eThYSSKrSSKrSSKJr SSKJr SSKJr SSK J r SSK J r \ "\ 5rSr"SS5r"S S 5r"S S \5r"S S\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS\5r"SS \5r"S!S"\5rg)#N)sparse)add_start_docstrings) get_loggerad Args: input_ids (`jnp.ndarray` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`PreTrainedTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) scores (`jnp.ndarray` of shape `(batch_size, config.vocab_size)`): Prediction scores of a language modeling head. These can be logits for each vocabulary when not using beam search or log softmax for each vocabulary token when using beam search kwargs (`Dict[str, Any]`, *optional*): Additional logits processor specific kwargs. Return: `jnp.ndarray` of shape `(batch_size, config.vocab_size)`: The processed prediction scores. c\rSrSrSr\"\5S\RS\RS\R4Sj5r Sr g) FlaxLogitsProcessor3zSAbstract base class for all logit processors that can be applied during generation. input_idsscoresreturnc2[URS35e)z"Flax method for processing logits.H is an abstract class. Only classes inheriting this class can be called.NotImplementedError __class__selfr r s c/var/www/auris/envauris/lib/python3.13/site-packages/transformers/generation/flax_logits_process.py__call__FlaxLogitsProcessor.__call__6!"~~f g  N __name__ __module__ __qualname____firstlineno____doc__r!LOGITS_PROCESSOR_INPUTS_DOCSTRINGjnpndarrayr__static_attributes__rrrrr3s>];< #++ s{{ s{{ = rrc\rSrSrSr\"\5S\RS\RS\R4Sj5r Sr g) FlaxLogitsWarper>zjAbstract base class for all logit warpers that can be applied during generation with multinomial sampling.r r r c2[URS35e)zFlax method for warping logits.rrrs rrFlaxLogitsWarper.__call__ArrrNrrrrr%r%>s>t;< #++ s{{ s{{ = rr%c \rSrSrSr\"\5S\RS\RS\ S\R4Sj5r Sr g ) FlaxLogitsProcessorListIa This class can be used to create a list of [`FlaxLogitsProcessor`] or [`FlaxLogitsWarper`] to subsequently process a `scores` input tensor. This class inherits from list and adds a specific *__call__* method to apply each [`FlaxLogitsProcessor`] or [`FlaxLogitsWarper`] to the inputs. r r cur_lenr c ^UHn[R"UR5Rn[ U5S:as[ U4Sj[ UR55SS55(d3[S[ UR55SURS35eU"XU40TD6nMU"XU5nM U$)Nc3,># UH oT;v M g7fNr).0argkwargss r 3FlaxLogitsProcessorList.__call__..UsS4RS&=4Rsrz,Make sure that all the required parameters: z for z$ are passed to the logits processor.) inspect signaturer parameterslenalllistkeys ValueErrorr)rr r r,r3 processor function_argss ` rr FlaxLogitsProcessorList.__call__PsI#--i.@.@ALLM=!A%SD9K9K9M4Nqr4RSSS$FtML^L^L`GaFbbg$..//SU#9gHH"9g> rrN) rrrrrrr r!r"intrr#rrrr*r*IsL ;< #++ s{{ S _b_j_j = rr*c|\rSrSrSrS\4SjrS\RS\RS\ S\R4S jr S r g ) FlaxTemperatureLogitsWarper`z [`FlaxLogitsWarper`] for temperature (exponential scaling output probability distribution). Args: temperature (`float`): The value used to module the logits distribution. temperaturecb[U[5(aUS:d[SU35eXlg)Nrz:`temperature` has to be a strictly positive float, but is ) isinstancefloatr=rE)rrEs r__init__$FlaxTemperatureLogitsWarper.