Add OPT (#17088)
* First version - OPT model * Final changes - putting use cache to False * few changes - remove commented block * few changes - remove unecessary files * fix style issues * few changes - remove a test file - added the logits test * Update src/transformers/models/auto/tokenization_auto.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * add gen tests * few changes - rm mask filling example on docstring * few changes - remove useless args * some changes - more tests should pass now - needs to clean more - documentation still needs to be done * fix code quality * major changes - change attention architecture to BART-like - modify some tests - style fix * rm useless classes - remove opt for: - QA - cond generation - seq classif * Removed autodoc calls to non-existant classes TOkenizers are not implemented * Update src/transformers/__init__.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/__init__.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Update src/transformers/models/auto/modeling_tf_auto.py Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> * Replaced OPTTokeniser with GPT2 tokenizer * added GPT2Tokenizer.from_pretrained("patrickvonplaten/opt_gpt2_tokenizer") * Removed OPTTokenizer * make style * Make style replaces ``` ...).unsqueeze(``` by ``` >>>).unsqueeze(``` * make repo consistency * Removed PretrainedOPTModel * fix opt.mdx removed other heads * fix init, removed 3 heads * removed heads * finished cleaning head * removed seauence classif and question answering * removed unused imports * removed useless dummy object for QA, SC and CG * removed tests for removed useless dummy object for QA, SC and CG * Removed head_mask using encoder layers which don't exist * fixed test * fix line * added OPT to toctree * Updated model path with pushed weigths * fix model path * fixed code quality * fixed embeddings and generation tests * update paths * clean comments * removed OPTClassificationHead for sentence classification * renamed hidden layer * renamed num layers to standard num_hidden_layers * num_attention_heads fix * changes for 125m * add first version for 125m * add first version - flax * add new version * causal LM output * replace output type with BaseModelOutputWithPastAndCrossAttentions * revert working config from 150m to 350m * clean * removed decoder input ids * fixed embed dim * more embed_dim issues * make style + removed enc_dec test * update falx model * removed troublesome copy * added is_encoder_decoder=False to config * added set_input emb fuinction to model class * requires torch on embed test * use head mask instead of decoder head mask input param solves a test * 8 test remaining, update * Updated create_and_check_decoder_model_past_large_inputs * Make style * update op tokenizer with condition * make style * See if I can push * some clean up * remove linear head hack * save intermediate * save correct attention * add copied from from bart * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * fix part of the reviewss Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * same changes in naming / conversion * correct mask * more fixes * delete FlaxOPT and TfOPT * clean traces of Flax and Tf * fix mask * fixed positionnal embedding length when past key value is provoded * get 125m, 6.7b to work * Added do_layer_norm * solved mismatch in load dictionnary * clean up preapre opt input dict * fixed past key value as bool * fix previus * fixed return dict False tuple issue * All tests are passing * Make style * Ignore OPTDecoder non tested * make fix-copies * make repo consistency * small fix * removed uselss @torch.no_grad decorator * make styl;e * fix previous opt test * style * make style * added opt documentation * update OPT_PRETRAINED_MODEL_ARCHIVE_LIST * up * more fixes * model & config work * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * added comment on padding hack (+2) * cleaup * review update * docstring for missing arg * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * Update src/transformers/models/opt/__init__.py Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> * update pretrained map * update path and tests * make style * styling * make consistency * add gpt2 tok new * more tok fixes * Update src/transformers/models/auto/tokenization_auto.py * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update docs/source/en/model_doc/opt.mdx Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update tests/models/opt/test_modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update src/transformers/models/opt/modeling_opt.py Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> * Update based on reviews * Apply suggestions from code review Co-authored-by: Lysandre Debut <lysandre@huggingface.co> * make style * make tokenizer auto tests pass * apply Lysandre suggestion * finish tests * add some good tokenizer tests * improve docs slighly Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com> Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com> Co-authored-by: ArthurZucker <arthur.zucker@gmail.com> Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com> Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
This commit is contained in:
Younes Belkada
@@ -294,6 +294,7 @@ Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih.
|
||||
1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
|
||||
1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
|
||||
1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
|
||||
1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
|
||||
1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
|
||||
1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
|
||||
1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
|
||||
|
||||
@@ -273,6 +273,7 @@ Flax, PyTorch, TensorFlow 설치 페이지에서 이들을 conda로 설치하는
|
||||
1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
|
||||
1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
|
||||
1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
|
||||
1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
|
||||
1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
|
||||
1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
|
||||
1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
|
||||
|
||||
@@ -297,6 +297,7 @@ conda install -c huggingface transformers
|
||||
1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (来自 Microsoft Research) 伴随论文 [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) 由 Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu 发布。
|
||||
1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (来自 Google AI) 伴随论文 [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) 由 Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel 发布。
|
||||
1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (来自 the University of Wisconsin - Madison) 伴随论文 [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) 由 Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh 发布。
|
||||
1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (来自 Meta AI) 伴随论文 [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) 由 Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al 发布。
|
||||
1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (来自 Google) 伴随论文 [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) 由 Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu 发布。
|
||||
1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (来自 Deepmind) 伴随论文 [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) 由 Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira 发布。
|
||||
1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (来自 VinAI Research) 伴随论文 [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) 由 Dat Quoc Nguyen and Anh Tuan Nguyen 发布。
|
||||
|
||||
@@ -309,6 +309,7 @@ conda install -c huggingface transformers
|
||||
1. **[MPNet](https://huggingface.co/docs/transformers/model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
|
||||
1. **[MT5](https://huggingface.co/docs/transformers/model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
|
||||
1. **[Nyströmformer](https://huggingface.co/docs/transformers/model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
|
||||
1. **[OPT](https://huggingface.co/docs/transformers/master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
|
||||
1. **[Pegasus](https://huggingface.co/docs/transformers/model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
|
||||
1. **[Perceiver IO](https://huggingface.co/docs/transformers/model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
|
||||
1. **[PhoBERT](https://huggingface.co/docs/transformers/model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
|
||||
|
||||
@@ -270,6 +270,8 @@
|
||||
title: Nyströmformer
|
||||
- local: model_doc/openai-gpt
|
||||
title: OpenAI GPT
|
||||
- local: model_doc/opt
|
||||
title: OPT
|
||||
- local: model_doc/gpt2
|
||||
title: OpenAI GPT2
|
||||
- local: model_doc/gptj
|
||||
|
||||
@@ -115,6 +115,7 @@ The library currently contains JAX, PyTorch and TensorFlow implementations, pret
|
||||
1. **[MPNet](model_doc/mpnet)** (from Microsoft Research) released with the paper [MPNet: Masked and Permuted Pre-training for Language Understanding](https://arxiv.org/abs/2004.09297) by Kaitao Song, Xu Tan, Tao Qin, Jianfeng Lu, Tie-Yan Liu.
