rh`ddlZddlZddlZddlZddlmZddlmZmZm Z m Z m Z ddl Z ddlZddlmZddlmZddlmZddlmZddlmZdd lmZmZmZmZmZm Z dd l!m"Z"dd l#m$Z$m%Z%m&Z&m'Z'dd l(m)Z)dd l*m+Z+ddl,m-Z-ddl.m/Z/m0Z0ddl1m2Z2ddl3m4Z4ddl5m6Z6m7Z7m8Z8m9Z9m:Z:m;Z;ddlZ>ddl?m@Z@mAZAmBZBddlCmDZDmEZEmFZFmGZGmHZHmIZImJZJmKZKmLZLddl!mMZMmNZNmOZOmPZPmQZQddlRmSZSddlTmUZUmVZVmWZWmXZXmYZYmZZZm[Z[m\Z\m]Z]m^Z^m_Z_m`Z`maZambZbmcZcmdZdmeZemfZfmgZgmhZhmiZimjZjmkZkmlZlmmZmmnZnddlompZpmqZqmrZrmsZsmtZtmuZumvZverddlwmxZxddlymzZzddl{m|Z|e;je~Ze7rddlmZmZgdZeGd d!e6ZeGd"d#e6ZeGd$d%e6ZeGd&d'e6ZeZeZeZeZeZeZeZeZeZeZe eefZe eefZe eefZe eefZe eefZe eefZe eefZe eefZGd(d)eSZd*ZdDd+Zd,ej:d-ej:d.ej:d/ej<d0ed1ed2ej:fd3Zd4ee6d5e"d2e6fd6Zd7ed8 d9dfd:ej:d;ej:ded2ej:f d?Zd@eedAeeej:fdBej:d2ej:fdCZy)EN) dataclass) TYPE_CHECKINGAnyCallableOptionalUnion) file_exists)version)nn) functional)Cache DynamicCacheEncoderDecoderCacheHybridChunkedCacheOffloadedCacheOffloadedHybridCache)PretrainedConfig)check_python_requirementsget_cached_module_fileget_class_in_moduleresolve_trust_remote_code)is_deepspeed_zero3_enabled)is_fsdp_managed_module)create_masks_for_generate)CausalLMOutputWithPastSeq2SeqLMOutput)isin_mps_friendly)ExtensionsTrie) ModelOutputis_accelerate_availableis_hqq_availableis_optimum_quanto_availableis_torchdynamo_exportinglogging)DisjunctiveConstraintPhrasalConstraint) BeamScorerBeamSearchScorerConstrainedBeamSearchScorer) AssistantVocabTranslatorCacheAssistedCandidateGenerator-AssistedCandidateGeneratorDifferentTokenizersCandidateGeneratorEarlyExitCandidateGeneratorPromptLookupCandidateGenerator%UniversalSpeculativeDecodingGenerator_prepare_attention_mask_prepare_token_type_ids) NEED_SETUP_CACHE_CLASSES_MAPPINGQUANT_BACKEND_CLASSES_MAPPING CompileConfigGenerationConfigGenerationMode)ContinuousMixin)#EncoderNoRepeatNGramLogitsProcessor'EncoderRepetitionPenaltyLogitsProcessorEpsilonLogitsWarperEtaLogitsWarperExponentialDecayLengthPenaltyForcedBOSTokenLogitsProcessorForcedEOSTokenLogitsProcessorHammingDiversityLogitsProcessorInfNanRemoveLogitsProcessorLogitNormalizationLogitsProcessorListMinLengthLogitsProcessor!MinNewTokensLengthLogitsProcessorMinPLogitsWarperNoBadWordsLogitsProcessorNoRepeatNGramLogitsProcessor PrefixConstrainedLogitsProcessor RepetitionPenaltyLogitsProcessorSequenceBiasLogitsProcessor$SuppressTokensAtBeginLogitsProcessorSuppressTokensLogitsProcessorTemperatureLogitsWarperTopKLogitsWarperTopPLogitsWarperTypicalLogitsWarper.UnbatchedClassifierFreeGuidanceLogitsProcessor)ConfidenceCriteriaEosTokenCriteriaMaxLengthCriteriaMaxTimeCriteriaStoppingCriteriaStoppingCriteriaListStopStringCriteria)PreTrainedModel)PreTrainedTokenizerBase) BaseStreamer)AlignDevicesHookadd_hook_to_module)past_key_values cache_paramsstatememspast_buckets_statesc.eZdZUdZej ed<dZee ejed<dZ ee ejed<dZ ee e ejed<dZ ee e ejed<dZee e e ejed<y) GenerateDecoderOnlyOutputaw Outputs of decoder-only generation models, when using non-beam methods. Args: sequences (`torch.LongTensor` of shape `(batch_size, sequence_length)`): The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early due to the `eos_token_id`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`. hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`. past_key_values (`tuple(tuple(torch.FloatTensor)))`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. sequencesNscoreslogits attentions hidden_statesra)__name__ __module__ __qualname____doc__torch LongTensor__annotations__rirtuple FloatTensorrjrkrlrap/var/www/html/ai-insurance-compliance-backend/venv/lib/python3.12/site-packages/transformers/generation/utils.pyrgrgs415FHU5,,- .515FHU5,,- .5<@JuU%6%6789@?CM8E%(9(9":;<CHLOXeE%0A0A*B$CDELrwrgceZdZUdZej ed<dZee ejed<dZ ee ejed<dZ ee ejed<dZ ee ejed<dZee e ejed<dZee e ejed <dZee e ejed <dZee e e ejed <y) GenerateEncoderDecoderOutputaI Outputs of encoder-decoder generation models, when using non-beam methods. Args: sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`): The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early due to the `eos_token_id`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Processed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size, config.vocab_size)`. encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`. cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`. decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, generated_length, hidden_size)`. past_key_values (`tuple(tuple(torch.FloatTensor)))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. rhNrirjencoder_attentionsencoder_hidden_statesdecoder_attentionscross_attentionsdecoder_hidden_statesra)rmrnrorprqrrrsrirrtrurjr{r|r}r~rrarvrwrxrzrzs!F15FHU5,,- .515FHU5,,- .5=Au'8'8!9:A@D8E%*;*;$<=DDHuU->->'?!@AHBFhuU5+<+<%=>?FGK8E%0A0A*B$CDKHLOXeE%0A0A*B$CDELrwrzc~eZdZUdZej ed<dZeejed<dZ ee ejed<dZ ee ejed<dZ eej ed<dZee e ejed<dZee e ejed <dZee e e ejed <y) GenerateBeamDecoderOnlyOutputa Outputs of decoder-only generation models, when using beam methods. Args: sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`): The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early due to the `eos_token_id`. sequences_scores (`torch.FloatTensor` of shape `(batch_size*num_return_sequences)`, *optional*, returned when `output_scores=True`): Final beam scores of the generated `sequences`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True`): Beam indices of generated token id at each generation step. `torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`. attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams, num_heads, generated_length, sequence_length)`. hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`. past_key_values (`tuple(tuple(torch.FloatTensor)))`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. rhNsequences_scoresrirj beam_indicesrkrlra)rmrnrorprqrrrsrrrurirtrjrrkrlrarvrwrxrrs@48hu001815FHU5,,- .515FHU5,,- .5/3L(5++,3<@JuU%6%6789@?CM8E%(9(9":;<CHLOXeE%0A0A*B$CDELrwrceZdZUdZej ed<dZeejed<dZ ee ejed<dZ ee ejed<dZ eej ed<dZee ejed<dZee ejed <dZee e ejed <dZee e ejed <dZee e ejed <dZee e e ejed <y) GenerateBeamEncoderDecoderOutputa Outputs of encoder-decoder generation models, when using beam methods. Args: sequences (`torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`): The generated sequences. The second dimension (sequence_length) is either equal to `max_length` or shorter if all batches finished early due to the `eos_token_id`. sequences_scores (`torch.FloatTensor` of shape `(batch_size*num_return_sequences)`, *optional*, returned when `output_scores=True`): Final beam scores of the generated `sequences`. scores (`tuple(torch.FloatTensor)` *optional*, returned when `output_scores=True`): Beam transition scores for each vocabulary token at each generation step. Beam transition scores consisting of log probabilities of tokens conditioned on log softmax of previously generated tokens in this beam. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. logits (`tuple(torch.FloatTensor)` *optional*, returned when `output_logits=True`): Unprocessed prediction scores of the language modeling head (scores for each vocabulary token before SoftMax) at each generation step. Tuple of `torch.FloatTensor` with up to `max_new_tokens` elements (one element for each generated token), with each tensor of shape `(batch_size*num_beams, config.vocab_size)`. beam_indices (`torch.LongTensor`, *optional*, returned when `output_scores=True`): Beam indices of generated token id at each generation step. `torch.LongTensor` of shape `(batch_size*num_return_sequences, sequence_length)`. encoder_attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer of the decoder) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. encoder_hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size*num_beams*num_return_sequences, sequence_length, hidden_size)`. decoder_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, num_heads, generated_length, sequence_length)`. cross_attentions (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_attentions=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size, num_heads, generated_length, sequence_length)`. decoder_hidden_states (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `output_hidden_states=True`): Tuple (one element for each generated token) of tuples (one element for each layer of the decoder) of `torch.FloatTensor` of shape `(batch_size*num_beams*num_return_sequences, generated_length, hidden_size)`. past_key_values (`tuple(tuple(torch.FloatTensor)))`, *optional*, returned when `use_cache=True`): Returns the model cache, used to speed up decoding. Different models have a different cache format, check the model's documentation. Usually, a [`~cache_utils.Cache`] instance. rhNrrirjrr{r|r}r~rra)rmrnrorprqrrrsrrrurirtrjrr{r|r}r~rrarvrwrxrr s/(T48hu001815FHU5,,- .515FHU5,,- .5/3L(5++,3=Au'8'8!9:A@D8E%*;*;$<=DDHuU->->'?!