mirror of
https://github.com/huggingface/transformers.git
synced 2025-08-02 19:21:31 +06:00
Add GPTBigCode model (Optimized GPT2 with MQA from Santacoder & BigCode) (#22575)
* Add model with cli tool * Remove unwanted stuff * Add new code * Remove inference runner * Style * Fix checks * Test updates * make fixup * fix docs * fix doc * fix test * hopefully fix pipeline tests * refactor * fix CIs * add comment * rename to `GPTBigCodeForCausalLM` * correct readme * make fixup + docs * make fixup * fixes * fixes * Remove pruning * Remove import * Doc updates * More pruning removal * Combine copies * Single MQA implementation, remove kv cache pre-allocation and padding * Update doc * Revert refactor to match gpt2 style * Merge back key and value caches, fix some type hints * Update doc * Fix position ids pith padding (PR 21080) * Add conversion script temporarily * Update conversion script * Remove checkpoint conversion * New model * Fix MQA test * Fix copies * try fix tests * FIX TEST!! * remove `DoubleHeadsModel` * add MQA tests * add slow tests * clean up * add CPU checker * final fixes * fixes - fix GPU issue - fixed slow tests - skip disk offload * fix final issue * Simplify and comment baddbmm fix * Remove unnecessary code * Transpose tweaks * Use beta=1 on cpu, improve tests --------- Co-authored-by: younesbelkada <younesbelkada@gmail.com>
This commit is contained in:
Joel Lamy-Poirier
GitHub
parent
656e869a45
commit
e0921c6b53
@ -350,6 +350,7 @@ Current number of checkpoints: ** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
|
||||
|
||||
@ -338,6 +338,7 @@ Número actual de puntos de control: ** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
|
||||
|
||||
@ -310,6 +310,7 @@ conda install -c huggingface transformers
|
||||
1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (ओपनएआई से) साथ में पेपर [लैंग्वेज मॉडल्स अनसुपरवाइज्ड मल्टीटास्क लर्नर्स हैं](https://blog.openai.com/better-language-models/) एलेक रैडफोर्ड*, जेफरी वू*, रेवन चाइल्ड, डेविड लुआन, डारियो एमोडी* द्वारा * और इल्या सुत्सकेवर** ने पोस्ट किया।
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (EleutherAI से) साथ वाला पेपर [kingoflolz/mesh-transformer-jax](https://github. com/kingoflolz/mesh-transformer-jax/) बेन वांग और अरन कोमात्सुजाकी द्वारा।
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (BigCode से) Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra. द्वाराअनुसंधान पत्र [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) के साथ जारी किया गया
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA से) साथ में कागज [GroupViT: टेक्स्ट सुपरविजन से सिमेंटिक सेगमेंटेशन इमर्जेस](https://arxiv .org/abs/2202.11094) जियारुई जू, शालिनी डी मेलो, सिफ़ी लियू, वोनमिन बायन, थॉमस ब्रेउएल, जान कौट्ज़, ज़ियाओलोंग वांग द्वारा।
|
||||
|
||||
@ -372,6 +372,7 @@ Flax、PyTorch、TensorFlowをcondaでインストールする方法は、それ
|
||||
1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (OpenAI から) Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** から公開された研究論文: [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/)
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (EleutherAI から) Ben Wang and Aran Komatsuzaki から公開されたレポジトリー [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/)
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (AI-Sweden から) Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren から公開された研究論文: [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf)
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (BigCode から) Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra. から公開された研究論文 [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988)
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) 坂本俊之(tanreinama)からリリースされました.
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (Microsoft から) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu から公開された研究論文: [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234).
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA から) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang から公開された研究論文: [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094)
|
||||
|
||||
@ -287,6 +287,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
|
||||
1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (OpenAI 에서) Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 의 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 논문과 함께 발표했습니다.
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (AI-Sweden 에서) Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren. 의 [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) 논문과 함께 발표했습니다.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (BigCode 에서 제공)은 Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.의 [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988)논문과 함께 발표했습니다.
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu 의 [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) 논문과 함께 발표했습니다.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (UCSD, NVIDIA 에서) Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 의 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 논문과 함께 발표했습니다.
|
||||
|
||||
@ -311,6 +311,7 @@ conda install -c huggingface transformers
|
||||
1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (来自 OpenAI) 伴随论文 [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) 由 Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever** 发布。
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (来自 EleutherAI) 伴随论文 [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) 由 Ben Wang and Aran Komatsuzaki 发布。
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (来自 BigCode) 伴随论文 [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) 由 Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra 发布。
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (来自 UCSD, NVIDIA) 伴随论文 [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) 由 Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang 发布。
|
||||
|
||||
@ -323,6 +323,7 @@ conda install -c huggingface transformers
|
||||
1. **[GPT-2](https://huggingface.co/docs/transformers/model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
|
||||
1. **[GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj)** (from EleutherAI) released with the paper [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
|
||||
1. **[GPT-Sw3](https://huggingface.co/docs/transformers/model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](https://huggingface.co/docs/transformers/main/model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
|
||||
1. **[GPTSAN-japanese](https://huggingface.co/docs/transformers/model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by 坂本俊之(tanreinama).
|
||||
1. **[Graphormer](https://huggingface.co/docs/transformers/model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](https://huggingface.co/docs/transformers/model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
|
||||
|
||||
@ -309,6 +309,8 @@
|
||||
title: GPT-J
|
||||
- local: model_doc/gpt2
|
||||
title: GPT2
|
||||
- local: model_doc/gpt_bigcode
|
||||
title: GPTBigCode
|
||||
- local: model_doc/gptsan-japanese
|
||||
title: GPTSAN Japanese
|
||||
- local: model_doc/gpt-sw3
|
||||
|
||||
@ -124,6 +124,7 @@ The documentation is organized into five sections:
|
||||
1. **[GPT-2](model_doc/gpt2)** (from OpenAI) released with the paper [Language Models are Unsupervised Multitask Learners](https://blog.openai.com/better-language-models/) by Alec Radford*, Jeffrey Wu*, Rewon Child, David Luan, Dario Amodei** and Ilya Sutskever**.
|
||||
1. **[GPT-J](model_doc/gptj)** (from EleutherAI) released in the repository [kingoflolz/mesh-transformer-jax](https://github.com/kingoflolz/mesh-transformer-jax/) by Ben Wang and Aran Komatsuzaki.
|
||||
1. **[GPT-Sw3](model_doc/gpt-sw3)** (from AI-Sweden) released with the paper [Lessons Learned from GPT-SW3: Building the First Large-Scale Generative Language Model for Swedish](http://www.lrec-conf.org/proceedings/lrec2022/pdf/2022.lrec-1.376.pdf) by Ariel Ekgren, Amaru Cuba Gyllensten, Evangelia Gogoulou, Alice Heiman, Severine Verlinden, Joey Öhman, Fredrik Carlsson, Magnus Sahlgren.
|
||||
1. **[GPTBigCode](model_doc/gpt_bigcode)** (from BigCode) released with the paper [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
|
||||
1. **[GPTSAN-japanese](model_doc/gptsan-japanese)** released in the repository [tanreinama/GPTSAN](https://github.com/tanreinama/GPTSAN/blob/main/report/model.md) by Toshiyuki Sakamoto(tanreinama).
