GitLab

* Ensure input_embeds and image_features are the same dtype in autocast

* Fix nans in half precision llava-next and fix autocasting behavior.

* Fix styling issues.

* fix randn newline instantiation

* fix broken slow llava test

* Fix llava next init.

* fix styling issues

* [run-slow]llava,llava_next

* fix styling issues

* Gemma: only display act. warning when necessary

This is a nit PR, but I was confused. I got the warning even after I
had changed `hidden_act` to `gelu_pytorch_tanh`, telling me that I
was using the "legacy" `gelu_pytorch_tanh`.

Another option is to keep the warning but change the message to say
something like "`hidden_act` is ignored, please use `hidden_activation`
instead. Setting Gemma's activation function to `gelu_pytorch_tanh`".

* Change message, and set `config.hidden_activation`

* Temporarily silence warnings in apply_chat_template until we can properly deprecate default chat templates

* make fixup

* Move the default chat template warning into apply_chat_template itself

* make fixup

* fix marian model coversion

* uncomment that line

* remove unnecessary code

* revert tie_weights, doesn't hurt

* move scaling to nn.Module

* let the test be here for now (need to fix)

* failing tests

* last failing models

* Revert commit 4c14817f38

* clean-up

* oops forgot

* codestyle

* raise NotImplemented when possible

* Update tests/test_modeling_common.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* skip tests in respective modeling files

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* fix

* add test

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

* fix doctest

* fix torch doctest

* make CI happy

* raise error

* make fixup

* added chat templating support for keydataset in generation pipeline

* fixed and improved test

* fix formatting test failures

* Fix tests

* Fix tests

* add_blip_get_multimodal_feautres

* Fix docstring error

* reimplement get_multimodal_features

* fix error

* recheck code quality

* add new necessary tests

* fix seq2seq data collator to respect the given padding strategy

further added tests for the seq2seq data collator in the style of the `data_collator_for_token_classification` (pt, tf, np)

* formatting and change bool equals "==" to "is"

* add missed return types in tests

* update numpy test as it can handle unequal shapes, not like pt or tf

* feat: support for dinov2

* feat: support for depth_anything

* feat: support for efficientformer

* feat: support for bert (is this right?)

* update: embedding split

* remove: empty string

* feat: support for align

* fix: copies

* fix: QAQBertEmbeddings

* fix: more consistency issues

* revert: support for effientformer

* feat: support for altclip

* feat: support for blip_text

* support for ChineseCLIP

* feat: support for depth anything

* feat: support for dpt

* feat: support for dpt

* feat: support for git

* feat: support for groupvit

* update: format

* fix: support for clip

* fix: consistency

* feat: support for pvt

* feat: support for vit_msn

* fix: consistency

* fix: other copies

* remove: device transfer

* revert: in-place add

* update: support for align

* update: support for bert

* update: support for Chinese CLIP

* revert: changes to efficientformer

* update: support for dpt

* update: support for efficientformer

* revert: changes to git

* revert: changes to groupvit

* revert: changes to roc_bert

* update: support for vit_msn

* revert: changes to dpt

* remove: extra space

* style: extra space

pass use_cache in kwargs

* Include safetensors

* Cleanup

* Reenable SDPA's FA2 during training with torch.compile

* fix Olmo's SDPA FA2 dispatching too

* update formatting

* improved SDPA comment

* formatting and explanatory comment

* is_causal if statement to one-liner

* Pass attn_implementation when using AutoXXX.from_config

* Fix

* Allow boolean FSDP options in fsdp_config

* Use lower() to be safe

* Fixed SegGptImageProcessor to handle 2D and 3D prompt mask inputs

* Added new test to check prompt mask equivalence

* New proposal

* Better proposal

* Removed unnecessary method

* Updated seggpt docs

* Introduced do_convert_rgb

* nits

* Decode b64encode and encodebytes strings

* Remove conditional encode -- image is always a string

Fix GroundingDINO, DPR after BET SDPA update

* Enable instantiating model with pretrained backbone weights

* Clarify pretrained import

* Use load_backbone instead

* Add backbone_kwargs to config

* Fix up

* Add tests

* Tidy up

* Enable instantiating model with pretrained backbone weights

* Update tests so backbone checkpoint isn't passed in

* Clarify pretrained import

* Update configs - docs and validation check

* Update src/transformers/utils/backbone_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Clarify exception message

* Update config init in tests

* Add test for when use_timm_backbone=True

* Use load_backbone instead

* Add use_timm_backbone to the model configs

* Add backbone_kwargs to config

* Pass kwargs to constructors

* Draft

* Fix tests

* Add back timm - weight naming

* More tidying up

* Whoops

* Tidy up

* Handle when kwargs are none

* Update tests

* Revert test changes

* Deformable detr test - don't use default

* Don't mutate; correct model attributes

* Add some clarifying comments

* nit - grammar is hard

---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Remove skipping logic now that set_epoch exists

