+ jgN^RIHt^RIHt^RIHtHtHt^RIt^RI H t H t H t H t ^RIHtHtHtHtHt^RIHt^RIHt^R IHt^R IHtHt^R IHtHtH t ^R I!H"t"]'d^R I#H$t$] PJ!]&4t']!RR]44t(!RR]4t)!RR]4t*R#))Callable) dataclass) TYPE_CHECKINGAnyOptionalN) DynamicCacheQuantizedCacheQuantizedLayer StaticCache)GenerateDecoderOnlyOutputGenerationConfigGenerationMixinLogitsProcessorLogitsProcessorList) ALL_STATIC_CACHE_IMPLEMENTATIONS)!LOGITS_PROCESSOR_INPUTS_DOCSTRING)StoppingCriteriaList)ALL_CACHE_NAMESGenerateNonBeamOutput)add_start_docstringsis_diffusers_availablelogging),VibeVoiceAcousticTokenizerConv1dPaddingCache) BaseStreamerc4a]tRt^.toRtRtV3RltRtVtR#)VibeVoiceGenerateOutputz Outputs of VibeVoiceForConditionalGeneration.generate. Args: audio (`list(torch.FloatTensor)` of length `batch_size`): The generated audio. NcX<V^8dQh/S[S[P,R,;R&#)Naudio)listtorch FloatTensor)format __classdict__s"s/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/vibevoice/generation_vibevoice.py __annotate__$VibeVoiceGenerateOutput.__annotate__.s" !! "T )0) __name__ __module__ __qualname____firstlineno____doc__r__annotate_func____static_attributes____classdictcell__r$s@r%rr.s-1Er(rcba]tRt^;toRtRV3RlRllt]!]4V3RlRl4tRt Vt R#) !VibeVoiceTokenConstraintProcessora Constrains token generation to only diffusion-related tokens during audio generation, as the role of the language model is to emit: - another audio-diffusion placeholder (which triggers the diffusion head to synthesize the next acoustic latent) - or an EOS token (which signals the end of the audio generation). The actual audio comes from the diffusion head, not from sampling the vocabulary. NcP<V^8dQhRS[S[,RS[RS[P/#)rvalid_token_ids vocab_sizedevice)r intr!r8)r#r$s"r%r&.VibeVoiceTokenConstraintProcessor.__annotate__Ds*22S 2s2ELL2r(c|\P!V3\R44pRWA&VPV4VnR#)z-infgN)r!fullfloatto logits_mask)selfr6r7r8r?s&&&& r%__init__*VibeVoiceTokenConstraintProcessor.__init__Ds1jj*f > '* $&>>&1r(ch<V^8dQhRS[PRS[PRS[P/#)r input_idsscoresreturn)r! LongTensorr")r#r$s"r%r&r:Ks5))%"2"2)E forward -> update model kwargs -> append the new token). c <V^8dQhRS[/#)rrF)r)r#r$s"r%r&%VibeVoiceGenerationMixin.__annotate__[s  8K r(cH<\SV`!V/VBpVPPVPPVPP VPP .pVP\W@PPVPR74V#)r8) super_get_logits_processorconfigaudio_bos_token_idaudio_eos_token_idaudio_token_id eos_token_idappendr4r7r8)r@argskwargs processors valid_tokens __class__s&*, r%rT.VibeVoiceGenerationMixin._get_logits_processor[sW2DCFC KK * * KK * * KK & & KK $ $  ;L++J`J`imitituvr(cJ<V^8dQhRS[R,RS[S[S[3,/#)rgeneration_configNrF)r tupledict)r#r$s"r%r&rPfs-5/5/!1D!85/ % &5/r(c P<VPRR4pVPRR4pVPRR4p\SV`!V3/VBwrT;'g\VRR4pVfVP V4pVf \ R4h\ VR4'd%\ VR4'd\ VR 4'g#\ R \V4P R 24hWqn W1n T;'g\VRR4;'g^ Vn VPfR Vn W3#) ab This method overrides [~generation.utils.GenerationMixin._prepare_generation_config]. It extracts VibeVoice-specific parameters for the generation config. VibeVoice-specific parameters