|
||||
1. **[MT5](model_doc/mt5)** (from Google AI) released with the paper [mT5: A massively multilingual pre-trained text-to-text transformer](https://arxiv.org/abs/2010.11934) by Linting Xue, Noah Constant, Adam Roberts, Mihir Kale, Rami Al-Rfou, Aditya Siddhant, Aditya Barua, Colin Raffel.
|
||||
1. **[Nyströmformer](model_doc/nystromformer)** (from the University of Wisconsin - Madison) released with the paper [Nyströmformer: A Nyström-Based Algorithm for Approximating Self-Attention](https://arxiv.org/abs/2102.03902) by Yunyang Xiong, Zhanpeng Zeng, Rudrasis Chakraborty, Mingxing Tan, Glenn Fung, Yin Li, Vikas Singh.
|
||||
1. **[OPT](master/model_doc/opt)** (from Meta AI) released with the paper [OPT: Open Pre-trained Transformer Language Models](https://arxiv.org/abs/2205.01068) by Susan Zhang, Stephen Roller, Naman Goyal, Mikel Artetxe, Moya Chen, Shuohui Chen et al.
|
||||
1. **[Pegasus](model_doc/pegasus)** (from Google) released with the paper [PEGASUS: Pre-training with Extracted Gap-sentences for Abstractive Summarization](https://arxiv.org/abs/1912.08777) by Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu.
|
||||
1. **[Perceiver IO](model_doc/perceiver)** (from Deepmind) released with the paper [Perceiver IO: A General Architecture for Structured Inputs & Outputs](https://arxiv.org/abs/2107.14795) by Andrew Jaegle, Sebastian Borgeaud, Jean-Baptiste Alayrac, Carl Doersch, Catalin Ionescu, David Ding, Skanda Koppula, Daniel Zoran, Andrew Brock, Evan Shelhamer, Olivier Hénaff, Matthew M. Botvinick, Andrew Zisserman, Oriol Vinyals, João Carreira.
|
||||
1. **[PhoBERT](model_doc/phobert)** (from VinAI Research) released with the paper [PhoBERT: Pre-trained language models for Vietnamese](https://www.aclweb.org/anthology/2020.findings-emnlp.92/) by Dat Quoc Nguyen and Anh Tuan Nguyen.
|
||||
@@ -231,6 +232,7 @@ Flax), PyTorch, and/or TensorFlow.
|
||||
| Nystromformer | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
| OpenAI GPT | ✅ | ✅ | ✅ | ✅ | ❌ |
|
||||
| OpenAI GPT-2 | ✅ | ✅ | ✅ | ✅ | ✅ |
|
||||
| OPT | ❌ | ❌ | ✅ | ❌ | ❌ |
|
||||
| Pegasus | ✅ | ✅ | ✅ | ✅ | ✅ |
|
||||
| Perceiver | ✅ | ❌ | ✅ | ❌ | ❌ |
|
||||
| PLBart | ✅ | ❌ | ✅ | ❌ | ❌ |
|
||||
|
||||
47
docs/source/en/model_doc/opt.mdx
Normal file
47
docs/source/en/model_doc/opt.mdx
Normal file
@@ -0,0 +1,47 @@
|
||||
<!--Copyright 2022 The HuggingFace Team. All rights reserved.
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with
|
||||
the License. You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on
|
||||
an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the
|
||||
specific language governing permissions and limitations under the License.
|
||||
-->
|
||||
|
||||
# OPT
|
||||
|
||||
## Overview
|
||||
|
||||
The OPT model was proposed in [Open Pre-trained Transformer Language Models](https://arxiv.org/pdf/2205.01068) by Meta AI.
|
||||
OPT is a series of open-sourced large causal language models which perform similar in performance to GPT3.
|
||||
|
||||
|
||||
The abstract from the paper is the following:
|
||||
|
||||
*Large language models, which are often trained for hundreds of thousands of compute days, have shown remarkable capabilities for zero- and few-shot learning. Given their computational cost, these models are difficult to replicate without significant capital. For the few that are available through APIs, no access is granted to the full model weights, making them difficult to study. We present Open Pre-trained Transformers (OPT), a suite of decoder-only pre-trained transformers ranging from 125M to 175B parameters, which we aim to fully and responsibly share with interested researchers. We show that OPT-175B is comparable to GPT-3, while requiring only 1/7th the carbon footprint to develop. We are also releasing our logbook detailing the infrastructure challenges we faced, along with code for experimenting with all of the released models.*
|
||||
|
||||
Tips:
|
||||
- OPT has the same architecture as [`BartDecoder`].
|
||||
- Contrary to GPT2, OPT adds the EOS token `</s>` to the beginning of every prompt. **Note**: Make sure to pass `use_fast=False` when loading OPT's tokenizer with [`AutoTokenizer`] to get the correct tokenizer.
|
||||
|
||||
This model was contributed by [Arthur Zucker](https://huggingface.co/ArthurZ), [Younes Belkada](https://huggingface.co/ybelkada), and [Patrick Von Platen](https://huggingface.co/patrickvonplaten).
|
||||
The original code can be found [here](https://github.com/facebookresearch/metaseq).