__init__is0+u--kAoYZeYfgh h&rr r r,r c"X R- nU$r0rErr r r,s rr$FlaxTemperatureLogitsWarper.__call__os*** rrLN) rrrrrrHrIr!r"rArr#rrrrCrC`sC'E' #++s{{SUXU`U`rrCc\rSrSrSr\"S5*S4S\S\S\4SjjrS \RS \RS \S \R4S jr Sr g)FlaxTopPLogitsWarperta [`FlaxLogitsWarper`] that performs top-p, i.e. restricting to top tokens summing to prob_cut_off <= prob_cut_off. Args: top_p (`float`): If set to < 1, only the smallest set of most probable tokens with probabilities that add up to `top_p` or higher are kept for generation. filter_value (`float`, *optional*, defaults to -inf): All filtered values will be set to this float value. min_tokens_to_keep (`int`, *optional*, defaults to 1): Minimum number of tokens that cannot be filtered. Inftop_p filter_valuemin_tokens_to_keepc[U[5(a US:dUS:a[SU35e[U[5(aUS:a[SU35eXlX lX0lg)Nrg?z.`top_p` has to be a float > 0 and < 1, but is rSz:`min_tokens_to_keep` has to be a positive integer, but is )rGrHr=rArTrUrV)rrTrUrVs rrIFlaxTopPLogitsWarper.__init__sj%''EAIMeWUV V,c227IA7MYZlYmno o ("4rr r r,r cF[R"X"RS5upE[R"X R 5n[ RRUSS9RSS9nXpR:n[R"US5nXRSS2S4RS5-nURSS2SUR24RS5n[R"XU5n [ RR!XY5Sn U $)NaxisrSrT)laxtop_kshaper! full_likerUjaxnnsoftmaxcumsumrTrollatsetrVwhere sort_key_val) rr r r, topk_scores topk_indices mask_scorescumulative_probs score_masktopk_next_scores next_scoress rrFlaxTopPLogitsWarper.__call__s$'IIfll26F$G! mmF,=,=> 66>>+B>?FFBFO% 2 XXj!, mmAqD)--d33  ]]1&?(?(?&?#?@DDTJ 99ZkJgg**<J2N r)rUrVrTN rrrrrrHrArIr!r"rr#rrrrPrPtsa =B%L=de5e555^a5#++s{{SUXU`U`rrPc\rSrSrSr\"S5*S4S\S\S\4SjjrS \RS \RS \S \R4S jr Sr g)FlaxTopKLogitsWarpera [`FlaxLogitsWarper`] that performs top-k, i.e. restricting to the k highest probability elements. Args: top_k (`int`): The number of highest probability vocabulary tokens to keep for top-k-filtering. filter_value (`float`, *optional*, defaults to -inf): All filtered values will be set to this float value. min_tokens_to_keep (`int`, *optional*, defaults to 1): Minimum number of tokens that cannot be filtered. rRrSr^rUrVc[U[5(aUS::a[SU35e[X5UlX lg)Nrz6`top_k` has to be a strictly positive integer, but is )rGrAr=maxr^rU)rr^rUrVs rrIFlaxTopKLogitsWarper.__init__s<%%%!UV[U\]^ ^3 (rr r r,r c URupE[R"XE-UR5n[ UR URS5n[ R "X'5up[R"[R"U5U-SS2S4XG45R5n UR5n U R5U -n URU RU 5nURXE5n U $)NrZ) r_r!fullrUminr^r] broadcast_toarangeflattenrfrgreshape)rr r r, batch_size vocab_sizenext_scores_flattopkrjrkshifttopk_scores_flattopk_indices_flatrps rrFlaxTopKLogitsWarper.__call__s!' 88J$;T=N=NO4::v||B/0$'IIf$;!   #**Z"8:"Eq$w!OR\Qcdlln&..0(002U:+../@AEEFVW&..zF r)rUr^Nrrrrrrtrtsa ;@,bc)c))\_) #++ s{{ S UXU`U` rrtc|\rSrSrSrS\4SjrS\RS\RS\S\R4S jr S r g ) !FlaxForcedBOSTokenLogitsProcessorz [`FlaxLogitsProcessor`] that enforces the specified token as the first generated token. Args: bos_token_id (`int`): The id of the token to force as the first generated token. bos_token_idcXlgr0r)rrs rrI*FlaxForcedBOSTokenLogitsProcessor.