include: - `noise_scheduler`: A custom noise scheduler instance. Optional: if not provided, a default one is built from `noise_scheduler_class`/`noise_scheduler_config` on the generation config (requires `diffusers`). - `monitor_progress`: Whether to display a progress bar tracking audio generation. Defaults to `False`. - `num_diffusion_steps`: Number of diffusion steps for audio latent sampling. Defaults to `10`. monitor_progressFnum_diffusion_stepsNnoise_schedulera2VibeVoice generation requires a `noise_scheduler`. Either pass one to `generate(...)`, e.g. `diffusers.DPMSolverMultistepScheduler(beta_schedule='squaredcos_cap_v2', prediction_type='v_prediction')`, or set `noise_scheduler_class` (and optionally `noise_scheduler_config`) on the model's generation config. set_timestepsstep timestepszThe provided noise_scheduler (z) is not compatible with VibeVoice generation. It must implement `set_timesteps` and `step` methods, and have a `timesteps` attribute.g?) poprS_prepare_generation_configgetattr_build_default_noise_scheduler ValueErrorhasattrtyper*rhrfrgguidance_scale) r@rbr\rfrgnoise_scheduler_kwarg model_kwargsrhr_s &&, r%rm3VibeVoiceGenerationMixin._prepare_generation_configfsL "::&8%@$jj)>E & +SUY#Z ` `^` -  + + 3/2  , ..r(c <V^8dQhRS[/#)rrbr )r#r$s"r%r&rPs33:J3r(c\VRR4p\4'g\RV R24h^RIp\W!4p\TRRR R R /4pT!R /TB# \d\ RT R24hi;i) noise_scheduler_classDPMSolverMultistepSchedulerz(The default VibeVoice noise scheduler (`z`) requires `diffusers`. Install it with `pip install diffusers`, or pass a custom `noise_scheduler` instance to `generate(...)`.Nz Could not find noise scheduler `z` in `diffusers`. Set `noise_scheduler_class` on the generation config to a valid `diffusers` scheduler, or pass a custom `noise_scheduler`.noise_scheduler_config beta_schedulesquaredcos_cap_v2prediction_type v_predictionr))rnr ImportError diffusersAttributeErrorrp)rbscheduler_class_namerscheduler_classscheduler_configs& r%ro7VibeVoiceGenerationMixin._build_default_noise_schedulers&'8:QSpq%'':;O:PQpp   %iFO #  $ 13Dn U  2!122 23G2HIqq  s AA1c p<V^8dQhRS[RS[RS[PRS[S[P S[3,/#)r batch_sizerbr8rF)r9r r!r8rcrGrd)r#r$s"r%r&rPsG;9;9;9,;9 ;9 u% & ;9r(c  R\P!V^3VPP\PVR7R\P !V^3\PVR7/pVP !V3/VBwrVVPRW`PPR7wrvVPVRVR7VR,pVPR4RJ;'dVPRJp VPR 4RJ;'dVPRJp VPVV V RVR,VP^,R 7pVPVVRVVP^, R R 7W3#) a! Set up the unconditional branch used for classifier-free guidance (CFG). This mirrors the positive branch's own setup in `generate`/`_sample` (resolve the generation config, prepare the model inputs and KV cache), applied to the negative branch. See the class docstring for the original implementation. Returns the initial negative `input_ids` and prepared `model_kwargs` for the negative pass. The negative sequence starts with a single `audio_bos_token_id` token and its KV cache is sized to match the positive generation's maximum