+ j0^RIHt^RIHt^RIt^RIt^RIHt^RI HuH t ^RI H t ^RIHt^RIHt^RIHt^RIHtHt^R IHt^R IHtHtHtHt^R IH t ^R I!H"t"H#t#H$t$^R I%H&t&^RI'H(t(H)t)H*t*^RI+H,t,^RI-H.t.H/t/H0t0H1t1H2t2^RI3H4t4H5t5H6t6H7t7H8t8^RI9H:t:]!RR7] !RR],444t;]]!RR]444t<]]!RR]444t=]]!RR]444t>!RR]04t?!RR ]&4t@!R!R"].4tA!R#R$]/4tB!R%R&]64tC!R'R(]54tD!R)R*]74tE!R+R,]44tF!R-R.]*4tG!R/R0])4tH!R1R2](4tI!R3R4]P4tK!R5R6]P4tL!R7R8]P4tM!R9R:]P4tN!R;R<]P4tO!R=R>]P4tP!R?R@]:4tQ]!RARB7!RCRD]Q44tR.REOtSR#)F)Callable) dataclassN)strict)initialization)Cache)PreTrainedConfig)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack) ModelOutputTransformersKwargsauto_docstringcan_return_tuple)maybe_autocast)CONFIG_MAPPING AutoConfig AutoModel)CLIPMLP) DacEncoderDacEncoderBlockDacResidualUnit) LlamaConfig)LlamaAttentionLlamaDecoderLayerLlamaRotaryEmbeddingapply_rotary_pos_embeager_attention_forward)"Qwen2_5OmniAntiAliasedActivation1dQwen2_5OmniDownSample1dQwen2_5OmniSnakeBetaQwen2_5OmniUpSample1dkaiser_sinc_filter1d)VoxtralPreTrainedModelzHKUSTAudio/xcodec2-hf) checkpointc.aa]tRt^8toRtRtR]/t^0tRt Rt Rt Rt Rt RtRtR t^t^t^ t^@tR t]!4t]!4t]!4t]!4t]!4t]!4t]!4t]!4tV3R lt] V3R lR l4t!] V3R lRl4t"V3Rlt#Rt$Vt%V;t&#) Xcodec2Configaa downsampling_ratios (`list[int]`, *optional*, defaults to `[2, 2, 4, 4, 5]`): Ratios for downsampling in the encoder. semantic_model_config (`Union[Dict, Wav2Vec2BertConfig]`, *optional*): An instance of the configuration object for the semantic (Wav2Vec2BertConfig) model. quantization_dim (`int`, *optional*, defaults to 2048): Dimension for the vector quantization codebook. quantization_levels (`list[int]`, *optional*, defaults to `[4, 4, 4, 4, 4, 4, 4, 4]`): Levels for the vector quantization codebook. Example: ```python >>> from transformers import Xcodec2Config, Xcodec2Model >>> # Initializing configuration >>> configuration = Xcodec2Config() >>> # Initializing a model (with random weights) from the configuration >>> model = Xcodec2Model(configuration) >>> # Accessing the model configuration >>> configuration = model.config ```xcodec2semantic_model_configNi>g?iiic p<\VP\4'd^VPPRR4VPR&\VPR,,!R/VPBVnM'VPf\R,!^R7Vn\ SV`!R/VBR#) model_typez wav2vec2-bertN)num_hidden_layers) isinstancer(dictgetrsuper __post_init__)selfkwargs __class__s&,l/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/xcodec2/modular_xcodec2.pyr1Xcodec2Config.__post_init__ns d00$ 7 77;7Q7Q7U7UVbds7tD & &| 4)78R8RS_8`)a*,,*D & ' ' /)7)H[])^D & ''c <V^8dQhRS[/#returnint)format __classdict__s"r5 __annotate__Xcodec2Config.