* mvp
* added test (a few models need fixes)
* fix a few test cases
* test nits
* harder test 😈
* revert changes in stablelm
* test with improved condition
* add todo
* tmp commit
* merged with main
* nits
* add todo
* final corrections
* add docs for generation compilation
* docs nits
* add tip
* PR suggestions
* add more details to the compilation docs
* fix cache positions
* cache is now init in generate; update docs
* tag test as flaky
* docs
* post rebase make fixup and other nits
* remove unintended changes
* whisper (encoder-decoder) not supported
* move token default updates to ; add tests for token defaults
* push changes
* manual rebase
* chameleon doesn't support this
* fix test_static_cache_mha_mqa_gqa (broken in another PR)
* docs: dynamic is better with end-to-end compilation
* token healing impl + trie with extensions
* make fixup
* prefix-robust space tokenization
* examples readme and requirements
* make fixup
* allow input prompt and model
* redundant defaults
* Specialized Trie
* make fixup
* updated tests with new inherited Tree
* input ids to auto device_map
* rm unused import
* Update src/transformers/generation/utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* naming convention
* Revert "naming convention"
This reverts commit dd39d9c5b7a969e2d8a8d2a8e54f121b82dc44f0.
* naming convention
* last -hopefully- changes
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* clean-up
* Update src/transformers/cache_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/cache_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* Update src/transformers/cache_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* fixup
* Update tests/quantization/quanto_integration/test_quanto.py
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
* Update src/transformers/generation/configuration_utils.py
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
* more suggestions
* mapping if torch available
* run tests & add 'support_quantized' flag
* fix jamba test
* revert, will be fixed by another PR
* codestyle
* HQQ and versatile cache classes
* final update
* typo
* make tests happy
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
Co-authored-by: Younes Belkada <49240599+younesbelkada@users.noreply.github.com>
* stash commit (will discard all of this)
* stash commit
* First commit - needs a lot of testing!
* Add a test
* Fix imports and make the tests actually test something
* Tests pass!
* Rearrange test
* Add comments (but it's still a bit confusing)
* Stop storing the tokenizer
* Comment fixup
* Fix for input_ids with a single sequence
* Update tests to test single sequences
* make fixup
* Fix incorrect use of isin()
* Expand tests to catch more cases
* Expand tests to catch more cases
* make fixup
* Fix length calculation and update tests
* Handle Ġ as a space replacement too
* Update src/transformers/generation/stopping_criteria.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
* Add optimizations from Joao's suggestion
* Remove TODO
* Update src/transformers/generation/stopping_criteria.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
* Update tests/generation/test_stopping_criteria.py
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
* make fixup
* Rename some variables and remove some debugging clauses for clarity
* Add tests for the sub-methods
* Clarify one test slightly
* Add stop_strings to GenerationConfig
* generate() supports stop_string arg, asks for tokenizer if not provided
* make fixup
* Cleanup code and rename variables for clarity
* Update tokenizer error
* Update tokenizer passing, handle generation on GPU
* Slightly more explanation cleanup
* More comment cleanup
* Factor out the token cleanup so it's more obvious what we're doing, and we can change it later
* Careful with that cleanup!
* Cleanup + optimizations to _get_matching_positions
* More minor performance tweaks
* Implement caching and eliminate some expensive ops (startup time: 200ms -> 9ms)
* Remove the pin_memory call
* Parallelize across all stop strings!
* Quick fix for tensor devices
* Update embeddings test for the new format
* Fix test imports
* Manual patching for BERT-like tokenizers
* Return a bool vector instead of a single True/False
* Better comment
* Better comment
* Add tests from @zucchini-nlp
* Amy's list creation nit
* tok_list -> token_list
* Push a big expanded docstring (should we put it somewhere else?)
* Expand docstrings
* Docstring fixups
* Rebase
* make fixup
* Make a properly general method for figuring out token strings
* Fix naming throughout the functions
* Move cache, refactor, fix tests
* Add comment
* Remove finished TODO
* Remove finished TODO
* make fixup
* Update src/transformers/generation/stopping_criteria.py
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
* Update and shorten docstring
* Update tests to be shorter/clearer and test specific cases
---------
Co-authored-by: Joao Gante <joaofranciscocardosogante@gmail.com>
Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>
* Add jamba arch
* apply "make fix-copies" changes
* fix link to model in JambaConfig docstring
* Add n_ctx in modeling file because repo-consistency wants that
* Add jamba to flash attention and sdpa documentation
* mamba dt_proj quant fix now works for LoRA as well
* override test_left_padding_compatibility and use a more permissive tolerance. left padding numerical difference are accentuated by mamba layers
* add jamba to tokenization auto
* fix comments of shape (PR #24 in the model page: https://huggingface.co/ai21labs/Jamba-v0.1/discussions/24)
* simple PR fixes
* remove unnecessary kwargs from JambaAttentionDecoderLayer and JambaMambaDecoderLayer
* remove the LoRA hack for the mamba dt_proj bias. It was solved in huggingface/peft#1530 (https://github.com/huggingface/peft/pull/1530)
* Add copied comment on JambaMLP (it's the same as MixtralMLP)
* remove padding_mask warnings. It's not supported anymore
* fix docstring. Float instead of int
* A few more minor PR fixes
* (1) lowercase names for mamba layernorms (2) remove _apply_inner_layernorms and do it directly in the forward pass
* Return None attention weights from mamba layers. Append to all attentions only if not None.