length. rDdtyper8attention_maskN)ru bos_token_idTrR max_length min_length)rbhas_default_max_lengthhas_default_min_lengthmodel_input_name inputs_tensorinput_ids_length#_previous_max_negative_cache_length)max_length_attr_name)r!r<rUrVlongonesrm_prepare_model_inputs_prepare_special_tokensgetrr_prepare_generated_lengthshape_prepare_cache_for_generation) r@rrbr8negative_kwargsnegative_generation_confignegative_model_kwargs_negative_input_idsnegative_has_default_max_lengthnegative_has_default_min_lengths &&&& r%_prepare_negative_generation5VibeVoiceGenerationMixin._prepare_negative_generations& ZOT[[5S5S[`[e[entu ejj*a SYZ  =A<[<[ = !0= 9"'+&@&@ 4;;CaCa'A' #1 $$%?f$U,[9    - 5 k k:T:_:_gk:k (    - 5 k k:T:_:_gk:k (&*%C%C8#B#B()+6/55a8 &D& " ** & !   & 1 1A 5!F + "88r(cB<V^8dQhRS[RS[RS[RS[RS[RS[RR/#) rrbrugeneration_modermax_cache_lengthrrFN)r rdrr9str)r#r$s"r%r&rPscBLBL+BLBL BL  BL  BL"BL BLr(c lRpVPV4pVeAVPe\RV R24h\V\4'd \R4hR#VP RJdR#VP\ 9dyVPe\WQP4p\VPVP4V,p VPVPV VVPVR7W'&R#VPR8XdlVPe VPM/p V PR VPP!R R 74V P#R R 4p \%RR V /V BW'&R#R VPP!R R 7/p VPR8XdR V R&\'R/V BW'&R#)a This method overrides [~generation.utils.GenerationMixin._prepare_cache_for_generation]. The base implementation memorizes the longest static cache length requested so far on the fixed `self._previous_max_cache_length` attribute, so that later calls are served by a cache of that same shape (avoiding a `torch.compile` recompilation). VibeVoice runs a second, independent generation pass for the classifier-free guidance negative branch, whose cache is much shorter than the positive branch's (the negative branch starts from a single token instead of the full prompt). The added `max_length_attr_name` argument lets the caller pick which attribute holds that memo (`"_previous_max_cache_length"` for the positive branch, `"_previous_max_negative_cache_length"` for the negative one), so the negative cache is not over-allocated to the positive branch's length. Only the code paths exercised by VibeVoice are kept: decoder-only dynamic and static caches. past_key_valuesNzMPassing both `cache_implementation` (used to initialize certain caches) and `zB` (a Cache object) is unsupported. Please use only one of the two.z]Passing a tuple of `past_key_values` is not supported anymore. Please use a `Cache` instance.F)cache_implementationr max_cache_lenprefill_chunk_sizer quantizedrUTdecoderbackendquanto offloaded offloadingr))rrrp isinstancerc use_cacherrmax num_beamsnum_return_sequences_prepare_static_cacher cache_config setdefaultrUget_text_configrlr r) r@rbrurrrr cache_nameuser_defined_cachecache_batch_sizerrdynamic_cache_kwargss &&&&&&& r%r6VibeVoiceGenerationMixin._prepare_cache_for_generations,' *--j9  ) 55A cdncopTT,e44 s   & &% /   1 15U U!..:#&'79X9X#Y "#4#>#>@Q@f@fgjtt '+'A'A%6%K%K+.