|
||||
|
||||
|
||||
## OPTConfig
|
||||
|
||||
[[autodoc]] OPTConfig
|
||||
|
||||
## OPTModel
|
||||
|
||||
[[autodoc]] OPTModel
|
||||
- forward
|
||||
|
||||
|
||||
## OPTForCausalLM
|
||||
|
||||
[[autodoc]] OPTForCausalLM
|
||||
- forward
|
||||
|
||||
@@ -247,6 +247,7 @@ _import_structure = {
|
||||
"NystromformerConfig",
|
||||
],
|
||||
"models.openai": ["OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "OpenAIGPTConfig", "OpenAIGPTTokenizer"],
|
||||
"models.opt": ["OPTConfig"],
|
||||
"models.pegasus": ["PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP", "PegasusConfig", "PegasusTokenizer"],
|
||||
"models.perceiver": ["PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP", "PerceiverConfig", "PerceiverTokenizer"],
|
||||
"models.phobert": ["PhobertTokenizer"],
|
||||
@@ -1323,6 +1324,14 @@ else:
|
||||
"load_tf_weights_in_openai_gpt",
|
||||
]
|
||||
)
|
||||
_import_structure["models.opt"].extend(
|
||||
[
|
||||
"OPT_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"OPTForCausalLM",
|
||||
"OPTModel",
|
||||
"OPTPreTrainedModel",
|
||||
]
|
||||
)
|
||||
_import_structure["models.pegasus"].extend(
|
||||
["PegasusForCausalLM", "PegasusForConditionalGeneration", "PegasusModel", "PegasusPreTrainedModel"]
|
||||
)
|
||||
@@ -2373,7 +2382,6 @@ else:
|
||||
"FlaxBartPreTrainedModel",
|
||||
]
|
||||
)
|
||||
|
||||
_import_structure["models.beit"].extend(
|
||||
[
|
||||
"FlaxBeitForImageClassification",
|
||||
@@ -2382,6 +2390,7 @@ else:
|
||||
"FlaxBeitPreTrainedModel",
|
||||
]
|
||||
)
|
||||
|
||||
_import_structure["models.bert"].extend(
|
||||
[
|
||||
"FlaxBertForCausalLM",
|
||||
@@ -2718,6 +2727,7 @@ if TYPE_CHECKING:
|
||||
from .models.mt5 import MT5Config
|
||||
from .models.nystromformer import NYSTROMFORMER_PRETRAINED_CONFIG_ARCHIVE_MAP, NystromformerConfig
|
||||
from .models.openai import OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OpenAIGPTConfig, OpenAIGPTTokenizer
|
||||
from .models.opt import OPTConfig
|
||||
from .models.pegasus import PEGASUS_PRETRAINED_CONFIG_ARCHIVE_MAP, PegasusConfig, PegasusTokenizer
|
||||
from .models.perceiver import PERCEIVER_PRETRAINED_CONFIG_ARCHIVE_MAP, PerceiverConfig, PerceiverTokenizer
|
||||
from .models.phobert import PhobertTokenizer
|
||||
@@ -3630,6 +3640,7 @@ if TYPE_CHECKING:
|
||||
OpenAIGPTPreTrainedModel,
|
||||
load_tf_weights_in_openai_gpt,
|
||||
)
|
||||
from .models.opt import OPT_PRETRAINED_MODEL_ARCHIVE_LIST, OPTForCausalLM, OPTModel, OPTPreTrainedModel
|
||||
from .models.pegasus import (
|
||||
PegasusForCausalLM,
|
||||
PegasusForConditionalGeneration,
|
||||
|
||||
@@ -87,6 +87,7 @@ from . import (
|
||||
mt5,
|
||||
nystromformer,
|
||||
openai,
|
||||
opt,
|
||||
pegasus,
|
||||
perceiver,
|
||||
phobert,
|
||||
|
||||
@@ -98,6 +98,7 @@ CONFIG_MAPPING_NAMES = OrderedDict(
|
||||
("megatron-bert", "MegatronBertConfig"),
|
||||
("mpnet", "MPNetConfig"),
|
||||
("bart", "BartConfig"),
|
||||
("opt", "OPTConfig"),
|
||||
("blenderbot", "BlenderbotConfig"),
|
||||
("reformer", "ReformerConfig"),
|
||||
("longformer", "LongformerConfig"),
|
||||
@@ -190,6 +191,7 @@ CONFIG_ARCHIVE_MAP_MAPPING_NAMES = OrderedDict(
|
||||
("blenderbot-small", "BLENDERBOT_SMALL_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("bert", "BERT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("bart", "BART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("opt", "OPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("blenderbot", "BLENDERBOT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("mbart", "MBART_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
("openai-gpt", "OPENAI_GPT_PRETRAINED_CONFIG_ARCHIVE_MAP"),
|
||||
@@ -301,6 +303,7 @@ MODEL_NAMES_MAPPING = OrderedDict(
|
||||
("mbart", "mBART"),
|
||||
("megatron-bert", "MegatronBert"),
|
||||
("bart", "BART"),
|
||||
("opt", "OPT"),
|
||||
("reformer", "Reformer"),
|
||||
("longformer", "Longformer"),
|
||||
("roberta", "RoBERTa"),
|
||||
|
||||
@@ -93,6 +93,7 @@ MODEL_MAPPING_NAMES = OrderedDict(
|
||||
("xlm-roberta-xl", "XLMRobertaXLModel"),
|
||||
("xlm-roberta", "XLMRobertaModel"),
|
||||
("bart", "BartModel"),
|
||||
("opt", "OPTModel"),
|
||||
("longformer", "LongformerModel"),
|
||||
("roberta", "RobertaModel"),
|
||||
("data2vec-text", "Data2VecTextModel"),
|
||||
@@ -261,6 +262,7 @@ MODEL_FOR_CAUSAL_LM_MAPPING_NAMES = OrderedDict(
|
||||
("xlm-prophetnet", "XLMProphetNetForCausalLM"),
|
||||
("prophetnet", "ProphetNetForCausalLM"),
|
||||
("bart", "BartForCausalLM"),
|
||||
("opt", "OPTForCausalLM"),
|
||||
("mbart", "MBartForCausalLM"),
|
||||
("pegasus", "PegasusForCausalLM"),
|
||||
("marian", "MarianForCausalLM"),
|
||||
|
||||
@@ -137,6 +137,7 @@ else:
|
||||
("openai-gpt", ("OpenAIGPTTokenizer", "OpenAIGPTTokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gpt2", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
|
||||
("gptj", ("GPT2Tokenizer", "GPT2TokenizerFast" if is_tokenizers_available() else None)),
|
||||
("opt", ("GPT2Tokenizer", None)),
|
||||
("transfo-xl", ("TransfoXLTokenizer", None)),
|
||||
(
|
||||
"xlnet",
|
||||
|
||||
@@ -162,20 +162,26 @@ class GPT2Tokenizer(PreTrainedTokenizer):
|
||||
unk_token="<|endoftext|>",
|
||||
bos_token="<|endoftext|>",
|
||||
eos_token="<|endoftext|>",
|
||||
pad_token=None,
|
||||
add_prefix_space=False,
|
||||
add_bos_token=False,
|
||||
**kwargs
|
||||
):
|
||||
bos_token = AddedToken(bos_token, lstrip=False, rstrip=False) if isinstance(bos_token, str) else bos_token
|
||||
eos_token = AddedToken(eos_token, lstrip=False, rstrip=False) if isinstance(eos_token, str) else eos_token
|
||||
unk_token = AddedToken(unk_token, lstrip=False, rstrip=False) if isinstance(unk_token, str) else unk_token
|
||||
pad_token = AddedToken(pad_token, lstrip=False, rstrip=False) if isinstance(pad_token, str) else pad_token
|
||||
super().__init__(
|
||||
errors=errors,
|
||||
unk_token=unk_token,
|
||||
bos_token=bos_token,
|
||||
eos_token=eos_token,
|
||||
pad_token=pad_token,
|
||||
add_prefix_space=add_prefix_space,
|
||||
add_bos_token=add_bos_token,
|
||||
**kwargs,
|
||||
)
|
||||
self.add_bos_token = add_bos_token
|
||||
|
||||
with open(vocab_file, encoding="utf-8") as vocab_handle:
|
||||
self.encoder = json.load(vocab_handle)
|
||||
@@ -242,6 +248,19 @@ class GPT2Tokenizer(PreTrainedTokenizer):
|
||||
self.cache[token] = word
|
||||
return word
|
||||
|
||||
def build_inputs_with_special_tokens(self, token_ids_0, token_ids_1=None):
|
||||
if self.add_bos_token:
|
||||
bos_token_ids = [self.bos_token_id]
|
||||
else:
|
||||
bos_token_ids = []
|
||||
|
||||
output = bos_token_ids + token_ids_0
|
||||
|
||||
if token_ids_1 is None:
|
||||
return output
|
||||
|
||||
return output + bos_token_ids + token_ids_1
|
||||
|
||||
def _tokenize(self, text):
|
||||
"""Tokenize a string."""