__annotate__zs66C6r7cT\\P!VP44#N)r=npproddownsampling_ratiosr2s&r5 hop_lengthXcodec2Config.hop_lengthys277433455r7c <V^8dQhRS[/#r9r<)r>r?s"r5r@rA~s##s#r7c(VP^,#))rHrGs&r5n_fftXcodec2Config.n_fft}s""r7cH<V^8dQh/S[;R&S[S[,S[S[R3,,;R&S[S[,R,;R&S[;R&S[;R&S[;R&S[S[,S[S[R3,,;R &S[;R &S[;R &S[;R &S[;R &S[;R&S[;R&S[;R&#)r:encoder_hidden_size.rFNr( sampling_rateactivation_dropoutquantization_dimquantization_levels hidden_sizeintermediate_sizenum_attention_headsnum_key_value_headsr+head_dimmax_position_embeddings)r=listtupler.rfloat)r>r?s"r5r@rA8s>!?@cU38_4FAB "22T9@CDEF#GH IJcU38_4OKLMN!OP!QR!STUVWX!'Yr7)r()r:r:rLrL)rLrLrLrLrLrLrLrL)'__name__ __module__ __qualname____firstlineno____doc__r*r sub_configsrPrFr(rQrRrSrTrUrVrWrXr+rYrZAttributeError vocab_size bos_token_id eos_token_idpretraining_tpmlp_bias use_cachebase_model_tp_planbase_model_pp_planr1propertyrHrM__annotate_func____static_attributes____classdictcell__ __classcell__r4r?s@@r5r&r&8s 2J*J7K!7F<@M # 7OK!!!H#'!J!#L!#L#%NH I')') (66##Mr7r&c@a]tRt^toRtRtRtRtRtV3Rlt Rt Vt R#) Xcodec2Outputa  audio_values (`torch.FloatTensor` of shape `(batch_size, 1, sequence_length)`, *optional*): Decoded audio waveform values in the time domain, obtained using the decoder part of Xcodec2. These represent the reconstructed audio signal. audio_codes (`torch.LongTensor` of shape `(batch_size, 1, codes_length)`, *optional*): Discrete code embeddings computed using `model.encode`. These are the quantized representations of the input audio used for further processing or generation. latents (`torch.Tensor` of shape `(batch_size, dimension, time_steps)`): Quantized continuous representation of input's embedding. audio_codes_mask (`torch.int32` of shape `(batch_size, 1, codes_length)`, *optional*): Downsampled `padding_mask` for indicating valid audio codes in `audio_codes`. Nc<V^8dQh/S[PR,;R&S[PR,;R&S[PR,;R&S[PR,;R&#)r:N audio_values audio_codeslatentsaudio_codes_mask)torch FloatTensor LongTensorTensor)r>r?s"r5r@Xcodec2Output.__annotate__sa ##d*1!"!!D(/#$\\D '%&llT)0'r7r,) r_r`rarbrcrwrxryrzrorprqr?s@r5rurus( .2L+/K#'G,0'r7ruc<a]tRt^toRtRtRtRtV3RltRt Vt R#)Xcodec2EncoderOutputaP audio_codes (`torch.LongTensor` of shape `(batch_size, 1, codes_length)`, *optional*): Discrete code embeddings computed using `model.encode`. These represent the compressed, quantized form of the input audio signal that can be used for storage, transmission, or generation. latents (`torch.Tensor` of shape `(batch_size, dimension, time_steps)`): Quantized continuous representation of input's embedding. audio_codes_mask (`torch.int32` of shape `(batch_size, 1, codes_length)`, *optional*): Downsampled `padding_mask` for indicating valid audio codes in `audio_codes`. Nc<V^8dQh/S[PR,;R&S[PR,;R&S[PR,;R&#)r:Nrxryrz)r{r}r~)r>r?s"r5r@!Xcodec2EncoderOutput.