[XLNet] Fix mems behavior (#8567)

* fix mems in xlnet

* fix use_mems

* fix use_mem_len

* fix use mems

* clean docs

* fix tf typo

* make xlnet tf for generation work

* fix tf test

* refactor use cache

* add use cache for missing models

* correct use_cache in generate

* correct use cache in tf generate

* fix tf

* correct getattr typo

* make sylvain happy

* change in docs as well

* do not apply to cookie cutter statements

* fix tf test

* make pytorch model fully backward compatible

docs/source/installation.md

@@ -97,6 +97,6 @@ You should check out our [swift-coreml-transformers](https://github.com/huggingf
 It contains a set of tools to convert PyTorch or TensorFlow 2.0 trained Transformer models (currently contains `GPT-2`, 
 `DistilGPT-2`, `BERT`, and `DistilBERT`) to CoreML models that run on iOS devices.
 
-At some point in the future, you'll be able to seamlessly move from pre-training or fine-tuning models in PyTorch or
+At some point in the future, you'll be able to seamlessly move from pretraining or fine-tuning models in PyTorch or
 TensorFlow 2.0 to productizing them in CoreML, or prototype a model or an app in CoreML then research its
 hyperparameters or architecture from PyTorch or TensorFlow 2.0. Super exciting!

docs/source/model_doc/bertgeneration.rst

@@ -10,7 +10,7 @@ Tasks <https://arxiv.org/abs/1907.12461>`__ by Sascha Rothe, Shashi Narayan, Ali
 
 The abstract from the paper is the following:
 
-*Unsupervised pre-training of large neural models has recently revolutionized Natural Language Processing. By
+*Unsupervised pretraining of large neural models has recently revolutionized Natural Language Processing. By
 warm-starting from the publicly released checkpoints, NLP practitioners have pushed the state-of-the-art on multiple
 benchmarks while saving significant amounts of compute time. So far the focus has been mainly on the Natural Language
 Understanding tasks. In this paper, we demonstrate the efficacy of pre-trained checkpoints for Sequence Generation. We

docs/source/model_doc/deberta.rst

@@ -20,8 +20,8 @@ disentangled attention mechanism, where each word is represented using two vecto
 position, respectively, and the attention weights among words are computed using disentangled matrices on their
 contents and relative positions. Second, an enhanced mask decoder is used to replace the output softmax layer to
 predict the masked tokens for model pretraining. We show that these two techniques significantly improve the efficiency
-of model pre-training and performance of downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half
-of the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9%
+of model pretraining and performance of downstream tasks. Compared to RoBERTa-Large, a DeBERTa model trained on half of
+the training data performs consistently better on a wide range of NLP tasks, achieving improvements on MNLI by +0.9%
 (90.2% vs. 91.1%), on SQuAD v2.0 by +2.3% (88.4% vs. 90.7%) and RACE by +3.6% (83.2% vs. 86.8%). The DeBERTa code and
 pre-trained models will be made publicly available at https://github.com/microsoft/DeBERTa.*
 

docs/source/model_doc/distilbert.rst

@@ -18,9 +18,9 @@ operating these large models in on-the-edge and/or under constrained computation
 remains challenging. In this work, we propose a method to pre-train a smaller general-purpose language representation
 model, called DistilBERT, which can then be fine-tuned with good performances on a wide range of tasks like its larger
 counterparts. While most prior work investigated the use of distillation for building task-specific models, we leverage
-knowledge distillation during the pre-training phase and show that it is possible to reduce the size of a BERT model by
+knowledge distillation during the pretraining phase and show that it is possible to reduce the size of a BERT model by
 40%, while retaining 97% of its language understanding capabilities and being 60% faster. To leverage the inductive
-biases learned by larger models during pre-training, we introduce a triple loss combining language modeling,
+biases learned by larger models during pretraining, we introduce a triple loss combining language modeling,
 distillation and cosine-distance losses. Our smaller, faster and lighter model is cheaper to pre-train and we
 demonstrate its capabilities for on-device computations in a proof-of-concept experiment and a comparative on-device
 study.*

docs/source/model_doc/electra.rst

@@ -12,14 +12,14 @@ identify which tokens were replaced by the generator in the sequence.
 
