* more precise name
* better docstrings
* Update src/transformers/cache_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Add YaRN and Dynamic-YaRN RoPE Scaling Methods
YaRN (Yet another RoPE extension method) combines the NTK-By-Parts
Interpolation and Attention Scaling methods, improving upon existing
RoPE interpolation methods for longer context window sizes.
Fine-tuned models maintain their original performance across benchmarks
while enabling efficient extrapolation and transfer learning for
quicker convergence, especially in compute-limited environments.
We implement YaRN and Dynamic-YaRN for the following list of models:
- LLaMA
- Falcon
- GPT-NeoX
- Olmo
- Persimmon
- Phi
- StableLM
- OpenLLaMA
New unit tests are added to assert YaRN's correct behavior on both
short and long sequence inputs.
For more details, please refer to https://arxiv.org/abs/2309.00071.
Co-authored-by: Miguel Almeida <miguel.pessanha.almeida@tecnico.ulisboa.pt>
* Refactor YaRN implementation for LLaMA
Iterate on YaRN implementation for LLaMA and remove diff from remaining
models for increased PR modularity.
This commit includes the following changes:
- Merge 'yarn_rope_scaling' and 'rope_scaling' dictionaries
- Remove unnecessary attributes ('extrapolation_factor' and 'finetuned')
from YaRN classes
- Inherit 'forward' method in YaRN classes from superclass
- Rename 'yarn' method to 'compute_yarn_scaling'
- Extend YaRN tests with further assertions
- Fix style inconsistencies
Co-authored-by: Miguel Monte e Freitas <miguelmontefreitas@tecnico.ulisboa.pt>
* Refactor Tensor Building Logic for YaRN
- Comply with the the tensor building logic introduced in #30743
- Add referencing to the optimized Attention Factor equation
- Remove Dynamic YaRN for a more agile deployment
Co-authored-by: mig-mfreitas <mig-mfreitas@users.noreply.github.com>
* remove unwanted file
---------
Co-authored-by: Miguel Almeida <miguel.pessanha.almeida@tecnico.ulisboa.pt>
Co-authored-by: mig-mfreitas <mig-mfreitas@users.noreply.github.com>
Co-authored-by: Joao Gante <joao@huggingface.co>
* Pass datasets trust_remote_code
* Pass trust_remote_code in more tests
* Add trust_remote_dataset_code arg to some tests
* Revert "Temporarily pin datasets upper version to fix CI"
This reverts commit b7672826ca.
* Pass trust_remote_code in librispeech_asr_dummy docstrings
* Revert "Pin datasets<2.20.0 for examples"
This reverts commit 833fc17a3e.
* Pass trust_remote_code to all examples
* Revert "Add trust_remote_dataset_code arg to some tests" to research_projects
* Pass trust_remote_code to tests
* Pass trust_remote_code to docstrings
* Fix flax examples tests requirements
* Pass trust_remote_dataset_code arg to tests
* Replace trust_remote_dataset_code with trust_remote_code in one example
* Fix duplicate trust_remote_code
* Replace args.trust_remote_dataset_code with args.trust_remote_code
* Replace trust_remote_dataset_code with trust_remote_code in parser
* Replace trust_remote_dataset_code with trust_remote_code in dataclasses
* Replace trust_remote_dataset_code with trust_remote_code arg
* add prefix space ignored in llama #29625
* adding test with add_prefix_space=False
* ruff
---------
Co-authored-by: Ita Zaporozhets <itazaporozhets@Itas-MBP.localdomain>
* Add MistralForTokenClassification
* Add tests and docs
* Add token classification for Mixtral and Qwen2
* Save llma for token classification draft
* Add token classification support for Llama, Gemma, Persimmon, StableLm and StarCoder2
* Formatting
* Add token classification support for Qwen2Moe model
* Add dropout layer to each ForTokenClassification model
* Add copied from in tests
* Update src/transformers/models/llama/modeling_llama.py
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
* Propagate suggested changes
* Style
---------
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
* push legacy to fast as well
* super strange
* Update src/transformers/convert_slow_tokenizer.py
* make sure we are BC
* fix Llama test
* nit
* revert
* more test
* style
* update
* small update w.r.t tokenizers
* nit
* don't split
* lol
* add a test for `add_prefix_space=False`
* fix gemma tokenizer as well
* update
* fix gemma
* nicer failures
* fixup
* update
* fix the example for legacy = False
* use `huggyllama/llama-7b` for the PR doctest
* nit
* use from_slow
* fix llama
* attempt to fix
* the actual fix that works with compilation!
* this?
* temporary update
* nit?
* dispatcg to memory efficient?
* update both models that have static cache support
* fix copies fix compile
* make sure fix
* fix cohere and gemma
* fix beams?
* nit
* slipped through the cracks
* nit
* nits
* update
* fix-copies
* skip failing tests
* nits
* use user_defined_symbols
* fixup
* nit
* add a very robust test
* make sure all models are tested with the `pretrained_tokenizer_to_test`
* should we make sure we test all of them?
* merge
* remove the id
* fix test
* update
* ousies
* oups
* fixup
* fix copies check
* remove `pretrained_tokenizer_to_test`
* remove control flow
* update gptneox
* update ....
