GitLab

* Script & Manual edition

* Update

Co-authored-by: fxmarty <9808326+fxmarty@users.noreply.github.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>

* Draft version of new KV Caching

This should allow Attention Sinks (https://github.com/tomaarsen/attention_sinks)
/ StreamingLLM (https://arxiv.org/abs/2309.17453) to be easily implemented
in a third-party or in transformers directly

* Address numerous PR suggestions

1. Move layer_idx from cache to ...Attention. Removes confusing set_layer_idx magic.
2. Always convert past_key_values to Cache instance at the start of ...Attention, removes all other isinstance calls.
3. Remove __bool__ and __getitem__ magic as they're confusing.
4. past_key_values.update(key, value, idx) now returns key, value.
5. Add use_legacy_cache flag, defaults to None, i.e. Falsey. This breaks generate for now, until 1) the cache is used is generate() or 2) use_legacy_cache is defaulted to True in generate() until we change it in another PR.
6. Separate key_cache and value_cache.

Some work is still needed to see if the SinkCache can conveniently be implemented with just one update method.

* Implement the SinkCache through backward+forward rotations

* Integrate (Sink)Cache with Llama FA2

* Set use_legacy_cache=True as default, allows for test passes

* Move from/to_legacy_cache to ...Model class

* Undo unnecessary newline change

* Remove copy utility from deprecated OpenLlama

* Match import style

* manual rebase with main

* Cache class working with generate (#1)

* Draft version of new KV Caching

This should allow Attention Sinks (https://github.com/tomaarsen/attention_sinks)
/ StreamingLLM (https://arxiv.org/abs/2309.17453

) to be easily implemented
in a third-party or in transformers directly

* Address numerous PR suggestions

1. Move layer_idx from cache to ...Attention. Removes confusing set_layer_idx magic.
2. Always convert past_key_values to Cache instance at the start of ...Attention, removes all other isinstance calls.
3. Remove __bool__ and __getitem__ magic as they're confusing.
4. past_key_values.update(key, value, idx) now returns key, value.
5. Add use_legacy_cache flag, defaults to None, i.e. Falsey. This breaks generate for now, until 1) the cache is used is generate() or 2) use_legacy_cache is defaulted to True in generate() until we change it in another PR.
6. Separate key_cache and value_cache.

Some work is still needed to see if the SinkCache can conveniently be implemented with just one update method.

* Integrate (Sink)Cache with Llama FA2

* Move from/to_legacy_cache to ...Model class

* Undo unnecessary newline change

* Match import style

* working generate

* Add tests; Simplify code; Apply changes to Mistral and Persimmon

* fix rebase mess

* a few more manual fixes

* last manual fix

* propagate changes to phi

* upgrade test

* add use_legacy_cache docstring; beef up tests

* reintroduce unwanted deletes

---------
Co-authored-by: Tom Aarsen <Cubiegamedev@gmail.com>

* move import

* add default to model_kwargs.get('use_legacy_cache')

* correct failing test

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* apply PR suggestions

* fix failing test

* Apply suggestions from code review
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>
Co-authored-by: Tom Aarsen <37621491+tomaarsen@users.noreply.github.com>

* PR comments

* tmp commit

* add docstrings

* more tests, more docstrings, add to docs

* derp

* tmp commit

* tmp dbg

* more dbg

* fix beam search bug

* cache can be a list of tuples in some models

* fix group beam search

* all but sinkcache integration tests

* fix sink cache and add hard integration test

* now also compatible with input_embeds input

* PR comments

* add Cache support to Phi+FA2

* make fixup

---------
Co-authored-by: Joao Gante <joao@huggingface.co>
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Rename index.mdx to index.md

* With saved modifs

* Address review comment

* Treat all files

* .mdx -> .md

* Remove special char

* Update utils/tests_fetcher.py
Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

---------
Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

* haha text go brrr (but in gradio)

* haha tokens go brrrr

* feat: add min new length logit processor

* test: add min new length logit processor

* docs: add MinNewTokensLengthLogitsProcessor

* feat: import MinNewTokensLengthLogitsProcessor

* fix: update pytorch dummy objects

* refactor & fix: rename attributes and var and get rid of dynamic attribute

* tests: align test with new interface

* docs: fix typo

* docs: minor clarification

* Empty-Commit

* empty commit

* run automated quality edits
Co-authored-by: Joao Gante <joao@huggingface.co>

* move generation_*.py src files into generation/*.py

* populate generation.__init__ with lazy loading

* move imports and references from generation.xxx.object to generation.object

Adding the state-of-the-art contrastive search decoding methods for the codebase of generation_utils.py (#19477)

* add: the contrastive search for generaton_utils

* add: testing scripts for contrastive search under examples/text-generation

* update the quality of codes

* revise the docstring; make the generation_contrastive_search.py scripts;

* revise the examples/pytorch/text-generation/run_generation_contrastive_search.py to the auto-APIs format

* revise the necessary documents

* fix: revise the docstring of generation_contrastive_search.py

* Fix the code indentation

* fix: revise the nits and examples in contrastive_search docstring.