|
||||
1. **[Graphormer](model_doc/graphormer)** (from Microsoft) released with the paper [Do Transformers Really Perform Bad for Graph Representation?](https://arxiv.org/abs/2106.05234) by Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, Tie-Yan Liu.
|
||||
1. **[GroupViT](model_doc/groupvit)** (from UCSD, NVIDIA) released with the paper [GroupViT: Semantic Segmentation Emerges from Text Supervision](https://arxiv.org/abs/2202.11094) by Jiarui Xu, Shalini De Mello, Sifei Liu, Wonmin Byeon, Thomas Breuel, Jan Kautz, Xiaolong Wang.
|
||||
@ -322,6 +323,7 @@ Flax), PyTorch, and/or TensorFlow.
|
||||
| GPT NeoX Japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
|
||||
| GPT-J | ❌ | ❌ | ✅ | ✅ | ✅ |
|
||||
| GPT-Sw3 | ✅ | ✅ | ✅ | ✅ | ✅ |
|
||||
| GPTBigCode | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
| GPTSAN-japanese | ✅ | ❌ | ✅ | ❌ | ❌ |
|
||||
| Graphormer | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
| GroupViT | ❌ | ❌ | ✅ | ✅ | ❌ |
|
||||
|
||||
65
docs/source/en/model_doc/gpt_bigcode.mdx
Normal file
65
docs/source/en/model_doc/gpt_bigcode.mdx
Normal file
@ -0,0 +1,65 @@
|
||||
<!--Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
-->
|
||||
|
||||
# GPTBigCode
|
||||
|
||||
## Overview
|
||||
|
||||
The GPTBigCode model was proposed in [SantaCoder: don't reach for the stars!](https://arxiv.org/abs/2301.03988) by BigCode. The listed authors are: Loubna Ben Allal, Raymond Li, Denis Kocetkov, Chenghao Mou, Christopher Akiki, Carlos Munoz Ferrandis, Niklas Muennighoff, Mayank Mishra, Alex Gu, Manan Dey, Logesh Kumar Umapathi, Carolyn Jane Anderson, Yangtian Zi, Joel Lamy Poirier, Hailey Schoelkopf, Sergey Troshin, Dmitry Abulkhanov, Manuel Romero, Michael Lappert, Francesco De Toni, Bernardo García del Río, Qian Liu, Shamik Bose, Urvashi Bhattacharyya, Terry Yue Zhuo, Ian Yu, Paulo Villegas, Marco Zocca, Sourab Mangrulkar, David Lansky, Huu Nguyen, Danish Contractor, Luis Villa, Jia Li, Dzmitry Bahdanau, Yacine Jernite, Sean Hughes, Daniel Fried, Arjun Guha, Harm de Vries, Leandro von Werra.
|
||||
|
||||
The abstract from the paper is the following:uery
|
||||
|
||||
*The BigCode project is an open-scientific collaboration working on the responsible development of large language models for code. This tech report describes the progress of the collaboration until December 2022, outlining the current state of the Personally Identifiable Information (PII) redaction pipeline, the experiments conducted to de-risk the model architecture, and the experiments investigating better preprocessing methods for the training data. We train 1.1B parameter models on the Java, JavaScript, and Python subsets of The Stack and evaluate them on the MultiPL-E text-to-code benchmark. We find that more aggressive filtering of near-duplicates can further boost performance and, surprisingly, that selecting files from repositories with 5+ GitHub stars deteriorates performance significantly. Our best model outperforms previous open-source multilingual code generation models (InCoder-6.7B and CodeGen-Multi-2.7B) in both left-to-right generation and infilling on the Java, JavaScript, and Python portions of MultiPL-E, despite being a substantially smaller model. All models are released under an OpenRAIL license at [this https URL.](https://huggingface.co/bigcode)*
|
||||
|
||||
The model is a an optimized [GPT2 model](https://huggingface.co/docs/transformers/model_doc/gpt2) with support for Multi-Query Attention.
|
||||
|
||||
## Technical details
|
||||
|
||||
The main differences compared to GPT2.
|
||||
- Added support for Multi-Query Attention.
|
||||
- Use `gelu_pytorch_tanh` instead of classic `gelu`.
|
||||
- Avoid unnecessary synchronizations (this has since been added to GPT2 in #20061, but wasn't in the reference codebase).
|
||||
- Use Linear layers instead of Conv1D (good speedup but makes the checkpoints incompatible).
|
||||
- Merge `_attn` and `_upcast_and_reordered_attn`. Always merge the matmul with scaling. Rename `reorder_and_upcast_attn`->`attention_softmax_in_fp32`
|
||||
- Cache the attention mask value to avoid recreating it every time.
|
||||
- Use jit to fuse the attention fp32 casting, masking, softmax, and scaling.
|
||||
- Combine the attention and causal masks into a single one, pre-computed for the whole model instead of every layer.
|
||||
- Merge the key and value caches into one (this changes the format of layer_past/ present, does it risk creating problems?)
|
||||
- Use the memory layout (self.num_heads, 3, self.head_dim) instead of `(3, self.num_heads, self.head_dim)` for the QKV tensor with MHA. (prevents an overhead with the merged key and values, but makes the checkpoints incompatible with the original gpt2 model).