|
||||
bpe_tokens = []
|
||||
|
||||
46
src/transformers/models/opt/__init__.py
Normal file
46
src/transformers/models/opt/__init__.py
Normal file
@@ -0,0 +1,46 @@
|
||||
# flake8: noqa
|
||||
# There's no way to ignore "F401 '...' imported but unused" warnings in this
|
||||
# module, but to preserve other warnings. So, don't check this module at all.
|
||||
|
||||
# Copyright 2022 The HuggingFace Team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
from typing import TYPE_CHECKING
|
||||
|
||||
from ...utils import _LazyModule, is_tokenizers_available, is_torch_available
|
||||
|
||||
|
||||
_import_structure = {
|
||||
"configuration_opt": ["OPT_PRETRAINED_CONFIG_ARCHIVE_MAP", "OPTConfig"],
|
||||
}
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
_import_structure["modeling_opt"] = [
|
||||
"OPT_PRETRAINED_MODEL_ARCHIVE_LIST",
|
||||
"OPTForCausalLM",
|
||||
"OPTModel",
|
||||
"OPTPreTrainedModel",
|
||||
]
|
||||
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from .configuration_opt import OPT_PRETRAINED_CONFIG_ARCHIVE_MAP, OPTConfig
|
||||
|
||||
if is_torch_available():
|
||||
from .modeling_opt import OPT_PRETRAINED_MODEL_ARCHIVE_LIST, OPTForCausalLM, OPTModel, OPTPreTrainedModel
|
||||
|
||||
else:
|
||||
import sys
|
||||
|
||||
sys.modules[__name__] = _LazyModule(__name__, globals()["__file__"], _import_structure, module_spec=__spec__)
|
||||
139
src/transformers/models/opt/configuration_opt.py
Normal file
139
src/transformers/models/opt/configuration_opt.py
Normal file
@@ -0,0 +1,139 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2022 The Metaseq Authors and The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
""" OPT model configuration"""
|
||||
from ...configuration_utils import PretrainedConfig
|
||||
from ...utils import logging
|
||||
|
||||
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
OPT_PRETRAINED_CONFIG_ARCHIVE_MAP = {
|
||||
"facebook/opt-125m": "https://huggingface.co/facebook/opt-125m/blob/main/config.json",
|
||||
"facebook/opt-350m": "https://huggingface.co/facebook/opt-350m/blob/main/config.json",
|
||||
"facebook/opt-1.3b": "https://huggingface.co/facebook/opt-1.3b/blob/main/config.json",
|
||||
"facebook/opt-2.7b": "https://huggingface.co/facebook/opt-2.7b/blob/main/config.json",
|
||||
"facebook/opt-6.7b": "https://huggingface.co/facebook/opt-6.7b/blob/main/config.json",
|
||||
"facebook/opt-13b": "https://huggingface.co/facebook/opt-13b/blob/main/config.json",
|
||||
}
|
||||
|
||||
|
||||
class OPTConfig(PretrainedConfig):
|
||||
r"""
|
||||
This is the configuration class to store the configuration of a [`OPTModel`]. It is used to instantiate a OPT 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 OPT
|
||||
[facebook/opt-350m](https://huggingface.co/facebook/opt-350m) 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 50272):
|
||||
Vocabulary size of the OPT model. Defines the number of different tokens that can be represented by the
|
||||
`inputs_ids` passed when calling [`OPTModel`]
|
||||
hidden_size (`int`, *optional*, defaults to 768):
|
||||
Dimensionality of the layers and the pooler layer.
|
||||
num_hidden_layers (`int`, *optional*, defaults to 12):
|
||||
Number of decoder layers.
|
||||
ffn_dim (`int`, *optional*, defaults to 3072):
|
||||
Dimensionality of the "intermediate" (often named feed-forward) layer in decoder.
|
||||
num_attention_heads (`int`, *optional*, defaults to 12):
|
||||
Number of attention heads for each attention layer in the Transformer decoder.
|
||||
activation_function (`str` or `function`, *optional*, defaults to `"relu"`):
|
||||
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
|
||||
`"relu"`, `"silu"` and `"gelu_new"` are supported.
|
||||
max_position_embeddings (`int`, *optional*, defaults to 2048):
|
||||
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).
|
||||
do_layer_norm_before (`bool`, *optional*, defaults to `True`):
|
||||
Whether to perform layer normalization before the attention block.
|
||||
word_embed_proj_dim (`int`, *optional*):
|
||||
`word_embed_proj_dim` can be set to down-project word embeddings, *e.g.* `opt-350m`. Defaults to
|
||||
`hidden_size`.
|
||||
dropout (`float`, *optional*, defaults to 0.1):
|
||||
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
|
||||
attention_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for the attention probabilities.
|
||||
activation_dropout (`float`, *optional*, defaults to 0.0):
|
||||
The dropout ratio for activations inside the fully connected layer.
|
||||
layerdrop: (`float`, *optional*, defaults to 0.0):
|
||||
The LayerDrop probability. See the [LayerDrop paper](see https://arxiv.org/abs/1909.11556) for more
|
||||
details.
|
||||
init_std (`float`, *optional*, defaults to 0.02):
|
||||
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
|
||||
use_cache (`bool`, *optional*, defaults to `True`):
|
||||
Whether or not the model should return the last key/values attentions (not used by all models).