* fix the copyright

* delete generation_contrastive_search.py

* revise the logic in contrastive_search

* update the intergration test and the docstring

* run the tests over

* add the slow decorate to the contrastive_search intergrate test

* add more test

* do the style, quality, consistency checks

* Allow from transformers import TypicalLogitsWarper

* Added TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

* Allow from transformers import TypicalLogitsWarper

Added TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

Allow from transformers import TypicalLogitsWarper

* refactor TF beam search

* refactored generate can now properly use attention masks

* add force bos/eos logit processors

* Reorganize doc for multilingual support

* Fix style

* Style

* Toc trees

* Adapt templates

* Big file_utils cleanup

* This one still needs to be treated separately

* added classes to get started with constrained beam search

* in progress, think i can directly force tokens now but not yet with the round robin

* think now i have total control, now need to code the bank selection

* technically works as desired, need to optimize and fix design choices leading to undersirable outputs

* complete PR #1 without disjunctive decoding

* removed incorrect tests

* Delete k.txt

* Delete test.py

* Delete test.sh

* revert changes to test scripts

* genutils

* full implementation with testing, no disjunctive yet

* shifted docs

* passing all tests realistically ran locally

* removing accidentally included print statements

* fixed source of error in initial PR test

* fixing the get_device() vs device trap

* fixed documentation docstrings about constrained_beam_search

* fixed tests having failing for Speech2TextModel's floating point inputs

* fix cuda long tensor

* added examples and testing for them and founx & fixed a bug in beam_search and constrained_beam_search

* deleted accidentally added test halting code with assert False

* code reformat

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py

* fixing based on comments on PR

* took out the testing code that should but work fails without the beam search moditification ; style changes

* fixing comments issues

* docstrings for ConstraintListState

* typo in PhrsalConstraint docstring

* docstrings improvements

* finished adding what is sort of an opinionated implementation of disjunctive generation, but it revealed errors in inner beam search logic during testing.

* fixed bug found in constrained beam search that used beam_idx that were not global across all the batches

* disjunctive constraint working 100% correctly

* passing all tests

* Accidentally included mlruns

* Update src/transformers/generation_beam_constraints.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update src/transformers/generation_beam_constraints.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* complete overhaul of type complexities and other nits

* strict type checks in generate()

* fixing second round of feedback by narsil

* fixed failing generation test because of type check overhaul

* generation test fail fix

* fixing test fails
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Add TF logits wrappers 

* Add sample method

* add tests for TF logit wrappers

* TF generate sample tests now run on CPU
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* TF generate start refactor

* Add tf tests for sample generate

* re-organize

* boom boom

* Apply suggestions from code review

* re-add

* add all code

* make random greedy pass

* make encoder-decoder random work

* further improvements

* delete bogus file

* make gpt2 and t5 tests work

* finish logits tests

* correct logits processors

* correct past / encoder_outputs drama

* refactor some methods

* another fix

* refactor shape_list

* fix more shape list

* import shape
_list

* finish docs

* fix imports

* make style

* correct tf utils

* Fix TFRag as well

* Apply Lysandre's and Sylvais suggestions

* Update tests/test_generation_tf_logits_process.py
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* Update src/transformers/tf_utils.py
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* remove cpu according to gante

* correct logit processor
Co-authored-by: Matt <Rocketknight1@users.noreply.github.com>

* added classes to get started with constrained beam search

* in progress, think i can directly force tokens now but not yet with the round robin

* think now i have total control, now need to code the bank selection

* technically works as desired, need to optimize and fix design choices leading to undersirable outputs

* complete PR #1 without disjunctive decoding

* removed incorrect tests

* Delete k.txt

* Delete test.py

* Delete test.sh

* revert changes to test scripts

* genutils

* full implementation with testing, no disjunctive yet

* shifted docs

* passing all tests realistically ran locally

* removing accidentally included print statements

* fixed source of error in initial PR test

* fixing the get_device() vs device trap

* fixed documentation docstrings about constrained_beam_search

* fixed tests having failing for Speech2TextModel's floating point inputs

* fix cuda long tensor

* added examples and testing for them and founx & fixed a bug in beam_search and constrained_beam_search

* deleted accidentally added test halting code with assert False

* code reformat

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Update tests/test_generation_utils.py

* fixing based on comments on PR

* took out the testing code that should but work fails without the beam search moditification ; style changes

* fixing comments issues

* docstrings for ConstraintListState

* typo in PhrsalConstraint docstring

* docstrings improvements
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

* Fix bad examples

* Add black formatting to style_doc

* Use first nonempty line

* Put it at the right place

* Don't add spaces to empty lines

* Better templates

* Deal with triple quotes in docstrings

* Result of style_doc

* Enable mdx treatment and fix code examples in MDXs

* Result of doc styler on doc source files

* Last fixes

* Break copy from

* Convert all tutorials and guides

* Convert all remaining rst to mdx

* Track and fix bad links