|
||||
|
||||
You can read more about the optimizations in the [original pull request](https://github.com/huggingface/transformers/pull/22575)
|
||||
|
||||
## GPTBigCodeConfig
|
||||
|
||||
[[autodoc]] GPTBigCodeConfig
|
||||
|
||||
|
||||
## GPTBigCodeModel
|
||||
|
||||
[[autodoc]] GPTBigCodeModel
|
||||
- forward
|
||||
|
||||
## GPTBigCodeForCausalLM
|
||||
|
||||
[[autodoc]] GPTBigCodeForCausalLM
|
||||
- forward
|
||||
|
||||
|
||||
## GPTBigCodeForSequenceClassification
|
||||
|
||||
[[autodoc]] GPTBigCodeForSequenceClassification
|
||||
- forward
|
||||
|
||||
## GPTBigCodeForTokenClassification
|
||||
|
||||
[[autodoc]] GPTBigCodeForTokenClassification
|
||||
- forward
|
||||
@ -34,7 +34,7 @@ Choose one of the following architectures:
|
||||
|
||||
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
|
||||
|
||||
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MVP](../model_doc/mvp), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
|
||||
[BART](../model_doc/bart), [BERT](../model_doc/bert), [Bert Generation](../model_doc/bert-generation), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BioGpt](../model_doc/biogpt), [Blenderbot](../model_doc/blenderbot), [BlenderbotSmall](../model_doc/blenderbot-small), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CodeGen](../model_doc/codegen), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [GIT](../model_doc/git), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT NeoX](../model_doc/gpt_neox), [GPT NeoX Japanese](../model_doc/gpt_neox_japanese), [GPT-J](../model_doc/gptj), [LLaMA](../model_doc/llama), [Marian](../model_doc/marian), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MVP](../model_doc/mvp), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Pegasus](../model_doc/pegasus), [PLBart](../model_doc/plbart), [ProphetNet](../model_doc/prophetnet), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [Speech2Text2](../model_doc/speech_to_text_2), [Transformer-XL](../model_doc/transfo-xl), [TrOCR](../model_doc/trocr), [XGLM](../model_doc/xglm), [XLM](../model_doc/xlm), [XLM-ProphetNet](../model_doc/xlm-prophetnet), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod)
|
||||
|
||||
<!--End of the generated tip-->
|
||||
|
||||
|
||||
@ -28,7 +28,7 @@ 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), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPT Neo](../model_doc/gpt_neo), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
[ALBERT](../model_doc/albert), [BART](../model_doc/bart), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BigBird-Pegasus](../model_doc/bigbird_pegasus), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [CTRL](../model_doc/ctrl), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [GPT Neo](../model_doc/gpt_neo), [GPT-J](../model_doc/gptj), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LED](../model_doc/led), [LiLT](../model_doc/lilt), [LLaMA](../model_doc/llama), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [mBART](../model_doc/mbart), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [MVP](../model_doc/mvp), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [OpenAI GPT](../model_doc/openai-gpt), [OPT](../model_doc/opt), [Perceiver](../model_doc/perceiver), [PLBart](../model_doc/plbart), [QDQBert](../model_doc/qdqbert), [Reformer](../model_doc/reformer), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [TAPAS](../model_doc/tapas), [Transformer-XL](../model_doc/transfo-xl), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
|
||||
|
||||
<!--End of the generated tip-->
|
||||
|
||||
@ -28,7 +28,7 @@ The task illustrated in this tutorial is supported by the following model archit
|
||||
|
||||
<!--This tip is automatically generated by `make fix-copies`, do not fill manually!-->
|
||||
|
||||
[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
[ALBERT](../model_doc/albert), [BERT](../model_doc/bert), [BigBird](../model_doc/big_bird), [BioGpt](../model_doc/biogpt), [BLOOM](../model_doc/bloom), [CamemBERT](../model_doc/camembert), [CANINE](../model_doc/canine), [ConvBERT](../model_doc/convbert), [Data2VecText](../model_doc/data2vec-text), [DeBERTa](../model_doc/deberta), [DeBERTa-v2](../model_doc/deberta-v2), [DistilBERT](../model_doc/distilbert), [ELECTRA](../model_doc/electra), [ERNIE](../model_doc/ernie), [ErnieM](../model_doc/ernie_m), [ESM](../model_doc/esm), [FlauBERT](../model_doc/flaubert), [FNet](../model_doc/fnet), [Funnel Transformer](../model_doc/funnel), [GPT-Sw3](../model_doc/gpt-sw3), [OpenAI GPT-2](../model_doc/gpt2), [GPTBigCode](../model_doc/gpt_bigcode), [I-BERT](../model_doc/ibert), [LayoutLM](../model_doc/layoutlm), [LayoutLMv2](../model_doc/layoutlmv2), [LayoutLMv3](../model_doc/layoutlmv3), [LiLT](../model_doc/lilt), [Longformer](../model_doc/longformer), [LUKE](../model_doc/luke), [MarkupLM](../model_doc/markuplm), [MEGA](../model_doc/mega), [Megatron-BERT](../model_doc/megatron-bert), [MobileBERT](../model_doc/mobilebert), [MPNet](../model_doc/mpnet), [Nezha](../model_doc/nezha), [Nyströmformer](../model_doc/nystromformer), [QDQBert](../model_doc/qdqbert), [RemBERT](../model_doc/rembert), [RoBERTa](../model_doc/roberta), [RoBERTa-PreLayerNorm](../model_doc/roberta-prelayernorm), [RoCBert](../model_doc/roc_bert), [RoFormer](../model_doc/roformer), [SqueezeBERT](../model_doc/squeezebert), [XLM](../model_doc/xlm), [XLM-RoBERTa](../model_doc/xlm-roberta), [XLM-RoBERTa-XL](../model_doc/xlm-roberta-xl), [XLNet](../model_doc/xlnet), [X-MOD](../model_doc/xmod), [YOSO](../model_doc/yoso)
|
||||
|
||||
<!--End of the generated tip-->
|
||||
|
||||
|
||||
@ -298,6 +298,7 @@ _import_structure = {
|
||||
"models.git": ["GIT_PRETRAINED_CONFIG_ARCHIVE_MAP", "GitConfig", "GitProcessor", "GitVisionConfig"],
|
||||
"models.glpn": ["GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP", "GLPNConfig"],
|
||||
"models.gpt2": ["GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPT2Config", "GPT2Tokenizer"],
|
||||
"models.gpt_bigcode": ["GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTBigCodeConfig"],
|
||||
"models.gpt_neo": ["GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTNeoConfig"],
|
||||
"models.gpt_neox": ["GPT_NEOX_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTNeoXConfig"],
|
||||
"models.gpt_neox_japanese": ["GPT_NEOX_JAPANESE_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTNeoXJapaneseConfig"],
|
||||
@ -1641,6 +1642,16 @@ else:
|
||||
"load_tf_weights_in_gpt2",
|
||||
]
|
||||
)
|
||||
_import_structure["models.gpt_bigcode"].extend(
|
||||
[
|
||||
"GPT_BIGCODE_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"GPTBigCodeForCausalLM",
|
||||
"GPTBigCodeForSequenceClassification",
|
||||
"GPTBigCodeForTokenClassification",
|
||||
"GPTBigCodeModel",
|
||||
"GPTBigCodePreTrainedModel",
|
||||
]
|
||||
)
|
||||
_import_structure["models.gpt_neo"].extend(
|
||||
[
|
||||
"GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
@ -3982,6 +3993,7 @@ if TYPE_CHECKING:
|
||||
from .models.git import GIT_PRETRAINED_CONFIG_ARCHIVE_MAP, GitConfig, GitProcessor, GitVisionConfig
|
||||
from .models.glpn import GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP, GLPNConfig
|
||||
from .models.gpt2 import GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP, GPT2Config, GPT2Tokenizer
