* Change the way tracing happens, enabling dynamic axes out of the box
* Update the tests and modeling xlnet
* Add the non recoding of leaf modules to avoid recording more values for the methods to record than what will be seen at tracing time (which would otherwise desynchronize the recorded values and the values that need to be given to the proxies during tracing, causing errors).
* Comments and making tracing work for gpt-j and xlnet
* Refactore things related to num_choices (and batch_size, sequence_length)
* Update fx to work on PyTorch 1.10
* Postpone autowrap_function feature usage for later
* Add copyrights
* Remove unnecessary file
* Fix issue with add_new_model_like
* Apply suggestions
* Symbolic trace dynamic axes support for BERT like models (albert, bert, distilbert, mobilebert, electra, megatron-bert)
* Sanity checks before tracing that make sure the model to trace is supported
* Adapted to PyTorch 1.9
Co-authored-by: Michael Benayoun <michael@huggingface.co>
* add past_key_values
* add use_cache option
* make mask before cutting ids
* adjust position_ids according to past_key_values
* flatten past_key_values
* fix positional embeds
* fix _reorder_cache
* set use_cache to false when not decoder, fix attention mask init
* add test for caching
* add past_key_values for Roberta
* fix position embeds
* add caching test for roberta
* add doc
* make style
* doc, fix attention mask, test
* small fixes
* adress patrick's comments
* input_ids shouldn't start with pad token
* use_cache only when decoder
* make consistent with bert
* make copies consistent
* add use_cache to encoder
* add past_key_values to tapas attention
* apply suggestions from code review
* make coppies consistent
* add attn mask in tests
* remove copied from longformer
* apply suggestions from code review
* fix bart test
* nit
* simplify model outputs
* fix doc
* fix output ordering
* Support BERT relative position embeddings
* Fix typo in README.md
* Address review comment
* Fix failing tests
* [tiny] Fix style_doc.py check by adding an empty line to configuration_bert.py
* make fix copies
* fix configs of electra and albert and fix longformer
* remove copy statement from longformer
* fix albert
* fix electra
* Add bert variants forward tests for various position embeddings
* [tiny] Fix style for test_modeling_bert.py
* improve docstring
* [tiny] improve docstring and remove unnecessary dependency
* [tiny] Remove unused import
* re-add to ALBERT
* make embeddings work for ALBERT
* add test for albert
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Put models in subfolders
* Styling
* Fix imports in tests
* More fixes in test imports
* Sneaky hidden imports
* Fix imports in doc files
* More sneaky imports
* Finish fixing tests
* Fix examples
* Fix path for copies
* More fixes for examples
* Fix dummy files
* More fixes for example
* More model import fixes
* Is this why you're unhappy GitHub?
* Fix imports in conver command
* Use the CI to identify failing tests
* Remove from all examples and tests
* More default switch
* Fixes
* More test fixes
* More fixes
* Last fixes hopefully
* Use the CI to identify failing tests
* Remove from all examples and tests
* More default switch
* Fixes
* More test fixes
* More fixes
* Last fixes hopefully
* Run on the real suite
* Fix slow tests
* add training tests
* correct longformer
* fix docs
* fix some tests
* fix some more train tests
* remove ipdb
* fix multiple edge case model training
* fix funnel and prophetnet
* clean gpt models
* undo renaming of albert
* first draft
* show design proposition for new generate method
* up
* make better readable
* make first version
* gpt2 tests pass
* make beam search for gpt2 work
* add first encoder-decoder code
* delete typo
* make t5 work
* save indermediate
* make bart work with beam search
* finish beam search bart / t5
* add default kwargs
* make more tests pass
* fix no bad words sampler
* some fixes and tests for all distribution processors
* fix test
* fix rag slow tests
* merge to master
* add nograd to generate
* make all slow tests pass
* speed up generate
* fix edge case bug
* small fix
* correct typo
* add type hints and docstrings
* fix typos in tests
* add beam search tests
* add tests for beam scorer
* fix test rag
* finish beam search tests
* move generation tests in seperate file
* fix generation tests
* more tests
* add aggressive generation tests
* fix tests
* add gpt2 sample test
* add more docstring
* add more docs
* finish doc strings
* apply some more of sylvains and sams comments
* fix some typos
* make fix copies
* apply lysandres and sylvains comments
* final corrections on examples
* small fix for reformer
* Feed forward chunking for Distilbert & Albert
* Added ff chunking for many other models
* Change model signature
* Added chunking for XLM
* Cleaned up by removing some variables.
* remove test_chunking flag
Co-authored-by: patrickvonplaten <patrick.v.platen@gmail.com>
* Chunked feed forward for Bert
This is an initial implementation to test applying feed forward chunking for BERT.
Will need additional modifications based on output and benchmark results.
* Black and cleanup
* Feed forward chunking in BertLayer class.
* Isort
* add chunking for all models
* fix docs
* Fix typo
Co-authored-by: patrickvonplaten <patrick.v.platen@gmail.com>
* improve unit tests
this is a sample of one test according to the request in https://github.com/huggingface/transformers/issues/5973
before I apply it to the rest
* batch 1
* batch 2
* batch 3
* batch 4
* batch 5
* style
* non-tf template
* last deletion of check_loss_output
* Kill model archive maps
* Fixup
* Also kill model_archive_map for MaskedBertPreTrainedModel
* Unhook config_archive_map
* Tokenizers: align with model id changes
* make style && make quality
* Fix CI
There's an inconsistency right now where:
- we load some models into CACHE_DIR
- and some models in the default cache
- and often, in both for the same models
When running the RUN_SLOW tests, this takes a lot of disk space, time, and bandwidth.
I'd rather always use the default cache
PyTorch < 1.3 requires multiplication operands to be of the same type.
This was violated when using default attention mask (i.e.,
attention_mask=None in arguments) given BERT in the decoder mode.
In particular, this was breaking Model2Model and made tutorial
from the quickstart failing.
I suspect the wrapper classes were created in order to prevent the
abstract base class (TF)CommonModelTester from being included in test
discovery and running, because that would fail.
I solved this by replacing the abstract base class with a mixin.
Code changes are just de-indenting and automatic reformattings
performed by black to use the extra line space.
This construct isn't used anymore these days.
Running python tests/test_foo.py puts the tests/ directory on
PYTHONPATH, which isn't representative of how we run tests.
Use python -m unittest tests/test_foo.py instead.