 The abstract from the paper is the following:
 
-*Masked language modeling (MLM) pre-training methods such as BERT corrupt the input by replacing some tokens with
-[MASK] and then train a model to reconstruct the original tokens. While they produce good results when transferred to
+*Masked language modeling (MLM) pretraining methods such as BERT corrupt the input by replacing some tokens with [MASK]
+and then train a model to reconstruct the original tokens. While they produce good results when transferred to
 downstream NLP tasks, they generally require large amounts of compute to be effective. As an alternative, we propose a
-more sample-efficient pre-training task called replaced token detection. Instead of masking the input, our approach
+more sample-efficient pretraining task called replaced token detection. Instead of masking the input, our approach
 corrupts it by replacing some tokens with plausible alternatives sampled from a small generator network. Then, instead
 of training a model that predicts the original identities of the corrupted tokens, we train a discriminative model that
 predicts whether each token in the corrupted input was replaced by a generator sample or not. Thorough experiments
-demonstrate this new pre-training task is more efficient than MLM because the task is defined over all input tokens
+demonstrate this new pretraining task is more efficient than MLM because the task is defined over all input tokens
 rather than just the small subset that was masked out. As a result, the contextual representations learned by our
 approach substantially outperform the ones learned by BERT given the same model size, data, and compute. The gains are
 particularly strong for small models; for example, we train a model on one GPU for 4 days that outperforms GPT (trained

docs/source/model_doc/flaubert.rst

@@ -19,7 +19,7 @@ representations (Dai and Le, 2015; Peters et al., 2018; Howard and Ruder, 2018;
 heterogeneous French corpus. Models of different sizes are trained using the new CNRS (French National Centre for
 Scientific Research) Jean Zay supercomputer. We apply our French language models to diverse NLP tasks (text
 classification, paraphrasing, natural language inference, parsing, word sense disambiguation) and show that most of the
-time they outperform other pre-training approaches. Different versions of FlauBERT as well as a unified evaluation
+time they outperform other pretraining approaches. Different versions of FlauBERT as well as a unified evaluation
 protocol for the downstream tasks, called FLUE (French Language Understanding Evaluation), are shared to the research
 community for further reproducible experiments in French NLP.*
 

docs/source/model_doc/gpt.rst

@@ -14,7 +14,7 @@ The abstract from the paper is the following:
 *Natural language understanding comprises a wide range of diverse tasks such as textual entailment, question answering,
 semantic similarity assessment, and document classification. Although large unlabeled text corpora are abundant,
 labeled data for learning these specific tasks is scarce, making it challenging for discriminatively trained models to
-perform adequately. We demonstrate that large gains on these tasks can be realized by generative pre-training of a
+perform adequately. We demonstrate that large gains on these tasks can be realized by generative pretraining of a
 language model on a diverse corpus of unlabeled text, followed by discriminative fine-tuning on each specific task. In
 contrast to previous approaches, we make use of task-aware input transformations during fine-tuning to achieve
 effective transfer while requiring minimal changes to the model architecture. We demonstrate the effectiveness of our

docs/source/model_doc/layoutlm.rst

@@ -6,19 +6,19 @@ Overview
 
 The LayoutLM model was proposed in the paper `LayoutLM: Pre-training of Text and Layout for Document Image
 Understanding <https://arxiv.org/abs/1912.13318>`__ by Yiheng Xu, Minghao Li, Lei Cui, Shaohan Huang, Furu Wei, and
-Ming Zhou. It's a simple but effective pre-training method of text and layout for document image understanding and
+Ming Zhou. It's a simple but effective pretraining method of text and layout for document image understanding and
 information extraction tasks, such as form understanding and receipt understanding.
 
