GitLab

Extend Transformers Trainer Class to Enable CPU AMP and Integrate Intel Extension for PyTorch (#17138)

* init PR

* fix import ipex

* minor fix on bf16

* refine optimizer

* refine args notes

* refine code

* refine ipex optimize args

* refine half_precision_backend

* black format

* isort format

* isort format files

* flake8 format

* doc builder format

* refine codes

* remove jit and optim bits

* black preview format

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

* refine code

* refine notes

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

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

* code refine

* add ipex ut

* add performance cpu doc

* link to the cpu doc from main perf doc

* install ipex into CI's docker

* Update perf_train_cpu.mdx

* Update docs/source/en/perf_train_cpu.mdx
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* Update perf_train_cpu.mdx

* Update perf_train_cpu.mdx
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Stas Bekman <stas@stason.org>
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* Update docker file
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* feat: initial implementation of data2vec segmentation model in TF.

* chore: minor corrections to make the segmenter work.

* chore: removed unncessary files.

* chore: add tests and other modifications.

* fix: loss computation for segmentation.

* chore: remove unused variable.

* chore: formatting.

* added a dummy adaptive pooling layer.

* removed unnecessary file.

* potentially add identifiers to layer names.

* fix: layer naming.

* chore: removed unnecessary print.

* Skipping unneeded test

* chore: add logging to debug tolerance.

* fix: segmentation tests for tfdata2vecvision

* chore: make style.

* fix: layer names, assertion to be resolved.

* Bumping test tolerance a bit

* chore: bump the tol in PT test.
Co-authored-by: matt <rocketknight1@gmail.com>

* Stricter pt-to-tf checks; Update docker image for related tests

* check all attributes in the output
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* added cbs to notebooks, made copy-paste error fix in generation_utils

* initial push for mctc model

* mctc feature extractor done

* added processor, tokenizer and their tests for MCTC. Have added an MCTC modeling test, adjusting model code accordingly.

* added processor, tokenizer and their tests for MCTC. Have added an MCTC modeling test, adjusting model code accordingly.

* passing attention, now struggling to figure out how attention masks make sense here

* works when excluding attention masks. ask later how one would integrate attention maskshere

* bizarre configuration error (model prefix comes first in config dict json and messes up the order)

* all passing but bizzarre config dict ordering issue when to_dict

* passing all major tests

* feature extraction, processor, tokenizer added & tests passing

* style & consistency & other logistical fixes

* copy paste fix

* model after feature extraction working

* commiting final feature extraction results; need to fix normalization

* feature extraction passing tests; probably should add tests on the specific flashlight-copied functions?

* delete print ; format code a bit

* fixing tests

* passing major tests

* fixing styles

* completed tokenization test with real example; not sure if these values are entirely correct.

* last test fixes from local

* reverting accidentally included custom setup configs

* remove load tf weights; fix config error

* testing couldnt import featureextractor

* fix docs

* fix docs

* resolving comments

* style fixes

* style fixes

* Update to MCTCConv1dSubSampler
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* relposemb fixes

* conv1d name issue; expecting config fail with paraentheses

* fix config issue

* fix config issue

* fix config issue

* change everything to MCTCT

* fixing naming change errors

* archive list

* copyrights and docs

* copyrights and docs

* copyrights and docs

* merge resolution

* move tests, fix to changed optionaldependency structure

* test directories changed

* fixing tests

* how to avoid tf tests?

* how to avoid tf tests?

* tests passing locally

* allow mctctprocessor imported any env

* allow mctctprocessor imported any env

* fixed second round of feedback, need to fix docs

* doc changes not being applied

* all fixed

* style fix

* feedback fixes

* fix copies and feature extraction style fix

* Update tests/models/visual_bert/test_modeling_visual_bert.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* copy paste huggingface:main visual bert

* added eof newline to visual bert; all tests are passing otherwise

* fix slow tests by adding attention mask

* change model id to speechbrain

* make fix-copies

* fix readme unwanted deletes

* fixing readmes, make fix-copies

* consistent M-CTC-T naming

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

* all fixed but variable naming

* adjust double quotes

* fixed variable names

* copyright and mr quilter

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

* correct slow tests

* make fix-copies

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

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

* m-ctc-t not mctct
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* Quicktour Portuguese Translation

Translated quicktour.mdx until line 161

* Finished translating quicktour.mdx

Ready to upload and adjust eventual .mdx or translation mistakes.