* nits
* Actually let's just break. Otherwise we are silently failing which imo is not optimal
* version BC
* fix tests
* fix eager causal
* nit
* add a test
* style
* nits
* nits
* more nits for the test
* update and fix
* make sure cuda graphs are not skipped
* read token is needed for meta llama
* update!
* fiixup
* compile test should be slow
* fix thet fix copies
* stle 🫠
* add add_dummy_prefix_space option to slow
* checking kwargs might be better. Should be there for all spm tokenizer IMO
* nits
* fix copies
* more copied
* nits
* add prefix space
* nit
* nits
* Update src/transformers/convert_slow_tokenizer.py
* fix inti
* revert wrong styling
* fix
* nits
* style
* updates
* make sure we use slow tokenizer for conversion instead of looking for the decoder
* support llama ast well
* update llama tokenizer fast
* nits
* nits nits nits
* update the doc
* update
* update to fix tests
* skip unrelated tailing test
* Update src/transformers/convert_slow_tokenizer.py
* add proper testing
* test decode as well
* more testing
* format
* fix llama test
* Apply suggestions from code review
* This is a test commit
* testing commit
* final commit with some changes
* Removed copy statement
* Fixed formatting issues
* Fixed error added past_key_values in the forward method
* Fixed a trailing whitespace. Damn the formatting rules are strict
* Added the copy statement
* Fix typos and grammar mistakes in docs and examples
* Fix typos in docstrings and comments
* Fix spelling of `tokenizer` in model tests
* Remove erroneous spaces in decorators
* Remove extra spaces in Markdown link texts
* skip bf16 test if not supported by device
* fix
* fix bis
* use is_torch_bf16_available_on_device
* use is_torch_fp16_available_on_device
* fix & use public llama
* use 1b model
* fix flacky test
---------
Co-authored-by: Your Name <you@example.com>
* add sdpa
* wip
* cleaning
* add ref
* yet more cleaning
* and more :)
* wip llama
* working llama
* add output_attentions=True support
* bigcode sdpa support
* fixes
* gpt-bigcode support, require torch>=2.1.1
* add falcon support
* fix conflicts falcon
* style
* fix attention_mask definition
* remove output_attentions from attnmaskconverter
* support whisper without removing any Copied from statement
* fix mbart default to eager renaming
* fix typo in falcon
* fix is_causal in SDPA
* check is_flash_attn_2_available in the models init as well in case the model is not initialized through from_pretrained
* add warnings when falling back on the manual implementation
* precise doc
* wip replace _flash_attn_enabled by config.attn_implementation
* fix typo
* add tests
* style
* add a copy.deepcopy on the config in from_pretrained, as we do not want to modify it inplace
* obey to config.attn_implementation if a config is passed in from_pretrained
* fix is_torch_sdpa_available when torch is not installed
* remove dead code
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/models/bart/modeling_bart.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* remove duplicate pretraining_tp code
* add dropout in llama
* precise comment on attn_mask
* add fmt: off for _unmask_unattended docstring
* precise num_masks comment
* nuke pretraining_tp in LlamaSDPAAttention following Arthur's suggestion
* cleanup modeling_utils
* backward compatibility
* fix style as requested
* style
* improve documentation
* test pass
* style
* add _unmask_unattended tests
* skip meaningless tests for idefics
* hard_check SDPA requirements when specifically requested
* standardize the use if XXX_ATTENTION_CLASSES
* fix SDPA bug with mem-efficient backend on CUDA when using fp32
* fix test
* rely on SDPA is_causal parameter to handle the causal mask in some cases
* fix FALCON_ATTENTION_CLASSES
* remove _flash_attn_2_enabled occurences
* fix test
* add OPT to the list of supported flash models
* improve test
* properly test on different SDPA backends, on different dtypes & properly handle separately the pad tokens in the test
* remove remaining _flash_attn_2_enabled occurence
* Update src/transformers/modeling_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/modeling_attn_mask_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update docs/source/en/perf_infer_gpu_one.md
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* remove use_attn_implementation
* fix docstring & slight bug
* make attn_implementation internal (_attn_implementation)
* typos
* fix tests
* deprecate use_flash_attention_2=True
* fix test
* add back llama that was removed by mistake
* fix tests
* remove _flash_attn_2_enabled occurences bis
* add check & test that passed attn_implementation is valid
* fix falcon torchscript export
* fix device of mask in tests
* add tip about torch.jit.trace and move bt doc below sdpa
* fix parameterized.expand order
* move tests from test_modeling_attn_mask_utils to test_modeling_utils as a relevant test class is already there
* update sdpaattention class with the new cache
* Update src/transformers/configuration_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/models/bark/modeling_bark.py
* address review comments
* WIP torch.jit.trace fix. left: test both eager & sdpa
* add test for torch.jit.trace for both eager/sdpa
* fix falcon with torch==2.0 that needs to use sdpa
* fix doc
* hopefully last fix
* fix key_value_length that has no default now in mask converter
* is it flacky?
* fix speculative decoding bug
* tests do pass
* fix following #27907
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
