GitLab

* Fix initialization for missing parameters in `from_pretrained` under ZeRO-3

* Test initialization for missing parameters under ZeRO-3

* Add more tests

* Only enable deepspeed context for per-module level parameters

* Enable deepspeed context only once

* Move class definition inside test case body

* add deepspeed scheduled test for amd

* fix image

* add dockerfile

* add comment

* enable tests

* trigger

* remove trigger for this branch

* trigger

* change runner env to trigger the docker build image test

* use new docker image

* remove test suffix from docker image tag

* replace test docker image with original image

* push new image

* Trigger

* add back amd tests

* fix typo

* add amd tests back

* fix

* comment until docker image build scheduled test fix

* remove deprecated deepspeed build option

* upgrade torch

* update docker & make tests pass

* Update docker/transformers-pytorch-deepspeed-amd-gpu/Dockerfile

* fix

* tmp disable test

* precompile deepspeed to avoid timeout during tests

* fix comment

* trigger deepspeed tests with new image

* comment tests

* trigger

* add sklearn dependency to fix slow tests

* enable back other tests

* final update

---------
Co-authored-by: Felix Marty <felix@hf.co>
Co-authored-by: Félix Marty <9808326+fxmarty@users.noreply.github.com>
Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>

* fix failing DeepSpeed CI tests due to `safetensors` being default

* debug

* remove debug statements

* resolve comments

* Update test_deepspeed.py

* fix the deepspeed tests

* resolve comment

deepspeed resume from ckpt fixes and adding support for deepspeed optimizer and HF scheduler (#25863)