@AHBFhuU5+<+<%=>?FGK8E%0A0A*B$CDKHLOXeE%0A0A*B$CDELrwrc$eZdZdZ ddeeeejfdee de fdZ de jdee jd ee jdee je jffd Zde jdee jd ee jdee je jffd Z dde jd eed ee jdee jd ee jf dZ ddee j(dee j(deeee j(fdee j(eeeee j(fffdZ ddee j(dee j(deeee j(fde jfdZde j(dedeeefde jfdZde j(deededeeeffdZ ddededeee j(fde j(dee j:dee jeee j(fff dZe ddede dee jdee jeeefffdZ dd e!deeefde d!edeeeff d"Z"dede jde j(d#d$d%e#d&d'd(d'dede$fd)Z% dded*eed+e jd,ee ee j(ge&efd%ee#deedeeeefd-ee j(d.ee j(de#fd/Z' dded0ee(d1ed'de(fd2Z)d3ee#e(fd4ee#e(fdee#e(ffd5Z* dd6e j(d7ee j(d8ee j(d9e de j(f d:Z+d;Z,deeeffd<Z-d=Z.d>Z/ ddeed?ee d@edeeeffdAZ0dBZ1deeeeeefffdCZ2dDededEede j:def dFZ3e4de fdGZ5deded#d$dedHede j:de fdIZ6de fdJZ7 ddedKee deee j:effdLZ8dedede fdMZ9e jt ddee j(deed%ee#d0ee(d,ee ee j(ge&efdNee d#ed$dOedPd-ee j(d.ee j(d?ee dQeedee;e jffdRZde jdVeee&efd%e#d0e(dedNe dOdPdee?e jffdWZ@e jtde jd%e#d0e(dedNe dOedPdee?e jffdXZAde jd%e#d0e(dedNe dOedPdee?e jffdYZBedZe j(de j(fd[ZCedZe j(ded\ede j(fd]ZDedZe j(d8e j(de j(fd^ZEed_e j(d`e j(dae j(dbe j(dceddedeedfee efdgeFfdhZGed_e j(dbe j(die j(dfee effdjZHdke j(dle j(dme j(dcedeedne doed\edpededee j(e j(e j(ffdqZIdre j(dse j(dte j(die j(d\edee j(e j(e j(ff duZJd6e j(dse j(dae j(dre j(d8e j(dte j(d_e j(dbe j(die j(dve j(d\edcedeedgeFdfee efdee j(e j(e j(e j(ff dwZKde jd%e#d0e(dedNe deeLe jff dxZMde jdyeNd%e#d0e(dedNe f dzZOde jd{ePd%e#d0e(dedNe deeLe jffd|ZQde jd}e$d%e#d0e(dedNe dOedPdee?e jffd~ZRde jdefdZSy)GenerationMixina A class containing all functions for auto-regressive text generation, to be used as a mixin in model classes. Inheriting from this class causes the model to have special generation-related behavior, such as loading a `GenerationConfig` at initialization time or ensuring `generate`-related tests are run in `transformers` CI. A model class should inherit from `GenerationMixin` to enable calling methods like `generate`, or when it has defined a custom `generate` method that relies on `GenerationMixin`, directly or indirectly, which approximately shares the same interface to public methods like `generate`. Three examples: - `LlamaForCausalLM` should inherit from `GenerationMixin` to enable calling `generate` and other public methods in the mixin; - `BlipForQuestionAnswering` has a custom `generate` method that approximately shares the same interface as `GenerationMixin.generate` (it has a few extra arguments, and the same output). That function also calls `GenerationMixin.generate` indirectly, through an inner model. As such, `BlipForQuestionAnswering` should inherit from `GenerationMixin` to benefit from all generation-related automation in our codebase; - `BarkModel` has a custom `generate` method and one of its inner models calls `GenerationMixin.generate`. However, its `generate` does not share the same interface as `GenerationMixin.generate`. In this case, `BarkModel` should NOT inherit from `GenerationMixin`, as it breaks the `generate` interface. The class exposes [`~generation.GenerationMixin.generate`], which can be used for: - *greedy decoding* if `num_beams=1` and `do_sample=False` - *contrastive search* if `penalty_alpha>0` and `top_k>1` - *multinomial sampling* if `num_beams=1` and `do_sample=True` - *beam-search decoding* if `num_beams>1` and `do_sample=False` - *beam-search multinomial sampling* if `num_beams>1` and `do_sample=True` - *diverse beam-search decoding* if `num_beams>1` and `num_beam_groups>1` - *constrained beam-search decoding* if `constraints!=None` or `force_words_ids!=None` - *assisted decoding* if `assistant_model` or `prompt_lookup_num_tokens` is passed to `.generate()` To learn more about decoding strategies refer to the [text generation strategies guide](../generation_strategies). Npretrained_model_name_or_pathtrust_remote_codereturnc tjj|}d}|r@tjjtjj|dsd}nt |dsd}|st d|dd|d}t |||| |t|fd d i|t|fd di|}td |}|S) at Loads and returns a custom generate function, given a model repo. Args: pretrained_model_name_or_path (`str` or `os.PathLike`): Can be either: - A string, the *model id* of a pretrained model hosted inside a model repo on huggingface.co. - A path to a *directory* containing model weights saved using [`~PreTrainedModel.save_pretrained`], e.g., `./my_model_directory/`. trust_remote_code (`bool`, *optional*): Whether or not to allow for custom models defined on the Hub in their own modeling files. This option should only be set to `True` for repositories you trust and in which you have read the code, as it will execute code present on the Hub on your local machine. **kwargs: Additional keyword arguments for remote code loading. Raises: OSError: If `pretrained_model_name_or_path` does not contain a `custom_generate` subdirectory. Returns: A callable that can be used to generate text. Tzcustom_generate/generate.pyF`zw` does not contain a `custom_generate` subdirectory with a `generate.py` file, can't load the custom generate function.zThe repository `zS` contains custom generation code that will override the default `generate` method.)has_local_codehas_remote_code error_messagerequirements_filez custom_generate/requirements.txt module_filegenerate) ospathexistsjoinr OSErrorrrrr) selfrrkwargs is_local_codehas_custom_generate_folderrmodulecustom_generate_functions rxload_custom_generatez$GenerationMixin.load_custom_generates:'DE %)" 77>>"'',,/LNk"lm-2*<>[\-2*)123OO <=>- -  "  )( --'   " ) =_ ci ( ) 7T X^ $7z6#J ''rw input_ids inputs_embedscache_positionc^tr|j|||S|/|jddk(r|dd|jd df}||fS||d|jdk\r|dd|jd df}||fS|jd|jdk7r |dd|f}||fS)a Generic cache-dependent input preparation The code is put in a separate function to allow granular unit testing as it needs a different implementation to be exportable. If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens - Exception 1: when passing input_embeds, input_ids may be missing entries - Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here - Exception 3: with synced GPUs cache_position may go out of bounds, but we only want dummy token in that case. - Exception 4: If input_embeds are passed then slice it through `cache_position`, to keep only the unprocessed tokens and generate the first token for each sequence. Later use the generated Input ids for continuation. The current implementation does not rely on ``self`` and could be a class method. It is left as a standard method to be easily rewritten. Nr&r)r$,_cache_dependant_input_preparation_exportingshape)rrrrs rx"_cache_dependant_input_preparationz2GenerationMixin._cache_dependant_input_preparations* $ %DDYP]_mn n  $);q)@)!n.B.B1.E-E-G*GHMi''  %r"iooa&88!!n&:&:1&=%=%?"?@Ii''__Q >#7#7#: :!!^"34Ii''rwc| |dd|f}||fSdddtj|jddk(fdfd|||g\}}||fS) z This method implements method ``_cache_dependant_input_preparation`` with :func:`torch.cond` to make it exportable with :func:`torch.export.export`. The code is put in a separate function to allow granular unit testing. NcP|dd|jd dfjSNrrclone)rrs rxbranch_1zNGenerationMixin._cache_dependant_input_preparation_exporting..branch_1s,$Q)=)=a)@(@(B%BCIIKKrwcP|dd|jd dfjSrrrrs rxbranch_2zNGenerationMixin._cache_dependant_input_preparation_exporting..branch_2s, ^%9%9!%<$<$>!>?EEGGrwc0|dd|fjSNrrs rxbranch_3zNGenerationMixin._cache_dependant_input_preparation_exporting..branch_3s N!2399;;rwr&rc6|||jfSrr)rrrrs rxzNGenerationMixin._cache_dependant_input_preparation_exporting.. s ?!)Frwch|tj|d|jdk\fd||gfS)Nrr&cttj|jd|jdk7d||gS)Nr&rc"|jSrrrs rxrzrGenerationMixin._cache_dependant_input_preparation_exporting......s yGXrwrqcondr)rrrs rxrz`GenerationMixin._cache_dependant_input_preparation_exporting....s> % $-OOA$6.:N:Nq:Q$Q$,%X%.$? !"rwr)rrrrrs rxrzNGenerationMixin._cache_dependant_input_preparation_exporting..sB% *2.)//!2DD$'7 Frwr)rrrrrrrs @@@rxrz:5( $M98i''rwraattention_maskc ^ i}||d<|||d<|j|||\}}|jjrdnd}|jjsQ|&t||jdk(r d||<||d<nL|j t j||<d|d<n#|j t j||<|jjr|nd} |jjr|jd dn|}|jjrd nd } |jjrd nd } ||j| ~| ttj|jjjvr<|j!j#d dz } | j%|dk(d| || <dD]~} |j| }||a|jd|djdn||jd}|dd| df}|j t j}||| <t'|t(r[|j*rN|K|j,dk(r;|jjs|d|dj\}}}n||jdd\}}t/||j0|}t3|dr|j5nd}t/|dd}|| t/|dd}|p|jdd}|j| d} t/|dt6}||jt j8||f|j:|||| |}n2||||j=|j:|||j|}|||| <| | |d <d|jj>vr|j@rt jB|jDd}|d ddd f}t jFt jHdi||j#djKdgd}tM|jOdz}|jQ\}}|dk(rt jR|fi|n-|jUt jBjKd}t jFt jHdi||j#dgd}tM|jO}|jW|jU|jD|jU|jD|||jYD]\} }!| |vs |!|| <|jdd|S)a Prepare the model inputs for generation. It includes operations like computing the 4D attention mask or slicing inputs given the existing cache. See the forward pass in the model documentation for expected arguments (different models might have different requirements for e.g. `past_key_values`). This function should work as is for most LLMs. rNradecoder_input_idsrr&r) memory_formatdecoder_attention_maskrdecoder_position_ids position_idsrr)rtoken_type_idsrr get_decoder5_prepare_4d_causal_attention_mask_with_cache_positionrrdtype)config input_embedsrrrarr)sequence_length target_lengthrr batch_sizerraflashrdevice) cu_seq_lens_q cu_seq_lens_k max_length_q max_length_klabels)r&)-rris_encoder_decoderlenrrrqcontiguous_formatpopgetsetinspect signatureforward parameterskeyslongcumsum masked_fill_ isinstanceris_compileablendimgetattrbase_model_prefixhasattrrremptyrget_max_cache_shape_attn_implementation_supports_attention_backendint32rcatzerosaddintmaxsizeonestoupdateitems)"rrrarrrr model_inputs input_ids_keyencoder_attention_maskattention_mask_keyposition_ids_keyrmodel_input_name model_inputcurrent_input_lengthrr_ base_modeldecodercausal_mask_creation_functionr tensor_kwsposrrbsseq_lenq_lenrrkeyvalues" rxprepare_inputs_for_generationz-GenerationMixin.prepare_inputs_for_generation&s$ )7 %&  &.=L* +'+'N'N=.( $M9 04{{/M/M+S^ {{--(S-@MDWDWXYDZ-Z.2 ]+0= _-/8ooELcLco.d ]+04 _-*3//H_H_/*`L '48;;3Q3QW[:>++:X:XFJJ/ 6^l :>9W9W5]m59[[5S5S1Yg  & +,4 C(9(9$,,(G(R(R(W(W(Y$ZZ)..077;a?L  % %n&91 ='3F# $![ =  **%56K&".(++O<H%_5;;A>)-8>>qA) #.a2F1F1G.G"HK"-"3"3%BYBY"3"ZK1< -. =  ...*##q(;;11l?6S6_1=o1N1T1T. OQ.:=.I.O.OPRQR.S+ O!t'='=tDJ29*m2Tj,,.Z^G,3SUY- )-49L07TVZ1- -4!-!1!12BD!I+//0@$G 07>Y[t0u-!>;;!&j/-JRVR\R\!]#1#1$3!-#1 ""?"$3"1"E"E"G**#1);;$3 "  %/=L+ , ! --CL) * dkk66 64;[;[#(;;$++FJ~.q"u5C!IIu{{'C 'CSZZPQ]EVEVWXEY&Z\]^Mswwy>A-L#..*KB4;qLEJJr0Z0cffU[[FYF]F]^_F`E!IIu{{'C 'CU\\RS_&UWXYMuyy{+L   +..t{{;+..t{{;))  !,,. *JC,&$) S! * 4(rwinputs bos_token_id model_kwargsc h|jjrFt|dr:|jj|jk7r|jj}n |j}|j Dcic]\}}|||k7s||}}}|j |d}||td|d|d|d|d ||}|dk(rd |vr|d |j d n|jjsd ttj|jjjv}|s#td |jjd |j!||| |d<|d d }}n| td |d d }}|j!