__init__s(rr r r,r c[R"UR[S5*5nS[R"US- 5- n[R "XTR SS2UR4RS5U5nU$NinfrSr) r!rzr_rHbool_rhrfrrgrr r r, new_scores apply_penaltys rr*FlaxForcedBOSTokenLogitsProcessor.__call__siXXfllU5\M: CIIgk22 =--4;L;L8L*M*Q*QRS*TV\] rrN rrrrrrArIr!r"rr#rrrrrsC)S)#++s{{SUXU`U`rrc\rSrSrSrS\S\4SjrS\RS\RS\S \R4S jr S r g ) !FlaxForcedEOSTokenLogitsProcessoraI [`FlaxLogitsProcessor`] that enforces the specified token as the last generated token when `max_length` is reached. Args: max_length (`int`): The maximum length of the sequence to be generated. eos_token_id (`int`): The id of the token to force as the last generated token when `max_length` is reached. max_length eos_token_idcXlX lgr0)rr)rrrs rrI*FlaxForcedEOSTokenLogitsProcessor.__init__s $(rr r r,r c.[R"UR[S5*5nS[R"X0R - S-5- n[R "XTRSS2UR4RS5U5nU$r) r!rzr_rHrrrhrfrrgrs rr*FlaxForcedEOSTokenLogitsProcessor.__call__sqXXfllU5\M: CIIg&?!&CDD =--4;L;L8L*M*Q*QRS*TV\] r)rrNrrrrrrsJ)3)c)#++s{{SUXU`U`rrc\rSrSrSrS\S\4SjrS\RS\RS\S \R4S jr S r g ) FlaxMinLengthLogitsProcessora [`FlaxLogitsProcessor`] enforcing a min-length by setting EOS probability to 0. Args: min_length (`int`): The minimum length below which the score of `eos_token_id` is set to `-float("Inf")`. eos_token_id (`int`): The id of the *end-of-sequence* token. min_lengthrc[U[5(aUS:a[SU35e[U[5(aUS:a[SU35eXlX lg)Nrz2`min_length` has to be a positive integer, but is z4`eos_token_id` has to be a positive integer, but is )rGrAr=rr)rrrs rrI%FlaxMinLengthLogitsProcessor.__init__s^*c**j1nQR\Q]^_ _,,, q0@ST`Sabc c$(rr r r,r cS[R"X0R- SS5- n[R"XBRSS2UR 4R [S5*5U5nU$)NrSrr)r!cliprrhrfrrgrHrr r r,rs rr%FlaxMinLengthLogitsProcessor.__call__s\CHHW%>1EE =))At7H7H4H*I*M*MuUZ|m*\^de r)rrNrrrrrrsJ)3)c)#++s{{SUXU`U`rrc,\rSrSrSrSrS\4SjrSrg)(FlaxSuppressTokensAtBeginLogitsProcessoria [`FlaxLogitsProcessor`] suppressing a list of tokens as soon as the `generate` function starts generating using `begin_index` tokens. This should ensure that the tokens defined by `begin_suppress_tokens` are not sampled at the beginning of the generation. Args: begin_suppress_tokens (`List[int]`): Tokens to not sample. begin_index (`int`): Index where the tokens are suppressed. c0[U5UlX lgr0)r;begin_suppress_tokens begin_index)rrrs rrI1FlaxSuppressTokensAtBeginLogitsProcessor.__init__s%)*?%@"&rr,cS[R"X0R- 5- n[R"XBRSS2UR 4R [S5*5U5nU$)NrSr)r!rrrhrfrrgrHrs rr1FlaxSuppressTokensAtBeginLogitsProcessor.__call__!s]CIIg0@0@&@AA =))At7Q7Q4Q*R*V*VX]^cXdWd*egmn r)rrN) rrrrrrIrArr#rrrrrs '3rrc|\rSrSrSrS\4SjrS\RS\RS\ S\R4S jr S r g ) !FlaxSuppressTokensLogitsProcessori)z [`FlaxLogitsProcessor`] suppressing a list of tokens at each decoding step. The processor will set their log probs to be `-inf` so they are not sampled. Args: suppress_tokens (`list`): Tokens to not sample. suppress_tokensc$[U5Ulgr0)r;r)rrs rrI*FlaxSuppressTokensLogitsProcessor.__init__3s#O4rr r r,r cnURSUR4R[S5*5nU$)N.r)rfrrgrHrMs rr*FlaxSuppressTokensLogitsProcessor.