 The abstract from the paper is the following:
 
 *Pre-training techniques have been verified successfully in a variety of NLP tasks in recent years. Despite the
-widespread use of pre-training models for NLP applications, they almost exclusively focus on text-level manipulation,
+widespread use of pretraining models for NLP applications, they almost exclusively focus on text-level manipulation,
 while neglecting layout and style information that is vital for document image understanding. In this paper, we propose
 the \textbf{LayoutLM} to jointly model interactions between text and layout information across scanned document images,
 which is beneficial for a great number of real-world document image understanding tasks such as information extraction
 from scanned documents. Furthermore, we also leverage image features to incorporate words' visual information into
 LayoutLM. To the best of our knowledge, this is the first time that text and layout are jointly learned in a single
-framework for document-level pre-training. It achieves new state-of-the-art results in several downstream tasks,
+framework for document-level pretraining. It achieves new state-of-the-art results in several downstream tasks,
 including form understanding (from 70.72 to 79.27), receipt understanding (from 94.02 to 95.24) and document image
 classification (from 93.07 to 94.42).*
 

docs/source/model_doc/lxmert.rst

@@ -19,7 +19,7 @@ Encoder Representations from Transformers) framework to learn these vision-and-l
 build a large-scale Transformer model that consists of three encoders: an object relationship encoder, a language
 encoder, and a cross-modality encoder. Next, to endow our model with the capability of connecting vision and language
 semantics, we pre-train the model with large amounts of image-and-sentence pairs, via five diverse representative
-pre-training tasks: masked language modeling, masked object prediction (feature regression and label classification),
+pretraining tasks: masked language modeling, masked object prediction (feature regression and label classification),
 cross-modality matching, and image question answering. These tasks help in learning both intra-modality and
 cross-modality relationships. After fine-tuning from our pretrained parameters, our model achieves the state-of-the-art
 results on two visual question answering datasets (i.e., VQA and GQA). We also show the generalizability of our

docs/source/model_doc/mbart.rst

@@ -13,7 +13,7 @@ The MBart model was presented in `Multilingual Denoising Pre-training for Neural
 Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
 
 According to the abstract, MBART is a sequence-to-sequence denoising auto-encoder pretrained on large-scale monolingual
-corpora in many languages using the BART objective. mBART is one of the first methods for pre-training a complete
+corpora in many languages using the BART objective. mBART is one of the first methods for pretraining a complete
 sequence-to-sequence model by denoising full texts in multiple languages, while previous approaches have focused only
 on the encoder, decoder, or reconstructing parts of the text.
 

docs/source/model_doc/prophetnet.rst

@@ -17,7 +17,7 @@ the next token.
 
 The abstract from the paper is the following:
 
-*In this paper, we present a new sequence-to-sequence pre-training model called ProphetNet, which introduces a novel
+*In this paper, we present a new sequence-to-sequence pretraining model called ProphetNet, which introduces a novel
 self-supervised objective named future n-gram prediction and the proposed n-stream self-attention mechanism. Instead of
 the optimization of one-step ahead prediction in traditional sequence-to-sequence model, the ProphetNet is optimized by
 n-step ahead prediction which predicts the next n tokens simultaneously based on previous context tokens at each time
@@ -25,7 +25,7 @@ step. The future n-gram prediction explicitly encourages the model to plan for t
 overfitting on strong local correlations. We pre-train ProphetNet using a base scale dataset (16GB) and a large scale
 dataset (160GB) respectively. Then we conduct experiments on CNN/DailyMail, Gigaword, and SQuAD 1.1 benchmarks for
 abstractive summarization and question generation tasks. Experimental results show that ProphetNet achieves new
-state-of-the-art results on all these datasets compared to the models using the same scale pre-training corpus.*
+state-of-the-art results on all these datasets compared to the models using the same scale pretraining corpus.*
 