|
||||
from .models.gpt_bigcode import GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTBigCodeConfig
|
||||
from .models.gpt_neo import GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTNeoConfig
|
||||
from .models.gpt_neox import GPT_NEOX_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTNeoXConfig
|
||||
from .models.gpt_neox_japanese import GPT_NEOX_JAPANESE_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTNeoXJapaneseConfig
|
||||
@ -5133,6 +5145,14 @@ if TYPE_CHECKING:
|
||||
GPT2PreTrainedModel,
|
||||
load_tf_weights_in_gpt2,
|
||||
)
|
||||
from .models.gpt_bigcode import (
|
||||
GPT_BIGCODE_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
GPTBigCodeForCausalLM,
|
||||
GPTBigCodeForSequenceClassification,
|
||||
GPTBigCodeForTokenClassification,
|
||||
GPTBigCodeModel,
|
||||
GPTBigCodePreTrainedModel,
|
||||
)
|
||||
from .models.gpt_neo import (
|
||||
GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
GPTNeoForCausalLM,
|
||||
|
||||
@ -82,6 +82,7 @@ from . import (
|
||||
git,
|
||||
glpn,
|
||||
gpt2,
|
||||
gpt_bigcode,
|
||||
gpt_neo,
|
||||
gpt_neox,
|
||||
gpt_neox_japanese,
|
||||
|
||||
@ -91,6 +91,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
|
||||
("glpn", "GLPNConfig"),
|
||||
("gpt-sw3", "GPT2Config"),
|
||||
("gpt2", "GPT2Config"),
|
||||
("gpt_bigcode", "GPTBigCodeConfig"),
|
||||
("gpt_neo", "GPTNeoConfig"),
|
||||
("gpt_neox", "GPTNeoXConfig"),
|
||||
("gpt_neox_japanese", "GPTNeoXJapaneseConfig"),
|
||||
@ -272,6 +273,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
|
||||
("git", "GIT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("glpn", "GLPN_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("gpt2", "GPT2_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("gpt_bigcode", "GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("gpt_neo", "GPT_NEO_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("gpt_neox", "GPT_NEOX_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("gpt_neox_japanese", "GPT_NEOX_JAPANESE_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
@ -455,6 +457,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
|
||||
("glpn", "GLPN"),
|
||||
("gpt-sw3", "GPT-Sw3"),
|
||||
("gpt2", "OpenAI GPT-2"),
|
||||
("gpt_bigcode", "GPTBigCode"),
|
||||
("gpt_neo", "GPT Neo"),
|
||||
("gpt_neox", "GPT NeoX"),
|
||||
("gpt_neox_japanese", "GPT NeoX Japanese"),
|
||||
|
||||
@ -89,6 +89,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
|
||||
("glpn", "GLPNModel"),
|
||||
("gpt-sw3", "GPT2Model"),
|
||||
("gpt2", "GPT2Model"),
|
||||
("gpt_bigcode", "GPTBigCodeModel"),
|
||||
("gpt_neo", "GPTNeoModel"),
|
||||
("gpt_neox", "GPTNeoXModel"),
|
||||
("gpt_neox_japanese", "GPTNeoXJapaneseModel"),
|
||||
@ -224,6 +225,7 @@ MODEL_FOR_PRETRAINING_MAPPING_NAMES = OrderedDict(
|
||||
("funnel", "FunnelForPreTraining"),
|
||||
("gpt-sw3", "GPT2LMHeadModel"),
|
||||
("gpt2", "GPT2LMHeadModel"),
|
||||
("gpt_bigcode", "GPTBigCodeForCausalLM"),
|
||||
("gptsan-japanese", "GPTSanJapaneseForConditionalGeneration"),
|
||||
("ibert", "IBertForMaskedLM"),
|
||||
("layoutlm", "LayoutLMForMaskedLM"),
|
||||
@ -293,6 +295,7 @@ MODEL_WITH_LM_HEAD_MAPPING_NAMES = OrderedDict(
|
||||
("git", "GitForCausalLM"),
|
||||
("gpt-sw3", "GPT2LMHeadModel"),
|
||||
("gpt2", "GPT2LMHeadModel"),
|
||||
("gpt_bigcode", "GPTBigCodeForCausalLM"),
|
||||
("gpt_neo", "GPTNeoForCausalLM"),
|
||||
("gpt_neox", "GPTNeoXForCausalLM"),
|
||||
("gpt_neox_japanese", "GPTNeoXJapaneseForCausalLM"),
|
||||
@ -362,6 +365,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
|
||||
("git", "GitForCausalLM"),
|
||||
("gpt-sw3", "GPT2LMHeadModel"),
|
||||
("gpt2", "GPT2LMHeadModel"),
|
||||
("gpt_bigcode", "GPTBigCodeForCausalLM"),
|
||||
("gpt_neo", "GPTNeoForCausalLM"),
|
||||
("gpt_neox", "GPTNeoXForCausalLM"),
|
||||
("gpt_neox_japanese", "GPTNeoXJapaneseForCausalLM"),
|
||||
@ -653,6 +657,7 @@ MODEL_FOR_SEQUENCE_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
|
||||
("funnel", "FunnelForSequenceClassification"),
|
||||
("gpt-sw3", "GPT2ForSequenceClassification"),
|
||||
("gpt2", "GPT2ForSequenceClassification"),
|
||||
("gpt_bigcode", "GPTBigCodeForSequenceClassification"),
|
||||
("gpt_neo", "GPTNeoForSequenceClassification"),
|
||||
("gptj", "GPTJForSequenceClassification"),
|
||||
("ibert", "IBertForSequenceClassification"),
|
||||
@ -800,6 +805,7 @@ MODEL_FOR_TOKEN_CLASSIFICATION_MAPPING_NAMES = OrderedDict(
|
||||
("funnel", "FunnelForTokenClassification"),
|
||||
("gpt-sw3", "GPT2ForTokenClassification"),
|
||||
("gpt2", "GPT2ForTokenClassification"),
|
||||
("gpt_bigcode", "GPTBigCodeForTokenClassification"),
|
||||
("ibert", "IBertForTokenClassification"),
|
||||
("layoutlm", "LayoutLMForTokenClassification"),
|
||||
("layoutlmv2", "LayoutLMv2ForTokenClassification"),
|
||||
|
||||
@ -156,6 +156,7 @@ else:
|
||||
("git", ("BertTokenizer", "BertTokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt-sw3", ("GPTSw3Tokenizer" if is_sentencepiece_available() else None, None)),
|
||||
("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt_bigcode", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt_neo", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt_neox", (None, "GPTNeoXTokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt_neox_japanese", ("GPTNeoXJapaneseTokenizer", None)),
|
||||
|
||||
65
src/transformers/models/gpt_bigcode/__init__.py
Normal file
65
src/transformers/models/gpt_bigcode/__init__.py
Normal file
@ -0,0 +1,65 @@
|
||||
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from ...utils import (
|
||||
OptionalDependencyNotAvailable,
|
||||
_LazyModule,
|
||||
is_torch_available,
|
||||
)
|
||||
|
||||
|
||||
_import_structure = {
|
||||
"configuration_gpt_bigcode": ["GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP", "GPTBigCodeConfig"],
|
||||
}
|
||||
|
||||
try:
|
||||
if not is_torch_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
_import_structure["modeling_gpt_bigcode"] = [
|
||||
"GPT_BIGCODE_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"GPTBigCodeForSequenceClassification",
|
||||
"GPTBigCodeForTokenClassification",
|
||||
"GPTBigCodeForCausalLM",
|
||||
"GPTBigCodeModel",
|
||||
"GPTBigCodePreTrainedModel",
|
||||
]
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .configuration_gpt_bigcode import GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP, GPTBigCodeConfig
|
||||
|
||||
try:
|
||||
if not is_torch_available():
|
||||
raise OptionalDependencyNotAvailable()
|
||||
except OptionalDependencyNotAvailable:
|
||||
pass
|
||||
else:
|
||||
from .modeling_gpt_bigcode import (
|
||||
GPT_BIGCODE_PRETRAINED_MODEL_ARCHIVE_LIST,
|
||||
GPTBigCodeForCausalLM,
|
||||
GPTBigCodeForSequenceClassification,
|
||||
GPTBigCodeForTokenClassification,
|
||||
GPTBigCodeModel,
|
||||
GPTBigCodePreTrainedModel,
|
||||
)
|
||||
|
||||
|
||||
else:
|
||||
import sys
|
||||
|
||||
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
||||
145
src/transformers/models/gpt_bigcode/configuration_gpt_bigcode.py
Normal file
145
src/transformers/models/gpt_bigcode/configuration_gpt_bigcode.py
Normal file
@ -0,0 +1,145 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2023 The BigCode team and HuggingFace Inc. team.
