class(inputs_dict, model_class))
added_hidden_states = 1
self.assertEqual(out_len + added_hidden_states, len(outputs))
self_attentions = outputs.attentions
self.assertEqual(len(self_attentions), expected_num_attentions)
self.assertListEqual(
list(self_attentions[0].shape[-3:]),
[self.model_tester.num_attention_heads, encoder_seq_length, encoder_key_length],
)
def test_left_padding_compatibility(self):
r"""
Overriding the test_left_padding_compatibility test as the mamba layers accentuate the numerical differences
effect of the left padding discussed in the issue in the note. Using a more permissive tolerance value.
"""
import inspect
# NOTE: left-padding results in small numerical differences. This is expected.
# See https://github.com/huggingface/transformers/issues/25420#issuecomment-1775317535
# First, filter out models that don't support left padding - generative and decoder-only.
# Jamba is a decoder-only architecture
decoder_only_classes = self.all_generative_model_classes
# Then, test left-padding
def _prepare_model_kwargs(input_ids, attention_mask, signature):
model_kwargs = {"input_ids": input_ids, "attention_mask": attention_mask}
if "position_ids" in signature:
position_ids = torch.cumsum(attention_mask, dim=-1) - 1
position_ids.masked_fill_(attention_mask == 0, 1)
model_kwargs["position_ids"] = position_ids
if "cache_position" in signature:
cache_position = torch.arange(input_ids.shape[-1], device=torch_device)
model_kwargs["cache_position"] = cache_position
return model_kwargs
for model_class in decoder_only_classes:
config, input_ids, attention_mask, _ = self._get_input_ids_and_config()
model = model_class(config).to(torch_device).eval()
signature = inspect.signature(model.forward).parameters.keys()
# Without padding
model_kwargs = _prepare_model_kwargs(input_ids, attention_mask, signature)
next_logits_wo_padding = model(**model_kwargs).logits[:, -1, :]
# With left-padding (length 32)
pad_size = (input_ids.shape[0], 32)
padding = torch.ones(pad_size, dtype=input_ids.dtype, device=torch_device) * config.pad_token_id
padded_input_ids = torch.cat((padding, input_ids), dim=1)
padded_attention_mask = torch.cat((torch.zeros_like(padding), attention_mask), dim=1)
model_kwargs = _prepare_model_kwargs(padded_input_ids, padded_attention_mask, signature)
next_logits_with_padding = model(**model_kwargs).logits[:, -1, :]
# They should result in very similar logits
self.assertTrue(torch.allclose(next_logits_wo_padding, next_logits_with_padding, atol=3e-3))
@require_flash_attn
@require_torch_gpu
@require_bitsandbytes
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_fp32_ln(self):
r"""
Overriding the test_flash_attn_2_fp32_ln test as the Jamba model, like Mixtral, doesn't support
right padding + use cache with FA2
"""
for model_class in self.all_generative_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
dummy_input = inputs_dict[model.main_input_name]