* remove some leftover jamba archive lists
* Better separation between expert vs non-expert layers. non-expert layers return None as router_logits, and it is not concatenated to all_router_logits returned from JambaModel
* no need to take router_logits at config.expert_layer_offset anymore. result.router_logits now holds results only for expert layers
* Add Jamba paper on READMEs
* (1) rename n_ctx -> max_position_embeddings (2) don't use it in the modeling file since it's not needed (set it as an exception to check_config_attributes)
* Add copied from comment
* remove the code path for apply_inner_layernorms=False. Jamba always has the inner mamba layernorms
* clearer docstring for _convert_to_standard_cache
* style fixes
* Change calc_logits_for_entire_prompt (bool) to num_logits_to_keep (int). Adapt assisted decoding code tp use it. Also small change in low memory beam search decoding path to support this new int value in model_inputs
* rename test so it still overrides what its meant to override
* draft
* oups
* nit
* remove more complexe logic
* fix names used in config
* fix fix fix
* style
* fix some more failing tests
* generate did not init the cache 🙃
* more small nits
* typo
* config.mamba_expand * config.hidden_size for the intermediate size of the mamba shapes
* fix init of pkv with torch.tensor()
* empty tensor
* fix some init issues
* stupid changes required by generate because it does not even support it's own DynamicCache class
* more fixes
* fix general assisted gen cache_position bug
* tests passing
* Add offsets and periods as SPECIAL_CASES_TO_ALLOW in check_config_attributes.py
* fix reorder_cache to reorder mamba states and override some more functions in HybridMambaAttentionDynamicCache
* no need to override test_past_key_values_format() and _check_past_key_values_for_generate() in tests anymore
* fix docstrings and typehints for past_key_values
* style fixes
* fix docs
* change typehint due to copy from Mixtral
* forgot import
* import order
* Add configuration_jamba and modeling_jamba to not_doctested because the model is too big to download (in docstring of JambaForCausalLM.forward)
* Add integration test with tiny tandom Jamba model on hub
* fix flash attention cache shapes
* bring back forgotten hidden states
* rename HybridMambaAttentionDynamicCache.seqlen_offset to has_previous_state (and make bool) and bugfix - it should be set to True after a finished forward pass of the entire model
* align integration test after modeling fixes
* bugfix - mamba can use precomputed states only of forward pass is on a single token
* bugfix - mamba can use precomputed states only if they match the batch size
* typo
* remove making _prepare_4d_causal_attention_mask a leaf function
* stop using past_seq_len.get_seq_length(). Use cache positions instead. Adjust test (test_decoder_model_past_with_large_inputs) accordingly
---------
Co-authored-by: Arthur Zucker <arthur.zucker@gmail.com>
Co-authored-by: Joao Gante <joao@huggingface.co>
* fix bug and add tests
* nit
* otherway to get the cur len instead of attention mask
* more places where this might have been broken
* nit
* oups
* inputs_embeds vs input_embeds
* test generated outptus
* style
* nit
* fix
* skip failing biogpt
* left-padding test revisited
* Apply suggestions from code review
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
---------
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
output_logits option behaves like output_scores, but returns the raw, unprocessed prediction logit scores,
ie. the values before they undergo logit processing and/or warping. The latter happens by default for the
regular output scores.
It's useful to have the unprocessed logit scores in certain circumstances. For example, unprocessed logit scores
are very useful with causallm models when one wants to determine the probability of a certain answer, e.g.
when asking a question with a yes/no answer. In that case getting the next-token probabilities of both "yes" and
"no" (and/or their relative ratio) is of interest for classification. The reason for getting these _before_ logit
processing and/or warping is b/c a) that can change the probabilities or b) reject the tokens of interest / reduce
the number of tokens to just 1.
For an example use-case see paper TabLLM: Few-shot Classification of Tabular Data with Large Language Models
by Stefan Hegselmann, Alejandro Buendia, Hunter Lang, Monica Agrawal, Xiaoyi Jiang, and David Sontag.
https://arxiv.org/abs/2210.10723
In addition:
- added dedicated unit test: tests/generation/test_utils/test_return_unprocessed_logit_scores
which tests return of logics with output_logits=True in generation.
- set output_logits=True in all other generation unit tests, that also have output_scores=True.
Implemented @gante's and @amyeroberts review feedback
Co-authored-by: kx79wq <max.baak@ing.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>
* skip 4 tests
* nits
* style
* wow it's not my day
* skip new failing tests
* style
* skip for NLLB MoE as well
* skip `test_assisted_decoding_sample` for everyone