|
||||
|
||||
Example:
|
||||
|
||||
```python
|
||||
>>> from transformers import OPTModel, OPTConfig
|
||||
|
||||
>>> # Initializing a OPT facebook/opt-large style configuration
|
||||
>>> configuration = OPTConfig()
|
||||
|
||||
>>> # Initializing a model from the facebook/opt-large style configuration
|
||||
>>> model = OPTModel(configuration)
|
||||
|
||||
>>> # Accessing the model configuration
|
||||
>>> configuration = model.config
|
||||
```"""
|
||||
model_type = "opt"
|
||||
keys_to_ignore_at_inference = ["past_key_values"]
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
vocab_size=50272,
|
||||
hidden_size=768,
|
||||
num_hidden_layers=12,
|
||||
ffn_dim=3072,
|
||||
max_position_embeddings=2048,
|
||||
do_layer_norm_before=True,
|
||||
word_embed_proj_dim=None,
|
||||
dropout=0.1,
|
||||
attention_dropout=0.0,
|
||||
activation_dropout=0.0,
|
||||
num_attention_heads=12,
|
||||
activation_function="relu",
|
||||
layerdrop=0.0,
|
||||
init_std=0.02,
|
||||
use_cache=True,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
eos_token_id=2,
|
||||
**kwargs
|
||||
):
|
||||
super().__init__(
|
||||
pad_token_id=pad_token_id,
|
||||
bos_token_id=bos_token_id,
|
||||
eos_token_id=eos_token_id,
|
||||
**kwargs,
|
||||
)
|
||||
self.vocab_size = vocab_size
|
||||
self.max_position_embeddings = max_position_embeddings
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.word_embed_proj_dim = word_embed_proj_dim if word_embed_proj_dim is not None else hidden_size
|
||||
self.ffn_dim = ffn_dim
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.dropout = dropout
|
||||
self.attention_dropout = attention_dropout
|
||||
self.activation_dropout = activation_dropout
|
||||
self.activation_function = activation_function
|
||||
self.init_std = init_std
|
||||
self.layerdrop = layerdrop
|
||||
self.use_cache = use_cache
|
||||
self.do_layer_norm_before = do_layer_norm_before
|
||||
@@ -0,0 +1,90 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2022 The 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.
|
||||
"""Convert OPT checkpoint."""
|
||||
|
||||
|
||||
import argparse
|
||||
from pathlib import Path
|
||||
|
||||
import torch
|
||||
|
||||
from transformers import OPTConfig, OPTModel
|
||||
from transformers.utils import logging
|
||||
|
||||
|
||||
logging.set_verbosity_info()
|
||||
logger = logging.get_logger(__name__)
|
||||
|
||||
|
||||
def load_checkpoint(checkpoint_path):
|
||||
"""Checkpoint path should end in model.pt"""
|
||||
sd = torch.load(checkpoint_path, map_location="cpu")
|
||||
if "model" in sd.keys():
|
||||
sd = torch.load(checkpoint_path, map_location="cpu")["model"]
|
||||
|
||||
# pop unnecessary weights
|
||||
keys_to_delete = [
|
||||
"decoder.version",
|
||||
"decoder.layer_norm.weight",
|
||||
"decoder.layer_norm.bias",
|
||||
"decoder.output_projection.weight",
|
||||
]
|
||||
for key in keys_to_delete:
|
||||
if key in sd:
|
||||
sd.pop(key)
|
||||
|
||||
keys_to_rename = {
|
||||
"decoder.project_in_dim.weight": "decoder.project_in.weight",
|
||||
"decoder.project_out_dim.weight": "decoder.project_out.weight",
|
||||
}
|
||||
for old_key, new_key in keys_to_rename.items():
|
||||
if old_key in sd:
|
||||
sd[new_key] = sd.pop(old_key)
|
||||
|
||||
return sd
|
||||
|
||||
|
||||
@torch.no_grad()
|
||||
def convert_opt_checkpoint(checkpoint_path, pytorch_dump_folder_path, config=None):
|
||||
"""
|
||||
Copy/paste/tweak model's weights to our BERT structure.
|
||||
"""
|
||||
state_dict = load_checkpoint(checkpoint_path)
|
||||
|
||||
if config is not None:
|
||||
config = OPTConfig.from_pretrained(config)
|
||||
else:
|
||||
config = OPTConfig()
|
||||
|
||||
model = OPTModel(config).half().eval()
|
||||
model.load_state_dict(state_dict)
|
||||
|
||||
# Check results
|
||||
Path(pytorch_dump_folder_path).mkdir(exist_ok=True)
|
||||
model.save_pretrained(pytorch_dump_folder_path)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
parser = argparse.ArgumentParser()
|
||||
# Required parameters
|
||||
parser.add_argument(
|
||||
"--fairseq_path",
|
||||
type=str,
|
||||
help="path to fairseq checkpoint in correct format. You can find all checkpoints in the correct format here: https://huggingface.co/models?other=opt_metasq",
|
||||
)
|
||||
parser.add_argument("--pytorch_dump_folder_path", default=None, type=str, help="Path to the output PyTorch model.")
|
||||
parser.add_argument("--hf_config", default=None, type=str, help="Define HF config.")
|
||||
args = parser.parse_args()
|
||||
convert_opt_checkpoint(args.fairseq_path, args.pytorch_dump_folder_path, config=args.hf_config)
|
||||
1006
src/transformers/models/opt/modeling_opt.py
Normal file
1006
src/transformers/models/opt/modeling_opt.py
Normal file
File diff suppressed because it is too large
Load Diff
@@ -2997,6 +2997,30 @@ def load_tf_weights_in_openai_gpt(*args, **kwargs):
|
||||
requires_backends(load_tf_weights_in_openai_gpt, ["torch"])
|
||||
|
||||
|
||||
OPT_PRETRAINED_MODEL_ARCHIVE_LIST = None
|
||||
|
||||
|
||||
class OPTForCausalLM(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class OPTModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class OPTPreTrainedModel(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
def __init__(self, *args, **kwargs):
|
||||
requires_backends(self, ["torch"])
|
||||
|
||||
|
||||
class PegasusForCausalLM(metaclass=DummyObject):
|
||||
_backends = ["torch"]
|
||||
|
||||
|
||||
@@ -175,6 +175,78 @@ class GPT2TokenizationTest(TokenizerTesterMixin, unittest.TestCase):
|
||||
padding="max_length",
|
||||
)
|
||||
|
||||
def test_padding_if_pad_token_set_slow(self):
|
||||
tokenizer = GPT2Tokenizer.from_pretrained(self.tmpdirname, pad_token="<pad>")
|
||||
|
||||
# Simple input
|
||||
s = "This is a simple input"
|
||||
s2 = ["This is a simple input looooooooong", "This is a simple input"]