|||}|||fScc}}w)zT This function extracts the model-specific `inputs` for generation. encoderNz `inputs`: z` were passed alongside z0 which is not allowed. Make sure to either pass z or z=...rrzAYou passed `inputs_embeds` to `.generate()`, but the model class z doesn't have its forwarding implemented. See the GPT2 implementation for an example (https://github.com/huggingface/transformers/pull/21405), and feel free to open a PR with it!)r zMYou passed `inputs_embeds` and `input_ids` to `.generate()`. Please pick one.)rrrrmain_input_namerr ValueErrorrrrr rr __class__rm*_maybe_initialize_input_ids_for_generation) rr r r  input_namekv inputs_kwarghas_inputs_embeds_forwardings rx_prepare_model_inputsz%GenerationMixin._prepare_model_inputss KK * *i( ,,0D0DD55J--J)5););)=bARSWaRa1b b$'' D9  #(:VH$#%%d&H&HITTYY[C0,4$[\`\j\j\s\s[tuxx-1,[,[L|-\- [)&2/%BO %$%tuu%1/%BO @@Wcdz<//[cs <F. F.cp||S|jd}|jjrR|P|jj dd}t j |t j|jdzSd}|jD]-}t|t js|jd}nd|vr2t j |dft j|jS| td t j |dft j|j|zS) z3Initializes input ids for generation, if necessary.Nencoder_outputsrrir&rrzB`bos_token_id` has to be defined when no `input_ids` are provided.)rrrlast_hidden_staterrqrrrvaluesrTensorrr)rr r r rrrr s rxrz:GenerationMixin._maybe_initialize_input_ids_for_generations   M&**+<= ;; ) )o.I#55::zQGenerationMixin._prepare_encoder_decoder_kwargs_for_generation..`sI!x**1-Isrr output_attentionsoutput_hidden_states return_dictrrv) get_encoderrr4r5r`r_rr'rrrrrr?r@r) rrr rr rirrelevant_prefixr=r encoder_kwargsencoder_signatureencoder_accepts_wildcards ` rx._prepare_encoder_decoder_kwargs_for_generationz>GenerationMixin._prepare_encoder_decoder_kwargs_for_generationJsy""$ 4 )w +26  /"7,7P7P &') s*!D9 D?,D?rdecoder_start_token_idrc|d|vr|jd}nd|vr|dk7r|jd}nd}| |j}|jdk(rC|jd|k7rt d|d|jd|j dd}n+t j|dft j| |z}||}||fSd |jjjvsI|jjd k(r5d |jjjjvr ||fS|jjd vr ||fS|dddf|dddfk7jj!r\t j"||gd }d|vr?|d}t j"t j$|ddddf|fd }||d<||fS)zGPrepares `decoder_input_ids` for generation with encoder-decoder modelsNrrr&rz1`decoder_start_token_id` expected to have length z but got rrdonutzvision-encoder-decoder)whisperdimr)rrrrrviewrqrrrrmlowerr model_typerallitemr ones_like)rrrr rHrrrs rx)_prepare_decoder_input_ids_for_generationz9GenerationMixin._prepare_decoder_input_ids_for_generationss(  #(;|(K , 0 01D E  L (-=-L , 0 0 =  $  >[[F ! & &! +%++A.*< G |S\]s]y]yz{]|\}~&<%@%@Q%G " J?%**VLOee #  $ 6 ,!,..%//557 7 KK " "&> >7dkkNaNaNlNlNrNrNtCt !,..[[ # #{ 2 !,.. 1%)?1)EE J J L Q Q S % +ACT*U[] ^ '<7)56N)O&).__%;._expand_dict_for_generationsd% d++&s+7">##6 E*8*=*O*OP[ab*O*cN3'  d" !rwrrLrzMIf `is_encoder_decoder` is True, make sure that `encoder_outputs` is defined.)rXrr)rUrrr rZs` rx_expand_inputs_for_generationz-GenerationMixin._expand_inputs_for_generations ! l* * "  !33KQ3GI2<@  12: !pqq.I,WhJi.jL* +,&&rwoutputsnum_new_tokensctD] }||vs|dvrd}n|}t||||<nd|vr7|d}tj||dddfj dgd|d<|sDd|vr|d}tj||j |j ddfgd|d<nCd |vr?|d } tj| | j | j ddfgd|d <|jd d r|d dd|z|d <|S|jd } tj| ddz| d|zdz| j j| j} tj| | f|d <|S)N)rerdrarrrLrrr&rr8Trr) ALL_CACHE_NAMESrrqr unsqueezenew_onesrrrarangerrr) rr\r rr]possible_cache_name cache_namerrrpast_positions new_positionss rx#_update_model_kwargs_for_generationz3GenerationMixin._update_model_kwargs_for_generations$3  "g-&*II!2J!4J+27!?16#^%<%R>RST>UWX=Y%Z[ac2 -. (<7)56N)O&9>+-C-L-LNdNjNjklNmopMq-rs: 56   K .-9:J-KBC-PSa-aL) * *--.>?N!LLr"Q&r(:^(Ka(OWeWkWkb&&' .3YY 7V-WL) *rwassistant_modelr\logits_processortarget_tokenizerr]assistant_tokenizerc td|||fD} |jt||||||} | S|j9t |j |j|j |j} | S| r|jdurctj|||jjj|d} d|j_t!|||||||||  } | S|jdurt#|||||||| } | St%d t'|jj(t+||||||} | S) zU Returns the candidate generator to be used in `assisted_generation` c3$K|]}|du ywrrv)r;rs rxr>z;GenerationMixin._get_candidate_generator..s"sQ1D="sN)rrhr r rri)r*num_output_tokensmax_matching_ngram_size max_lengthT)rhassistant_prune_lm_head) rrhr r rrirjrkatm_translatorF)rrhr r rrirjrkz7Invalid value for `do_sample`: expected a boolean, got )rQassistant_early_exitr0prompt_lookup_num_tokensr1r&rprq do_sampler,get_translatorrget_text_config vocab_sizer repetition_penaltyr2r.rtypermr-) rr rrrhrirjrkr different_tokenizerscandidate_generatorrss rx_get_candidate_generatorz(GenerationMixin._get_candidate_generators #"s?L\^q:r"ss  1 1 ="=# $"3)+!1 # z#"k 7 7 C"@.@@"3"L"L(9(Q(Q,77 # h#"]" **d2!>!M!M$'KK//1<<$3,0 "HL11D&K'$3&7!-"/%5%5(;#1 '#F#"1#,,5&S'$3&7!-"/%5%5(; '#.#"!MdSdSnSnNoNxNxMyz#=# /"3)+!1 # #"rwinput_ids_seq_lengthencoder_input_idsprefix_allowed_tokens_fnnegative_prompt_idsnegative_prompt_attention_maskc Zt} |g}|jB|jdk7r3| jt|j||| |j|j %| jt |j |jJ|jdkDr;| jt|j|j|j|jh|jdk7rYt|jdk(r'| jt|j| ntj d t"|j$4|j$dk7r%| jt'|j$ |j(3|j(d kDr$| jt+|j(|j,g|j,d kDrXt|jdk(r&| jt/|j,|ntj d t"|j0/| jt3|j0|j4|j6Mt9|dd@|j6d kDr1| jt;|j6|j4||j<Nt9|ddA|j<d kDr2| jt?||j<|j4||2| jtA||j|jz|jB$| jtE|jB|jF1| jtI|jJ|jF||jLdur| jtO|jP0| jtS|jP|j4||jT&| jtW|jT||jXA|} |dkDs |jB| n| dz} | jt[|jX| ||j]| |} |j^rQ|jdkDrwta|j4tbrt|j4dz} nFta|j4tdjfr|j4jd dz} nd} nd} |jh3|jhdk7r$| jtk|jh|jl5|jld k7r&| jto|jl| |jp5|jpdkr&| jts|jp| |jt&| jtw|jt| |jx5|jxdkr&| jt{|jx| |j|>d|j|cxkrdkr)nn&| jt|j|| |j?d|jcxkrdkr*nn'| jt|j| ||jM| j|jj|jjj||jdur| jt| S)z This class returns a [`LogitsProcessorList`] list object that contains all relevant [`LogitsProcessor`] instances used to modify the scores of the language model head. Nr&)unconditional_idsunconditional_attention_maskr8 sequence_bias)diversity_penalty num_beamsnum_beam_groups?r )penaltyrzyPassing `encoder_repetition_penalty` requires some form of `input_ids` to be passed to `generate`, ignoring the argument.)rrz{Passing `encoder_no_repeat_ngram_size` requires some form of `input_ids` to be passed to `generate`, ignoring the argument.r&rT)top_kmin_tokens_to_keep)top_pr)min_pr)massr)epsilonr)rrr)IrEguidance_scaleappendrTr8rrMrrBrrencoder_repetition_penaltyrrr<warningswarn UserWarningrzrLno_repeat_ngram_sizerJencoder_no_repeat_ngram_sizer; bad_words_idsrIr& min_lengthrrFmin_new_tokensrGrKforced_bos_token_idr@forced_eos_token_idrArqremove_invalid_valuesrC exponential_decay_length_penaltyr?suppress_tokensrObegin_suppress_tokensrN_merge_criteria_processor_listrvrlistrqr temperaturerPrrQrrRrrH typical_prSepsilon_cutoffr= eta_cutoffr>watermarking_configconstruct_processorrrxryrenormalize_logitsrD) rr rrrrirr rr processors begin_indexrs rx_get_logits_processorz%GenerationMixin._get_logits_processorRs2")*  #!   + + 7%44&91O/99    * * 6   9HYHgHgh i  . . :?P?b?beh?h   /&7&I&I/99$5$E$E   8 8 D!<<C$**+q0!!; 1 L L*; 9  / / ;@Q@d@dhk@k   >GXGkGkl m  1 1 =BSBhBhklBl   :;L;a;ab c  : : F!>>B$**+q0!!7)FF) 9  * * 6   )%33%77   ( ( 4)+>EQ!,,q0   (%00%77!   , , 8)+>EQ!0014   1(%44%77!   $ /   0,%//3D3T3TT   0 0 <   -%99   0 0 <   -%00%99!   2 2d :   9; <  = = I   -%FF%77(   , , 8   -%55!   2 2 >.K)1,0A0U0U0] 1_    4%;;! 88EUV   & &!**Q./AA4H),->-P-P)QTU)U& 1 C CU\\R):)L)L)R)RST)UXY)Y&)*&%&"!,,8=N=Z=Z^a=a!!"9:K:W:W"XY &&27H7N7NRS7S!!$+<+B+BWij!&&27H7N7NQT7T!!$+<+B+BWij!&&2!!$+<+B+BWij!**6;L;V;VY\;\!!'->-H-H]op!//;FWFfFf@lil@l!!' 1 @ @Ug !++7CBSB^B^1` at generate-time. normalize_logits (`bool`, *optional*, defaults to `False`): Whether to normalize the logits (which, for legacy reasons, may be unnormalized). Return: `torch.Tensor`: A `torch.Tensor` of shape `(batch_size*num_return_sequences, sequence_length)` containing the transition scores (logits) Examples: ```python >>> from transformers import GPT2Tokenizer, AutoModelForCausalLM >>> import numpy as np >>> tokenizer = GPT2Tokenizer.from_pretrained("gpt2") >>> model = AutoModelForCausalLM.from_pretrained("openai-community/gpt2") >>> tokenizer.pad_token_id = tokenizer.eos_token_id >>> inputs = tokenizer(["Today is"], return_tensors="pt") >>> # Example 1: Print the scores for each token generated with Greedy Search >>> outputs = model.generate(**inputs, max_new_tokens=5, return_dict_in_generate=True, output_scores=True) >>> transition_scores = model.compute_transition_scores( ... outputs.sequences, outputs.scores, normalize_logits=True ... ) >>> # input_length is the length of the input prompt for decoder-only models, like the GPT family, and 1 for >>> # encoder-decoder models, like BART or T5. >>> input_length = 1 if model.config.is_encoder_decoder else inputs.input_ids.shape[1] >>> generated_tokens = outputs.sequences[:, input_length:] >>> for tok, score in zip(generated_tokens[0], transition_scores[0]): ... # | token | token string | log probability | probability ... print(f"| {tok:5d} | {tokenizer.decode(tok):8s} | {score.numpy():.3f} | {np.exp(score.numpy()):.2%}") | 262 | the | -1.414 | 24.33% | 1110 | day | -2.609 | 7.36% | 618 | when | -2.010 | 13.40% | 356 | we | -1.859 | 15.58% | 460 | can | -2.508 | 8.14% >>> # Example 2: Reconstruct the sequence scores from Beam Search >>> outputs = model.generate( ... **inputs, ... max_new_tokens=5, ... num_beams=4, ... num_return_sequences=4, ... return_dict_in_generate=True, ... output_scores=True, ... ) >>> transition_scores = model.compute_transition_scores( ... outputs.sequences, outputs.scores, outputs.beam_indices, normalize_logits=False ... ) >>> # If you sum the generated tokens' scores and apply the length penalty, you'll get the sequence scores. >>> # Tip 1: recomputing the scores is only guaranteed to match with `normalize_logits=False`. Depending on the >>> # use case, you might want to recompute it with `normalize_logits=True`. >>> # Tip 2: the output length does NOT include the input length >>> output_length = np.sum(transition_scores.numpy() < 0, axis=1) >>> length_penalty = model.generation_config.length_penalty >>> reconstructed_scores = transition_scores.sum(axis=1) / (output_length**length_penalty) >>> print(np.allclose(outputs.sequences_scores, reconstructed_scores)) True ```Nrrr&rL)rqrbrrNrrexpandrstackreshape transposerrxryr r log_softmaxrsumrrgather) rrhrirrbeam_indices_maskmax_beam_lengthbeam_sequence_indicescut_idxindicestransition_scoress rxcompute_transition_scoresz)GenerationMixin.compute_transition_scoressh   <<q (:;@@QGJJ9K[K[\L'..r3v;?LV$,,S["=GG1M ^^B (C(C(E(P(PRXR^R^_aRbcFXX((44V4CF^^B R(89F)1,05577<z6GenerationMixin._validate_assistant..s3X\ T*go6L6Ld.SSs7:zThe main model and the assistant don't have compatible encoder-dependent input shapes. Ensure you load the assistant with the correct encoder-decoder class, e.g. `AutoModelForSpeechSeq2Seq` for Whisper.zc(see https://huggingface.co/docs/transformers/en/generation_strategies#universal-assisted-decoding)z`assistant_tokenizer` is not required when the main and assistant models use the same tokenizer. Please omit `assistant_tokenizer` from `generate()` .z~The main and assistant models have different tokenizers. Please provide `tokenizer` and `assistant_tokenizer` to `generate()` )rrdirrQrrxry)rrhrrkattributes_to_checkr are_equal doc_references`` rx_validate_assistantz#GenerationMixin._validate_assistantsB  "  ;; ) )/2H2H2[2["b 478N8N4O"oDSW[nSn4"o "o`sI J r  ;; & & ( 3 37M7M7]7]7_7j7j j". lmzl{{|}/  $7$? UVcUddef%@'#ps C0C0c|jjrdD]}|j|dg}tt j |j j}d|vsd|vr5|tt j |jjz}|jjrt||jd}t|dd}|| t|dd}|7tt j |jj}||z}t|dd}|| t|dd}|Htt j |jj} || D chc]} d|  c} z}|jD]\}} |  ||vs|j| |rtd|d ycc} w) zXValidates model kwargs for generation. Generate argument typos will also be caught here.)rNrr rrr6z8The following `model_kwargs` are not used by the model: zG (note: typos in the generate arguments will also show up in this list))rrrrrrr rrrrrrr) rr runused_model_args model_argsrrencoder_model_argsrdecoder_model_argsxr s rx_validate_model_kwargsz&GenerationMixin._validate_model_kwargss ;; ) ), ,  d+ ,**4+M+MNYYZ  z !^z%A #g// =HHI IJ ;; ) ) t'='=tDJdIt4G:#9!*i>"%():):7??)K)V)V%W"00 dIt4G:#9!*i>"%():):7??)K)V)V%W"7IJ!!~JJ &,,. .JC S %:!((- . JK\J]^FF   Ks< Gc |rC|j7|jdk(r(tjd|jdt||jk\r9|j j rdnd}td|d|d |jd d }|r|d |jd z }|jQ|j|jkDr8tjd|jd|jd|zt|j[|j|z}||jkDrd i|}||fS) z Prepares the base generation config, then applies any generation configuration options from kwargs. This function handles retrocompatibility with respect to configuration files. FNra?You have modified the pretrained model configuration to control generation. This is a deprecated strategy to control generation and will be removed in v5. Please use and modify the model generation configuration (see https://huggingface.co/docs/transformers/generation_strategies#default-text-generation-configuration )Tz4.50.0rtransformers_versionrzX`generation_config` default values have been modified to match model-specific defaults: z=. If this is not desired, please set these values explicitly.rv) r _from_model_config_original_object_hashhashrr&_get_non_default_generation_parametersr8from_model_configrrrcopydeepcopyr parser base_version__dict__rr:rsetattrrrvrrr r*r)rHr)rr rrusing_model_generation_confignew_generation_configmodel_base_versionmodified_values global_default_generation_configmodel_generation_configrmodel_gen_config_valueglobal_default_valuecustom_gen_config_valuer s rx_prepare_generation_configz*GenerationMixin._prepare_generation_configs).%  $&&99**@@DI_I_D`` JJLMPQQ(8(J(J4;;(W%(D,B,BBMMB$ .CD* $ 6 6 ,0 )!MM*;<,")w}}T=S=S=h=h/i/v/v!w !T)"*/AW]]S[E\/\"$3C3E0*.*@*@'3J3S3S3Y3Y3[ P/C/~~c*c5K.K +23SUXZ^+_(.56Gd.S+/3GG26JJ/E, 138NO P%00C727%/%-#o2F2J''r*++hj %119595K5K5X5X%2$119595K5K5X5X%2$119595K5K5X5X%2$;;C?C?U?U?l?l%<0(//9&9  ,..rwcd|vr|d|Sd|vrY|jjsCtj|dddddftjj ddz }nd|vrY|jjrCtj|dddddftjj ddz }n8tj |tj|j ddz }d}|jd ^|d }d}t|ts|ddjd }n,t|d r |j|j}||d}||d<|S) zbCalculates `cache_position` for the pre-fill stage based on `input_ids` and optionally past lengthrNrrrr&decoder_inputs_embedsrrar get_seq_length) rrrqrSint64rrrrrrrr)r seq_lengthrr r past_lengthcaches rx_get_initial_cache_positionz+GenerationMixin._get_initial_cache_position ss | + =M0N0Z  l *4;;3Q3Q"__\/-J1aQR7-S[`[f[fgnnopqtuuN $ 49W9W -D EaAg NV[VaVabiijklopp #ZZ %++fU\\]^_bccN   - . : !23EKeU+#Ahqk//2  01e6J6J6L6X#224 +KL9N)7 %&rwcd}t|drt|jjdhk(s(t|jjddhk(rd}n0|jjDcgc] }|dvs| c}d}|jj Dcic]\}}||dvr|n|}}}|y|j j j}t|dk(r&d|vr"tjt||dSi}t|d rd}|jD]\}} | |jus|}n| td |D cgc] } || vs|  } } t| |k\r-t|D]} | D]} d | d | d vs|| || < n~ ||vr#tjt|||}nVd |vr|jd dd}ntd |d Q|D]'} t|D]} d | d | d vs|| || <')t|D]} | |vstd| d|Scc}wcc}}wcc} w)z Returns the device map for each decoder layer, to allocate the cache on the right device. Inspired from `dispatch_model` in accelerate. Nr3cpudisk)rrrr&rzw`model.get_decoder()` is not returning a named module of the model. This is unexpected, please open an issue on GitHub.rz Decoder name z" not found in execution device mapzlayer z! has not been mapped to a device.)rrr3rrrrxnum_hidden_layersrdictfromkeysrange named_modulesr RuntimeErrorrsplit)rexecution_device_map main_devicednamerrlayer_device_map decoder_namer module_namedecoder_mapped_modulesidxlayers rx$_get_layer_device_map_for_cache_initz4GenerationMixin._get_layer_device_map_for_cache_init%s $ 4 )4%%,,./E7:c$BTBTB[B[B]>^chjpbq>q# *.*<*<*C*C*EbQRaIaqbcde %)$6$6$<$<$>$ D&V%>kFJ$ $ '!KK779KK # $ )b4H.H=='8!9;OPR;ST T 4 'L $ 2 2 4  fT--//#'L #"/ 0D& +|WbGb & "& )*.?? !23"C'=" se1:K=)::4H4U,S1!""#';;+/==?P9QSghtSu+v( ,'3':':3'B1'E *]<.Hj+kll.  !23C3%qzwa[00DU0K(-  *+ TC**"VC50Q#RSS T Cc$:&s= II.I!. I'8I'cache_implementation max_cache_lenc|dk(rdt|jddvrd}t|}|jjxs|j ddu}t |dr$|r|j jn |j }|d k(r t|jj|}t |d xsGt| xs8|j|k7xs't|ttfxs|j|k} |rFt |dr:| xs6|j jj|dd j d k7} | rt |jd r|jj"} n |j$} |j'} |jj)||| || d } |dvr| j+d|j,i|di| |_|rI| j/} |dd j d | d<t1|j |di| |_|j S|j j3|j S)z Sets a cache for `generate`, that will persist across calls. A new cache will only be initialized a new `generate` call requires a larger cache or uses a different batch size. Returns the resulting cache object. hybridllama4rPrhybrid_chunkedrN_cachesliding_windowrr&_pre_quantization_dtype)rmax_batch_sizer.rrr&)staticr0offloaded_statictp_sizer.rv)rrr5rrrr3self_attention_cacherr4rr6rrr.cross_attention_cacherr5rr,rxrr9rrreset)rr-rr.rr  cache_clsrequires_cross_attention_cachecache_to_checkneed_new_cache cache_dtyper& cache_kwargsrDs rx _get_cachezGenerationMixin._get_cachess] 8 +GDKKQ]_a.>?P.QY].] ' 4 "A_T[[==eiepepN #3 3 : :MJMh' ' <~y99 <,, : <!35I J < ++m;  *gdH.Er;;44BBlSdFefgFhFnFnopFqq  t{{$=>"kkAA "jj #HHJ ++557",!.$ $4 L$'OO##Y $=>#3l3DK-!-!2!2!42>?P2QRS2T2Z2Z[\2]/1$++y?Z>?Z[ {{ KK   {{rwcJj xrtfddDS)a Return `True` if current model can use a `DynamicCache` instance when initializing the `past_key_values`. This adds exception for some models like `Mamba` models which use their own caches and do not need to initialize the Cache in advance in order to save memory (because no back and forth `to_legacy_cache` and `from_legacy_cache` will be performed for mamba-based models). c3VK|] }|jjv"ywr)rmrO)r;special_model_nameclss rxr>zBGenerationMixin._supports_default_dynamic_cache..s+, " cll&8&8&: :, s&))reformerminimaxxlnetlfm2) _is_statefulrQ)rGs`rx_supports_default_dynamic_cachez/GenerationMixin._supports_default_dynamic_caches1### , ', )  rwmax_cache_lengthcTtfddD}|sdnd}jjxs|jddu} |j|} | k|jt d|dt | tr?jr/| stj| ntj| ||<y|jd uryjs5|j(tjd |jd ty|6|j*t j#d |jd d|_|jxs't%jj'ddd|_|jf|jt(vrk|jdk(rj*s t dj-|jt/|j0|j2|z|||||<y|jdk(rjjsjs t d|j4 |j4nddi} t6| d} | ddk(rt9s t;d| ddk(rt=s t;d| di| ||<y|jdk(rt?||<y|jdk(rt||<yy| s tnttt||<y)z Prepares the cache for generation (if applicable), given `generate`'s parameterization. If a cache is instantiated, writes it to `model_kwargs`, under the name expected by the model. c3jK|]*}|jjjv,ywr)rrmrO)r; class_namers rxr>z@GenerationMixin._prepare_cache_for_generation..s*tPZjDNN,C,C,I,I,KKts03)mambafalconh1rarbrNzMPassing both `cache_implementation` (used to initialize certain caches) and `zB` (a Cache object) is unsupported. Please use only one of the two.FzThis model does not support `Cache` instances, it only supports the legacy cache format (tuple of tuples). `cache_implementation` (set to z) will be ignored.zRAn assistant model is provided, using a dynamic cache instead of a cache of type='z'.T)rr-r7zThis model does not support `cache_implementation='static'`. Please check the following issue: https://github.com/huggingface/transformers/issues/28981)r-rr.rr  quantizedzThis model does not support the quantized cache. If you want your model to support quantized cache, please open an issue and tag @zucchini-nlp.backendquantozYou need to install optimum-quanto in order to use KV cache quantization with optimum-quanto backend. Please install it via with `pip install optimum-quanto`HQQzYou need to install `HQQ` in order to use KV cache quantization with HQQ backend. Please install it via with `pip install hqq` offloadeddynamicrv) r'rrrr-rrrtrMrfrom_legacy_cacherr8rrrrrrrxr5_can_compile_fullgraphrCrrnum_return_sequences cache_configr6r# ImportErrorr"r) rr r rhrrNris_hybrid_cacherdr>user_defined_cacher] cache_classs ` rx_prepare_cache_for_generationz-GenerationMixin._prepare_cache_for_generations~t^stt.=&>  KK * * ]l.>.>?P.QY].] '*--j9  ) 55A cdncopTT,e49]9]9_:!223EF,>>?QRZ(   & &% / 335 55A BBSBhBhAij    &+<+Q+Q+]   %::;2? 6:  21B1W1W2 [b KK ' ' ' 57Mt\ .  1 1 = 559YY$99XEdNiNi$Z,0??):)O)O"#4#>#>@Q@f@fgjtt"2!!- ,;, Z(#77;F;;119]9]9_$M)55A&22#X. N>P%H ,7+F+F Z("77;F+9+; Z("779D+7> Z(E6(H  $rwcdttj|jjj vS)z Return True if the current model supports the keyword argument `logits_to_keep` in forward() to save memory. Checking it in this way allows to avoid using a new model attribute. logits_to_keep)rrrrrr)rs rx_supports_logits_to_keepz(GenerationMixin._supports_logits_to_keepCs2  3w'8'8'F'Q'Q'V'V'X#YYYrwkwargs_has_attention_maskcd fd }||j|}||j|}||j|}||j|}jj r||n|}| |j dk(r|jd}|9|7||stjd|d}tjd|djj r | td|4t|| jr||stjd |Atj|s|dkjrtjd |d ||_||_||_||_y)a Prepares the special tokens for generation, overwriting the generation config with their processed versions converted to tensor. Note that `generation_config` is changed in place and stops being serializable after this method is called. That is no problem if called within `generate` (`generation_config` is a local copy that doesn't leave the function). However, if called outside `generate`, consider creating a copy of `generation_config` first. Nc||S||n j}t|tjr|j |Stj ||tj S)Nrr)rrrqrrtensorr)tokenrrs rx_tensor_or_nonez@GenerationMixin._prepare_special_tokens.._tensor_or_noneZsQ} %1Vt{{F%.xx''<<fEJJG GrwrrzThe attention mask and the pad token id were not set. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.z)Setting `pad_token_id` to `eos_token_id`:z for open-end generation.z\`decoder_start_token_id` or `bos_token_id` has to be defined for encoder-decoder generation.r"zThe attention mask is not set and cannot be inferred from input because pad token is same as eos token. As a consequence, you may observe unexpected behavior. Please pass your input's `attention_mask` to obtain reliable results.z;`eos_token_id` should consist of positive integers, but is zq. Your generation will not stop until the maximum length is reached. Depending on other flags, it may even crash.r)r r*r)rHrrrr`rrrrr'rrqis_floating_point_bos_token_tensorr&r%_decoder_start_token_tensor) rr rfrrlbos_token_tensoreos_token_tensorpad_token_tensordecoder_start_token_tensors ` rx_prepare_special_tokensz'GenerationMixin._prepare_special_tokensJs  H++<+I+IRXY*+<+I+IRXY*+<+I+IRXY%45F5]5]fl%m" ;; ) ).H.T*Zj '  ',<,A,AQ,F/99!<   #(8(D(4=Vq 02  NNFGWFXXqr s ;; ) ).H.Pn   (!+;K[\``b(4=V##C  '  # #$4 5:JQ:N9S9S9U NNMN^M_`rr /?+.>+.>+8R5rwct|jry|jjdk(xs/t|jduxr|jj }t |jdtxr|dj}|xr|xr |j}t|dd||jjz}t|drGt|jj!}d|vxrt#|dkD}|| z}d |v}|| z}|j|st$j'd |S) zp Determines whether to trigger auto-compilation of the model's forward pass at generation time. FcudaNra hf_quantizerr3rr&rzsYou have set `compile_config`, but we are unable to meet the criteria for compilation. Compilation will be skipped.)disable_compilerr{boolcompile_config_compile_all_devicesrrrrr[rrwrrr3rrrr) rr r valid_hardwareusing_compilable_cache can_compileall_model_deviceshas_cpu_offloadhas_disk_offloads rx_valid_auto_compile_criteriaz,GenerationMixin._valid_auto_compile_criteriasS  , ,))V3 t  , ,D 8 r=N=]=]=r=r8  |''(9:E B u|TeGfGuGu %_)?_DD_D_  4 . : 4,,;; ;K 4 ) #D$6$6$=$=$? @ #'88WSAR=SVW=WO . .K &)::  // /K  + + 7   #  rw synced_gpusstreamerr^custom_generatec | jdd}| dhd}tjDcic] \}}||vs ||}}}|j| |j| fd|i| }|d=d|i|S| jdd}| jdd}|j | fi| \}|j |j|j||||0txs t|xrtjdkD}||n t}||n t}dtt!j"|j$j&j)v}d |v}|j+dddu}|j-|j.|\}}}|j0d }|j2}|j5|| |j6j8sjj:^|dkDrYt=|j0d k(rAt?j@|ddd fj:k(d kDrtBjEd|j6j8s |dk(rd_#|s|r|r|jI|||d<n-|r+|dk(r&t=|dj0d kDr tKd|j6j8rd |vr|jM|||}|j6j8r.|jO|||jP|j2\}}n|dk(r|n|jd}jRr|jU||}||jW|jY|j0d}| j+dduxrjZdu} | j+dduxrj\du}!|j_| |!||||jar d|vrd|d<|jc|| jZdz }"|j0d|k7r-|dk(r(|j6j8s|"|j0dz }"|je||||"|jg|}#|jhdkDr tKd|j2jj|j2jjk7r`tmjnd|j2jjd|j2jjd|j2jjdtp|js|||||j2|| |  }$|jtd=||d| }%jF|d<|#tvjxk(rĉjzdkDrtKd jzd!|dkDr tKd"|ds tKd#j|d$vr tKd%|j~r"tKd&|jj|j|||||||'}&|j|f|&|$|%||d(|}'n|#tvjk(rk|stBjd)|j~r"tKd*|jj|j|fj|$|%||d+|}'nK|#tvjk(rp|stBjd,|ds tKd-|j~r"tKd.|jj|j|f|$|%||d/|}'n|#tvjtvjfvrR|jd=|jz|j6j8d0|\}}|j|f|$|%||d/|}'nT|#tvjtvjfvrQ|jd=|jh|j6j8d0|\}}|j|f|$|%|d1|}'n|#tvjk(rtBjd2t|jh|j2jjjzjjZ3}(|jd=|jh|j6j8d0|\}}|j||(f|$|%|d1|}'n|#tvjk(rtBjd4g})j j})jfd5}*tjtrt=jd k(r|*jD]}+t|+d trct|+trt=|+d k(r|*td6|+Dr|*td7|+Dr|*t|+},nIt|+trt=|+d k(r|*td8|+Dr|*t|+},|)j|,t|)|jh|j2jjjzjZ9}-|jd=|jh|j6j8d0|\}}|j|f|-|$|%|d:|}'jdurAt'd;r5t|'jd<|'jj|'_e'Scc}}w)>a Generates sequences of token ids for models with a language modeling head. Most generation-controlling parameters are set in `generation_config` which, if not passed, will be set to the model's default generation configuration. You can override any `generation_config` by passing the corresponding parameters to generate(), e.g. `.generate(inputs, num_beams=4, do_sample=True)`. For an overview of generation strategies and code examples, check out the [following guide](../generation_strategies). Parameters: inputs (`torch.Tensor` of varying shape depending on the modality, *optional*): The sequence used as a prompt for the generation or as model inputs to the encoder. If `None` the method initializes it with `bos_token_id` and a batch size of 1. For decoder-only models `inputs` should be in the format of `input_ids`. For encoder-decoder models *inputs* can represent any of `input_ids`, `input_values`, `input_features`, or `pixel_values`. generation_config ([`~generation.GenerationConfig`], *optional*): The generation configuration to be used as base parametrization for the generation call. `**kwargs` passed to generate matching the attributes of `generation_config` will override them. If `generation_config` is not provided, the default will be used, which has the following loading priority: 1) from the `generation_config.json` model file, if it exists; 2) from the model configuration. Please note that unspecified parameters will inherit [`~generation.GenerationConfig`]'s default values, whose documentation should be checked to parameterize generation. logits_processor (`LogitsProcessorList`, *optional*): Custom logits processors that complement the default logits processors built from arguments and generation config. If a logit processor is passed that is already created with the arguments or a generation config an error is thrown. This feature is intended for advanced users. stopping_criteria (`StoppingCriteriaList`, *optional*): Custom stopping criteria that complements the default stopping criteria built from arguments and a generation config. If a stopping criteria is passed that is already created with the arguments or a generation config an error is thrown. If your stopping criteria depends on the `scores` input, make sure you pass `return_dict_in_generate=True, output_scores=True` to `generate`. This feature is intended for advanced users. prefix_allowed_tokens_fn (`Callable[[int, torch.Tensor], list[int]]`, *optional*): If provided, this function constraints the beam search to allowed tokens only at each step. If not provided no constraint is applied. This function takes 2 arguments: the batch ID `batch_id` and `input_ids`. It has to return a list with the allowed tokens for the next generation step conditioned on the batch ID `batch_id` and the previously generated tokens `inputs_ids`. This argument is useful for constrained generation conditioned on the prefix, as described in [Autoregressive Entity Retrieval](https://huggingface.co/papers/2010.00904). synced_gpus (`bool`, *optional*): Whether to continue running the while loop until max_length. Unless overridden, this flag will be set to `True` if using `FullyShardedDataParallel` or DeepSpeed ZeRO Stage 3 with multiple GPUs to avoid deadlocking if one GPU finishes generating before other GPUs. Otherwise, defaults to `False`. assistant_model (`PreTrainedModel`, *optional*): An assistant model that can be used to accelerate generation. The assistant model must have the exact same tokenizer. The acceleration is achieved when forecasting candidate tokens with the assistant model is much faster than running generation with the model you're calling generate from. As such, the assistant model should be much smaller. streamer (`BaseStreamer`, *optional*): Streamer object that will be used to stream the generated sequences. Generated tokens are passed through `streamer.put(token_ids)` and the streamer is responsible for any further processing. negative_prompt_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): The negative prompt needed for some processors such as CFG. The batch size must match the input batch size. This is an experimental feature, subject to breaking API changes in future versions. negative_prompt_attention_mask (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Attention_mask for `negative_prompt_ids`. use_model_defaults (`bool`, *optional*): When it is `True`, unset parameters in `generation_config` will be set to the model-specific default generation configuration (`model.generation_config`), as opposed to the global defaults (`GenerationConfig()`). If unset, models saved starting from `v4.50` will consider this flag to be `True`. custom_generate (`str`, *optional*): A string containing the name of a huggingface.co repository. If provided, the custom `generate` function defined in that reposity's `custom_generate/generate.py` file will be executed instead of the standard `generate` method. Note that the logic is for generation is entirely defined in that repository, and the return type may be different from the standard `generate` method. kwargs (`dict[str, Any]`, *optional*): Ad hoc parametrization of `generation_config` and/or additional model-specific kwargs that will be forwarded to the `forward` function of the model. If the model is an encoder-decoder model, encoder specific kwargs should not be prefixed and decoder specific kwargs should be prefixed with *decoder_*. Return: [`~utils.ModelOutput`] or `torch.LongTensor`: A [`~utils.ModelOutput`] (if `return_dict_in_generate=True` or when `config.return_dict_in_generate=True`) or a `torch.LongTensor`. If the model is *not* an encoder-decoder model (`model.config.is_encoder_decoder=False`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateDecoderOnlyOutput`], - [`~generation.GenerateBeamDecoderOnlyOutput`] If the model is an encoder-decoder model (`model.config.is_encoder_decoder=True`), the possible [`~utils.ModelOutput`] types are: - [`~generation.GenerateEncoderDecoderOutput`], - [`~generation.GenerateBeamEncoderDecoderOutput`] rN>rrrrglobal_keys_to_excludemodelrrkr&rrrrr rzA decoder-only architecture is being