GitLab

* Integrate DeBERTa v2(the 1.5B model surpassed human performance on SuperGLUE); Add DeBERTa v2 900M,1.5B models;

* DeBERTa-v2

* Fix v2 model loading issue (#10129)

* Doc members

* Update src/transformers/models/deberta/modeling_deberta.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Address Sylvain's comments

* Address Patrick's comments
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Style
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Fix AMP and XLA

* Remove useless var

* Fix AMP

* Apply style

* Remove unused import

* Fix XLA

* Rework cast

* Apply style

* Fix AMP

* Trigger CI

* Rework cast

* Fix AMP

* Rework cast

* Apply style

* implement --fp16_full_eval

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* style

* add test
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* memory tracker metrics

* go back to eval for somewhat consistency

* handle no-gpu case

* deal with stackable eval calls

* restore callback order

* style

* simplify the API

* add test

* docs

* consistently use eval_ prefix

* improve docs

* Update src/transformers/trainer_utils.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* rename method

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

* rework savedmodel slow test

* Improve savedmodel tests

* Remove useless content

* Fix XLA and AMP

* Fix AMP and XLA

* Apply style

* Apply Patrick's comment

* Fix XLA and AMP

* Apply style

* Remove useless cast

* Fix Flaubert and XLM

* Remove useless cast

* Tiny fix

* Tiny fix

* Update BART

* Update Blenderbot

* Update BlenderbotSmall

* Update Marian

* Update MBart

* Update MBart

* Update Pegasus

* Update template

* Fix Marian and Pegasus

* Apply style

* Default initializer

* Default initializer

* Default initializer

* Remove int32 casts

* Fix template

* Remove more cast

* Fix head_mask and decoder_head_mask in TFT5 models

* Enable test_headmasking both fot TFT5 tester
and TFT5EncoderOnly tester
Co-authored-by: patrickvonplaten <patrick.v.platen@gmail.com>

Co-authored-by: Quentin Lhoest <lhoest.q@gmail.com>
Co-authored-by: Quentin Lhoest <lhoest.q@gmail.com>

* add tokenizer for mBART-50

* update tokenizers

* make src_lang and tgt_lang optional

* update tokenizer test

* add setter

* update docs

* update conversion script

* update docs

* update conversion script

* update tokenizer

* update test

* update docs

* doc

* address Sylvain's suggestions

* fix test

* fix formatting

* nits

* Add check-ops script

* Finish to implement check_tf_ops and start the test

* Make the test mandatory only for BERT

* Update tf_ops folder

* Remove useless classes

* Add the ONNX test for GPT2 and BART

* Add a onnxruntime slow test + better opset flexibility

* Fix test + apply style

* fix tests

* Switch min opset from 12 to 10

* Update src/transformers/file_utils.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Fix GPT2

* Remove extra shape_list usage

* Fix GPT2

* Address Morgan's comments
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

Fixes #10168

* Conversion from slow to fast for BPE spm vocabs contained an error.

- There is only 1 test currently (tokenizers + slow) that used the modified path
and it's reformer, which does not contain any ids modification so the
bug was silent for now.
- The real issue is that vocab variable was overloaded by
SentencePieceExtractor, leading to Slow specific vocab oddities to be
completely ignored
- The bug was reported here https://github.com/huggingface/transformers/issues/9518
- Ran the complete tokenization test suite with slow without error
(`RUN_SLOW=1 pytest -sv tests/test_tokenization_*`)

* Remove rebase error.

* Adding the fixture.

* [hf_api] delete deprecated methods and tests

cc @lhoestq

* Update test_hf_api.py

* save intermediate

* finish batch the same as fairseq

* add normalization

* fix batched input

* add better comment

* Update src/transformers/models/wav2vec2/modeling_wav2vec2.py

* add nice docstring

* add tokenizer tests

* make all slow tests pass

* finish PR

* correct import

* add forced logits processors

* delete adjust_logits method

* add forced_eos_token_id argument in config

* add tests for forced logits processors

* update gen utils tests

* add forced option to tf generate

* remove adjust_logits method from tf models

* update adjust_logits for marian

* delete _force_token_id_to_be_generated method

* style

* import warnings

* pass max_length to _get_logits_processor

* set forced_eos_token_id to None

* set forced attributes in conf utils

* typo

* fix rag generate

* add forced_eos_token_id in rag config

* remove force_bos_token_to_be_generated from BartConfig

* remove _force_token_ids_generation from FSMT

* nit

* fix negative constant

* apply suggestions from code review

* Fix test

* Remove commented test

* Fix name

* Apply style

* Fix check copies

* Remove prints

* Restore boolean

* Fix reshape

* Add head masking to TF LED

* Add head_mask to Longformer + one doc piece to LED

* Fix integration tests

* add wav2vec2CTC and deprecate for maskedlm

* remove from docs

* claiming this issue

* Integration test for BertGeneration(Encoder and Decoder)

* fix code quality

* Refacto BERT

* Restore all the concerned models

* Remove print

* Update template

* Apply Sylvain's and Morgan's comments

* Fix cast

* Put the cast inside call

* Remove cond in ebds

* Fix funnel

* Restore previous dot product (attention_scores) computation

* Add ConvBERT and BART

* Make all the S2S models ONNX compliant

* Fix test

* Fix check copies

* Disable temporarily too slow tests

* Fix style

* Fix template

* Cleaning up `ConversationalPipeline` to support more than DialoGPT.

Currently ConversationalPipeline was heavily biased towards DialoGPT
,which is the default model for this pipeline.

This PR proposes changes to put back the modifications specific to
DialoGPT into tokenizer-specific behavior wherever possible, by
creating `_build_conversation_input_ids` function that takes
conversation as input, and returns a list of ints corresponding
to the tokens. It feels natural to put here because all models
have probably different strategies to build input_ids from the
full conversation and it's the tokenizer's job to transform strings
into tokens (and vice-versa)

If `_build_conversation_input_ids` is missing, previous behavior is
used so we don't break anything so far (except for blenderbot where it's a fix).

This PR also contains a fix for too long inputs. There used
to be dead code for trying to limit the size of incoming input.
The introduced fixed is that we limit
within `_build_conversation_input_ids` to `tokenizer.model_max_length`.
It corresponds to the intent of the removed dead code and is actually
better because it corresponds to `model_max_length` which is different
from `max_length` (which is a default parameter for `generate`).

- Removed `history` logic from the Conversation as it's not relevant
anymore because tokenization logic has been moved to tokenizer.
And tokenizer cannot save any cache, and conversation cannot know
what is relevant or not.
Also it's not usable from `blenderbot` because the input_ids are
not append only (EOS tokens is always at the end).

- Added `iter_texts` method on `Conversation` because all
the code was literred with some form of this iteration of
past/generated_responses.

* Removing torch mention in types.

* Adding type checking to `_build_conversation_input_ids`.

* Fixing import in strings.