|
||||
#
|
||||
# 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.
|
||||
""" GPTBigCode configuration"""
|
||||
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
GPT_BIGCODE_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
"bigcode/gpt_bigcode-santacoder": "https://huggingface.co/bigcode/gpt_bigcode-santacoder/resolve/main/config.json",
|
||||
}
|
||||
|
||||
|
||||
class GPTBigCodeConfig(PretrainedConfig):
|
||||
"""
|
||||
This is the configuration class to store the configuration of a [`GPTBigCodeModel`]. It is used to instantiate a
|
||||
GPTBigCode 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 GPTBigCode
|
||||
[gpt_bigcode](https://huggingface.co/gpt_bigcode) architecture.
|
||||
|
||||
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
|
||||
documentation from [`PretrainedConfig`] for more information.
|
||||
|
||||
|
||||
Args:
|
||||
vocab_size (`int`, *optional*, defaults to 50257):
|
||||
Vocabulary size of the GPT-2 model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`GPTBigCodeModel`].
|
||||
n_positions (`int`, *optional*, defaults to 1024):
|
||||
The maximum sequence length that this model might ever be used with. Typically set this to something large
|
||||
just in case (e.g., 512 or 1024 or 2048).
|
||||
n_embd (`int`, *optional*, defaults to 768):
|
||||
Dimensionality of the embeddings and hidden states.
|
||||
n_layer (`int`, *optional*, defaults to 12):
|
||||
Number of hidden layers in the Transformer encoder.
|
||||
n_head (`int`, *optional*, defaults to 12):
|
||||
Number of attention heads for each attention layer in the Transformer encoder.
|
||||
n_inner (`int`, *optional*, defaults to None):
|
||||
Dimensionality of the inner feed-forward layers. `None` will set it to 4 times n_embd
|
||||
activation_function (`str`, *optional*, defaults to `"gelu_pytorch_tanh"`):
|
||||
Activation function, to be selected in the list `["relu", "silu", "gelu", "tanh", "gelu_new",
|
||||
"gelu_pytorch_tanh"]`.
|
||||
resid_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
||||
embd_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout ratio for the embeddings.
|
||||
attn_pdrop (`float`, *optional*, defaults to 0.1):
|
||||
The dropout ratio for the attention.
|
||||
layer_norm_epsilon (`float`, *optional*, defaults to 1e-5):
|
||||
The epsilon to use in the layer normalization layers.
|
||||
initializer_range (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
scale_attn_weights (`bool`, *optional*, defaults to `True`):
|
||||
Scale attention weights by dividing by sqrt(hidden_size)..
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models).
|
||||
attention_softmax_in_fp32 (`bool`, *optional*, defaults to `True`):
|
||||
Whether to call the fused softmax in float32.
|
||||
scale_attention_softmax_in_fp32 (`bool`, *optional*, defaults to `True`):
|
||||
Whether to scale the attention softmax in float32.
|
||||
attention_type (`bool`, *optional*, defaults to `True`):
|
||||
Whether to use Multi-Query Attion (`True`) or Multi-Head Attention (`False`).
|
||||
Example:
|
||||
|
||||
```python
|
||||
>>> from transformers import GPTBigCodeConfig, GPTBigCodeModel
|
||||
|
||||
>>> # Initializing a GPTBigCode configuration
|
||||
>>> configuration = GPTBigCodeConfig()
|
||||
|
||||
>>> # Initializing a model (with random weights) from the configuration
|
||||
>>> model = GPTBigCodeModel(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
|
||||
model_type = "gpt_bigcode"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
attribute_map = {
|
||||
"hidden_size": "n_embd",
|
||||
"max_position_embeddings": "n_positions",
|
||||
"num_attention_heads": "n_head",
|
||||
"num_hidden_layers": "n_layer",
|
||||
}
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=50257,
|
||||
n_positions=1024,
|
||||
n_embd=768,
|
||||
n_layer=12,
|
||||
n_head=12,
|
||||
n_inner=None,
|
||||
activation_function="gelu_pytorch_tanh",
|
||||
resid_pdrop=0.1,
|
||||
embd_pdrop=0.1,
|
||||
attn_pdrop=0.1,
|
||||
layer_norm_epsilon=1e-5,
|
||||
initializer_range=0.02,
|
||||
scale_attn_weights=True,
|
||||
use_cache=True,
|
||||
bos_token_id=50256,
|
||||
eos_token_id=50256,
|
||||
attention_softmax_in_fp32=True,
|
||||
scale_attention_softmax_in_fp32=True,
|
||||
multi_query=True,
|
||||
**kwargs,
|
||||
):
|
||||
self.vocab_size = vocab_size
|
||||
self.n_positions = n_positions
|
||||
self.n_embd = n_embd
|
||||
self.n_layer = n_layer
|
||||
self.n_head = n_head
|
||||
self.n_inner = n_inner
|
||||
self.activation_function = activation_function
|
||||
self.resid_pdrop = resid_pdrop
|
||||
self.embd_pdrop = embd_pdrop
|
||||
self.attn_pdrop = attn_pdrop
|
||||
self.layer_norm_epsilon = layer_norm_epsilon
|
||||
self.initializer_range = initializer_range
|
||||
self.scale_attn_weights = scale_attn_weights
|
||||
self.use_cache = use_cache
|
||||
self.attention_softmax_in_fp32 = attention_softmax_in_fp32
|
||||
self.scale_attention_softmax_in_fp32 = scale_attention_softmax_in_fp32
|
||||
self.multi_query = multi_query
|
||||
|
||||
self.bos_token_id = bos_token_id
|
||||
self.eos_token_id = eos_token_id
|
||||
|
||||
super().__init__(bos_token_id=bos_token_id, eos_token_id=eos_token_id, **kwargs)
|
||||
1068
src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
Normal file
1068
src/transformers/models/gpt_bigcode/modeling_gpt_bigcode.py
Normal file
File diff suppressed because it is too large
Load Diff
@ -3153,6 +3153,44 @@ def load_tf_weights_in_gpt2(*args, **kwargs):
|
||||
requires_backends(load_tf_weights_in_gpt2, ["torch"])
|
||||
|
||||
|
||||
GPT_BIGCODE_PRETRAINED_MODEL_ARCHIVE_LIST = None
|
||||
|
||||
|
||||
class GPTBigCodeForCausalLM(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class GPTBigCodeForSequenceClassification(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class GPTBigCodeForTokenClassification(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class GPTBigCodeModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class GPTBigCodePreTrainedModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
GPT_NEO_PRETRAINED_MODEL_ARCHIVE_LIST = None
|
||||
|
||||
|
||||
|
||||
0
tests/models/gpt_bigcode/__init__.py
Normal file
0
tests/models/gpt_bigcode/__init__.py
Normal file
617
tests/models/gpt_bigcode/test_modeling_gpt_bigcode.py
Normal file
617
tests/models/gpt_bigcode/test_modeling_gpt_bigcode.py
Normal file
@ -0,0 +1,617 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2023 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
import math