__annotate__sI!!D(/\\D ' llT)0!r7r,) r_r`rarbrcrxryrzrorprqrs@r5rrs# ,0K#'G,0!r7rc4a]tRt^toRtRtV3RltRtVtR#)Xcodec2DecoderOutputa) audio_values (`torch.FloatTensor` of shape `(batch_size, 1, segment_length)`, *optional*): Decoded audio waveform values in the time domain, obtained by converting the discrete codes back into continuous audio signals. This represents the reconstructed audio that can be played back. NcH<V^8dQh/S[PR,;R&#)r:Nrw)r{r|)r>r?s"r5r@!Xcodec2DecoderOutput.__annotate__s##d*1r7r,) r_r`rarbrcrwrorprqrs@r5rrs.2Lr7rc]tRt^tRtR#)Xcodec2RotaryEmbeddingr,Nr_r`rarbrpr,r7r5rrr7rc>aa]tRt^toV3RlV3RlltRtVtV;t#) Xcodec2MLPc <V^8dQhRS[/#r:configr&)r>r?s"r5r@Xcodec2MLP.__annotate__sWW}Wr7c<\SV`V4\P!VPVP RR7Vn\P!VP VPRR7VnR#)F)biasN)r0__init__nnLinearrUrVfc1fc2r2rr4s&&r5rXcodec2MLP.__init__sR  99V//1I1IPUV99V55v7I7IPUVr7)rrr_r`rarbrrprqrrrss@@r5rrsWWWr7rcTaa]tRt^toV3RlV3RlltRV3RlRlltRtVtV;t#)Xcodec2Attentionc&<V^8dQhRS[RS[/#)r:r layer_idxr&r=)r>r?s"r5r@Xcodec2Attention.__annotate__s}r7c4<\SV`W4RVnR#)FN)r0r is_causalr2rrr4s&&&r5rXcodec2Attention.__init__s +r7c<V^8dQhRS[PRS[S[PS[P3,R,RS[PR,RS[R,RS[S[,RS[S[PS[P3,/#)r: hidden_statesposition_embeddingsNattention_maskpast_key_valuesr3r;)r{r~r\rr r )r>r?s"r5r@rs))))||))#5<<#=>E)) t+ ))  )) +, )) u||U\\) *))r7c VPRRp.VORNVPN5pVPV4PV4P ^^4pVP V4PV4P ^^4p VP V4PV4P ^^4p Vwr\WW^R7wrVeVPWVP4wr\P!VPP\4p T !VVV V V3RVP'gRM VP RVP"/VBwrVP$!.VORN5!P'4pVP)V4pW3#)N) unsqueeze_dimdropoutscaling)shaperYq_projview transposek_projv_projrupdaterr get_interfacer_attn_implementationrtrainingattention_dropoutrreshape contiguouso_proj)r2rrrrr3 input_shape hidden_shape query_states key_states value_statescossinattention_interface attn_output attn_weightss&&&&&, r5forwardXcodec2Attention.forwards$))#2. 88b8$--8 {{=166|DNNqRST [[/44\BLLQPQR {{=166|DNNqRST &$8 RUjk#l  &'6'='=jX\XfXf'g $J(?(M(M KK , ,.E) %8      % $}}}C$2H2H % LL %  % ! "));;;;FFH kk+. ((r7)r)NNN r_r`rarbrrrprqrrrss@@r5rrs))))))r7rc]tRt^tRtR#)Xcodec2DecoderLayerr,Nrr,r7r5rrrr7rc]tRt^tRtR#)Xcodec2SnakeBetar,Nrr,r7r5rrrr7rc&a]tRt^toRtRtVtR#)Xcodec2DownSample1dc8VP^,p\P!WPVP3RR7p\P !VVP PVP4PVRR4VPVR7pV#) replicatemodestridegroupsr) rFpadpad_left pad_rightconv1dfiltertodtypeexpandr)r2rchannelsouts&& r5rXcodec2DownSample1d.forwardsw &&q)mmmT^^-LS^_ hh  KKNN=.. / 6 6xR H;;   r7r,Nr_r`rarbrrprqrs@r5rrs  r7rc&a]tRtRtoRtRtVtR#)Xcodec2UpSample1di c VP^,p\P!WPVP3RR7pVP\P!VVP P VP4PVRR4VPVR7,pVRVPVP)13,pV#)rrrr.r) rrrratioconv_transpose1drrrrrrr)r2rrs&& r5rXcodec2UpSample1d.forward s &&q)mhh-A T  Q%7%7  KKNN=.. / 6 6xR H;; &  &c4==DNN?