 The Authors' code can be found `here <https://github.com/microsoft/ProphetNet>`__.
 

docs/source/model_doc/t5.rst

@@ -17,7 +17,7 @@ The abstract from the paper is the following:
 task, has emerged as a powerful technique in natural language processing (NLP). The effectiveness of transfer learning
 has given rise to a diversity of approaches, methodology, and practice. In this paper, we explore the landscape of
 transfer learning techniques for NLP by introducing a unified framework that converts every language problem into a
-text-to-text format. Our systematic study compares pre-training objectives, architectures, unlabeled datasets, transfer
+text-to-text format. Our systematic study compares pretraining objectives, architectures, unlabeled datasets, transfer
 approaches, and other factors on dozens of language understanding tasks. By combining the insights from our exploration
 with scale and our new "Colossal Clean Crawled Corpus", we achieve state-of-the-art results on many benchmarks covering
 summarization, question answering, text classification, and more. To facilitate future work on transfer learning for

docs/source/model_doc/xlmprophetnet.rst

@@ -19,7 +19,7 @@ just the next token. Its architecture is identical to ProhpetNet, but the model
 
 The abstract from the paper is the following:
 
-*In this paper, we present a new sequence-to-sequence pre-training model called ProphetNet, which introduces a novel
+*In this paper, we present a new sequence-to-sequence pretraining model called ProphetNet, which introduces a novel
 self-supervised objective named future n-gram prediction and the proposed n-stream self-attention mechanism. Instead of
 the optimization of one-step ahead prediction in traditional sequence-to-sequence model, the ProphetNet is optimized by
 n-step ahead prediction which predicts the next n tokens simultaneously based on previous context tokens at each time
@@ -27,7 +27,7 @@ step. The future n-gram prediction explicitly encourages the model to plan for t
 overfitting on strong local correlations. We pre-train ProphetNet using a base scale dataset (16GB) and a large scale
 dataset (160GB) respectively. Then we conduct experiments on CNN/DailyMail, Gigaword, and SQuAD 1.1 benchmarks for
 abstractive summarization and question generation tasks. Experimental results show that ProphetNet achieves new
-state-of-the-art results on all these datasets compared to the models using the same scale pre-training corpus.*
+state-of-the-art results on all these datasets compared to the models using the same scale pretraining corpus.*
 
 The Authors' code can be found `here <https://github.com/microsoft/ProphetNet>`__.
 

docs/source/model_summary.rst

@@ -527,7 +527,7 @@ Pegasus
 <https://arxiv.org/pdf/1912.08777.pdf>`_, Jingqing Zhang, Yao Zhao, Mohammad Saleh and Peter J. Liu on Dec 18, 2019.
 
 Sequence-to-sequence model with the same encoder-decoder model architecture as BART. Pegasus is pre-trained jointly on
-two self-supervised objective functions: Masked Language Modeling (MLM) and a novel summarization specific pre-training
+two self-supervised objective functions: Masked Language Modeling (MLM) and a novel summarization specific pretraining
 objective, called Gap Sentence Generation (GSG).
 
   * MLM: encoder input tokens are randomly replaced by a mask tokens and have to be predicted by the encoder (like in
@@ -609,7 +609,7 @@ MT5
 `mT5: A massively multilingual pre-trained text-to-text transformer <https://arxiv.org/abs/2010.11934>`_, Linting Xue
 et al.
 
-The model architecture is same as T5. mT5's pre-training objective includes T5's self-supervised training, but not T5's
+The model architecture is same as T5. mT5's pretraining objective includes T5's self-supervised training, but not T5's
 supervised training. mT5 is trained on 101 languages.
 
 The library provides a version of this model for conditional generation.
@@ -630,8 +630,8 @@ MBart
 `Multilingual Denoising Pre-training for Neural Machine Translation <https://arxiv.org/abs/2001.08210>`_ by Yinhan Liu,
 Jiatao Gu, Naman Goyal, Xian Li, Sergey Edunov Marjan Ghazvininejad, Mike Lewis, Luke Zettlemoyer.
 