|
||||
import unittest
|
||||
|
||||
from parameterized import parameterized
|
||||
|
||||
from transformers import GPTBigCodeConfig, is_torch_available
|
||||
from transformers.testing_utils import require_torch, slow, torch_device
|
||||
|
||||
from ...generation.test_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, floats_tensor, ids_tensor, random_attention_mask
|
||||
from ...test_pipeline_mixin import PipelineTesterMixin
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import (
|
||||
GPT2TokenizerFast,
|
||||
GPTBigCodeForCausalLM,
|
||||
GPTBigCodeForSequenceClassification,
|
||||
GPTBigCodeForTokenClassification,
|
||||
GPTBigCodeModel,
|
||||
)
|
||||
from transformers.models.gpt_bigcode.modeling_gpt_bigcode import GPTBigCodeAttention
|
||||
|
||||
|
||||
class GPTBigCodeModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=14,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_token_type_ids=True,
|
||||
use_input_mask=True,
|
||||
use_labels=True,
|
||||
use_mc_token_ids=True,
|
||||
vocab_size=99,
|
||||
hidden_size=32,
|
||||
num_hidden_layers=5,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=37,
|
||||
hidden_act="relu",
|
||||
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,
|
||||
multi_query=True,
|
||||
scope=None,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_token_type_ids = use_token_type_ids
|
||||
self.use_input_mask = use_input_mask
|
||||
self.use_labels = use_labels
|
||||
self.use_mc_token_ids = use_mc_token_ids
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.intermediate_size = intermediate_size
|
||||
self.hidden_act = hidden_act
|
||||
self.hidden_dropout_prob = hidden_dropout_prob
|
||||
self.attention_probs_dropout_prob = attention_probs_dropout_prob
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.type_vocab_size = type_vocab_size
|
||||
self.type_sequence_label_size = type_sequence_label_size
|
||||
self.initializer_range = initializer_range
|
||||
self.num_labels = num_labels
|
||||
self.num_choices = num_choices
|
||||
self.scope = None
|
||||
self.bos_token_id = vocab_size - 1
|
||||
self.eos_token_id = vocab_size - 2
|
||||
self.pad_token_id = vocab_size - 3
|
||||
self.multi_query = multi_query
|
||||
|
||||
def get_large_model_config(self):
|
||||
return GPTBigCodeConfig.from_pretrained("bigcode/gpt_bigcode-santacoder")
|
||||
|
||||
def prepare_config_and_inputs(
|
||||
self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
|
||||
):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
input_mask = None
|
||||
if self.use_input_mask:
|
||||
input_mask = random_attention_mask([self.batch_size, self.seq_length])
|
||||
|
||||
token_type_ids = None
|
||||
if self.use_token_type_ids:
|
||||
token_type_ids = ids_tensor([self.batch_size, self.seq_length], self.type_vocab_size)
|
||||
|
||||
mc_token_ids = None
|
||||
if self.use_mc_token_ids:
|
||||
mc_token_ids = ids_tensor([self.batch_size, self.num_choices], 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(
|
||||
gradient_checkpointing=gradient_checkpointing,
|
||||
scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
|
||||
reorder_and_upcast_attn=reorder_and_upcast_attn,
|
||||
)
|
||||
|
||||
head_mask = ids_tensor([self.num_hidden_layers, self.num_attention_heads], 2)
|
||||
|
||||
return (
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
head_mask,
|
||||
token_type_ids,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
)
|
||||
|
||||
def get_config(
|
||||
self, gradient_checkpointing=False, scale_attn_by_inverse_layer_idx=False, reorder_and_upcast_attn=False
|
||||
):
|
||||
return GPTBigCodeConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
n_embd=self.hidden_size,
|
||||
n_layer=self.num_hidden_layers,
|
||||
n_head=self.num_attention_heads,
|
||||
n_inner=self.intermediate_size,
|
||||
activation_function=self.hidden_act,
|
||||
resid_pdrop=self.hidden_dropout_prob,
|
||||
attn_pdrop=self.attention_probs_dropout_prob,
|
||||
n_positions=self.max_position_embeddings,
|
||||
type_vocab_size=self.type_vocab_size,
|
||||
initializer_range=self.initializer_range,
|
||||
use_cache=True,
|
||||
bos_token_id=self.bos_token_id,
|
||||
eos_token_id=self.eos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
gradient_checkpointing=gradient_checkpointing,
|
||||
scale_attn_by_inverse_layer_idx=scale_attn_by_inverse_layer_idx,
|
||||
reorder_and_upcast_attn=reorder_and_upcast_attn,
|
||||
attention_softmax_in_fp32=False,
|
||||
scale_attention_softmax_in_fp32=False,
|
||||
multi_query=self.multi_query,
|
||||
)
|
||||
|
||||
def get_pipeline_config(self):
|
||||
config = self.get_config()
|
||||
config.vocab_size = 300
|
||||
return config
|
||||
|
||||
def prepare_config_and_inputs_for_decoder(self):
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
head_mask,
|
||||
token_type_ids,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = self.prepare_config_and_inputs()
|
||||
|
||||
encoder_hidden_states = floats_tensor([self.batch_size, self.seq_length, self.hidden_size])
|
||||
encoder_attention_mask = ids_tensor([self.batch_size, self.seq_length], vocab_size=2)
|
||||
|
||||
return (
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
head_mask,
|
||||
token_type_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
encoder_hidden_states,
|
||||
encoder_attention_mask,
|
||||
)
|
||||
|
||||
def create_and_check_gpt_bigcode_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
|
||||
model = GPTBigCodeModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
result = model(input_ids, token_type_ids=token_type_ids, head_mask=head_mask)
|
||||
result = model(input_ids, token_type_ids=token_type_ids)
|
||||
result = model(input_ids)
|
||||
|
||||
self.parent.assertEqual(result.last_hidden_state.shape, (self.batch_size, self.seq_length, self.hidden_size))
|
||||
self.parent.assertEqual(len(result.past_key_values), config.n_layer)
|
||||
|
||||
def create_and_check_gpt_bigcode_model_past(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
|
||||
model = GPTBigCodeModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, token_type_ids=token_type_ids, use_cache=True)
|
||||
outputs_use_cache_conf = model(input_ids, token_type_ids=token_type_ids)
|
||||
outputs_no_past = model(input_ids, token_type_ids=token_type_ids, use_cache=False)
|
||||
|
||||
self.parent.assertTrue(len(outputs) == len(outputs_use_cache_conf))
|
||||
self.parent.assertTrue(len(outputs) == len(outputs_no_past) + 1)
|
||||
|
||||
output, past = outputs.to_tuple()