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
# NOTE: Jamba does not support right padding + use_cache with FA2.
dummy_attention_mask[:, -1] = 1
model = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
low_cpu_mem_usage=True,
load_in_4bit=True,
)
for _, param in model.named_parameters():
# upcast only layer norms
if (param.dtype == torch.float16) or (param.dtype == torch.bfloat16):
param.data = param.data.to(torch.float32)
_ = model(dummy_input)
# with attention mask
_ = model(dummy_input, attention_mask=dummy_attention_mask)
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_padding_right(self):
r"""
Overriding the test_flash_attn_2_generate_padding_right test as the Jamba model, like Mixtral, doesn't support
right padding + use cache with FA2
"""
import torch
for model_class in self.all_generative_model_classes:
config, _ = self.model_tester.prepare_config_and_inputs_for_common()
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
model = model_class.from_pretrained(tmpdirname, torch_dtype=torch.float16, low_cpu_mem_usage=True).to(
torch_device
)
dummy_input = torch.LongTensor([[0, 2, 3, 4], [0, 2, 3, 4]]).to(torch_device)
dummy_attention_mask = torch.LongTensor([[1, 1, 1, 1], [1, 1, 1, 0]]).to(torch_device)
model.generate(dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False)
model = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
low_cpu_mem_usage=True,
).to(torch_device)
with self.assertRaises(ValueError):
_ = model.generate(
dummy_input, attention_mask=dummy_attention_mask, max_new_tokens=1, do_sample=False
)
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_generate_use_cache(self):
r"""
Overriding the test_flash_attn_2_generate_use_cache test as the Jamba model, like Mixtral, doesn't support
right padding + use cache with FA2
"""
import torch
max_new_tokens = 30
for model_class in self.all_generative_model_classes:
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
dummy_input = inputs_dict[model_class.main_input_name]
if dummy_input.dtype in [torch.float32, torch.bfloat16]:
dummy_input = dummy_input.to(torch.float16)
# make sure that all models have enough positions for generation
if hasattr(config, "max_position_embeddings"):
config.max_position_embeddings = max_new_tokens + dummy_input.shape[1] + 1
model = model_class(config)
with tempfile.TemporaryDirectory() as tmpdirname:
model.save_pretrained(tmpdirname)
dummy_attention_mask = inputs_dict.get("attention_mask", torch.ones_like(dummy_input))
# NOTE: Jamba does not support right padding + use_cache with FA2.
dummy_attention_mask[:, -1] = 1
model = model_class.from_pretrained(
tmpdirname,
torch_dtype=torch.float16,
attn_implementation="flash_attention_2",
low_cpu_mem_usage=True,
).to(torch_device)
# Just test that a large cache works as expected
_ = model.generate(
dummy_input,
attention_mask=dummy_attention_mask,
max_new_tokens=max_new_tokens,
do_sample=False,
use_cache=True,
)
@require_flash_attn
@require_torch_gpu
@pytest.mark.flash_attn_test
@slow
def test_flash_attn_2_inference_equivalence_right_padding(self):
r"""
Overriding the test_flash_attn_2_inference_padding_right test as the Jamba model, like Mixtral, doesn't support
right padding + use cache with FA2
"""
self.skipTest("Jamba flash attention does not support right padding")
@unittest.skip("Jamba has its own special cache type")
@parameterized.expand([(1, False), (1, True), (4, False)])
def test_new_cache_format(self, num_beams, do_sample):
pass
@require_torch
class JambaModelIntegrationTest(unittest.TestCase):
model = None
tokenizer = None
@classmethod
def setUpClass(cls):
model_id = "ai21labs/Jamba-tiny-random"
cls.model = JambaForCausalLM.from_pretrained(model_id, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True)
cls.tokenizer = AutoTokenizer.from_pretrained(model_id)
@slow
def test_simple_generate(self):
self.model.to(torch_device)
input_ids = self.tokenizer("Hey how are you doing on this lovely evening?", return_tensors="pt")[
"input_ids"
].to(torch_device)
out = self.model.generate(input_ids, do_sample=False, max_new_tokens=10)
output_sentence = self.tokenizer.decode(out[0, :])