#4#G#G%9 (B(L $ 3 3{ B=N=[=[=g,99moL  # #Hdkk.I.IRV.I.W X"&&y(;G'5'Vg'V'VL $%-dkk.I.IRV.I.W#X  55D59$\2'3'K6J'KL $r(c L<V^8dQhRS[RS[RS[RS[R,RS[RS[/#)rrrrrNrrF)rr9r )r#r$s"r%r&rP8sK,,!,, ,  $J , " , ,r(c 2RV9p\WR4p\W74p\VPP RR7VVR7p VeCVP 4p V e/V wrV P VV V VPVPR7\WV4V #)a/ This method overrides [~generation.utils.GenerationMixin._prepare_static_cache]. Same behavior as the base method (a fresh `StaticCache` per call, sized to the longest length requested so far so that `torch.compile` does not recompile on later calls), except the memorized length is stored under the attribute named by `max_length_attr_name` instead of always `self._previous_max_cache_length`. This lets the CFG negative branch track its own (much shorter) length. Only the decoder-only path used by VibeVoice is kept. Returns the resulting cache object. rTr)rUrr)r num_headshead_dimrr8) rnrr rUr_get_static_cache_init_shapeearly_initializationrr8setattr) r@rrrrr offload_cacheprevious_max_leneffective_lengthcache init_shaperrs &&&&&& r%r.VibeVoiceGenerationMixin._prepare_static_cache8s&$';; "4rB}?;;..t.<*$   )::r8) r@rrrrreset_attention_mask last_input_idmask_4dlayer layer_mask_4ds &&&& r%%_reset_negative_cache_for_audio_start>VibeVoiceGenerationMixin._reset_negative_cache_for_audio_startws//?@$//?&'U#27++ * *2 .0D3 ./+51 $)KK %//-A_A_"`bo% 5! ! $ $%6 7 C*//Aq!randnrrU audio_config hidden_sizerilenrkrepeatsplitrj prev_sampler)r@r rrhrgrsdiffusion_head_devicernoisy_audio_latenthalftimestepcombinedepscond_eps uncond_eps guided_epss&&&&&& r%_sample_audio_latent-VibeVoiceGenerationMixin._sample_audio_latents!%TZZ%>%>%I%I%K L S SII1FANQQRgh "[[);T[[=U=U=a=abeefop%%:M%N%&!+'11Hyy"4Ud";=OPUQU=V!W]^_H**++//(..*;<??IU^,C$);;sa#@ H#8M&NNJ))Z4!r8latent_bias_factorallr!zerosrr audio_towerdecode)r@r&rrr' scaled_latent padded_latents&&&&& r%_decode_audio_latent-VibeVoiceGenerationMixin._decode_audio_latents%zz'G'G'J'J   (  JJ ) ) , ,\-@-@ AB !!##!KK M]5H5H5LMPP$$m&9&9M-:. ))Mzz%%,,   TZZ33:: ;(-  r(c<V^8dQhRS[PRS[RS[RS[RS[RS[R,RS[S[P,/#) rrDlogits_processorstopping_criteriarb synced_gpusstreamerrrF)r!rGrrr boolrr)r#r$s"r%r&rPspr#r###r#.r#0 r# , r#  r#>*r# !1!1 1r#r(c aVPpVPp VPp VPp VPp VP p \ ;QJdRV4F 'gK RM RM !RV44pVPpV 'd V 'dRMRpV 'd V 'dRMRpV 'd V 'dRMRpV 'd V 'dRMRpVP^,pRp\P!V\PVPR7pVeVPVP4pVPSV4'dVPVP 4M VP"pRpVP%VVSVP&'*R7pVP(pVP*pVP,pV'd\.P1R4RRRppp\3V4U u.uFp .NK p!p VP^,p"VP5VWAPR7wp#p$VPSV4'dVP7VP 4M VP"p%VPR,p&SR ,P9RR 7p'\P:!VP<V', \P>!V'VP<V&, 44p(V'd2\@PB!\EV(PG44R R R 7p)MRp)VPIVWQPR7'EdV)eV)PK^4V'dSR,'d^MRp*VPL!V3RV*/SBp+V+PORR4V+PORR4VV+R&VPQ4;_uu_4V!R/V+BRR/BpRRR4RpVPSXSVPTPVR7oV'd V'dKVPXR,PR\PZVPR7p,V!VV,4p-V 'd]V 'd VV-3, pV 'd VV,3, pV 'dVVP\3, pV 'dVVP^3, p\P`!V-RR 7p.V'dV.V,V^V, ,,p.\Pb!VV.R,.RR 7pVe VPeV.Pg44VPi4!V.4Pk^4p/VPm4V.VPTPn8H,p0VPqV0V#V$4VPm4V.VPTPr8H,p1V1P 4'EdVPuV1V.VV#V$V%4wp2p#p$VPvV1RR3,p3VPyV3V2VVVPz4p4VP}V4V1VV4p5V5P~p\V1PRR7PR44F=wp6p7V!V7P4,PV5PV6,4K? VPPV5PVRR7p8V8PV1PV8PP4,p9VPPV44p:VPPV94p;V:V;PV:P4,pp=V 'dRp?