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
||||
next_token_types = ids_tensor([self.batch_size, 1], self.type_vocab_size)
|
||||
|
||||
# append to next input_ids and token_type_ids
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, token_type_ids=next_token_type_ids)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, token_type_ids=next_token_types, past_key_values=past)[
|
||||
"last_hidden_state"
|
||||
]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# 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_gpt_bigcode_model_attention_mask_past(
|
||||
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
|
||||
):
|
||||
model = GPTBigCodeModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
# create attention mask
|
||||
attn_mask = torch.ones(input_ids.shape, dtype=torch.long, device=torch_device)
|
||||
half_seq_length = self.seq_length // 2
|
||||
attn_mask[:, half_seq_length:] = 0
|
||||
|
||||
# first forward pass
|
||||
output, past = model(input_ids, attention_mask=attn_mask).to_tuple()
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size)
|
||||
|
||||
# change a random masked slice from input_ids
|
||||
random_seq_idx_to_change = ids_tensor((1,), half_seq_length).item() + 1
|
||||
random_other_next_tokens = ids_tensor((self.batch_size, 1), config.vocab_size).squeeze(-1)
|
||||
input_ids[:, -random_seq_idx_to_change] = random_other_next_tokens
|
||||
|
||||
# append to next input_ids and attn_mask
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
attn_mask = torch.cat(
|
||||
[attn_mask, torch.ones((attn_mask.shape[0], 1), dtype=torch.long, device=torch_device)],
|
||||
dim=1,
|
||||
)
|
||||
|
||||
# get two different outputs
|
||||
output_from_no_past = model(next_input_ids, attention_mask=attn_mask)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, past_key_values=past, attention_mask=attn_mask)["last_hidden_state"]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = ids_tensor((1,), output_from_past.shape[-1]).item()
|
||||
output_from_no_past_slice = output_from_no_past[:, -1, random_slice_idx].detach()
|
||||
output_from_past_slice = output_from_past[:, 0, random_slice_idx].detach()
|
||||
|
||||
# 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_gpt_bigcode_model_past_large_inputs(
|
||||
self, config, input_ids, input_mask, head_mask, token_type_ids, *args
|
||||
):
|
||||
model = GPTBigCodeModel(config=config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, token_type_ids=token_type_ids, attention_mask=input_mask, use_cache=True)
|
||||
|
||||
output, past = outputs.to_tuple()
|
||||
|
||||
# create hypothetical next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_token_types = ids_tensor([self.batch_size, 3], self.type_vocab_size)
|
||||
next_mask = ids_tensor((self.batch_size, 3), vocab_size=2)
|
||||
|
||||
# append to next input_ids and token_type_ids
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
next_token_type_ids = torch.cat([token_type_ids, next_token_types], dim=-1)
|
||||
next_attention_mask = torch.cat([input_mask, next_mask], dim=-1)
|
||||
|
||||
output_from_no_past = model(
|
||||
next_input_ids, token_type_ids=next_token_type_ids, attention_mask=next_attention_mask
|
||||
)["last_hidden_state"]
|
||||
output_from_past = model(
|
||||
next_tokens, token_type_ids=next_token_types, attention_mask=next_attention_mask, past_key_values=past
|
||||
)["last_hidden_state"]
|
||||
self.parent.assertTrue(output_from_past.shape[1] == next_tokens.shape[1])
|
||||
|
||||
# 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()
|
||||
|
||||
# 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_lm_head_model(self, config, input_ids, input_mask, head_mask, token_type_ids, *args):
|
||||
model = GPTBigCodeForCausalLM(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
|
||||
self.parent.assertEqual(result.loss.shape, ())
|
||||
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
||||
|
||||
def create_and_check_forward_and_backwards(
|
||||
self, config, input_ids, input_mask, head_mask, token_type_ids, *args, gradient_checkpointing=False
|
||||
):
|
||||
model = GPTBigCodeForCausalLM(config)
|
||||
model.to(torch_device)
|
||||
if gradient_checkpointing:
|
||||
model.gradient_checkpointing_enable()
|
||||
|
||||
result = model(input_ids, token_type_ids=token_type_ids, labels=input_ids)
|
||||
self.parent.assertEqual(result.loss.shape, ())
|
||||
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
|
||||
result.loss.backward()
|
||||
|
||||
def create_and_check_gpt_bigcode_for_sequence_classification(
|
||||
self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
|
||||
):
|
||||
config.num_labels = self.num_labels
|
||||
model = GPTBigCodeForSequenceClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids, labels=sequence_labels)
|
||||
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.num_labels))
|
||||
|
||||
def create_and_check_gpt_bigcode_for_token_classification(
|
||||
self, config, input_ids, input_mask, head_mask, token_type_ids, mc_token_ids, sequence_labels, *args
|
||||
):
|
||||
config.num_labels = self.num_labels
|
||||
model = GPTBigCodeForTokenClassification(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
result = model(input_ids, attention_mask=input_mask, token_type_ids=token_type_ids)
|
||||
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.num_labels))
|
||||
|
||||
def create_and_check_gpt_bigcode_weight_initialization(self, config, *args):
|
||||
model = GPTBigCodeModel(config)
|
||||
model_std = model.config.initializer_range / math.sqrt(2 * model.config.n_layer)
|
||||
for key in model.state_dict().keys():
|
||||
if "c_proj" in key and "weight" in key:
|
||||
self.parent.assertLessEqual(abs(torch.std(model.state_dict()[key]) - model_std), 0.001)
|
||||
self.parent.assertLessEqual(abs(torch.mean(model.state_dict()[key]) - 0.0), 0.01)
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config_and_inputs = self.prepare_config_and_inputs()
|
||||
|
||||
(
|
||||
config,
|
||||
input_ids,
|
||||
input_mask,
|
||||
head_mask,
|
||||
token_type_ids,
|
||||
mc_token_ids,
|
||||
sequence_labels,
|
||||
token_labels,
|
||||
choice_labels,
|
||||
) = config_and_inputs
|
||||
|
||||
inputs_dict = {
|
||||
"input_ids": input_ids,
|
||||
"token_type_ids": token_type_ids,
|
||||
"head_mask": head_mask,
|
||||
}
|
||||
|
||||
return config, inputs_dict
|
||||
|
||||
|
||||
@require_torch
|
||||
class GPTBigCodeMQAModelTest(ModelTesterMixin, GenerationTesterMixin, PipelineTesterMixin, unittest.TestCase):