-The model architecture and pre-training objective is same as BART, but MBart is trained on 25 languages and is intended
-for supervised and unsupervised machine translation. MBart is one of the first methods for pre-training a complete
+The model architecture and pretraining objective is same as BART, but MBart is trained on 25 languages and is intended
+for supervised and unsupervised machine translation. MBart is one of the first methods for pretraining a complete
 sequence-to-sequence model by denoising full texts in multiple languages,
 
 The library provides a version of this model for conditional generation.
@@ -658,7 +658,7 @@ ProphetNet
 `ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training, <https://arxiv.org/abs/2001.04063>`__ by
 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang, Ming Zhou.
 
-ProphetNet introduces a novel *sequence-to-sequence* pre-training objective, called *future n-gram prediction*. In
+ProphetNet introduces a novel *sequence-to-sequence* pretraining objective, called *future n-gram prediction*. In
 future n-gram prediction, the model predicts the next n tokens simultaneously based on previous context tokens at each
 time step instead instead of just the single next token. The future n-gram prediction explicitly encourages the model
 to plan for the future tokens and prevent overfitting on strong local correlations. The model architecture is based on
@@ -683,8 +683,8 @@ XLM-ProphetNet
 `ProphetNet: Predicting Future N-gram for Sequence-to-Sequence Pre-training, <https://arxiv.org/abs/2001.04063>`__ by
 Yu Yan, Weizhen Qi, Yeyun Gong, Dayiheng Liu, Nan Duan, Jiusheng Chen, Ruofei Zhang, Ming Zhou.
 
-XLM-ProphetNet's model architecture and pre-training objective is same as ProphetNet, but XLM-ProphetNet was
-pre-trained on the cross-lingual dataset `XGLUE <https://arxiv.org/abs/2004.01401>`__.
+XLM-ProphetNet's model architecture and pretraining objective is same as ProphetNet, but XLM-ProphetNet was pre-trained
+on the cross-lingual dataset `XGLUE <https://arxiv.org/abs/2004.01401>`__.
 
 The library provides a pre-trained version of this model for multi-lingual conditional generation and fine-tuned
 versions for headline generation and question generation, respectively.

docs/source/task_summary.rst

@@ -305,7 +305,7 @@ Language modeling is the task of fitting a model to a corpus, which can be domai
 transformer-based models are trained using a variant of language modeling, e.g. BERT with masked language modeling,
 GPT-2 with causal language modeling.
 
-Language modeling can be useful outside of pre-training as well, for example to shift the model distribution to be
+Language modeling can be useful outside of pretraining as well, for example to shift the model distribution to be
 domain-specific: using a language model trained over a very large corpus, and then fine-tuning it to a news dataset or
 on scientific papers e.g. `LysandreJik/arxiv-nlp <https://huggingface.co/lysandre/arxiv-nlp>`__.
 

src/transformers/configuration_utils.py

@@ -55,8 +55,6 @@ class PretrainedConfig(object):
             Whether or not the model should return all hidden-states.
         output_attentions (:obj:`bool`, `optional`, defaults to :obj:`False`):
             Whether or not the model should returns all attentions.
-        use_cache (:obj:`bool`, `optional`, defaults to :obj:`True`):
-            Whether or not the model should return the last key/values attentions (not used by all models).
         return_dict (:obj:`bool`, `optional`, defaults to :obj:`True`):
             Whether or not the model should return a :class:`~transformers.file_utils.ModelOutput` instead of a plain
             tuple.
@@ -168,7 +166,6 @@ class PretrainedConfig(object):
         self.return_dict = kwargs.pop("return_dict", True)
         self.output_hidden_states = kwargs.pop("output_hidden_states", False)
         self.output_attentions = kwargs.pop("output_attentions", False)
-        self.use_cache = kwargs.pop("use_cache", True)  # Not used by all models
         self.torchscript = kwargs.pop("torchscript", False)  # Only used by PyTorch models
         self.use_bfloat16 = kwargs.pop("use_bfloat16", False)
         self.pruned_heads = kwargs.pop("pruned_heads", {})