\ ;QJd#V3Rl\4F 'gK RM RM!V3Rl\44'd#\V3Rl\44p@SV@,p?\VVVVVV?V>R7#V>#uup i +'giEL;iuup=i) z This method overrides [~generation.utils.GenerationMixin._sample]. To ease maintenance, modifications are marked with the comment "VibeVoice specific". c3:"TFp\VR4xK R#5i)rYN)rq).0criterias& r% 3VibeVoiceGenerationMixin._sample..s'lZkh.(I(IZksTFNr)is_first_iterationzVibeVoice generation does not support sampling-based token selection. Tokens will be selected using argmax regardless of do_sample=True.rRrrzGenerating audioz tokens)totaldescunitrr input_values padding_maskrrrr)copyrr8)as_tupler)c3,<"TF qS9xK R#5irIr)r= cache_keyrus& r%r?r@sNo ,osc38<"TFqS9gK VxK R#5irIr)rJs& r%r?r@s iOy\hOhOs ) sequencesrElogits attentions hidden_statesrrr)r)rrrr)P_pad_token_tensoroutput_attentionsoutput_hidden_states output_scores output_logitsreturn_dict_in_generateany do_samplerr!rrr8r>_valid_auto_compile_criteriaget_compiled_callrrK_prefill is_assistantrhrfrglogger warning_oncerangerrsumminrrrtqdmr9r_has_unfinished_sequencesupdaterrl_optimize_model_for_decoderrUrrNfloat32rOrPargmaxrputcpuget_input_embeddingsrr:rVrrXrrr#rsr3r* enumeratenonzeroritemrZrrsemantic_tokenizer_encoderlatentsmulti_modal_projectorsemantic_connectorcloseendrrr)Ar@rDr6r7rbr8r9ru pad_token_idrRrSrTrUrVhas_eos_stopping_criteriarXrE raw_logitsdecoder_attentionsdecoder_hidden_statesrthis_peer_finishedunfinished_sequences model_forwardprefill_consumedoutputsrhrfrgr'semantic_cacherr audio_chunkscur_lenrrrinitial_lengthinitial_length_per_samplemax_step_per_sample progress_barr model_inputsnext_token_logitsnext_token_scoresrnext_inputs_embedsrrrr r& audio_outputi sample_idxsemantic_outputssemantic_featuresacoustic_embedsemantic_embeddiffusion_embedschunksgenerated_audiorrKsA&&&&&&&l r%_sample VibeVoiceGenerationMixin._samples):: -??0EE)77 )77 "3"K"K$'C'lZk'lCCC'lZk'l$l!%// 0M3 RD $;@QRX\'>CW^b__Q' "$zz*EJJyO_O_`  #'??9+;+;-M-M N #,$01A$B$F$F2$F$N!#ii  ( (+D D OO57H7S7SVd7d e  "<+QUQ\Q\QoQoBo#p  6 67KM_av w2668K4;;KeKe%>~rST?T%U"#88&&#'%44    $88~Wacqr !-!;!;%.~/E/Eu/E/U/Z/Z[]/^%_MAz !23::<;M;Ma;PQ&`$(::#H#H &&"0"$I$ %5$<$<^=N=NO_OgOgOnOn=o$p!!%!A!A,!O!%!>!>?P!Q#1N4E4EnF[F[4\#\ 5E5H5HI[IbIb5c">2!1!?!?.M qLG$8;LYX^;_:_#_ !5!9!9!;q!@   #       LLNXddWcV599V4DHWcd #EsNoNsssNoNNN iO ii $Y/*#!-3 %% # "]7X766Zes8 c  c c%$c% c" )rr)_previous_max_cache_length)FN)r*r+r,r-r.rTrm staticmethodrorrrrrrr#r3rr0r1 __classcell__)r_r$s@@r%rNrNOs   5/5/n332;9;9zBLBLH,,\,,"$$L$M$ML66>  0r#r#r#r(rN)+collections.abcr dataclassesrtypingrrrr! cache_utilsrr r r generationr r rrrgeneration.configuration_utilsrgeneration.logits_processrgeneration.stopping_criteriargeneration.utilsrrutilsrrrBvibevoice_acoustic_tokenizer.modeling_vibevoice_acoustic_tokenizerrgeneration.streamersr get_loggerr*r]rr4rNr)r(r%rs%!// TTOJ@FJJ 4   H %  17 1  1))(] #] #r(