+J&JK r7r,Nrrs@r5rr s  r7rcBaa]tRtRtoRV3RlV3RllltRtVtV;t#)Xcodec2AntiAliasedActivation1dic2<V^8dQhRS[RS[RS[RS[/#)r:up_ratio down_ratioup_kernel_sizedown_kernel_sizer<)r>r?s"r5r@+Xcodec2AntiAliasedActivation1d.__annotate__s;LLL L  L  Lr7cp<\SV`VVVVVR7\W$4Vn\ W54VnR#)) activationrrrrN)r0rrupsampler downsample)r2rrrrrr4s&&&&&&r5r'Xcodec2AntiAliasedActivation1d.__init__s@ !!)-  *(C -jKr7)rr)r:r: rrrss@@r5rrsLLLr7rc2aa]tRtRtoV3RltRtVtV;t#)Xcodec2ResidualUniti-c<\SV`W4\\V4R7Vn\\V4R7VnR#)rN)r0rrrsnake1snake2)r2 dimensiondilationr4s&&&r5rXcodec2ResidualUnit.__init__.s6 -4@PQZ@[\ 4@PQZ@[\ r7)rrrrss@@r5rr-s]]r7rcBaa]tRtRtoRV3RlV3RllltRtVtV;t#)Xcodec2EncoderBlocki4c,<V^8dQhRS[RS[RS[/#)r:rr stride_indexr)r>r?s"r5r@ Xcodec2EncoderBlock.__annotate__5s'bb}bcbSbr7c<\SV`WV4VP^V,,p\\ V^,4R7VnR#r:rN)r0rrPrrr)r2rrr rr4s&&&& r5rXcodec2EncoderBlock.__init__5s? 6..L@ 4@PQZ^_Q_@`a r7r)rrrrss@@r5rr4sbbbr7rc>aa]tRtRtoV3RlV3RlltRtVtV;t#)Xcodec2Encoderi;c <V^8dQhRS[/#rr)r>r?s"r5r@Xcodec2Encoder.__annotate__<s[[}[r7c<\SV`V4VP^\VP4,,p\ \ V4R7VnR#r )r0rrPlenrFrrr)r2rd_modelr4s&& r5rXcodec2Encoder.__init__<sC  ,,qC8R8R4S/SS4@PQX@YZ r7rrrss@@r5rr;s[[[r7rcPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)Xcodec2ResNetBlockiBc <V^8dQhRS[/#rr)r>r?s"r5r@Xcodec2ResNetBlock.__annotate__Cskk}kr7c"<\SV`4\P!^ VPRRR7Vn\P !4Vn\P!VPVP^^^R7Vn \P!^ VPRRR7Vn \P !4Vn VPVn \P!VPVP^^^R7Vn R#) ư>T) num_groups num_channelsepsaffine) kernel_sizerpaddingN)r0rr GroupNormrUnorm1SiLU activation1Conv1dconv1norm2 activation2rRconv2rs&&r5rXcodec2ResNetBlock.__init__Cs \\Rf>P>PVZcgh 779YYv1163E3EST]^hij \\Rf>P>PVZcgh 779"(";";YYv1163E3EST]^hij r7cN<V^8dQhRS[PRS[P/#r:rr;r{r~)r>r?s"r5r@rMs# : :U\\ :ell :r7cVP^^4pTpVPV4pVPV4pVPV4pVP V4pVP V4p\ PPWPVPR7pVPV4pW,P^^4#)r)pr) rr%r'r)r*r+r functionalrrRrr,r2rresiduals&& r5rXcodec2ResNetBlock.forwardMs%//15   =1 ((7  =1  =1 ((7  --m?V?Vaeanan-o  =1 (33Aq99r7)r'r+rRr)r,r%r*rrss@@r5rrBs kk : : :r7rcaa]tRtRtoRtV3RlV3RlltR RltV3RlRltRV3RlR lltV3R lR lt R t Vt V;t #)Xcodec2FiniteScalarQuantizationiZa  Finite Scalar Quantization (FSQ) module that quantizes continuous latent representations into discrete codes. Original code: https://github.com/lucidrains/vector-quantize-pytorch/blob/353d46027888dfb140c3c65a67a7356f1492d71d/vector_quantize_pytorch/finite_scalar_quantization.py#L64 Original modeling uses `ResidualFSQ` with a single quantizer: https://huggingface.co/HKUSTAudio/xcodec2/blob/main/vq/codec_decoder_vocos.py#L389 But we can directly use FSQ since a main feature of Xcodec2 is that it uses a single codebook. c <V^8dQhRS[/#rr)r>r?s"r5r@,Xcodec2FiniteScalarQuantization.