GitLab

* GPT2DoubleHeadsModel made parallelizeable

* GPT2DoubleHeadsModel added as parallelizeable onto the GPT2 test suite

Documentation was referring to slow tokenizer class while it should be the fast tokenizer.

* enable auto tokenizer for mbart50 tokenizers

* fix imports

* Pass encoder_outputs into generate()

* Remove an if-statement

* Reformat

* Minimize changes to generate()

* Comment on input_ids

* fix: #10628 expanduser path in TrainingArguments

* docs: explain why we expand paths in TrainingArguments

* Style
Co-authored-by: Sylvain Gugger <sylvain.gugger@gmail.com>

* Add auto_wrap option in fairscale integration

* Style

* TF model exists for Blenderbot 400M

* Marian

* RAG

* [WIP] Adding new parameter to `generate`:  `max_time`.

Generation by tokens number is sometimes a bit clunky because we don't
know how many tokens are good enough or even how many tokens are in
the payload (for pipelines users for instance). This leads to hard
to understand behavior.

This PR proposes a new argument `max_time` which is a float of seconds
for the allowed time for `generate` to run on.
Ideally combinations of `max_tokens=None`, `max_time=2` could be used to
generate as many tokens as possible within time budget.

NB: Another possible approach consists of passing a callback to `generate`
  putting the caller in charge of the actual decision of when to stop
  generating tokens. It opens the door to 'which args should we pass'
  to this callback. It's hard to imagine other use-cases for this
  early stopping behavior than time (that are not already covered by
  parameters of generate)

* Revamp with StoppingCriteria

* Removing deprecated mentions.

* Forgot arguments to stopping criteria.

* Readding max_length it's not just used as a stopping criteria.

* Default value for `stopping_criteria`.

* Address @patrickvonplaten comments.

- More docstrings
- Actual doc
- Include in global namespace
- Remove TF work.

* Put back `max_length` (deprecation different PR).

* Doc quality.

* Fixing old behavior without `stopping_criteria` but with `max_length`.

Making sure we don't break that in the future.

* Adding more tests for possible inconsistencies between

`max_length` and `stopping_criteria`.

* Fixing the torch imports.

* fix typing error for TrainingArguments Optional[bool]

* updating equality check for Optional[bool]

* Fixed broken link

* fixed max length violation
Co-authored-by: WybeKoper <WybeKoper@users.noreply.github.com>

* add deberta to pretraining mapping

* add deberta_v2 to PRETRAINING_MAPPING

* PoC

* Fix slow tests for the PT1.8 Embedding problem

* Allow to pass kwargs to model's from_pretrained when using pipeline.

* Disable the use of past_keys_values for GPT2 when exporting to ONNX.

* style

* Remove comment.

* Appease the documentation gods

* Fix style
Co-authored-by: Lysandre <lysandre.debut@reseau.eseo.fr>

* Adds a @require_torch to a test that requires it

* Tokenizer too

* Style

* Remove special path for custom vocab files

* Update src/transformers/tokenization_utils_base.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Expand error message
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* S2S + M2M100 should be available in tokenization_auto

* Requires sentencepiece

* SentencePiece for S2T as well :)

* add conversion script

* add wav2vec2 xslr models

* finish

* Update docs/source/model_doc/xlsr_wav2vec2.rst
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

correct spell error: 'nether'

* renamed logging to hf_logging

* changed logging from hf_logging to logging and loggin to native_logging

* removed everything trying to fix import Trainer error

* adding imports again

* added custom add_handler function to logging.py

* make style

* added remove_handler

* added another conditional to assert

* Move tokenizer files in each repo

* Fix mBART50 tests

* Fix mBART tests

* Fix Marian tests

* Update templates

* s2t

* fix config

* conversion script

* fix import

* add tokenizer

* fix tok init

* fix tokenizer

* first version working

* fix embeds

* fix lm head

* remove extra heads

* fix convert script

* handle encoder attn mask

* style

* better enc attn mask

* override _prepare_attention_mask_for_generation

* handle attn_maks in encoder and decoder

* input_ids => input_features

* enable use_cache

* remove old code

* expand embeddings if needed

* remove logits bias

* masked_lm_loss => loss

* hack tokenizer to support feature processing

* fix model_input_names

* style

* fix error message

* doc

* remove inputs_embeds

* remove input_embeds

* remove unnecessary docstring

* quality

* SpeechToText => Speech2Text

* style

* remove shared_embeds

* subsample => conv

* remove Speech2TextTransformerDecoderWrapper

* update output_lengths formula

* fix table

* remove max_position_embeddings

* update conversion scripts

* add possibility to do upper case for now

* add FeatureExtractor and Processor

* add tests for extractor

* require_torch_audio => require_torchaudio

* add processor test

* update import

* remove classification head

* attention mask is now 1D

* update docstrings

* attention mask should be of type long

* handle attention mask from generate

* alwyas return attention_mask

* fix test

* style

* doc

* Speech2TextTransformer => Speech2Text

* Speech2TextTransformerConfig => Speech2TextConfig

* remove dummy_inputs

* nit

* style

* multilinguial tok

* fix tokenizer

* add tgt_lang setter

* save lang_codes

* fix tokenizer

* add forced_bos_token_id to tokenizer

* apply review suggestions

* add torchaudio to extra deps

* add speech deps to CI

* fix dep

* add libsndfile to ci

* libsndfile1

* add speech to extras all

* libsndfile1 -> libsndfile1

* libsndfile

* libsndfile1-dev

* apt update

* add sudo to install

* update deps table

* install libsndfile1-dev on CI

* tuple to list

* init conv layer

* add model tests

* quality

* add integration tests

* skip_special_tokens

* add speech_to_text_transformer in toctree

* fix tokenizer

* fix fp16 tests

* add tokenizer tests

* fix copyright

* input_values => input_features

* doc

* add model in readme

* doc

* change checkpoint names

* fix copyright

* fix code example

* add max_model_input_sizes in tokenizer

* fix integration tests

* add do_lower_case to tokenizer

* remove clamp trick

* fix "Add modeling imports here"

* fix copyrights

* fix tests

* SpeechToTextTransformer => SpeechToText

* fix naming

* fix table formatting

* fix typo

* style

* fix typos

* remove speech dep from extras[testing]

* fix copies

* rename doc file,

* put imports under is_torch_available

* run feat extract tests when torch is available

* dummy objects for processor and extractor

* fix imports in tests

* fix import in modeling test

* fxi imports

* fix torch import

* fix imports again

* fix positional embeddings

* fix typo in import

* adapt new extractor refactor

* style

* fix torchscript test

* doc

* doc

* Apply suggestions from code review
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* fix docs, copied from, style

* fix docstring

* handle imports

* remove speech from all extra deps

* remove s2t from seq2seq lm mapping

* better names

* skip training tests

* add install instructions

* List => Tuple

* doc

* fix conversion script

* fix urls

* add instruction for libsndfile

* fix fp16 test
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Add new GLUE example with no Trainer.

* Style

* Address review comments

Fixes an issue in `text-classification` where MNLI eval/test datasets are not being preprocessed. (#10621)

* Fix MNLI tests

* Linter fix

* Fix tests of TrainerCallback

* Update tests/test_trainer_callback.py
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Hotfix fairscale FSDP

* Evaluation works

* Save on process zero