Handling ONNX models with external data (#586)
* attempt at fixing saving onnx model with external data
* styling
* fix: `cache_dir` wasn't used when loading from transformers
* separate onnx_cache_dir argument from model's cache_dir
* we can now load large ONNX models by specifying external's data directory
* Fix saving external data for large models (seq2seq)
* fix saving external data for all ORT models
* make style
* typing
Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>
* apply suggestions
* make style
* apply suggestion
Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>
* export onnx models to separate subfolders when multiple models
* this should get us correct file_names but not correct subfolder!!
* export_models is only used for multiple submodels
* we can now load seq2seq model from local dir (multiple submodels)
* fix save_pretrained
* * `infer_onnx_filename` now returns a path
* refactor `_from_pretrained`
* try saving to single file
* didn't work, reverting "try saving to single file"
This reverts commit 0fdfd873ca6b55db1083365f69f783a7d9d4be49.
* add test for seq2seq model with external data
* quick fix
* try saving to a single file again
* Revert "quick fix"
This reverts commit 89e64eda59a2b0790087a354e65a468e4d51bee6.
* quick fix test
* save external data in a single file
* save_pretrained now moves model instead of copying from temp directory
* Revert "save_pretrained now moves model instead of copying from temp directory"
This reverts commit b315f8566591c1acecae1f90652660ff5ac0169f.
* add push to hub test
* add FORCE_ONNX_EXTERNAL_DATA env and faster test to push to hub
* quick fix
* we can now save and load large seq2seq models to hub + added test
* we no longer save to subfolders, as we use a singla file for external data
* make style
* apply same fixes to `modeling_decoder.py`
* apply same fixes to `modeling_ort.py`
* add tests
* fix auth token in tests
* add **kwargs to all `_save_pretrained`
* quick fix
* make style
* try reducing memory footprint when exporting onnx
* replace large seq2seq model with small on to make tests pass
* fix merge
* we no longer export models to subfolders. instead we regroup external data in a single data file
* util from last commit
* empty commit
* fix import
* add onnx utils
* fix import2
* better tests
* parameterized and skip order
Co-authored-by: Michael Benayoun <mickbenayoun@gmail.com>
Co-authored-by: Felix Marty <9808326+fxmarty@users.noreply.github.com>
