GitLab

Change logic

Fix step bugs in no trainer + load checkpoint + grad acc

Act on deprecation

* fix bug in group_texts function, that was inserting short batches

* fully exclude short batches and return empty dict instead

* fix style

* add low_cpu_mem_usage option in run_clm.py example which will benefit LLM loading
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* update all the example and README under language-modeling
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

---------
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* Respect documentation on passive log level

* Fix test and set log level in examples

* Add doc

* Result of black 23.1

* Update target to Python 3.7

* Switch flake8 to ruff

* Configure isort

* Configure isort

* Apply isort with line limit

* Put the right black version

* adapt black in check copies

* Fix copies

* simplify column_names in run_clm

* simplify column_names in run_mlm

* minor

* [run_clm example] add streaming dataset support

* unrefactor kwargs

* fix

* fix

* require datasets>=2.0.0

* port to mlm

* Adapt repository creation to latest hf_hub

* Update all examples

* Fix other tests, add Flax examples

* Address review comments

* [run_clm example] add torch_dtype option for model load.
for BLOOM 175B model. peak memory will reduce about 350G for inference. the weight of BLOOM in model hub is bfloat16
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* add other type in option

* fix style
Signed-off-by: Wang, Yi A <yi.a.wang@intel.com>

* Update reqs to include min gather_for_metrics Accelerate version

* Other reqs

* Init Update

* ClearML Callbacks integration

* update corrections

* args reporting updated

* {'tensorboard': False, 'pytorch': False}

* ClearML Tests added

* add clearml

* output_uri=True in Task.init

* reformatted integrations.py

* reformatted and fixed

* IF-ELSE statement issue on "has_clearml" resolved

* Add clearml in main callback docs

* Add additional clearml documentation

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

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

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

* Small change in comments

* Make style clearml

* Accept suggestion
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Victor Sonck <victor.sonck@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Only resize embeddings when necessary

* Add comment

* fixed bug which caused overwrite_cache to always be True (#18967).

* reformatting changes

* add accelerator.end_training()

Some trackers need this to end their runs.

* fixup and quality

* add space

* add space again ?!?

* `model.tie_weights()` should be applied after `accelerator.prepare`

Weight tying should be done after the model has been moved to XLA device as mentioned on PyTorch/XLA Troubleshooting guide [here](https://github.com/pytorch/xla/blob/master/TROUBLESHOOTING.md#xla-tensor-quirks)

* format code

* Update run_translation_no_trainer.py

found an error in selecting `no_decay` parameters and some small modifications when the user continues to train from a checkpoint

* fixs `no_decay` and `resume_step` issue

1. change `no_decay` list
2. if use continue to train their model from provided checkpoint, the `resume_step` will not be initialized properly if `args.gradient_accumulation_steps != 1`

* Added accelerate gradient accumulation wrapper to run_image_classification_no_trainer.py example script

* make fixup changes

* PR comments

* changed input to Acceletor based on PR comment, ran make fixup

* Added comment explaining the sync_gradients statement

* Fixed lr scheduler max steps

* Changed run_clm_no_trainer.py script to use accelerate gradient accum wrapper

* Fixed all scripts except wav2vec2 pretraining to use accelerate gradient accum wrapper

* Added accelerate gradient accum wrapper for wav2vec2_pretraining_no_trainer.py script

* make fixup and lr_scheduler step inserted back into run_qa_beam_search_no_trainer.py

* removed changes to run_wav2vec2_pretraining_no_trainer.py script and fixed using wrong constant in qa_beam_search_no_trainer.py script

* zero chance anyone's using that constant no?

* `transformers-cli login` => `huggingface-cli login`

* `transformers-cli repo create` => `huggingface-cli repo create`

* `make style`