used, but right-padding was detected! For correct generation results, please set `padding_side='left'` when initializing the tokenizer.rTrz1`attention_mask` passed to `generate` must be 2D.)rrr rHrrqr)r rrrrrrdzZ`streamer` cannot be used with beam search (yet!). Make sure that `num_beams` is set to 1.z~You are calling .generate() with the `input_ids` being on a device type different than your model's device. `input_ids` is on z, whereas the model is on z. You may experience unexpected behaviors or slower generation. Please make sure that you have put `input_ids` to the correct device by calling for example input_ids = input_ids.to('z ') before running `.generate()`.) r rrrrirr rr)r rrr8zFnum_return_sequences has to be 1 when doing assisted generate, but is rz6assisted generate is only supported for batch_size = 1z+assisted generate requires `use_cache=True`)r7r0r4z=assisted generate is not supported with Static cache classes`zCassisted generation is not supported with stateful models, such as )r rrrhrirjrkr )r}rirr rrzDoLa Decoding is scheduled to be moved to a `custom_generate` repository in v4.55.0. To prevent loss of backward compatibility, add `trust_remote_code=True` to your `generate` call.z=dola decoding is not supported with stateful models, such as ) dola_layersrirr rrzContrastive Search is scheduled to be moved to a `custom_generate` repository in v4.55.0. To prevent loss of backward compatibility, add `trust_remote_code=True` to your `generate` call.z,Contrastive search requires `use_cache=True`zBcontrastive search is not supported with stateful models, such as )rirr rrrrUr)rirr rzGroup Beam Search is scheduled to be moved to a `custom_generate` repository in v4.55.0. To prevent loss of backward compatibility, add `trust_remote_code=True` to your `generate` call.)rrrlength_penaltydo_early_stoppingnum_beam_hyps_to_keeprrqzConstrained Beam Search is scheduled to be moved to a `custom_generate` repository in v4.55.0. To prevent loss of backward compatibility, add `trust_remote_code=True` to your `generate` call.c6tdjd)Nzo`force_words_ids` has to either be a `list[list[list[int]]]` or `list[list[int]]` of positive integers, but is r)rforce_words_idsr srx typeerrorz+GenerationMixin.generate..typeerror s*$88I8Y8Y7ZZ[]rwc3>K|]}t|t ywr)rrr; token_idss rxr>z+GenerationMixin.generate.. sY9:i#>>Ysc3@K|]}td|Dyw)c3LK|]}t|t xs|dkywrNrrr;token_ids rxr>z5GenerationMixin.generate... s'jT\Z#%>!>!N(Q,!Nj"$N)r'rs rxr>z+GenerationMixin.generate.. s$ ) j`ijjsc3LK|]}t|t xs|dkywrrrs rxr>z+GenerationMixin.generate.. s'hS[Jx$= = MA Mhr) constraintsrrrrrrrq)constrained_beam_scorerrirr rrato_legacy_cacherv)grlocalsrrrrrrrrrdistget_world_sizerErZrrrrrrrrr rrrtrrr%rrqrrrr8r1rrGrTro token_healing heal_tokensputrrqrrrerrbget_generation_moderr{rrrrrr9ASSISTED_GENERATIONr\r-rLrrmr~_assisted_decodingDOLA_GENERATIONr_dola_decodingrCONTRASTIVE_SEARCH_contrastive_searchSAMPLE GREEDY_SEARCHr[_sample BEAM_SAMPLE BEAM_SEARCH _beam_searchGROUP_BEAM_SEARCHr*rearly_stoppingr_group_beam_searchCONSTRAINED_BEAM_SEARCHrrrrr'r'r(rr+_constrained_beam_searchreturn_legacy_cacherrrar).rr r rirrrrhrrrrrrrrrr generate_argumentsrrrkr accepts_attention_maskrequires_attention_maskrfrrrrrrrrrNgeneration_modeprepared_logits_processorprepared_stopping_criteriar}result beam_scorerfinal_constraintsrword_ids constraintrs. ` rxrzGenerationMixin.generates ^#JJ':DA  && "@Fx~~?O!ueSV^tSt#u*!u !u  % %f -'@t'@'@(3D(HN( $,M$M:LM MJJ{D1 $jj)>E*I$*I*I 1+ 5;+ '< ##L$5$5$78   )=PQ  57W;QRV;Wv]a]p]p]ruv]vK/?/K+QdQf1B1N-ThTj!1S9J9J4<<9X9c9c9h9h9j5k!k"3<"G$0$4$45Et$LTX$X!9=8R8R %22L9 5 '#((+ %% $$%68QZ`$a{{--"33?N ++,1IImArE26G6Y6YYZ]^^l{{--2Bo2U*.  '(-DI_-1-X-X0,.L) *';.3|DT7U7[7[3\_`3` !TUU ;; ) ).?|.SNN|-=?PL ;; ) )&*&T&T%!1)'8'T'T$++ 'U' #I|*:[)H lN^N^_jNkI  * *((I>I   LL )%??1-!'L!9T!A!nFWFbFbjnFn!'L!9T!A!nFWFbFbjnFn ::/#9#9-'- ;   ( ( */?|/S-.L) * ''(9;KMcd -77!;    "&6 6 O3KK22  3 3A 6 6  ** |_jJZ\b ,??P  %6%@%@1%Dl  ;;  y//44 4 MM@@I@P@P@U@U?VW++**+,TTXT_T_TdTdSef* *  %)$>$>/!1+%=- ''% 3+I%? % !&AT%@%@& /CT`i& ms& " %6$?$? [! n@@ @ 559 /DDEQHA~ !YZZ , !NOO 559__ !`aa  !YZ^ZhZhZqZqYrs #'"?"?"3#+ /!1!*$7)#@ # -T,, $7!:"<"3'!  F > > >$##w   STXTbTbTkTkSlm)T(( -99!:"<"3'!  F A A A$##w , !OPP   XY]YgYgYpYpXqr.T--!:"<"3'! F!6!68T8T U U&Hd&H&H'#-BB#';;#A#A' ' #I|"T\\!:"<"3'! F!;!;^=W=W X X&Hd&H&H'#-77#';;#A#A' ' #I|'T&&!:"<"3'   F @ @ @   s  +%+55$++0??"3"B"B&7&L&L 1 A A,77 K'Id&H&H'#-77#';;#A#A' ' #I|-T,,";"<"3' F F F F   s !#  ,,8$5$A$A! 00<##4#D#DdK,<<=BK 1 A A9H!(1+t4)(D9S]a=O%KYPXYY%K-5&K%:8%D )(D9S]a=O%Kh_ghh%K%6x%@ %,,Z8)9.'B-%+55$++0??"3"B"B&7&L&L,77 ' #'Id&H&H'#-77#';;#A#A' ' #I|3T22(?!:"<"3' F  1 1T 9 12..0ABN%+%;%;%K%K%MF " q "vs nnthis_peer_finishedc|r_tj|rdnd|}tj|tjj |j dk(ryy|ryy)z Returns whether there are still unfinished sequences in the device. The existence of unfinished sequences is fed through `this_peer_finished`. ZeRO stage 3-friendly. rrr)opFT)rqrjr all_reduceReduceOpSUMrR)rrrrthis_peer_finished_flags rx_has_unfinished_sequencesz)GenerationMixin._has_unfinished_sequences s^ ',ll:L3RU^d&e # OO3 8I8I J&++-4 rwc` tdjj}}tj }t d|}j |dDcgc]}|j}}|ddjj|j}tj||k(||} |jd k(r|S|ddd fj} jj!d d fd | D} t#t%| | D]\} \} } || }tj&||k(j)r5 |j+| Dcic]}j!|fd}}t-|dk(ru|| fxxdz cc<|j/||dd } |jd k(rt-|||k7dk(r||d <|j1|j3d ||| <|Scc}wcc}w)a Generates sequences of token ids for models with a language modeling head. Parameters: input_ids (`torch.LongTensor`): The sequence used as a prompt for the generation. tokenizer (`PreTrainedTokenizerBase`, *optional*): The tokenizer used to decode the input ids. Return: `torch.LongTensor` where each sequence has its tail token replaced with its appropriate extension. Nzs When generating with token healing, you must pass the model's tokenizer to the `tokenizer` argument of `generate`.r&)rr)T)skip_special_tokenspt)return_tensorspaddingrr c3`K|]%}j|jd'yw)rN)decodereplace)r;t space_tokrs rxr>z.GenerationMixin.heal_tokens.. s)SQY%%a(00i@Ss+.)prefixg$@rrr)rr r)r get_vocabr8 batch_decodestriprrrrqwherenumeltolistconvert_ids_to_tokensconvert_tokens_to_ids enumerateziprQrR extensionsrrrr`)rrrr r) vocab_trier r<promptstail_ids tail_toks batch_idxtail_idtail_tok batch_idsalt_tokseq_bias trimmed_idsrs ` @rxrzGenerationMixin.heal_tokens sr  *  &/%;%;Y=S=Sl #I$7$7$9: ,ALY'0&<&z1GenerationMixin._dola_decoding..~ 'lh.(I'lrvNr rFrr&low(rhighz?dola_layers must be either 'low', 'high' or a list of integers.zCDoLa is not supported for models that don't have output embeddings.rT)rAr?r@r)rrrrL num_samplesrarhrirjrkrlra)4rrrr%r?r@ output_scores output_logitsreturn_dict_in_generater'rvrrqrrrrrxrtie_word_embeddingsrstrrrget_output_embeddingsrr rjdetachrfloat32floatrlrg_dola_select_contrastr}rkr~rr r softmax multinomialsqueezeargmaxrrrrendrgr))rrrrirr rrr r)r?r@rrrhas_eos_stopping_criteriarvri raw_logitsr}r~rr cur_lengthunfinished_sequencesr final_layer start_layercandidate_premature_layersilm_headrr\final_layer_next_token_logits final_logitscandidate_premature_logitscandidate_premature_layernext_token_logitsnext_token_scoresprobs next_tokenss) rxrzGenerationMixin._dola_decodingB sd ;; ) )WX X):: -??0EE)77 )77 "3"K"K$''lZk'l$l!%// 0M3 RD $;@QRX\"9>O2VZ'>CW^b"+!!4 J$zz*EJJyO_O_`77 IDTDTVbc "kk113EE {{..K 1_K A KK k3 'K5,@kQ...9]*#b({K1,>!R(399;L)+ &-G *)HO))*CDQAXNI"\(()++DE *  CC#';;#A#ADL 1 5*,F ! !2 4 4Y5E5E F  0|j"}!|jA|||jj}|s-|jBd|| |jjd|\}"}|jd}#|#"tE|jFjHdt-|#d tJtjLfr|#d d jd |k7r"tE|jFjHd||!}$tNjPjS|$d }%tjT|%d| \}&}'|r|r||!fz }|r||$fz }| rY||jjr jVfn jXfz }|jjr||jZfz }| r3||jjr j6fn j8fz }~|s|d}(t-|(t\s*t-|(t0r,t-|(j^t\r|(ja| nYg})|(D]G}*g}+|*D]$},|+jc|,j)| d &|)jctK|+ItK|)}(|(|d<|rg}-te| D]}.|j4|'dd|.fjgddfi|}/|di|/dd| d}t-|dt\s0t-|dt0r6t-|dj^t\rd|d<|djid|-jc|tk|-|jjm}n1|j4|'jgddfi|}/|di|/dd| d}~/|jjr|j6d}0|j6}1n|j8d}0|j8}1|j:dddddfjo}2 j)| d }3tq|3|0|&|| | }4tjr||ju|jd dfgd }|4j=d}4tjvty|4D.5cgc] \}.}5|5|.| zzc}5}.}6|'tet{|'|4f}7tj|tj~|0jd | }0|0te||4ddf}0tjr| |0jdgd } d}8|1D]B}*tj|tj~|*| te||4ddf}*|8|*fz }8D|r>|j4|'dd|4fjgddfi|}9|di|9dddd}:|:d};nd};tD]}<|;xs t||| r|jZD]B}*tj|tj~|*| d te||4df}*|=|*fz }=D|jVD]B}*tj|tj~|*| d te||4df}*|>|*fz }>Dt|;|8|>xsd|=xsd}ngd}?| rQ|jXD]B}*tj|tj~|*| d te||4df}*|?|*fz }?Dt|;|8|?xsd}|jA|||jj}|r|r .|r|7|z| d|z zz}7tjr||7dddfgd }||j|7j||||z}|jd k(}|j+|||j"r ||j|r|jdt-|dt\s0t-|dt0r2t-|dj^t\r|djidnXg}#|dD]@}*g}@|*D]},@jc|,dddddf|#jctK@BtK|#|d<|jjr#t|||||||jd St||||||jd!S|Scc}5}.w)"a Generates sequences of token ids for models with a language modeling head using **contrastive search** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models. Parameters: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): The sequence used as a prompt for the generation. logits_processor (`LogitsProcessorList`): An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`] used to modify the prediction scores of the language modeling head applied at each generation step. stopping_criteria (`StoppingCriteriaList`): An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3). streamer (`BaseStreamer`, *optional*): Streamer object that will be used to stream the generated sequences. Generated tokens are passed through `streamer.put(token_ids)` and the streamer is responsible for any further processing. model_kwargs: Additional model specific keyword arguments will be forwarded to the `forward` function of the model. If model is an encoder-decoder model the kwargs should include `encoder_outputs`. Return: [`~generation.GenerateDecoderOnlyOutput`], [`~generation.GenerateEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. c34K|]}t|dywrrrs rxr>z6GenerationMixin._contrastive_search..B rrrvNrrkrlr rrrrrrLFrraTr8)rAr@r?rrrrrrzQ does not support caching and therefore **can't** be used for contrastive search.z| does not have a standard cache format and therefore **can't** be used for contrastive search without further modifications.)rMrr&r.)rarr}r~)rarlrk rhrirjr{r|r}r~rrar)Mr'r penalty_alphar%r?r@rrr low_memoryrrrrrqrrrrrSrXrrrrrr rrlrjrrrgr[rrrmrtrr r rtopkr}rkr~rr:batch_repeat_interleaverrrNcropstack_model_outputsrxr _ranking_fastrrarjrrrsplitrr`r_rbatch_select_indicesrrrrrr rzrg)Arrrirr rrr r rr r)r?r@rrr sequentialr rir}r~rr{r|rcur_lenrcosine_matrix_maskrrr\last_hidden_stateslogit_for_next_steprraprocessed_logit_for_next_step next_probs top_k_probs top_k_idspastnew_key_valuesr+rrR all_outputsrnext_model_inputs next_hiddenfull_hidden_statesrjcontext_hidden selected_idxr augmented_idxrnext_decoder_hidden_statesnext_model_inputselected_outputsnext_past_key_valuesrcnext_step_cross_attentionsnext_step_decoder_attentionsnext_step_attentionslayer_past_key_valuessA rxrz#GenerationMixin._contrastive_search s X%('lZk'l$l!!'')77 (:: -??0EE)77 )77 "3"K"K&11 4 RD /M$;@QRX\"9>O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf " (oobq1 G$zz*EJJyO_O_`77AQAQS_` #__YejjI ;; ) )'<7LIa!? /AA%QAO",,-?U^UeUe,f 12:<(9:UDW)>r)B& '.nnQAX&>&A&AU]]9;K;K'B'# $GG '+{{'E'E H  "'Id&H&H'"+$)+/;;+I+I'' 'OA|#/"2"23D"E"*$>>223422 #?1#5u||7LM&q)!,2215C$>>2234UU-=YH[,\ )../LRT.UJ%*ZZ e%L "K' #6"88J <>>F$&9=9W9W335^e^p^p]r&{{55(W-E-E,GG(');;99!668%335)#$56dL1t%89jIbIbdp>q007%'N!