|
||||
# TODO: Update the tests to use valid pretrained models.
|
||||
all_model_classes = (
|
||||
(
|
||||
GPTBigCodeModel,
|
||||
GPTBigCodeForCausalLM,
|
||||
GPTBigCodeForSequenceClassification,
|
||||
GPTBigCodeForTokenClassification,
|
||||
)
|
||||
if is_torch_available()
|
||||
else ()
|
||||
)
|
||||
all_generative_model_classes = (GPTBigCodeForCausalLM,) if is_torch_available() else ()
|
||||
fx_compatible = False
|
||||
test_missing_keys = False
|
||||
test_pruning = False
|
||||
test_torchscript = False
|
||||
multi_query = True
|
||||
pipeline_model_mapping = (
|
||||
{
|
||||
"feature-extraction": GPTBigCodeModel,
|
||||
"text-classification": GPTBigCodeForSequenceClassification,
|
||||
"text-generation": GPTBigCodeForCausalLM,
|
||||
"token-classification": GPTBigCodeForTokenClassification,
|
||||
"zero-shot": GPTBigCodeForSequenceClassification,
|
||||
}
|
||||
if is_torch_available()
|
||||
else {}
|
||||
)
|
||||
|
||||
# special case for DoubleHeads model
|
||||
def _prepare_for_class(self, inputs_dict, model_class, return_labels=False):
|
||||
inputs_dict = super()._prepare_for_class(inputs_dict, model_class, return_labels=return_labels)
|
||||
|
||||
return inputs_dict
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = GPTBigCodeModelTester(self, multi_query=self.multi_query)
|
||||
self.config_tester = ConfigTester(self, config_class=GPTBigCodeConfig, n_embd=37)
|
||||
|
||||
def tearDown(self):
|
||||
import gc
|
||||
|
||||
gc.collect()
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
@unittest.skip("MQA models does not support retain_grad")
|
||||
def test_retain_grad_hidden_states_attentions(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Contrastive search not supported due to non-standard caching mechanism")
|
||||
def test_contrastive_generate(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Contrastive search not supported due to non-standard caching mechanism")
|
||||
def test_contrastive_generate_dict_outputs_use_cache(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("CPU offload seems to be broken for some reason - tiny models keep hitting corner cases")
|
||||
def test_cpu_offload(self):
|
||||
pass
|
||||
|
||||
@unittest.skip("Disk offload seems to be broken for some reason - tiny models keep hitting corner cases")
|
||||
def test_disk_offload(self):
|
||||
pass
|
||||
|
||||
def test_gpt_bigcode_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_model(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_model_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_model_past(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_model_att_mask_past(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_model_attention_mask_past(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_model_past_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_lm_head_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_lm_head_model(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_sequence_classification_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_for_sequence_classification(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_token_classification_model(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_for_token_classification(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_gradient_checkpointing(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs, gradient_checkpointing=True)
|
||||
|
||||
def test_gpt_bigcode_scale_attn_by_inverse_layer_idx(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs(scale_attn_by_inverse_layer_idx=True)
|
||||
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_reorder_and_upcast_attn(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs(reorder_and_upcast_attn=True)
|
||||
self.model_tester.create_and_check_forward_and_backwards(*config_and_inputs)
|
||||
|
||||
def test_gpt_bigcode_weight_initialization(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_gpt_bigcode_weight_initialization(*config_and_inputs)
|
||||
|
||||
|
||||
@require_torch
|
||||
class GPTBigCodeMHAModelTest(GPTBigCodeMQAModelTest):
|
||||
# `parameterized_class` breaks with mixins, so we use inheritance instead
|
||||
multi_query = False
|
||||
|
||||
|
||||
@slow
|
||||
@require_torch
|
||||
class GPTBigCodeModelLanguageGenerationTest(unittest.TestCase):
|
||||
def test_generate_simple(self):
|
||||
model = GPTBigCodeForCausalLM.from_pretrained("bigcode/gpt_bigcode-santacoder").to(torch_device)
|
||||
tokenizer = GPT2TokenizerFast.from_pretrained("bigcode/gpt_bigcode-santacoder")
|
||||
|
||||
input_ids = tokenizer("def print_hello_world():", return_tensors="pt").input_ids.to(torch_device)
|
||||
|
||||
output_sequence = model.generate(input_ids)
|
||||
output_sentence = tokenizer.decode(output_sequence[0], skip_special_tokens=True)
|
||||
|
||||
expected_output = """def print_hello_world():\n print("Hello World!")\n\n\ndef print_hello_"""
|
||||
self.assertEqual(output_sentence, expected_output)
|
||||
|
||||
def test_generate_batched(self):
|
||||
tokenizer = GPT2TokenizerFast.from_pretrained("bigcode/gpt_bigcode-santacoder")
|
||||
tokenizer.pad_token = tokenizer.eos_token
|
||||
tokenizer.padding_side = "left"
|
||||
|
||||
model = GPTBigCodeForCausalLM.from_pretrained("bigcode/gpt_bigcode-santacoder").to(torch_device)
|
||||
|
||||
inputs = tokenizer(["def print_hello_world():", "def say_hello():"], return_tensors="pt", padding=True).to(
|
||||
torch_device
|
||||
)
|
||||
outputs = model.generate(**inputs)
|
||||
outputs = tokenizer.batch_decode(outputs, skip_special_tokens=True)
|
||||
|
||||
expected_output = [
|
||||
'def print_hello_world():\n print("Hello World!")\n\n\ndef print_hello_',
|
||||
'def say_hello():\n print("Hello, World!")\n\n\nsay_hello()',
|
||||
]
|
||||
self.assertListEqual(outputs, expected_output)
|
||||
|
||||
|
||||
@require_torch
|
||||
class GPTBigCodeMQATest(unittest.TestCase):
|
||||
def get_attention(self, multi_query):
|
||||
config = GPTBigCodeConfig.from_pretrained(
|
||||
"bigcode/gpt_bigcode-santacoder",
|
||||
multi_query=multi_query,
|
||||
attn_pdrop=0,
|
||||
resid_pdrop=0,
|
||||
)
|
||||
return GPTBigCodeAttention(config)
|
||||
|
||||
@parameterized.expand([(seed, is_train_mode) for seed in range(5) for is_train_mode in [True, False]])
|
||||
def test_mqa_reduces_to_mha(self, seed, is_train_mode=True):
|
||||
torch.manual_seed(seed)
|
||||
|
||||
# CREATE MQA AND MHA ATTENTIONS
|
||||
attention_mqa = self.get_attention(True)
|
||||
attention_mha = self.get_attention(False)
|
||||
|
||||
# ENFORCE MATCHING WEIGHTS
|
||||
num_heads = attention_mqa.num_heads
|
||||
embed_dim = attention_mqa.embed_dim
|
||||
head_dim = attention_mqa.head_dim
|
||||
|
||||
with torch.no_grad():
|
||||
mqa_q_weight = attention_mqa.c_attn.weight[:embed_dim, :].view(num_heads, 1, head_dim, embed_dim)
|
||||
mqa_kv_weight = attention_mqa.c_attn.weight[embed_dim:, :].view(1, 2, head_dim, embed_dim)
|
||||
mha_c_weight = torch.cat(
|
||||
[mqa_q_weight, mqa_kv_weight.expand(num_heads, 2, head_dim, embed_dim)], dim=1
|
||||
).view(3 * num_heads * head_dim, embed_dim)
|
||||
|
||||
mqa_q_bias = attention_mqa.c_attn.bias[:embed_dim].view(num_heads, 1, head_dim)
|
||||
mqa_kv_bias = attention_mqa.c_attn.bias[embed_dim:].view(1, 2, head_dim)
|
||||
mha_c_bias = torch.cat([mqa_q_bias, mqa_kv_bias.expand(num_heads, 2, head_dim)], dim=1).view(
|
||||
3 * num_heads * head_dim
|
||||
)
|
||||
|
||||
attention_mha.c_attn.weight.copy_(mha_c_weight)
|
||||
attention_mha.c_attn.bias.copy_(mha_c_bias)
|
||||
attention_mha.c_proj.weight.copy_(attention_mqa.c_proj.weight)
|
||||
attention_mha.c_proj.bias.copy_(attention_mqa.c_proj.bias)
|
||||
|
||||
# PUT THE MODEL INTO THE CORRECT MODE
|
||||
attention_mha.train(is_train_mode)
|
||||
attention_mqa.train(is_train_mode)
|
||||
|
||||
# RUN AN INPUT THROUGH THE MODELS
|
||||
num_tokens = 5
|
||||
hidden_states = torch.randn(1, num_tokens, embed_dim)
|
||||
attention_mha_result = attention_mha(hidden_states)[0]
|
||||
attention_mqa_result = attention_mqa(hidden_states)[0]
|
||||
|
||||
# CHECK THAT ALL OUTPUTS ARE THE SAME
|
||||
self.assertTrue(torch.allclose(attention_mha_result, attention_mqa_result, atol=1e-5))
|
||||
Loading…
Reference in New Issue
Block a user