|
||||
p = ("This is a simple input", "This is a pair")
|
||||
p2 = [
|
||||
("This is a simple input loooooong", "This is a simple input"),
|
||||
("This is a simple pair loooooong", "This is a simple pair"),
|
||||
]
|
||||
|
||||
pad_token_id = tokenizer.pad_token_id
|
||||
|
||||
out_s = tokenizer(s, padding="max_length", max_length=30, return_tensors="np")
|
||||
out_s2 = tokenizer(s2, padding=True, truncate=True, return_tensors="np")
|
||||
out_p = tokenizer(*p, padding="max_length", max_length=60, return_tensors="np")
|
||||
out_p2 = tokenizer(p2, padding=True, truncate=True, return_tensors="np")
|
||||
|
||||
# s
|
||||
# test single string max_length padding
|
||||
self.assertEqual(out_s["input_ids"].shape[-1], 30)
|
||||
self.assertTrue(pad_token_id in out_s["input_ids"])
|
||||
self.assertTrue(0 in out_s["attention_mask"])
|
||||
|
||||
# s2
|
||||
# test automatic padding
|
||||
self.assertEqual(out_s2["input_ids"].shape[-1], 33)
|
||||
# long slice doesn't have padding
|
||||
self.assertFalse(pad_token_id in out_s2["input_ids"][0])
|
||||
self.assertFalse(0 in out_s2["attention_mask"][0])
|
||||
# short slice does have padding
|
||||
self.assertTrue(pad_token_id in out_s2["input_ids"][1])
|
||||
self.assertTrue(0 in out_s2["attention_mask"][1])
|
||||
|
||||
# p
|
||||
# test single pair max_length padding
|
||||
self.assertEqual(out_p["input_ids"].shape[-1], 60)
|
||||
self.assertTrue(pad_token_id in out_p["input_ids"])
|
||||
self.assertTrue(0 in out_p["attention_mask"])
|
||||
|
||||
# p2
|
||||
# test automatic padding pair
|
||||
self.assertEqual(out_p2["input_ids"].shape[-1], 52)
|
||||
# long slice pair doesn't have padding
|
||||
self.assertFalse(pad_token_id in out_p2["input_ids"][0])
|
||||
self.assertFalse(0 in out_p2["attention_mask"][0])
|
||||
# short slice pair does have padding
|
||||
self.assertTrue(pad_token_id in out_p2["input_ids"][1])
|
||||
self.assertTrue(0 in out_p2["attention_mask"][1])
|
||||
|
||||
def test_add_bos_token_slow(self):
|
||||
bos_token = "$$$"
|
||||
tokenizer = GPT2Tokenizer.from_pretrained(self.tmpdirname, bos_token=bos_token, add_bos_token=True)
|
||||
|
||||
s = "This is a simple input"
|
||||
s2 = ["This is a simple input 1", "This is a simple input 2"]
|
||||
|
||||
bos_token_id = tokenizer.bos_token_id
|
||||
|
||||
out_s = tokenizer(s)
|
||||
out_s2 = tokenizer(s2)
|
||||
|
||||
self.assertEqual(out_s.input_ids[0], bos_token_id)
|
||||
self.assertTrue(all(o[0] == bos_token_id for o in out_s2.input_ids))
|
||||
|
||||
decode_s = tokenizer.decode(out_s.input_ids)
|
||||
decode_s2 = tokenizer.batch_decode(out_s2.input_ids)
|
||||
|
||||
self.assertEqual(decode_s.split()[0], bos_token)
|
||||
self.assertTrue(all(d.split()[0] == bos_token for d in decode_s2))
|
||||
|
||||
# tokenizer has no padding token
|
||||
def test_padding_different_model_input_name(self):
|
||||
pass
|
||||
|
||||
0
tests/models/opt/__init__.py
Normal file
0
tests/models/opt/__init__.py
Normal file
362
tests/models/opt/test_modeling_opt.py
Normal file
362
tests/models/opt/test_modeling_opt.py
Normal file
@@ -0,0 +1,362 @@
|
||||
# coding=utf-8
|
||||
# Copyright 2021, 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 OPT model. """
|
||||
|
||||
|
||||
import copy
|
||||
import tempfile
|
||||
import unittest
|
||||
|
||||
import timeout_decorator # noqa
|
||||
|
||||
from transformers import OPTConfig, is_torch_available, pipeline
|
||||
from transformers.testing_utils import require_sentencepiece, require_tokenizers, require_torch, slow, torch_device
|
||||
from transformers.utils import cached_property
|
||||
|
||||
from ...generation.test_generation_utils import GenerationTesterMixin
|
||||
from ...test_configuration_common import ConfigTester
|
||||
from ...test_modeling_common import ModelTesterMixin, ids_tensor
|
||||
|
||||
|
||||
if is_torch_available():
|
||||
import torch
|
||||
|
||||
from transformers import GPT2Tokenizer, OPTForCausalLM, OPTModel
|
||||
|
||||
|
||||
def prepare_opt_inputs_dict(
|
||||
config,
|
||||
input_ids,
|
||||
decoder_input_ids=None,
|
||||
attention_mask=None,
|
||||
decoder_attention_mask=None,
|
||||
head_mask=None,
|
||||
decoder_head_mask=None,
|
||||
):
|
||||
if attention_mask is None:
|
||||
attention_mask = input_ids.ne(config.pad_token_id)
|
||||
return {
|
||||
"input_ids": input_ids,
|
||||
"attention_mask": attention_mask,
|
||||
"head_mask": head_mask,
|
||||
}
|
||||
|
||||
|
||||
class OPTModelTester:
|
||||
def __init__(
|
||||
self,
|
||||
parent,
|
||||
batch_size=13,
|
||||
seq_length=7,
|
||||
is_training=True,
|
||||
use_labels=False,
|
||||
vocab_size=99,
|
||||
hidden_size=16,
|
||||
num_hidden_layers=2,
|
||||
num_attention_heads=4,
|
||||
intermediate_size=4,
|
||||
hidden_act="gelu",
|
||||
hidden_dropout_prob=0.1,
|
||||
attention_probs_dropout_prob=0.1,
|
||||
max_position_embeddings=20,
|
||||
eos_token_id=2,
|
||||
pad_token_id=1,
|
||||
bos_token_id=0,
|
||||
embed_dim=16,
|
||||
word_embed_proj_dim=16,
|
||||
):
|
||||
self.parent = parent
|
||||
self.batch_size = batch_size
|
||||
self.seq_length = seq_length
|
||||
self.is_training = is_training
|
||||
self.use_labels = use_labels
|
||||
self.vocab_size = vocab_size
|
||||
self.hidden_size = hidden_size
|
||||
self.num_hidden_layers = num_hidden_layers
|
||||
self.num_attention_heads = num_attention_heads
|
||||
self.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.eos_token_id = eos_token_id
|
||||
self.pad_token_id = pad_token_id
|
||||
self.bos_token_id = bos_token_id
|
||||
self.embed_dim = embed_dim
|
||||
self.word_embed_proj_dim = word_embed_proj_dim
|
||||
self.is_encoder_decoder = False
|
||||
|
||||
def prepare_config_and_inputs(self):
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size).clamp(
|
||||
3,
|
||||
)
|
||||
input_ids[:, -1] = self.eos_token_id # Eos Token
