* Remove n_ctx from configs
* Fix GPTJ and OpenAIGPT, both are acceptable breaking changes as there are no configs such that it breaks
* Remove unecessary n_positions from TFOpenAIGPT
* Make gradient_checkpointing a training argument
* Update src/transformers/modeling_utils.py
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* Update src/transformers/configuration_utils.py
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
* Fix tests
* Style
* document Gradient Checkpointing as a performance feature
* Small rename
* PoC for not using the config
* Adapt BC to new PoC
* Forgot to save
* Rollout changes to all other models
* Fix typo
Co-authored-by: Stas Bekman <stas00@users.noreply.github.com>
Co-authored-by: Stas Bekman <stas@stason.org>
Cleaner and more scalable implementation of symbolic tracing with torch.fx, and provides support for new architectures:
- ALBERT
- DistilBERT
- MobileBERT
- MegatronBERT
- GPT2
- GPT Neo
Co-authored-by: Michael Benayoun <michael@huggingface.co>
* [WIP] Adding new parameter to `generate`: `max_time`.
Generation by tokens number is sometimes a bit clunky because we don't
know how many tokens are good enough or even how many tokens are in
the payload (for pipelines users for instance). This leads to hard
to understand behavior.
This PR proposes a new argument `max_time` which is a float of seconds
for the allowed time for `generate` to run on.
Ideally combinations of `max_tokens=None`, `max_time=2` could be used to
generate as many tokens as possible within time budget.
NB: Another possible approach consists of passing a callback to `generate`
putting the caller in charge of the actual decision of when to stop
generating tokens. It opens the door to 'which args should we pass'
to this callback. It's hard to imagine other use-cases for this
early stopping behavior than time (that are not already covered by
parameters of generate)
* Revamp with StoppingCriteria
* Removing deprecated mentions.
* Forgot arguments to stopping criteria.
* Readding max_length it's not just used as a stopping criteria.
* Default value for `stopping_criteria`.
* Address @patrickvonplaten comments.
- More docstrings
- Actual doc
- Include in global namespace
- Remove TF work.
* Put back `max_length` (deprecation different PR).
* Doc quality.
* Fixing old behavior without `stopping_criteria` but with `max_length`.
Making sure we don't break that in the future.
* Adding more tests for possible inconsistencies between
`max_length` and `stopping_criteria`.
* Fixing the torch imports.
* Update past_key_values in gpt2 (#9391)
* Update generation_utils, and rename some items
* Update modeling_gpt2 to avoid an error in gradient_checkpointing
* Remove 'reorder_cache' from util and add variations to XLNet, TransfoXL, GPT-2
* Change the location of '_reorder_cache' in modeling files
* Add '_reorder_cache' in modeling_ctrl
* Fix a bug of my last commit in CTRL
* Add '_reorder_cache' to GPT2DoubleHeadsModel
* Manage 'use_cache' in config of test_modeling_gpt2
* Clean up the doc string
* Update src/transformers/models/gpt2/modeling_gpt2.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* Fix the doc string (GPT-2, CTRL)
* improve gradient_checkpointing_behavior
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* 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
* Fix passing token_type_ids during GPT2DoubleHeadsModel.generate() if used
and for GPT2LMHeadModel too
* Update tests to check token_type_ids usage in GPT2 models
* 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
* GPT2 gradient checkpointing
* find_unused_parameters removed if checkpointing
* find_unused_parameters removed if checkpointing
* Update src/transformers/configuration_gpt2.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
* Added a test for generation with checkpointing
* Update src/transformers/configuration_gpt2.py
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Sylvain Gugger <35901082+sgugger@users.noreply.github.com>
* cleanup torch unittests: part 2
* remove trailing comma added by isort, and which breaks flake
* one more comma
* revert odd balls
* part 3: odd cases
* more ["key"] -> .key refactoring
* .numpy() is not needed
* more unncessary .numpy() removed
* more simplification
* 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
* add first copy past test to tf 2 generate
* add tf top_k_top_p_filter fn
* add generate function for TF
* add generate function for TF
* implemented generate for all models expect transfoXL
* implemented generate for all models expect transfoXL
* implemented generate for all models expect transfoXL
* make style
* change permission of test file to correct ones
* delete ipdb
* delete ipdb
* fix bug and finish simple gpt2 integration test
* clean test file
* clean test file
* make style
* make style
* make style
* make style
* change import style
* change import style
* make style
* make style
* add decorators
* add decorators
* fix tf ctrl bug dim => axis in TF
* make style
* make style
* refactored test file
* refactored test file
* take out test_torch_tf_conversion if nothing is defined
* take out test_torch_tf_conversion if nothing is defined
* remove useless files
* remove useless files
* fix conflicts
* fix conflicts
* fix conflicts
* fix conflicts
* fix conflicts
* solve conflicts
* solve conflicts
* fix conflicts
* fix conflicts
* merge conflicts
* delete ipdb
* exposed top_k_top_p_filtering fns
* delete weirdly created w! file
* add comment to test tf common modeling
* fix conflicts
* fix conflicts
* make style
* merge conflicts
* make style
* change tf.tensor.shape to shape_list(tensor)