__call__6s13 4 445995<-H r)rN) rrrrrr;rIr!r"rArr#rrrrr)sC55#++s{{SUXU`U`rrct\rSrSrSrSrS\RS\RS\S\R4Sjr S r g ) FlaxForceTokensLogitsProcessori<a [`FlaxLogitsProcessor`] that takes a list of pairs of integers which indicates a mapping from generation indices to token indices that will be forced before sampling. The processor will set their log probs to 0 and all other tokens to `-inf` so that they are sampled at their corresponding index. Args: force_token_map (`list`): Map giving token ids and indices where they will be forced to be sampled. cH[U5n[R"[UR 55S-[R S9S-nUR 5H(up4UcM URURU5nM* [R "U5Ul g)NrSdtyperZ) dictr!onesrwr<int32itemsrfrgforce_token_array)rforce_token_maprindextokens rrI'FlaxForceTokensLogitsProcessor.__init__Gs/ HHc/*>*>*@&AA&EciiX[]]+113LE $5$8$8$?$C$CE$J!4"%+TRSnTRUn[R"TTRS9[ S5*-n[R "US4TRS9n[R"X4SU45nU$)NrrrrS) r_rr! ones_likerrHzerosr]dynamic_update_slice)generation_idxr current_tokenrupdatesr rs r _force_token=FlaxForceTokensLogitsProcessor.__call__.._force_tokenSstaJ 22>BMvV\\BeEl]RJiiQv||DG11*=GYZJ rrc>T$r0rr sr9FlaxForceTokensLogitsProcessor.__call__.._sFrcf>[R"TRTS:UU4SjU4Sj5$)Nrc>T"T5$r0r)rr,srrKFlaxForceTokensLogitsProcessor.__call__....ds W-rc>T$r0rrsrrrfsr)r]condr)rr,r rsrrras'CHH&&w/14- r)r]rrr_)rr r r,rs` ``@rr'FlaxForceTokensLogitsProcessor.__call__Rs@  t--33A6 6     r)rN) rrrrrrIr!r"rArr#rrrrr<s< >#++s{{SUXU`U`rrc$\rSrSrSrSrSrSrg)#FlaxWhisperTimeStampLogitsProcessorilaG Whisper specific Processor. This processor can be used to force a list of tokens. The processor will set their log probs to `inf` so that they are sampled at their corresponding index. Args: generate_config (`GenerateConfig`): The generate config used to generate the output. The following parameters are required: eos_token_id (`int`, *optional*, defaults to 50257): The id of the *end-of-sequence* token. no_timestamps_token_id (`int`, *optional*, defaults to 50363): The id of the `"<|notimestamps|>"` token. max_initial_timestamp_index (`int`, *optional*, defaults to 1): Used to set the maximum value of the initial timestamp. This is used to prevent the model from predicting timestamps that are too far in the future. cvURUlURUlURS-UlUS-UlUR(aU=RS- sl[ US5(aUR UlOURUlUR cURUlgg)NrSrmax_initial_timestamp_index)rno_timestamps_token_idtimestamp_beginris_multilingualhasattrrr)rgenerate_config model_configdecoder_input_lengths rrI,FlaxWhisperTimeStampLogitsProcessor.__init__}s+88&5&L&L#.EEI/!3  * *    !  ?$A B B/>/Z/ZD ,/;/F/FD ,  + + 3/;/F/FD , 4rc^^URSS2TR4R[S5*5nUU4Sjn[R "U5"X5n[ R"TTR:HSS5n[ R"TRSLUS5nTRTR-n[ R"UURSS2US-S24R[S5*5U5n[RRUSS9nU4Sjn[R "U5"Xr5nU$) Nrc r>[R"TTR- S:SS5n[R"UTS- TR:US5n[R"TTR- S:SS5n[R"UTS- TR:SU5n[R"U[R"US:URTRSR [ S5*5URSTRR [ S5*55U5$)NrSTFrrr)r!rhrrrfrgrHr) input_ids_kscores_klast_was_timestamppenultimate_was_timestampr,rs r handle_pairsBFlaxWhisperTimeStampLogitsProcessor.__call__..handle_pairss"!$Gd6F6F,F1+LdTY!Z !