self.assertEqual(
output_sentence,
"<|startoftext|>Hey how are you doing on this lovely evening? Canyon rins hugaughter glamour Rutgers Singh Hebrew cases Cats",
)
with torch.no_grad():
logits = self.model(input_ids=input_ids).logits
EXPECTED_LOGITS_NO_GRAD = torch.tensor(
[
0.0140, -0.2246, 0.0408, -0.1016, 0.0471, 0.2715, -0.1465, 0.1631,
-0.2949, -0.0297, 0.0250, -0.5586, -0.2139, -0.1426, -0.1602, 0.1309,
0.0703, 0.2236, 0.1729, -0.2285, -0.1152, -0.1177, -0.1367, 0.0289,
0.1245, 0.2363, 0.0442, 0.1094, -0.1348, -0.2295, 0.1494, -0.3945,
0.1777, -0.4570, -0.0408, 0.2412, 0.1562, -0.1943, 0.2373, -0.0593
]
, dtype=torch.float32) # fmt: skip
torch.testing.assert_close(logits[0, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD, rtol=1e-3, atol=1e-3)
@slow
def test_simple_batched_generate_with_padding(self):
self.model.to(torch_device)
inputs = self.tokenizer(
["Hey how are you doing on this lovely evening?", "Tell me a story"], padding=True, return_tensors="pt"
).to(torch_device)
out = self.model.generate(**inputs, do_sample=False, max_new_tokens=10)
output_sentences = self.tokenizer.batch_decode(out)
self.assertEqual(
output_sentences[0],
"<|startoftext|>Hey how are you doing on this lovely evening? Canyon rins hugaughter glamour Rutgers Singh Hebrew cases Cats",
)
self.assertEqual(
output_sentences[1],
"<|pad|><|pad|><|pad|><|pad|><|pad|><|pad|><|startoftext|>Tell me a storyptus Nets Madison El chamadamodern updximVaparsed",
)
with torch.no_grad():
logits = self.model(input_ids=inputs["input_ids"]).logits
EXPECTED_LOGITS_NO_GRAD_0 = torch.tensor(
[
0.0140, -0.2246, 0.0408, -0.1016, 0.0471, 0.2715, -0.1465, 0.1631,
-0.2949, -0.0297, 0.0250, -0.5586, -0.2139, -0.1426, -0.1602, 0.1309,
0.0703, 0.2236, 0.1729, -0.2285, -0.1152, -0.1177, -0.1367, 0.0289,
0.1245, 0.2363, 0.0442, 0.1094, -0.1348, -0.2295, 0.1494, -0.3945,
0.1777, -0.4570, -0.0408, 0.2412, 0.1562, -0.1943, 0.2373, -0.0593
]
, dtype=torch.float32) # fmt: skip
EXPECTED_LOGITS_NO_GRAD_1 = torch.tensor(
[
-0.1289, 0.2363, -0.4180, -0.0302, -0.0476, 0.0327, 0.2578, 0.0874,
0.1484, 0.2305, -0.1152, -0.1396, -0.1494, -0.1113, -0.0021, -0.2832,
0.2002, -0.2676, 0.0598, -0.1982, -0.2539, -0.1133, -0.1973, 0.2148,
0.0559, 0.1670, 0.1846, 0.1270, 0.1680, -0.1250, -0.2656, -0.2871,
0.2344, 0.2637, 0.0510, -0.1855, 0.2158, -0.1289, 0.1758, 0.0074
]
, dtype=torch.float32) # fmt: skip
torch.testing.assert_close(logits[0, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD_0, rtol=1e-3, atol=1e-3)
torch.testing.assert_close(logits[1, -1, :40].cpu(), EXPECTED_LOGITS_NO_GRAD_1, rtol=1e-3, atol=1e-3)

9
utils/check_config_attributes.py
View File

@ -32,6 +32,15 @@ transformers = direct_transformers_import(PATH_TO_TRANSFORMERS)
CONFIG_MAPPING = transformers.models.auto.configuration_auto.CONFIG_MAPPING
SPECIAL_CASES_TO_ALLOW = {
# 'max_position_embeddings' is not used in modeling file, but needed for eval frameworks like Huggingface's lighteval (https://github.com/huggingface/lighteval/blob/af24080ea4f16eaf1683e353042a2dfc9099f038/src/lighteval/models/base_model.py#L264).
# periods and offsers are not used in modeling file, but used in the configuration file to define `layers_block_type` and `layers_num_experts`.
"JambaConfig": [
"max_position_embeddings",
"attn_layer_offset",
"attn_layer_period",
"expert_layer_offset",
"expert_layer_period",
],
# used to compute the property `self.chunk_length`
"EncodecConfig": ["overlap"],
# used to compute the property `self.layers_block_type`

2
utils/not_doctested.txt
View File

@ -631,6 +631,8 @@ src/transformers/models/instructblip/configuration_instructblip.py
src/transformers/models/instructblip/convert_instructblip_original_to_pytorch.py
src/transformers/models/instructblip/modeling_instructblip.py
src/transformers/models/instructblip/processing_instructblip.py
src/transformers/models/jamba/configuration_jamba.py
src/transformers/models/jamba/modeling_jamba.py
src/transformers/models/jukebox/configuration_jukebox.py
src/transformers/models/jukebox/convert_jukebox.py
src/transformers/models/jukebox/modeling_jukebox.py