* Add _toctree.yml and fix nits

* Fixed pt-br mdx syntax problem

Closed <frameworkcontent> instance

* Changed </frameworkcontent> line

* Copied missing block from english version of quicktour.mdx

* Reviwed the entire file once again. It should be working now.
Co-authored-by: Omar U. Espejel <espejelomar@gmail.com>

* Support for deberta and deberta-v2

* Support for LXMert

* Support for Hubert

* Fix for pt1.11

* Trigger CI

* Add examples telemetry

* Alternative approach

* Add to all other examples

* Add to templates as well

* Put framework separately

* Same for TensorFlow

* Fix circular import in onnx.utils

* Add comment for test fetcher

* Here too

* Style

* update versions
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* Add method to call to_tf_dataset() with column inference

* Add test for dataset creation

* Add a default arg for data collator

* Fix test

* Fix call with non-dev version of datasets

* Test correct column removal too

* make fixup

* More tests to make sure we remove unwanted columns

* Fix test to avoid predicting on unbuilt models

* Fix test to avoid predicting on unbuilt models

* Fix test to remove unwanted head mask columns from inputs

* Stop pushing your debug breakpoints to the main repo of the $2bn company you work for

* Skip the test in convnext because no grouped conv support

* Drop bools from the dataset dict

* Make style

* Skip the training test for models whose input dicts don't give us labels

* Skip transformerXL in the test because it doesn't return a simple loss

* Skip TFTapas because of some odd NaN losses

* make style

* make fixup

* Add docstring

* fixup

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

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

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

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

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

* Remove breakpoint from tests

* Fix assert, add requires_backends

* Protect tokenizer import with if TYPE_CHECKING

* make fixup

* Add noqa, more fixup

* More rearranging for ~* aesthetics *~

* Adding defaults for shuffle and batch_size to match to_tf_dataset()

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

* [deepspeed] fix load_best_model test

* [deepspeed] add state reset on unittest tearDown

* added toctree.yml file

* first translation

* added pipeline_tutorial.mdx translation

added pipeline_tutorial.mdx
updated _toctree.yml

* updated pipeline_tutorial.mdx

* updated _toctree.yml

Updated preprocessing and training

* updated preprocessing.mdx

start translation

* Update _toctree.yml

* Delete preprocessing.mdx

* Update _toctree.yml

* updated _toctree.yml

* added preprocessing

* Update _toctree.yml

* updated _toctree.yml

* undo

* Revert "undo"

This reverts commit 5d38d768752dc80918bf60ada9d185f98b742520.

* Revert "Revert "undo""

This reverts commit 8aa0830b587f915ca7d154ebca282b782e82bd92.

* Add the Italian translation of the file installation.mdx and edit _toctree

* Add the Italian translation of the file installation.mdx and edit _toctree

* add tasks/token_classification pt doc structure

* add tasks/token_classification pt doc translation

* add tasks/token_classification pt doc translation

This PR updates our Expert Acceleration Program image with a new image featuring our experts.

This is similar to our Transformers/README.md image update that has proven to be successful.

* Clean imports to fix test_fetcher

* Add dependencies printer

* Update utils/tests_fetcher.py
Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>

* Fix Perceiver import
Co-authored-by: lewtun <lewis.c.tunstall@gmail.com>

* Fix activation and layers in MLP head

* Remove unused import

* Start adding perceiver support for ONNX

* Fix pad token bug for fast tokenizers

* Fix formatting

* Make get_preprocesor more opinionated (processor priority, otherwise tokenizer/feature extractor)

* Clean docs format

* Minor cleanup following @sgugger's comments

* Fix typo in docs

* Fix another docs typo

* Fix one more typo in docs

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

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

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

* Allow from transformers import TypicalLogitsWarper

* Added TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

Added TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

* Add gated-silu to t5 architecture to support UL2

* Fix error message

* formatting

* formatting again

* refactor

* fix classnames in _init_weights

* remove is_gated

* add test

* fix test

* Try without the test?

* Add back the test.

* Improve error message.
Co-authored-by: Daniel Hesslow <daniel@lighton.ai>

* [trainer/deepspeed] load_best_model

* to sync with DS PR #1947

* simplify

* rework load_best_model test

* cleanup

* bump deepspeed>=0.6.5
Co-authored-by: Olatunji Ruwase <olruwase@microsoft.com>

* Implemented loss for training AudioFrameClassification

* reported changes in wav2vec2 main class and used make copies to propagate

* running black for code formatting

* Script for README

* Fix copies

* Complete error message

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>