$GaK(D,@,@@+"  ), 7T=M=M3MQR2RTXZ_(` %(+ GaK(D,@,@@)) % 99" -1KK 4 4 67;;U5\MJKK 3$"3"3488%,G  rTFrSrZr[c4>[RRUTRSSS9n[R "USTR5n[R "X#:URSTRR[S5*5U5$)NrZr[r) rarb logsumexprr!rwrhrfrgrH) logprobs_krtimestamp_logprobmax_text_token_logprobrs rhandle_cumulative_probsMFlaxWhisperTimeStampLogitsProcessor.__call__..handle_cumulative_probss # 0 0D>e }M 2,' :&)ii4;K;K0KTSX&Y#&)ii  , ,D 80 ' # ++d.N.NN  ' IIa)++ , 0 0%, ?  66%%f2%6 128D r)rrrrrN)rrrrrrIrr#rrrrrls G"<rrc\rSrSrSrS\4SjrS\RS\S\4Sjr S \RS \R4S jr S\RS \RS\S \R4S jr Sr g) FlaxNoRepeatNGramLogitsProcessoria! [`FlaxLogitsProcessor`] that enforces no repetition of n-grams. See [Fairseq](https://github.com/pytorch/fairseq/blob/a07cb6f40480928c9e0548b737aadd36ee66ac76/fairseq/sequence_generator.py#L345). Args: ngram_size (`int`): All ngrams of size `ngram_size` can only occur once. ngram_sizecb[U[5(aUS::a[SU35eXlg)Nrz;`ngram_size` has to be a strictly positive integer, but is )rGrAr=r)rrs rrI)FlaxNoRepeatNGramLogitsProcessor.__init__s/*c**jAoZ[eZfgh h$rr rr,c ^^^ TRum nUTRS- - nUTRS- - nU UU4SjnT U-TRS-4n[RR ST U-U[ R "UTRS95n [ R"T U-5T U-:RS5n [R"X4T 4U4TR--S9$)z get a matrix of size (batch_size,) + (vocab_size,)*n (for n-grams) that represent the n-grams that occurred previously. The BCOO representation allow to store only the few non-zero entries, instead of the full (huge) matrix rSc  >UT-nUT-nURUR[R"U/[ TR 5Vs/sH n[R"T5X#U-4PM" sn-55$s snfr0)rfrgr!arrayranger)ivalbposjrr rs rbody_funFFlaxNoRepeatNGramLogitsProcessor.get_previous_ngrams..body_funsJAz/C66!9== BGtAWXAWAsyy+AQwJ7AWXY  Ys 'Brrfloat32)r_) r_rrar] fori_loopr!rrr}astyperBCOO) rr rr,seq_len seq_ngrams cur_ngramsr r_all_update_indicesdatars `` @rget_previous_ngrams4FlaxNoRepeatNGramLogitsProcessor.get_previous_ngramss (oo G! 34 ! 34  j($//A*=> WW.. zJ&#))E2Y   : 23j:6MMUUV_`{{D5j]j]]a]l]lMl=lmmr latest_tokensr c[R[RS55n[R"U"X55$)z\ Determines which tokens must be banned given latest tokens and the previously seen ngrams. cU[U5$r0)tuple)rprevious_ngramss rinner_fnIFlaxNoRepeatNGramLogitsProcessor.get_banned_tokens_mask..inner_fns#5#78 8r)rsparsifyrar bcoo_todense)rrrrs rget_banned_tokens_mask7FlaxNoRepeatNGramLogitsProcessor.get_banned_tokens_masks>   9   9""8M#KLLrr c^^^^UUUU4Sjn[RRTTRS- :UU4Sj5nU$)Nc >TRupTRTUT5n[R"TRSTRS- 4TR S9n[ RRU[ RRTSTTRS- - 4TRSTRS- 45S5nTRX25RS5n[R"U[S5*T5$)NrrSr)rrboolr)r_rr!rrrrar]r dynamic_slicer!rrhrH) _rrrbanned_tokens_indices_maskr,r r rs rtrue_fn:FlaxNoRepeatNGramLogitsProcessor.__call__..true_fn s"LLMA"66y*gVO IIyq'94??Q;N&OW`WfWfgMGG88%%7doo.A#BCiooVWFX[_[j[jmn[nEp M*.)D)D])d)k)klr)s &997%,O OrrSc>T$r0rrsrr;FlaxNoRepeatNGramLogitsProcessor.__call__..sQWr)rar]rr)rr r r,r)outputs```` rr)FlaxNoRepeatNGramLogitsProcessor.__call__ s: P P&w$//A*==X r)rN) rrrrrrArIr!r"rr!rr#rrrrrs%3% nS[[ nc nTW nD MCKK MUXU`U` M#++s{{SUXU`U`rr)r6rajax.laxr] jax.numpynumpyr!jax.experimentalrutilsr utils.loggingrrloggerr rr%r;r*rCrPrtrrrrrrrrrrrr6s  #(& H %!*    d."2()+)X + F(;,(;2#6</B2(;&-%8-`^*=^BS':Sr