src/transformers/data/datasets/language_modeling.py

@@ -229,7 +229,7 @@ class LineByLineWithSOPTextDataset(Dataset):
         # to `block_size` anyways, so short sequences are generally wasted
         # computation. However, we *sometimes*
         # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
-        # sequences to minimize the mismatch between pre-training and fine-tuning.
+        # sequences to minimize the mismatch between pretraining and fine-tuning.
         # The `target_seq_length` is just a rough target however, whereas
         # `block_size` is a hard limit.
         target_seq_length = max_num_tokens
@@ -425,7 +425,7 @@ class TextDatasetForNextSentencePrediction(Dataset):
         # to `block_size` anyways, so short sequences are generally wasted
         # computation. However, we *sometimes*
         # (i.e., short_seq_prob == 0.1 == 10% of the time) want to use shorter
-        # sequences to minimize the mismatch between pre-training and fine-tuning.
+        # sequences to minimize the mismatch between pretraining and fine-tuning.
         # The `target_seq_length` is just a rough target however, whereas
         # `block_size` is a hard limit.
         target_seq_length = max_num_tokens

src/transformers/generation_tf_utils.py

@@ -38,6 +38,7 @@ class TFGenerationMixin:
 
     def _use_cache(self, outputs, use_cache):
         """During generation, decide whether to pass the `past` variable to the next forward pass."""
+        use_cache = getattr(self.config, "use_cache", False)
         if len(outputs) <= 1 or use_cache is False:
             return False
         if hasattr(self.config, "mem_len") and self.config.mem_len == 0:
@@ -194,7 +195,6 @@ class TFGenerationMixin:
         min_length = min_length if min_length is not None else self.config.min_length
         do_sample = do_sample if do_sample is not None else self.config.do_sample
         early_stopping = early_stopping if early_stopping is not None else self.config.early_stopping
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
         num_beams = num_beams if num_beams is not None else self.config.num_beams
         temperature = temperature if temperature is not None else self.config.temperature
         top_k = top_k if top_k is not None else self.config.top_k
@@ -224,7 +224,6 @@ class TFGenerationMixin:
         assert isinstance(min_length, int) and min_length >= 0, "`min_length` should be a positive integer."
         assert isinstance(do_sample, bool), "`do_sample` should be a boolean."
         assert isinstance(early_stopping, bool), "`early_stopping` should be a boolean."
-        assert isinstance(use_cache, bool), "`use_cache` should be a boolean."
         assert isinstance(num_beams, int) and num_beams > 0, "`num_beams` should be a strictly positive integer."
         assert temperature > 0, "`temperature` should be strictly positive."
         assert isinstance(top_k, int) and top_k >= 0, "`top_k` should be a positive integer."

src/transformers/generation_utils.py

@@ -462,7 +462,6 @@ class GenerationMixin:
         pad_token_id = pad_token_id if pad_token_id is not None else self.config.pad_token_id
         bos_token_id = bos_token_id if bos_token_id is not None else self.config.bos_token_id
         eos_token_id = eos_token_id if eos_token_id is not None else self.config.eos_token_id
-        use_cache = use_cache if use_cache is not None else self.config.use_cache
 
         if input_ids is None:
             # init `input_ids` with bos_token_id

src/transformers/models/albert/modeling_albert.py

@@ -730,7 +730,7 @@ class AlbertModel(AlbertPreTrainedModel):
 
 @add_start_docstrings(
     """
-    Albert Model with two heads on top as done during the pre-training: a `masked language modeling` head and a
+    Albert Model with two heads on top as done during the pretraining: a `masked language modeling` head and a
     `sentence order prediction (classification)` head.
     """,
     ALBERT_START_DOCSTRING,