__annotate__cs>>}>r7c,<\SV`4\VP4VnVP 4wr#p\ P !VRR7Vn\ P !VRR7Vn\ P !VRR7Vn R#)F persistentN) r0rr[rT_compute_buffersrBufferlevelsbasiscodebook)r2rr@rArBr4s&& r5r(Xcodec2FiniteScalarQuantization.__init__csj #'(B(B#C "&"7"7"9xii59 YYu7  (u= r7c \P!VP\PVR7p\P!\P!^.VPRR,VR7^\PR7p\P !\ \P!VP44VR7PR4pWC,V,pV^,pWV, V, pW#V3#)zFCompute the levels, basis, and codebook buffers for the FSQ quantizer.)rdeviceNrE)dimrr) r{tensorrTint32cumprodaranger=rDrE unsqueeze)r2rEr@rAindices level_indices half_widthrBs&& r5r>0Xcodec2FiniteScalarQuantization._compute_buffersksd66ekkRXY LL!t77<r?s"r5r@r:ws#%,,r7c VPR4pWP,VP,pVP^,pW#, V, pV#)zP Convert integer codebook indices to normalized per-dimension codes in [-1, 1]. r)rLrAr@)r2rMrNrOcodess&& r5_indices_to_codes1Xcodec2FiniteScalarQuantization._indices_to_codeswsI##B' JJ.$++= [[A% +z9 r7cT<V^8dQhRS[PRS[RS[P/#)r:rr r;)r{r~r])r>r?s"r5r@r:s/DD5<<DeDu||Dr7c VP^, ^V,,^, p\P!VP^,^8HRR4pWC, P4pW,P 4V,V, #)a Constrain `hidden_states` to the valid quantization range for each dimension. Uses a scaled tanh to soft-clip values into the interval $[-(L-1)/2, (L-1)/2]$ (offset by 0.5 for even-level dimensions), where $L$ is the number of quantization levels. The small `eps` margin prevents values from saturating exactly at the boundary, which would zero out gradients. Args: hidden_states (`torch.Tensor`): Continuous input to be bounded. eps (`float`, *optional*, defaults to `1e-3`): Small margin added to the level range to avoid gradient saturation at boundaries. Returns: `torch.Tensor`: Bounded values in the valid quantization range. ?r)r@r{whereatanhtanh)r2rr  half_rangeoffsetshifts&&& r5bound%Xcodec2FiniteScalarQuantization.boundsk"kkAo!c'2Q6 T[[1_13<$++-%++- :VCCr7cx<V^8dQhRS[PRS[S[PS[P3,/#r/r{r~r\)r>r?s"r5r@r:s111U\\1eELL%,,R>R>W>W[`>`  % %   U C C)//1M)J JJ}5M#))+G!}%<$D$D$FFE&E -;K"ZZ/444<?? LGDxx'00D Cs ?B8E E" )rArBr@rTrC)gMbP?) r_r`rarbrcrr>rUr`rrprqrrrss@@r5r8r8Zs>>> 'DD,111r7r8cTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #)Xcodec2ISTFTHeadiz Head for converting decoder outputs to waveform via STFT projection and ISTFT. Uses custom "same" padding ISTFT from Vocos: https://github.com/gemelo-ai/vocos/blob/c859e3b7b534f3776a357983029d34170ddd6fc3/vocos/spectral_ops.py#L47 c <V^8dQhRS[/#rr)r>r?s"r5r@Xcodec2ISTFTHead.__annotate__s::}:r7c<\SV`4\P!VPVP ^,4VnVP VnVPVnVP VP, ^,Vn\P!VP 4p\P!VRR7Vn R#)r:Fr<N) r0rrrrUrMlinearrHr#r{ hann_windowr?window)r2rrzr4s&& r5rXcodec2ISTFTHead.