GitLab

* document Deepspeed-Inference and parallelformers

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

* [Deepspeed] warmup_ratio docs

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

* style

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

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

* Add _CHECKPOINT_FOR_DOC

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

* Add versioning system to fast tokenizer files

* Deal with offline mode

* Use staging env in tests

* Style

* Apply suggestions from code review
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Style
Co-authored-by: Lysandre Debut <lysandre@huggingface.co>

* Refer warmup_ratio when setting warmup_num_steps.

* Add a method to get number of warmup steps to TrainerArguments class.

* Fix.

* Fix.

* Expose get_config() on ModelTesters

* Typo

* [trainer] fix % 0

* sanity checks

* fix logging_strategy

* correction

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

* fix_torch_device_generate_test

* remove @

* correct longformer docs
Co-authored-by: Patrick von Platen <patrick@huggingface.co>

cc @patil-suraj

* Allow validation split when loading dataset from local file

* Flax clm & t5, enable validation split for datasets loaded from local file

This fixes the padded batch [issue](https://github.com/huggingface/transformers/issues/12282). The error was generated due to the maximum sequence length of the attention mask not matching the padded sequence length of the hidden_states. `np.allclose` now passes with a 1e-2 absolute tolerance.

This change fixes

* Enforce eval and save strategies are compatible when --load_best_model_at_end

* Update doc

* Fix typos

* Fix tests

* Fix dynamic_modules ImportError with Ray Tune

* Nit

* Detect mismatch by analyzing config

* Fix comment

* Fix import

* Update src/transformers/tokenization_utils_base.py
Co-authored-by: SaulLu <55560583+SaulLu@users.noreply.github.com>

* Revise based on reviews

* remove kwargs

* Fix exception

* Fix handling exception again

* Disable mismatch test in PreTrainedTokenizerFast
Co-authored-by: SaulLu <55560583+SaulLu@users.noreply.github.com>

* fix_torch_device_generate_test

* remove @

* finish

* correct script

* correct script

* preserve type of `additional_special_tokens` in `special_token_map`

* format

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

* Set model in eval mode when exporting to ONNX.

* Disable t5 for now.

* Disable T5 with past too.

* Style.

* [testing] details of how to start self-push workflow

* style

* fix

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

* fix_torch_device_generate_test

* remove @

* start adding tests

* correct wav2vec2 pretraining

* up

* up
Co-authored-by: Patrick von Platen <patrick@huggingface.co>