Tony Wu f33a0cebb3

Add ColPali to 🤗 transformers (#33736) ...

* feat: run `add-new-model-like`

* feat: add paligemma code with "copied from"

* feat: add ColPaliProcessor

* feat: add ColPaliModel

* feat: add ColPaliConfig

* feat: rename `ColPaliForConditionalGeneration` to `ColPaliModel`

* fixup modeling colpali

* fix: fix root import shortcuts

* fix: fix `modeling_auto` dict

* feat: comment out ColPali test file

* fix: fix typos from `add-new-model-like`

* feat: explicit the forward input args

* feat: move everything to `modular_colpali.py`

* fix: put back ColPaliProcesor

* feat: add auto-generated files

* fix: run `fix-copies`

* fix: remove DOCStRING constants to make modular converter work

* fix: fix typo + modular converter

* fix: add missing imports

* feat: no more errors when loading ColPaliModel

* fix: remove unused args in forward + tweak doc

* feat: rename `ColPaliModel` to `ColPaliForRetrieval`

* fix: apply `fix-copies`

* feat: add ColPaliProcessor to `modular_colpali`

* fix: run make quality + make style

* fix: remove duplicate line in configuration_auto

* feat: make ColPaliModel inehrit from PaliGemmaForConditionalGeneration

* fix: tweak and use ColPaliConfig

* feat: rename `score` to `post_process_retrieval`

* build: run modular formatter + make style

* feat: convert colpali weights + fixes

* feat: remove old weight converter file

* feat: add and validate tests

* feat: replace harcoded path to "vidore/colpali-v1.2-hf" in tests

* fix: add bfloat16 conversion in weight converter

* feat: replace pytest with unittest in modeling colpali test

* feat: add sanity check for weight conversion (doesn't work yet)

* feat: add shape sanity check in weigth converter

* feat: make ColPaliProcessor args explicit

* doc: add doc for ColPali

* fix: trying to fix output mismatch

* feat: tweaks

* fix: ColPaliModelOutput inherits from ModelOutput instead of PaliGemmaCausalLMOutputWithPast

* fix: address comments on PR

* fix: adapt tests to the Hf norm

* wip: try things

* feat: add `__call__` method to `ColPaliProcessor`

* feat: remove need for dummy image in `process_queries`

* build: run new modular converter

* fix: fix incorrect method override

* Fix tests, processing, modular, convert

* fix tokenization auto

* hotfix: manually fix processor -> fixme once convert modular is fixed

* fix: convert weights working

* feat: rename and improve convert weight script

* feat: tweaks

* fest: remove `device` input for `post_process_retrieval`

* refactor: remove unused `get_torch_device`

* Fix all tests

* docs: update ColPali model doc

* wip: fix convert weights to hf

* fix logging modular

* docs: add acknowledgements in model doc

* docs: add missing docstring to ColPaliProcessor

* docs: tweak

* docs: add doc for `ColPaliForRetrievalOutput.forward`

* feat: add modifications from colpali-engine v0.3.2 in ColPaliProcessor

* fix: fix and upload colapli hf weights

* refactor: rename `post_process_retrieval` to `score_retrieval`

* fix: fix wrong typing for `score_retrieval`

* test: add integration test for ColPali

* chore: rerun convert modular

* build: fix root imports

* Update docs/source/en/index.md

Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>

* fix: address PR comments

* wip: reduce the prediction gap in weight conversion

* docs: add comment in weight conversion script

* docs: add example for `ColPaliForRetrieval.forward`

* tests: change dataset path to the new one in hf-internal

* fix: colpali weight conversion works

* test: add fine-grained check for ColPali integration test

* fix: fix typos in convert weight script

* docs: move input docstring in a variable

* fix: remove hardcoded torch device in test

* fix: run the new modular refactor

* docs: fix python example for ColPali

* feat: add option to choose `score_retrieval`'s output dtype and device

* docs: update doc for `score_retrieval`

* feat: add `patch_size` property in ColPali model

* chore: run `make fix-copies`

* docs: update description for ColPali cookbooks

* fix: remove `ignore_index` methods

* feat: remove non-transformers specific methods

* feat: update `__init__.py` to new hf format

* fix: fix root imports in transformers

* feat: remove ColPali's inheritance from PaliGemma

* Fix CI issues

* nit remove prints

* feat: remove ColPali config and model from `modular_colpali.py`

* feat: add `ColPaliPreTrainedModel` and update modeling and configuration code

* fix: fix auto-removed imports in root `__init__.py`

* fix: various fixes

* fix: fix `_init_weight`

* temp: comment `AutoModel.from_config` for experiments

* fix: add missing `output_attentions` arg in ColPali's forward

* fix: fix `resize_token_embeddings`

* fix: make `input_ids` optional in forward

* feat: rename `projection_layer` to `embedding_proj_layer`

* wip: fix convert colpali weight script

* fix tests and convert weights from original repo

* fix unprotected import

* fix unprotected torch import

* fix style

* change vlm_backbone_config to vlm_config

* fix unprotected import in modular this time

* fix: load config from Hub + tweaks in convert weight script

* docs: move example usage from model docstring to model markdown

* docs: fix input docstring for ColPali's forward method

* fix: use `sub_configs` for ColPaliConfig

* fix: remove non-needed sanity checks in weight conversion script + tweaks

* fix: fix issue with `replace_return_docstrings` in ColPali's `forward`

* docs: update docstring for `ColPaliConfig`

* test: change model path in ColPali test

* fix: fix ColPaliConfig

* fix: fix weight conversion script

* test: fix expected weights for ColPali model

* docs: update ColPali markdown

* docs: fix minor typo in ColPaliProcessor

* Fix tests and add _no_split_modules

* add text_config to colpali config

* [run slow] colpali

* move inputs to torch_device in integration test

* skip test_model_parallelism

* docs: clarify quickstart snippet in ColPali's model card

* docs: update ColPali's model card

---------

Co-authored-by: yonigozlan <yoni.gozlan@huggingface.co>
Co-authored-by: Yoni Gozlan <74535834+yonigozlan@users.noreply.github.com>

2024-12-17 11:26:43 +01:00

0 lines

Python

Raw Blame History