%< "$)OD!LL)?)?1)?)MNO&--eEl; <!0D26 ./ u0A(J(J(J9UVXYUY?K_K_`bdeKf(wjv(w%"+$(-1*; G "'*;" @U@UVW@X+Y_`!a )+ &+ 7 EKKu$=>uZ?PR^`a?ab*uh6* 7 #E4#E#Eao.33B:$>J$ $($&$ $).&+ $ (88I'J$(,$+:o'+?+n77TgimCn(o2LA35HI"#7#L#Ll[(==mL%'N!5< "$)CD!LLmS.@)ABC&--eEl; <,1+@("'++ekk&%.H"I%PZJ[]iklJl"m "5"8"89I9I"J {{---/*/1,$!(!9!9? % EKKu!,L MeT^N_amorNr s2uh>2?")!;!;A % EKKu!,L MeT^N_amorNr s4@4A*$8*D'C'Kt%?%G4 (*$$!(!3!39 % EKKu!,L MeT^N_amorNr s,8,91$8"<3;t CC#';;#A#ADL 1)),@@l+<=|L|,=>@ST"<0A#B#W#WYef !2388<&(O!-.?!@M02-$)LD188c3B3k9JKL'..u5J/KL M 7z*GenerationMixin._sample.. rrrvNrrkrlr Fr0TOKENIZERS_PARALLELISMflash_attention_2rar) fullgraphzWhen using Flash Attention 2 and a static cache, you cannot use the option `CompileConfig(fullgraph=True)` as FA2 introduces graph breaks. We overrode the option with `fullgraph=False`.Trr?r@rArrrrLr&rrrr)8r%r?r@rrrr'rvrrrrrqrrrr__call__rrenvironrrrzr7rHrrget_compiled_callprefill_chunk_size_prefill_chunkingrr rrgrjrrr}rkr~rrlr r rrrr rrrrr rzrg)$rrrirr rrr r)r?r@rrrr rvrir r}r~rr{r|rr*rr model_forwardcompile_forward is_prefillrr\rrrrs$ rxrzGenerationMixin._sample sV):: -??0EE)77 )77 "3"K"K$''lZk'l$l!%// 0M3 RD $;@QRX\"9>O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf " (oobq1 G"$zz*EJJyO_O_`77AQAQS_`  ;;LJ[\ 36BJJ/ 0{{//3FF7  !235EuL%33;7Du7U%4&55??''fBG%44> 223D3S3STM  / / ;1411)=N_R^_LJJ,,-?U^UeUe,f=4==iX<XL   L]!46G Hce f   Rf!79M Nln o@@4@" 'I,IDI CC#';;#A#ADL 1!(q"ax 8 ; ;U]]clcscs ; t !1 [batch_size * num_beams, ...]rr&r Nrrrqr)rjrs rx_flatten_beam_dimz!GenerationMixin._flatten_beam_dim}s>V\\"}}VeAhq&9%:U12Y%FGGrwrcht|j}tj|||g|ddzS)z=[batch_size * num_beams, ...] -> [batch_size, num_beams, ...]r&NrR)rjrrrs rx_unflatten_beam_dimz#GenerationMixin._unflatten_beam_dims3V\\"}}Vj)% Use a heuristic that assumes the best score comes from the current length minus the decoder prompt length. -> See detailed discussion: https://github.com/huggingface/transformers/pull/20901#issuecomment-1369845565 2. `early_stopping == "never"`: -> Estimate the best score using either `max_length` or `cur_len`, depending on the sign of `length_penalty`. -> A positive length penalty favors longer sequences, so we use `max_length` in that case. NOTE: the canonical beam search implementation can be replicated with `early_stopping="never"` and `length_penalty=0.0`, which are NOT the default flags. The default behavior was empirically found to produce better sequences (prior to 2022), and changing it is BC breaking. neverrNr&TrMkeepdimrer)rqrrr') r[r\r]r^r*rqr_rrbest_hypothetical_lengthbest_possible_running_scoreworst_finished_scores rx_check_early_stop_heuristicz+GenerationMixin._check_early_stop_heuristicsJ W $#)='14F'F $'.1C'C $&9!RaR%&@D\^lDl&m#${{+;UYY{XYcg=hij=kmst2UYY '*> >BPT6   rw!next_token_hits_stopping_criteriactj|}tj||duz}tj|}||z|zS)zv Beam Search stopping condition -- halts the generation loop if any of these conditions becomes False T)rqr'rQ)r[r^ririmprovement_possibleexists_open_beamvalid_continuationss rx%_beam_search_has_unfinished_sequencesz5GenerationMixin._beam_search_has_unfinished_sequencessZ %yy)LM#YY'78Nd!$($6$67KMf$g!!%!3!34EG`!a*,19q!W}-||JxGLLRQRSV__ !:\!IH^!!Q2D(D"DE57PPPrwrwrzryc||jtjdzz}tj||d}|j ||}|j ||} |j ||} || | fS)z Given the top-K continuations, their scores, and whether they hit a stopping criteria, select the best non-finished beams to continue beam search in the next iteration. rdrur&)rrqrr"rZ) rrwrzryrirtopk_running_log_probsnext_topk_indicesrpr\rqs rx%_get_running_beams_for_next_iterationz5GenerationMixin._get_running_beams_for_next_iterations"02S2V2VW\WdWd2ehn2n!n!JJ'=KAN ../EGXY"001GIZ[#112KM^_ "57KKKrwtop_num_beam_maskc| | dddfz}|| dz| z |zz }tj|dd|duz}||jtjdzz }||jtjdzz }||dzz }tj||fd}tj||fd}tj||fd}tj||fd}tj || d}|j ||}|j ||}|j ||}|j ||}||||fS) z Updates the finished beams if (and only if) there are new completed sequences that have a higher score than the current finished sequences. Nr&rT)axiskeepdimsrdrLru)rqrQrrrr"rZ)rrhrzr]rwrryr[r^rirrr*r_rrdid_top_num_beams_just_finishedbeams_in_batch_are_fullmerged_sequences merged_scoresmerged_beam_indicesmerged_is_sent_finishedtopk_merged_indicess rx_update_finished_beamsz&GenerationMixin._update_finished_beams/s2+LN_`dfg`gNh*h'(GaK:L,LQ_+_`"')),<2PT"UYgkoYo"p144U]]CfLL??CCEMMRU[[[ ;;vEE !99i1G%HaP ;"?QG #ii7P(QWXY"'))-=?^,_ef"g#jj)DQG&&'79LM ((8KL ))*=?RS --.EGZ[+|5EEErwc |j}|j}|j} |j} |j} |j } |j } |j}|j}|j}|j}|j}|j}|jdd\}}||z}|jjdk(r|j j"}nX|jjdk(r|j%j&}n$|j j)j*}|}d}||jdnd}t-dd|z|z}t/j0t/j2|t.j4t/j6||z t.j4fd j9|j:}|j=||j:|}|j>}|r tAd | r| rd nd}| r| rd nd}| r| rd nd} | r| rd nd}!| r| rd nd}"| r| rd nd}#| rF|j jBr0| r|d jEd nd}$| r|d jEdnd}%| |xs|dnd}&t/jF|||f|&t.jH|j:}'|jK||||'ddddd|f<|'jMjO}(t/j6||ft.jP|j:})d|)ddddf<t/jF||fdt.jP|j:}*t/j6||ft.j4|j:}+t/j2|dft.j4|j:},t/j6||ft.j4|j:}-t/jF||||z fdt.jR|j:}.|.jMjO} |jU|||j:rO|jW|'ddddd|f}/|jX|/fi|}0|0j[| rd| ini|0j[| rd| ini|d i|0ddi}1|j]|1||j jB}|r|r|1j^dddddfj9dt.j`|j:}2tbjdjg|2d }3||/|3}3| r| r||2jOfz }| r| r||3jOfz }| rY|!|j jBr |1jhfn |1jjfz }!|j jBr|"|1jlfz }"| r3|#|j jBr |1jnfn |1jpfz }#~1|jK|3||}3|3|)dddddfz}3t/jr|3|||zf}3|ju|3|'|.||||||| \}4}5}6||jW|5ddddd|dzf|}-|jK|-||}-|jw|4|5|6|-|\}'})}.|jy|(|5|*|4| |6|,|+|-||||||\}(}*} }+|jEddR|jW|.d||z f}7t{|dr|j}|d|7|d<n|dj|7|dz}|j|,|)|*|+||||| },|j|,|+|-| }|jU|||j:rO|jW|(ddd|ddf}(|jW|*ddd|f}*|jW| ddd|ddf} | dzj5jd j-}8||8z}9|(ddd|9f}(| ddd|8f} | ra| sd}*|j jBr%t|(|*||| $%|!|"|#|jEd! St|(|*||| |!|#|jEd"S|(S)#a Generates sequences of token ids for models with a language modeling head using **beam search decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models. If it's the first time you're diving into Beam Search, we recommend you read the following blog post: https://huggingface.co/blog/how-to-generate (especially the beam search section). You can recompute the sequence scores from the individual scores using the `compute_transition_scores` function (https://huggingface.co/docs/transformers/main_classes/text_generation#transformers.GenerationMixin.compute_transition_scores) Parameters: input_ids (`torch.LongTensor` of shape `(batch_size*num_beams, sequence_length)`): The sequence used as a prompt for the generation. logits_processor (`LogitsProcessorList`): An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`] used to modify the prediction scores of the language modeling head applied at each generation step. stopping_criteria (`StoppingCriteriaList`: An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3). model_kwargs: Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is an encoder-decoder model the kwargs should include `encoder_outputs`. Return: [`generation.GenerateBeamDecoderOnlyOutput`], [`~generation.GenerateBeamEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateBeamDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateBeamEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. Nr MoshiDepthDecoderImageGPTForCausalImageModelingFrr&rrLz`low_memory=True` is not supported after the beam search refactor. Please check the discussion in #35802 *after the PR got merged*, and add a comment there if your questions are not yet answered.rvrrkrlr) fill_valuerrrrdrr?r@rATrr) rorprqr*r_rvrrrryr)rwrzryrir)rhrzr]rwrryr[r^rirrr*r_rrra._reorder_cache) r[r\r]r^r*rqr_rr rhrrirjrr{r|r}r~rrarhrrirjrrkrlra)Er%r&r?r@rrrrvrrrqrr\rrrmraudio_vocab_sizer out_featuresrxryrrqrrryrrrrr!rrrfullrrUrrrrrrSr rrgrjrr r rr}rkr~rrlrr~rrrr reorder_cacherhrnrrr):rrrirr rr r)r*r?r@rrrrvrrrqrr\batch_size_unflattenedr*rryr_r n_eos_tokensrrrr) all_scoresr rr}r~rr{r|output_fill_valuerprhr\r]r^r[rirqflat_running_sequencesr model_outputsrj log_probsrwrzrybeam_idxmax_generated_length output_lengths: rxrzGenerationMixin._beam_searchfs \):: (:: -??0EE)77 )77 "3"K"K%// *99*99&11 %// 0EE*3//"1*='+y8 >> " "&9 955J ^^ $ $(H H335BBJ446AAJ$" 1=0H|))!,a Aq</09< !II ZZ5:: 6 ]U^E^glgqgq8r s  "Y     77AQAQS_` '11 t  4 RD 3 RD 5-rd $;@QRX\"9>O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf "@L?WL;LO]_!JJ J /(++##   -1,D,DYPZ\e,f!Q.)%,,.446 $kk:y*A]f]m]mn%)AqrE"jj*i!8TQVQ\Q\eneueuv !;; I'>ejjYbYiYij/4jj*aPUPZPZclcscs.t+-2KK  #5::i>N>N- ) %zz J$8 9bPUP[P[dmdtdt ,224::< ,,-?U^UeUe,f%)%;%;TeXdeL   L]!46G Hce f   Rf!79M Nln o B<BTBM CC#';;#A#ADL 1#))!R(366D ^g^n^n6oF 11&b1AI()?KI' 6<<>"33J*}9??#4"66J$&;;99'99;+668& {{55(]-K-K,MM(');;99'<<>+99;)00J RI!