|
||||
|
||||
decoder_input_ids = ids_tensor([self.batch_size, self.seq_length], self.vocab_size)
|
||||
|
||||
config = self.get_config()
|
||||
inputs_dict = prepare_opt_inputs_dict(config, input_ids, decoder_input_ids)
|
||||
return config, inputs_dict
|
||||
|
||||
def get_config(self):
|
||||
return OPTConfig(
|
||||
vocab_size=self.vocab_size,
|
||||
hidden_size=self.hidden_size,
|
||||
num_hidden_layers=self.num_hidden_layers,
|
||||
num_attention_heads=self.num_attention_heads,
|
||||
ffn_dim=self.intermediate_size,
|
||||
dropout=self.hidden_dropout_prob,
|
||||
attention_dropout=self.attention_probs_dropout_prob,
|
||||
max_position_embeddings=self.max_position_embeddings,
|
||||
eos_token_id=self.eos_token_id,
|
||||
bos_token_id=self.bos_token_id,
|
||||
pad_token_id=self.pad_token_id,
|
||||
embed_dim=self.embed_dim,
|
||||
is_encoder_decoder=False,
|
||||
word_embed_proj_dim=self.word_embed_proj_dim,
|
||||
)
|
||||
|
||||
def get_pipeline_config(self):
|
||||
config = self.get_config()
|
||||
config.max_position_embeddings = 100
|
||||
return config
|
||||
|
||||
def prepare_config_and_inputs_for_common(self):
|
||||
config, inputs_dict = self.prepare_config_and_inputs()
|
||||
return config, inputs_dict
|
||||
|
||||
def create_and_check_decoder_model_past_large_inputs(self, config, inputs_dict):
|
||||
model = OPTModel(config=config).to(torch_device).eval()
|
||||
|
||||
input_ids = inputs_dict["input_ids"]
|
||||
attention_mask = inputs_dict["attention_mask"]
|
||||
head_mask = inputs_dict["head_mask"]
|
||||
|
||||
# first forward pass
|
||||
outputs = model(input_ids, attention_mask=attention_mask, head_mask=head_mask, use_cache=True)
|
||||
|
||||
output, past_key_values = outputs.to_tuple()
|
||||
|
||||
# create hypothetical multiple next token and extent to next_input_ids
|
||||
next_tokens = ids_tensor((self.batch_size, 3), config.vocab_size)
|
||||
next_attn_mask = ids_tensor((self.batch_size, 3), 2)
|
||||
|
||||
# append to next input_ids and
|
||||
next_input_ids = torch.cat([input_ids, next_tokens], dim=-1)
|
||||
next_attention_mask = torch.cat([attention_mask, next_attn_mask], dim=-1)
|
||||
|
||||
output_from_no_past = model(next_input_ids, attention_mask=next_attention_mask)["last_hidden_state"]
|
||||
output_from_past = model(next_tokens, attention_mask=next_attention_mask, past_key_values=past_key_values)[
|
||||
"last_hidden_state"
|
||||
]
|
||||
|
||||
# select random slice
|
||||
random_slice_idx = 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))
|
||||
|
||||
|
||||
@require_torch
|
||||
class OPTModelTest(ModelTesterMixin, GenerationTesterMixin, unittest.TestCase):
|
||||
all_model_classes = (OPTModel, OPTForCausalLM) if is_torch_available() else ()
|
||||
all_generative_model_classes = (OPTForCausalLM,) if is_torch_available() else ()
|
||||
is_encoder_decoder = False
|
||||
test_pruning = False
|
||||
test_missing_keys = False
|
||||
|
||||
def setUp(self):
|
||||
self.model_tester = OPTModelTester(self)
|
||||
self.config_tester = ConfigTester(self, config_class=OPTConfig)
|
||||
|
||||
def test_config(self):
|
||||
self.config_tester.run_common_tests()
|
||||
|
||||
def test_save_load_strict(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs()
|
||||
for model_class in self.all_model_classes:
|
||||
model = model_class(config)
|
||||
|
||||
with tempfile.TemporaryDirectory() as tmpdirname:
|
||||
model.save_pretrained(tmpdirname)
|
||||
model2, info = model_class.from_pretrained(tmpdirname, output_loading_info=True)
|
||||
self.assertEqual(info["missing_keys"], [])
|
||||
|
||||
def test_decoder_model_past_with_large_inputs(self):
|
||||
config_and_inputs = self.model_tester.prepare_config_and_inputs()
|
||||
self.model_tester.create_and_check_decoder_model_past_large_inputs(*config_and_inputs)
|
||||
|
||||
def test_inputs_embeds(self):
|
||||
config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
|
||||
|
||||
for model_class in (OPTModel,):
|
||||
model = model_class(config)
|
||||
model.to(torch_device)
|
||||
model.eval()
|
||||
|
||||
inputs = copy.deepcopy(self._prepare_for_class(inputs_dict, model_class))
|
||||
|
||||
if not self.is_encoder_decoder:
|
||||
input_ids = inputs["input_ids"]
|
||||
del inputs["input_ids"]
|
||||
else:
|
||||
encoder_input_ids = inputs["input_ids"]
|
||||
decoder_input_ids = inputs.get("decoder_input_ids", encoder_input_ids)
|
||||
del inputs["input_ids"]
|
||||
inputs.pop("decoder_input_ids", None)
|
||||
|
||||
wte = model.get_input_embeddings()
|
||||
if not self.is_encoder_decoder:
|
||||
inputs["inputs_embeds"] = wte(input_ids)
|
||||
else:
|
||||
inputs["inputs_embeds"] = wte(encoder_input_ids)
|
||||
inputs["decoder_inputs_embeds"] = wte(decoder_input_ids)
|
||||
|
||||
with torch.no_grad():
|
||||
model(**inputs)[0]
|
||||
|
||||
def test_generate_fp16(self):
|
||||
config, input_dict = self.model_tester.prepare_config_and_inputs()
|
||||
input_ids = input_dict["input_ids"]
|
||||
attention_mask = input_ids.ne(1).to(torch_device)
|
||||
model = OPTForCausalLM(config).eval().to(torch_device)
|
||||
if torch_device == "cuda":
|
||||
model.half()
|
||||
model.generate(input_ids, attention_mask=attention_mask)
|
||||
model.generate(num_beams=4, do_sample=True, early_stopping=False, num_return_sequences=3)
|
||||
|
||||
|
||||
def assert_tensors_close(a, b, atol=1e-12, prefix=""):
|
||||
"""If tensors have different shapes, different values or a and b are not both tensors, raise a nice Assertion error."""
|
||||
if a is None and b is None:
|
||||
return True
|
||||
try:
|
||||
if torch.allclose(a, b, atol=atol):
|
||||
return True
|
||||
raise
|
||||
except Exception:
|
||||
pct_different = (torch.gt((a - b).abs(), atol)).float().mean().item()
|
||||
if a.numel() > 100:
|
||||
msg = f"tensor values are {pct_different:.1%} percent different."