* Working version, clean

* Adding SDPA support for BERT

* Using the proper input name for testing model input in inference()

* Adding documentation for SDPA in BERT model page

* Use the stable link for the documentation

* Adding a gate to only call .contiguous() for torch < 2.2.0

* Additions and fixes to the documentation

* Minor updates to documentation

* Adding extra requirements needed for the contiguous() bug

* Adding "Adapted from" in plcae of the "Copied from"

* Add benchmark speedup tables to the documentation

* Minor fixes to the documentation

* Use ClapText as a replacemenet for Bert in the Copied-From

* Some more fixes for the fix-copies references

* Overriding the test_eager_matches_sdpa_generate in bert tests to not load with low_cpu_mem_usage

[test all]

* Undo changes to separate test

* Refactored SDPA self attention code for KV projections

* Change use_sdpa to attn_implementation

* Fix test_sdpa_can_dispatch_on_flash by preparing input (required for MultipleChoice models)

* Update modeling_utils/dtype_byte_size to handle float8 types

* Add a test for dtype_byte_size

* Format

* Fix bool

Fix the `bitsandbytes` error when some modules are not properly offloaded.

Update quantizer_eetq.py

ensure popular quant methods are supported

* Introduce saveable callbacks

* Add note

* Test for non-present and flag

* Support early stopping and refusing to train further

* Update docstring

* More saving

* Import oopsie

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Make it go through TrainerArguments

* Document

* Fix test

* Apply suggestions from code review
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Rework to allow for duplicates

* CLean

* Fix failing tests

---------
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Added WSD scheduler.

* Added tests.

* Fixed errors.

* Fix formatting.

* CI fixes.

* first modeling code

* make repository

* still WIP

* update model

* add tests

* add latest change

* clean docstrings and copied from

* update docstrings md and readme

* correct chroma function

* correct copied from and remove unreleated test

* add doc to toctree

* correct imports

* add convert script to notdoctested

* Add suggestion from Sanchit
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* correct get_uncoditional_inputs docstrings

* modify README according to SANCHIT feedback

* add chroma to audio utils

* clean librosa and torchaudio hard dependencies

* fix FE

* refactor audio decoder -> audio encoder for consistency with previous musicgen

* refactor conditional -> encoder

* modify sampling rate logics

* modify license at the beginning

* refactor all_self_attns->all_attentions

* remove ignore copy from causallm generate

* add copied from for from_sub_models

* fix make copies

* add warning if audio is truncated

* add copied from where relevant

* remove artefact

* fix convert script

* fix torchaudio and FE

* modify chroma method according to feedback-> better naming

* refactor input_values->input_features

* refactor input_values->input_features and fix import fe

* add input_features to docstrigs

* correct inputs_embeds logics

* remove dtype conversion

* refactor _prepare_conditional_hidden_states_kwargs_for_generation ->_prepare_encoder_hidden_states_kwargs_for_generation

* change warning for chroma length

* Update src/transformers/models/musicgen_melody/convert_musicgen_melody_transformers.py
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* change way to save wav, using soundfile

* correct docs and change to soundfile

* fix import

* fix init proj layers

* add draft training

* fix cross entropy

* clean loss computation

* fix labels

* remove line breaks from md

* fix issue with docstrings

* add FE suggestions

* improve is in logics and remove useless imports

* remove custom from_pretrained

* simplify docstring code

* add suggestions for modeling tests

* make style

* update converting script with sanity check

* remove encoder attention mask from conditional generation

* replace musicgen melody checkpoints with official orga

* rename ylacombe->facebook in checkpoints

* fix copies

* remove unecessary warning

* add shape in code docstrings

* add files to slow doc tests

* fix md bug and add md to not_tested

* make fix-copies

* fix hidden states test and batching

* update training code

* add training tests for melody

* add training for o.g musicgen

* fix copied from

* remove final todos

* make style

* fix style

* add suggestions from review

* add ref to the original loss computation code

* rename method + fix labels in tests

* make style

---------
Co-authored-by: Sanchit Gandhi <93869735+sanchit-gandhi@users.noreply.github.com>

* Use EAFP principle to prevent crash with third parties

* Remove leftover debugging code

* Add info-level logger message

* Fix SigLip classification doctest

* Remove extra line

* Update src/transformers/models/siglip/modeling_siglip.py

* fix codellama conversion

* nit

* Non blocking support

* Check for optimization

* Doc

* Check removing flag for torch

* LLM oops

* Getting there...

* More discoveries

* Change

* Clean up and prettify

* Logic check

* Not

save_safetensor=True is default as of release 4.35.0, which then
required TPU hotfix https://github.com/huggingface/transformers/pull/27799
(issue https://github.com/huggingface/transformers/issues/27578).
However, when the flag save_safetensor is set to False (compatibility mode),
moving the model to CPU causes generation of too many graphs
during checkpoint https://github.com/huggingface/transformers/issues/28438.
This PR disable moving of model to CPU when save_safetensor=False.