$71d $CCI i*i*>T1UVIQUPmPm&/"3%9#5#+#%%Qn Q MN24M1B&&'=aMgPQkM>Q'RS1 -150H0H1:}1 - LPKuKu-'=*C2S# LvL H 24HFJE`E`#'='-)*C4W!12S"3##5--FaF BI{L2B. 148D112FsGVhLhGh2ij4!126:6I6I,WhJiks6tL!23 !23AA(KkG262R2R4W$7'!1%#5--3S 3 /&*%O%O3 1 &" O,,-?U^UeUe,f`**9Q8M9M8Mq5P+QR ,,[S?S>SUV;V.WX ".!1 7 7 9>>1>EIIK*-AA a-/0 #A'<(<'<$<= " " {{--7'%0%%!-'9*?'9%5*?$0$4$45F$G  5'%0%%!-1"7$0$4$45F$G   rwrc X:;<|j}|j} |j} |j} |j} |j } |j }|j}|j}||z}t|j|z}|j}|j\}}|j||j|}|r8| r6t|Dcgc] }tdt||zD"c}|$j@dddddf}%| r1|$j@dddddfjCd|j}&t|D]T::|z}'tE|'|z|}(|(|'z })g}*t|D]1}+|*jGt|'|(D,cgc] },|+|z|,z c},3||*}-|$j@|*dddfjCt(jH|j }.tJjLjO|.d}/|/jd}0||-|/|!:}1|1||*jQdz}/|/jS|1}/| r|1%|*<|/j/||)|0z}/| | jd nd }2t)jT|/tWdd|2z|)zddd\}/}3t)jX|3|0d}4|3|0z}3< t[z5GenerationMixin._group_beam_search..s!P"!P N)Batch dimension of `input_ids` should be , but is rrvrrkrlrdrrFr&rr?r@rATrr)rrrL)current_tokensbeam_group_idxr rMlargestsortedfloor) rounding_mode)r)r*r group_indexr_next_beam_scoresnext_beam_tokensnext_beam_indicesc3BK|]}||fzywrrv)r;rrrrs rxr>z5GenerationMixin._group_beam_search..us/9WX ^4Xa[AXa[NR9srarr)r*rqrr_sequence_scoresrhrrr)>r%r&r?r@rrrrrr _beam_hypsrrrrrtrrrrrqrrrNrrrrr rrg zeros_likerjrrextendrr r rr` expand_asr"rdivrprocessrr}rkr~rrlrrris_donerQfinalizerqrr)=rrrrirr rr r)r*r?r@rrrrr num_sub_beamsrrbatch_beam_sizer*rrir r}r~rr{r|r]rr_rreordering_indicesrr\processed_scoreraw_logit_scoregroup_start_idx group_end_idx group_sizebatch_group_indicesrr*group_input_idsrrrynext_token_scores_processedrr next_indicesprocess_beam_indices beam_outputsbeam_next_tokensfinal_beam_indicessequence_outputsrrrs= @@@rxrz"GenerationMixin._group_beam_searchs V):: (:: -??0EE)77 )77 "3"K"K)) %55!_4 //0OC !!#,?? 77AQAQS_` "}Z_`oZpqUVE!PeMJ4N.O!PPqLL z !_ 4;I O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf " jj*i!8$ekkZ`a *+ A&'!&& Y(>'@A "&__Q/,,-?U^UeUe,f"[[i)?y_efN"'Z)-C5::^d!e >4==iX<XL   L]!46G Hce f   Rf!79M Nln o<\>!R(3K"L#*..B":"="=4PYP`P`"="a"'"8O "0="@ #Om$CY O *_< ')#!&z!2I'..@EoWd@efY.4f#,,?"@ %,NN3FA3M$N$Q$Q-- 0@0@%R%!%'MM$=$=%2%>%!/44R8 .>#%6~ft/+%@+NaBbBlBlmoBp$p!$5$?$?@[$\! ;VO$78%6$:$::zT^G^$_!9E8P|11!4VW 16%s1a,.>'?*'LRS]ajn2.!; %yyjPWX )J6 AM@Xs<'<^b$*22#% !-!-!5 .'9 3  4@@R3S /0#/0B#C '(;<*}389\abefrstfubv\w94L02A1J -."'))_Xq[-IK[KeKefhKi,jpr"s6Eae6L23  (Jg VV%&*,.##67WO d' 00F ?"44J$&9=9W9W335^e^p^p]r&{{55(W-E-E,GG(');;99!668%335)  9n.F.Fr.J"KQSTI  148D4!126:6I6I$%679K7L!23!!23AABTUkG""c*;Iv*N&O%)"w,,-?U^UeUe,fz7C6NSr2TX&//    %%(33+10   # 6: !23{{--7.{;%56G%H!%!1.!A'9*?'9%5*?$0$4$45F$G  5.{;%56G%H!%!1.!A1"7$0$4$45F$G  $K0 0QrVgs %^"^' rc | *+|j}|j} |j} |j} |j} |j } |j }t|j}|j}|jdd\}}|j||j|}||z|k7rtd||zd|d|r| rdnd}|r| rdnd}|r| rtdt|Dnd+|r| rdnd}|r| rdnd}|r| rdnd}|rF|j j"r0| r|dj%d nd}| r|dj%d nd}t'j(||ft&j*|j }d |ddd df<|j-||zf}d}|jd }|j/|||jrF|j0|fi|}|j3| rd| ini|j3| rd| ini|di|ddi}|j5|||j j"}|r|r|d z}|j6dddddfj9dt&j:|j}t<j>jA|d} ||| }!|!|dddfjC|!z} | jE}"|r| r|| fz }| r||fz }| rY||j j"r |jFfn |jHfz }|j j"r||jJfz }| r3||j j"r |jLfn |jNfz }| jd}#| j-|||#z} | | jdnd}$t'jP| tSdd |$z|zd dd\} }%|%|#z jU}&|%|#z}%|jW|| |%|&|"|| +| }'|'d}|'d}(|'d*t'jX|*ddf|(j[dgd}~|j%dd9t]|dr|j_|d*|d<n|dja*|r)| r't*+fd tt+D+|d z}|jbste|||rd}|j/|||jrF|jg||%&|| |jh+|! })|rv| sd|)d"<|j j"r.tk|)d#|)d"|||)d$||||j%d% Stm|)d#|)d"|||)d$|||j%d&S|)d#S)'a| Generates sequences of token ids for models with a language modeling head using **constrained beam search decoding** and can be used for text-decoder, text-to-text, speech-to-text, and vision-to-text models. Parameters: input_ids (`torch.LongTensor` of shape `(batch_size*num_beams, sequence_length)`): The sequence used as a prompt for the generation. constrained_beam_scorer (`ConstrainedBeamSearchScorer`): A derived instance of [`BeamScorer`] that defines how beam hypotheses are constructed, stored and sorted during generation, while satisfying a list of positive constraints. For more information, the documentation of [`ConstrainedBeamSearchScorer`] should be read. logits_processor (`LogitsProcessorList`): An instance of [`LogitsProcessorList`]. List of instances of class derived from [`LogitsProcessor`] used to modify the prediction scores of the language modeling head applied at each generation step. stopping_criteria (`StoppingCriteriaList`): An instance of [`StoppingCriteriaList`]. List of instances of class derived from [`StoppingCriteria`] used to tell if the generation loop should stop. generation_config ([`~generation.GenerationConfig`]): The generation configuration to be used as parametrization of the decoding method. synced_gpus (`bool`): Whether to continue running the while loop until max_length (needed to avoid deadlocking with `FullyShardedDataParallel` and DeepSpeed ZeRO Stage 3). model_kwargs: Additional model specific kwargs will be forwarded to the `forward` function of the model. If model is an encoder-decoder model the kwargs should include `encoder_outputs`. Return: [`~generation.GenerateBeamDecoderOnlyOutput`], [`~generation.GenerateBeamEncoderDecoderOutput`] or `torch.LongTensor`: A `torch.LongTensor` containing the generated tokens (default behaviour) or a [`~generation.GenerateBeamDecoderOnlyOutput`] if `model.config.is_encoder_decoder=False` and `return_dict_in_generate=True` or a [`~generation.GenerateBeamEncoderDecoderOutput`] if `model.config.is_encoder_decoder=True`. Nr rrrrvc3 K|]}dywrrvrs rxr>z;GenerationMixin._constrained_beam_search.. s5"5rrrkrlrrdr&Frr?r@rATrrrrLrr)r)r*rr_rrrrarc3<K|]}||fzywrrv)r;rrrs rxr>z;GenerationMixin._constrained_beam_search..s&$rTU\(1+%>(1+%O$rsrrrhrrr)7r%r&r?r@rrrrrrrrrrrtrrrrrqrrrNrr rrgrjrrr r rrrr}rkr~rrlr"rrrrr`rrrrrQrrqrr),rrrrirr rr r)r*r?r@rrrrrrr*rir r}r~rr{r|r]rr_rr\rrrscores_for_all_vocabryrrrrrrrrs, @@rxrz(GenerationMixin._constrained_beam_searchsX):: (:: -??0EE)77 )77 "3"K"K0;;< +55 #,??2A#6 77AQAQS_` z !_ 4;I O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf " kk:y"9U^UeUef ! AqrE!&& Y(>'@A "&__Q/,,-?U^UeUe,f=4==iX<XL   L]!46G Hce f   Rf!79M Nln o<\R>\>\+?! $5#:#:#< ' 022F #4"66J$&9=9W9W335^e^p^p]r&{{55(W-E-E,GG(');;99!668%335)+004J 1 6 6z9zCY Z 5A4L<--a0RSL-2ZZ!3q!l*:#;i#GQX\ei. * {(*4::? #$78H 9Xq[#9;K;U;UVX;Y"Z`bcI  148D4!126:6I6I,WhJiks6tL!23 !23AA(K&=$$rY^_bco_pYq$rr kG&..#6G SY6Z2[%)"U,,-?U^UeUe,fX3;;    %%(33%1<   # 6: !23{{--7.{;%56G%H!%!1.!A'9*?'9%5*?$0$4$45F$G  5.{;%56G%H!%!1.!A1"7$0$4$45F$G  $K0 0rwr}c *+|j} |j} |j} |j} |j} |j }|r| rdnd}|r| rdnd}|r| rdnd}|r| rdnd}|r| rdnd}|rF|j jr0| r|djdnd}| r|djdnd}|jdd\}}tj|tj|j}|j||j|}d}d }|j|||j r|jd }|j!|\}}|j#|j}||j#|j}|jd |jd z }||d}t%j$|}t'||jd |j j}t)||jd }d |vrQtj*|d tj,|||z|jtj fd|d <|j.|fi|} d| vr|d z| d<| j1| rd| ini| j1| rd| ini|di| }!|!j2dd| d z dfj#tj4|j**j7+t9|dkDrz5GenerationMixin._assisted_decoding..s#[AJq!Qw$7#[c36K|]}dd|ddfywrrv)r;rrs rxr>z5GenerationMixin._assisted_decoding..s'fq(9!Q'(B'fr)is_decoder_attentionrh heuristicrarr);rvr?r@rrrrrrrrqrrrrrget_candidatesrrr3r4rrbr rrjrrrr_speculative_samplingrrrr rrrrrar$update_candidate_strategyrgrt_split_model_outputsr~r}rkrrlrr rrhr num_assistant_tokens_schedulenum_assistant_tokensrzrg),rrr}rirr rrr rvr?r@rrrrir r}r~rr{r|rr*rris_first_iterationcandidate_input_idscandidate_logitscandidate_lengthis_done_candidatecandidate_kwargsrr\r valid_tokens n_matchesrselected_tokenscandidate_new_tokens new_cur_lennewly_added_lengthrrs, @@rxrz"GenerationMixin._assisted_decodingsb&// -??0EE)77 )77 "3"K"K0M3 RD $;@QRX\"9>O2VZ'>CW^b #t{{'E'EVg.?!@!D!D\!Rmq H\ ./33ODbf " (oobq1 G$zz*EJJyO_O_`77AQAQS_` "!,,-?U^UeUe,fooa(G5H4V4VW`4a 1 !1"5"8"8"E +#3#6#6t{{#C 288;iooa>PP  12Et L  $yy6 6 "5";";A"> @^@^   77GI\IbIbcdIef #335:YY()9: Wg8H.HQZQaQainisist 6 !12>4==>QfUefL</1AA1E -.   L]!46G Hce f   Rf!79M Nln o*\*G!,<+A(  A(F (## (  1 12 (!!1!12 ( u  %"2"22 3 (D=(##=(  1 12=(!!1!12 =( u  %"2"22 3 =(D,0595959 [##["%[!!1!12 [   1 12 [ !!1!12 [~*./3:> A0&A0u||,A0tC$567 A0 u||Xc]Dell1B,CC D A0J*./3:> `&`u||,`tC$567 `    `@"||","38n "    "H'||'#3- ' , ' c3h '^*. 9/9/9/3 ,- 9/ !& 9/ & 9/ uc5<<&7!88 99/v#(04 ' '  'E,,- ' uc3h/ 0 ' 'L$) //38n/! /  / c3h /bN#+N###N#|| N# + N# . N#4N#7N#N# N#f/3.2W[:> $156:AEc+c'smc!++ c #+8S%,,4Gc4R+S"T c ##67 c ctCH~.c&ell3c)1(>c cR:> %+%$$89%56 %  %N#/1EEF#.0DDE# "$88 9 #R04!& x!<<x!ell#x!u||, x!  x!  x!t/b*X6!rcgW/!)*:!;W/QYZ^Q_W/rvW/ % &W/r8L htCsTWxDX?Y6ZL \@$'@58@IL@V[VbVb@ @D   $y+yy+ y  y  y y yvZ$Z5959 PS+PS$,D>PSu||S012 PSd((Rb(gk(TU]]_*.8<:><@W[&*7;-16:AE-1)-R&R$$45R##67 R $$89 R #+8S%,,4Gc4R+S"T Rd^R""34R>*R&ell3R)1(>R%TNR"#R ~u/// 0RRhDt]b]i]inr&]aS))S6>?X6YS   SjR##R3S >*R. R 0 R , RR!R $e&6&66 7RhU]]_\##\.\0 \ , \  \>*\ $e&6&66 7\\| F##F.F0 F , F  F>*F $e&6&66 7FPH%,,H5<<HH JELLJcJcJV[VbVbJJ ell%,,5<<&, -2\\, "\\, \\,  ,, ,  ,  ,  , dCi(, , , \M-2\\M,,M,1<<MdCi( MM,4Q$||4Q!<<4Q$ll 4Q  4Q  4Q4Q4Q4Q4Q4Q u||U\\5<<7 84QlL L!& L$)<< L ,1<< L  L u||U\\5<<7 8L,3F<<3F!& 3F\\ 3F  3F ll 3F$)<<3F.3\\3F ,,3F,1<<3F!<<3F3F3F 3F3F dCi(!3F" u||U\\5<<E F#3FnN##N.N0 N , N  N !5#3#33 4N` e1##e1 e1. e1 0 e1 , e1e1N n1##n1"=n1. n1 0 n1 , n1n1 !5#3#33 4n1`H##H0H. H 0 H , HH>*H $e&6&66 7HT*5+;+;*P`*rwrcd|dd| df}|jd}|ddtj||fjdd}|jd}|ddtj||fjdd} | |z } tj| } | | k} | j ddkj } |r| |k(r| dz} |ddd| dzf}|| fS|jd}|dd| ddf}| |krF|dd| ddf}tj||z d}|j|j n|}tj|djddddf}| dkDr&tj|ddd| f|fd}|| fS|}|| fS)a Applies sampling as in the speculative decoding paper (https://huggingface.co/papers/2211.17192, algorithm 1). Returns the selected tokens, as well as the number of candidate matches. NOTE: Unless otherwise stated, the variable names match those in the paper. NrrLrr&)rr) rrqrbr rand_likerrrclampdiv_rr)rrrrrnew_candidate_input_idsqq_ir<p_iprobability_ratior_i is_acceptedrrgamma p_n_plus_1 q_n_plus_1p_primers rxrrs2!6F5F5G2GH   R (A Au||,-/F F G O OPQST UCr"A Au||,-/F F G O OPQST UCc  //+ ,C**K,&&2&.2779IY*:: Q .q/IM//AB &  ""!!&&q)q)Q' u 1i?+Jkk: #:CG LL ' G   g1 5 = =a @q I q= 99&=a)m&La%PVXYL  ""L  ""rwc*t|dk(r