|
||||
else:
|
||||
msg = f"{a} != {b}"
|
||||
if prefix:
|
||||
msg = prefix + ": " + msg
|
||||
raise AssertionError(msg)
|
||||
|
||||
|
||||
def _long_tensor(tok_lst):
|
||||
return torch.tensor(tok_lst, dtype=torch.long, device=torch_device)
|
||||
|
||||
|
||||
@require_torch
|
||||
@require_sentencepiece
|
||||
@require_tokenizers
|
||||
class OPTModelIntegrationTests(unittest.TestCase):
|
||||
@cached_property
|
||||
def default_tokenizer(self):
|
||||
return GPT2Tokenizer.from_pretrained("patrickvonplaten/opt_gpt2_tokenizer")
|
||||
|
||||
@slow
|
||||
def test_inference_no_head(self):
|
||||
model = OPTModel.from_pretrained("facebook/opt-350m").to(torch_device)
|
||||
input_ids = _long_tensor([[0, 31414, 232, 328, 740, 1140, 12695, 69, 46078, 1588, 2]])
|
||||
attention_mask = input_ids.ne(model.config.pad_token_id)
|
||||
with torch.no_grad():
|
||||
output = model(input_ids=input_ids, attention_mask=attention_mask).last_hidden_state
|
||||
expected_shape = torch.Size((1, 11, 1024))
|
||||
self.assertEqual(output.shape, expected_shape)
|
||||
expected_slice = torch.tensor(
|
||||
[[0.7144, 0.8143, -1.2813], [0.7144, 0.8143, -1.2813], [-0.0467, 2.5911, -2.1845]], device=torch_device
|
||||
)
|
||||
self.assertTrue(torch.allclose(output[:, :3, :3], expected_slice, atol=1e-3))
|
||||
|
||||
|
||||
@require_tokenizers
|
||||
@require_torch
|
||||
@slow
|
||||
class OPTEmbeddingsTest(unittest.TestCase):
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
self.path_model = "facebook/opt-350m"
|
||||
|
||||
def test_load_model(self):
|
||||
try:
|
||||
_ = OPTForCausalLM.from_pretrained(self.path_model)
|
||||
except BaseException:
|
||||
self.fail("Failed loading model")
|
||||
|
||||
def test_logits(self):
|
||||
model = OPTForCausalLM.from_pretrained(self.path_model)
|
||||
model = model.eval()
|
||||
tokenizer = GPT2Tokenizer.from_pretrained("patrickvonplaten/opt_gpt2_tokenizer")
|
||||
tokenizer.add_special_tokens({"pad_token": "<pad>"})
|
||||
|
||||
prompts = [
|
||||
"Today is a beautiful day and I want to",
|
||||
"In the city of",
|
||||
"Paris is the capital of France and",
|
||||
"Computers and mobile phones have taken",
|
||||
]
|
||||
input_ids = tokenizer(prompts, return_tensors="pt", padding=True).input_ids
|
||||
logits = model(input_ids)[0].mean(dim=-1)
|
||||
# logits_meta = torch.load(self.path_logits_meta)
|
||||
logits_meta = torch.Tensor(
|
||||
[
|
||||
[1.3851, -13.8923, -10.5229, -10.7533, -0.2309, -10.2384, -0.5365, -9.0947, -5.1670],
|
||||
[-4.7073, -10.6276, -3.9415, -21.5242, -0.2822, -0.2822, -0.2822, -0.2822, -0.2822],
|
||||
[0.6247, -3.4229, -8.9179, -1.4297, -14.1650, 1.4146, -9.0218, -0.2703, -0.2703],
|
||||
[6.4783, -1.9913, -10.7926, -2.3336, 1.5092, -0.9974, -6.8213, 1.3477, 1.3477],
|
||||
]
|
||||
)
|
||||
|
||||
assert torch.allclose(logits, logits_meta, atol=1e-4)
|
||||
|
||||
|
||||
@require_tokenizers
|
||||
@slow
|
||||
class OPTGenerationTest(unittest.TestCase):
|
||||
def setUp(self):
|
||||
super().setUp()
|
||||
self.all_model_path = ["facebook/opt-125m", "facebook/opt-350m"]
|
||||
|
||||
def test_generation(self):
|
||||
prompts = [
|
||||
"Today is a beautiful day and I want to",
|
||||
"In the city of",
|
||||
"Paris is the capital of France and",
|
||||
"Computers and mobile phones have taken",
|
||||
]
|
||||
NEXT_TOKENS = [3392, 764, 5, 81]
|
||||
GEN_OUTPUT = []
|
||||
|
||||
tokenizer = GPT2Tokenizer.from_pretrained("patrickvonplaten/opt_gpt2_tokenizer")
|
||||
for model in self.all_model_path:
|
||||
model = OPTForCausalLM.from_pretrained(self.path_model)
|
||||
model = model.eval()
|
||||
model.config.eos_token_id = tokenizer.eos_token_id
|
||||
|
||||
gen = pipeline("text-generation", model=model, tokenizer=tokenizer, return_tensors=True)
|
||||
|
||||
for prompt in prompts:
|
||||
len_input_sentence = len(tokenizer.tokenize(prompt))
|
||||
predicted_next_token = gen(prompt)[0]["generated_token_ids"][len_input_sentence]
|
||||
GEN_OUTPUT.append(predicted_next_token)
|
||||
self.assertListEqual(GEN_OUTPUT, NEXT_TOKENS)
|
||||
@@ -45,6 +45,7 @@ PRIVATE_MODELS = [
|
||||
# Being in this list is an exception and should **not** be the rule.
|
||||
IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
|
||||
# models to ignore for not tested
|
||||
"OPTDecoder", # Building part of bigger (tested) model.
|
||||
"DecisionTransformerGPT2Model", # Building part of bigger (tested) model.
|
||||
"SegformerDecodeHead", # Building part of bigger (tested) model.
|
||||
"PLBartEncoder", # Building part of bigger (tested) model.
|
||||
@@ -92,6 +93,7 @@ IGNORE_NON_TESTED = PRIVATE_MODELS.copy() + [
|
||||
"SeparableConv1D", # Building part of bigger (tested) model.
|
||||
"FlaxBartForCausalLM", # Building part of bigger (tested) model.
|
||||
"FlaxBertForCausalLM", # Building part of bigger (tested) model. Tested implicitly through FlaxRobertaForCausalLM.
|
||||
"OPTDecoderWrapper",
|
||||
]
|
||||
|
||||
# Update this list with test files that don't have a tester with a `all_model_classes` variable and which don't
|
||||
|
||||
Reference in New Issue
Block a user