* Add support for deepspeed optimizer and HF scheduler

* fix bug

* fix the import

* fix issue with deepspeed scheduler saving for hf optim + hf scheduler scenario

* fix loading of hf scheduler when loading deepspeed checkpoint

* fix import of `DeepSpeedSchedulerWrapper`

* add tests

* add the comment and skip the failing tests

* address comment

* move deepspeed to `lib_integrations.deepspeed`

* more refactor

* oops

* fix slow tests

* Fix docs

* fix docs

* addess feedback

* address feedback

* final modifs for PEFT

* fixup

* ok now

* trigger CI

* trigger CI again

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

* import from `integrations`

* address feedback

* revert removal of `deepspeed` module

* revert removal of `deepspeed` module

* fix conflicts

* ooops

* oops

* add deprecation warning

* place it on the top

* put `FutureWarning`

* fix conflicts with not_doctested.txt

* add back `bitsandbytes` module with a depr warning

* fix

* fix

* fixup

* oops

* fix doctests

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

* mixed precision support via accelerate

* fix issues

* fix for the sharded ddp case

* fix flax and tf failing tests

* `refactor the place to create `Accelerator` object

* move ddp prep to accelerate

* fix 😅

* resolving comments

* move fsdp handling to accelerate

* fixex

* fix saving

* shift torch dynamo handling to accelerate

* shift deepspeed integration and save & load utils to accelerate

* fix accelerate launcher support

* oops

* fix 🐛

* save ckpt fix

* Trigger CI

* nasty 🐛 😅

* as deepspeed needs grad_acc fixes, transfer grad_acc to accelerate

* make tests happy

* quality ✨

* loss tracked needs to account for grad_acc

* fixing the deepspeed tests

* quality ✨

* 😅😅😅

* tests 😡

* quality ✨



* Trigger CI

* resolve comments and fix the issue with the previous merge from branch

* Trigger CI

* accelerate took over deepspeed integration

---------
Co-authored-by: Stas Bekman <stas@stason.org>

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

* [deepspeed] offload + non-cpuadam optimizer exception

* flip

* revert min version

* [deepspeed tests] fix issues introduced by #21700

* fix

* fix

[tests] report_to none

* 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

* Refine BF16 check in CPU/GPU

* Fixes

* Renames

* [deepspeed] fix load_best_model test

* [deepspeed] add state reset on unittest tearDown

* [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>

* model zoo take 2

* add deberta

* new param for zero2

* doc update

* doc update

* add layoutlm

* bump deepspeed

* add deberta-v2, funnel, longformer

* new models

* style

* add t5_v1

* update TAPAS status

* reorg problematic models

* move doc to another PR

* style

* fix checkpoint check test

* making progress on more models running

* cleanup

* new version

* cleanup

* [trainer / deepspeed] fix hyperparameter_search

* require optuna

* style

* oops

* add dep in the right place

* create deepspeed-testing dep group

* Trigger CI

* Split file_utils in several submodules

* Fixes

* Add back more objects

* More fixes

* Who exactly decided to import that from there?

* Second suggestion to code with code review

* Revert wront move

* Fix imports

* Adapt all imports

* Adapt all imports everywhere

* Revert this import, will fix in a separate commit

* [WIP] add support for bf16 mode

* prep for bf16

* prep for bf16

* fix; zero2/bf16 is ok

* check bf16 is available

* test fixes

* enable zero3_bf16

* config files

* docs

* split stage_dtype; merge back to non-dtype-specific config file

* fix doc

* cleanup

* cleanup

* bfloat16 => bf16 to match the PR changes

* s/zero_gather_fp16_weights_on_model_save/zero_gather_16bit_weights_on_model_save/; s/save_fp16_model/save_16bit_model/

* test fixes/skipping

* move

* fix

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

* backticks

* cleanup

* cleanup

* cleanup

* new version

* add note about grad accum in bf16
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>

* fix deepspeed tests

* style

* more fixes

* Per-folder tests reorganization
Co-authored-by: sgugger <sylvain.gugger@gmail.com>
Co-authored-by: Stas Bekman <stas@stason.org>

* [deepspeed] fix a bug in a test

* consistency

* [deepspeed] fix load_best_model_at_end

* try with pull_request_target

* revert: try with pull_request_target

* style

* add test

* cleanup

* [deepspeed] zero inference

* only z3 makes sense for inference

* fix and style

* docs

* rework

* fix test

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

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

* defer to DS_TEST_PORT if set

* style
Co-authored-by: Stas Bekman <stas@stason.org>

* Use DS callable API to allow hf_scheduler + ds_optimizer

* Preserve backward-compatibility

* Restore backward compatibility

* Tweak arg positioning

* Tweak arg positioning

* bump the required version

* Undo indent

* Update src/transformers/trainer.py

* style
Co-authored-by: Stas Bekman <stas@stason.org>
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>

* zero_to_fp32 tests

* args change

* remove unnecessary work

* use transformers.trainer_utils.get_last_checkpoint

* document the new features

* cleanup

* wip

* fix fsmt

* add bert

* cleanup

* add xlm-roberta

* electra works

* cleanup

* sync

* split off the model zoo tests

* cleanup

* cleanup

* cleanup

* cleanup

* reformat

* cleanup

* casing

* deepspeed>=0.4.3

* adjust distilbert

* 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>

* bug fixes and a rename

* add extended DDP test

* wip

* wip - but working with https://github.com/microsoft/DeepSpeed/pull/1044

* cleanup

* workaround

* working 5/8 modes

* solve fp32 distributed zero3

* style

* sync

* sync

* rework

* deprecation

* cleanup

* https://github.com/microsoft/DeepSpeed/pull/1044

 pr was merged

* clean up

* add a guide

* more prose

* more prose

* fix

* more prose

* sub_group_size was too big

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

* refactor

* bug fix

* make the true check explicit

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

* replace deprecated config

* sub_group_size was too big

* complete deprecation removal

* wip

* add mismatch validation + test

* renames

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

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

* add nvme skip rule

* fix

* move code and docs

* style

* moved

* restore

* decouple DeepSpeedConfigHF from Trainer

* add LoggingLevel ctx manager; add new test

* cleanup

* add docs

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

* implemented suggested renames

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