__init__s ii 2 2FLL14DE \\  ++ T__4: ""6<<0ii59 r7cN<V^8dQhRS[PRS[P/#r/r0)r>r?s"r5r@rvs#""""U\\""ell""r7cTVPV4pVP^R R7wr4VP4pVP4p\P!V4P RR7p\P !W44p\PPWPPR RR7P^^4pW`PR ,,pVP^,pV^, VP,VP,p\P!V^V3^VP3^VP3R7R^^VP VP )13,p \P!VPP#4P%^VR 4P^^4^V3^VP3^VP3R7P'4VP VP )p V P RR 7p W, p V P)^4#) r:rigY@)maxbackward)rGnormNNN) output_sizer"rgdy=)minr)NrN)rxchunkr]r{expclamppolarfftirfftrMrrzrrHrfoldr#squarersqueezerL) r2r stft_pred magnitudephasespectrogram_complex time_frames num_framesraudiowindow_envelopes && r5rXcodec2ISTFTHead.forwardsKK . $??1"?5 OO%  IIi(..3.7 #kk);iioo&9::2T^o_iijkmno !KK $>> (..q1 !A~84::E  K(DJJt'  Q4<<4<<-/ / 1&& KK   ' ':r : D DQ JK(DJJt'  ')DLLDLL= 2*//E/:'q!!r7)rHrxrMr#rz r_r`rarbrcrrrprqrrrss@@r5rtrts#::""""""r7rtcbaa]tRtRtoV3RlV3RlltV3RlRltV3RlRltRtVtV;t #) Xcodec2Quantizeric <V^8dQhRS[/#rr)r>r?s"r5r@Xcodec2Quantizer.__annotate__s__}_r7c(<\SV`4\V4Vn\P !VP \VP44Vn \P !\VP4VP 4Vn R#rC) r0rr8 quantizerrrrSrrT project_in project_outrs&&r5rXcodec2Quantizer.__init__sa 8@))F$;$;SA[A[=\]99S)C)C%DfF]F]^r7cN<V^8dQhRS[PRS[P/#rRr0)r>r?s"r5r@rs#''%,,'5<<'r7cVPR4pVPPV,pVPV4#rr)rrrBr)r2rMrTs&& r5 from_codesXcodec2Quantizer.from_codess5//"%''0&&r7cx<V^8dQhRS[PRS[S[PS[P3,/#r/rc)r>r?s"r5r@rs1&&U\\&eELL%,, (()9)9.)IJ ##B'%%r7)rrr) r_r`rarbrrrrprqrrrss@@r5rrs*__ '' &&&r7rcTaa]tRtRtoRtV3RlV3RlltV3RlRltRtVtV;t #)Xcodec2DecoderizVVocos-based decoder with ResNet, Transformer, and ISTFT head for audio reconstruction.c <V^8dQhRS[/#rr)r>r?s"r5r@Xcodec2Decoder.__annotate__s - -} -r7c &<\SV`4\P!VPVP P,VP4Vn\P!VPVP^^R7Vn\P!\V4\V4.4Vn VPVn \VR7Vn\P!\VP 4Uu.uFp\#W4NK up4Vn\P!\V4\V4.4Vn\P(!VPRR7Vn\-V4VnR#uupi))r"r#)rr)r N)r0rrrrUr(fcr(embed ModuleListr prior_netrWr rotary_embranger+rlayerspost_net LayerNormrrtheadrs&& r5rXcodec2Decoder.__init__s! ))F..1M1M1Y1YY[a[m[mnYYv1163E3EST^_` (:6(BDVW]D^'_`#)#=#= 0?mmEJ6KcKcEd eEd  3Ed e   '9&'ACUV\C]&^_ LL!3!3> $V, fsFcN<V^8dQhRS[PRS[P/#r/r0)r>r?s"r5r@r s#33U\\3 3r7c VPV4pVP^^4pVPV4pVP^^4pVPF pV!V4pK \P !VP VPR7P^4pVPW4pVPFpV!V3RV/VBpK VPF pV!V4pK VPVPV44#)rrFr)rrrrr{rKrWrErLrrrrr)r2rr3layer position_idsrs&&, r5rXcodec2Decoder.forward s . %//15  =1 %//15 ^^E!-0M$||D$<$<]EYEYZddefg "oomJ[[E!-cEXc\bcM!]]E!-0M#yy=122r7) rrrrrrWrrrrrss@@r5rrs!` - -333r7rcPaa]tRtRtoV3RlV3RlltV3RlRltRtVtV;t#)Xcodec2SemanticAdapteri%c <V^8dQhRS[/#rr)r>r?s"r5r@#Xcodec2SemanticAdapter.__annotate__&s  } r7c<\SV`4\P!VPP VPP ^^RR7Vn\P!4Vn\P!VPP VPP ^^RR7Vn \P!4Vn \P!VPP VPP ^^RR7Vn \P!VPP VPP ^^RR7Vn R#)F) in_channels out_channelsr"r#rT)r"r#rN) r0rrr(r(rUr)ReLUact1r,act2conv3conv4rs&&r5rXcodec2SemanticAdapter.__init__&s  YY44@@55AA   GGI YY  ( ( 4 4  ( ( 4 4   GGI YY  ( ( 4 4  ( ( 4 4   YY44@@55AA   r7cN<V^8dQhRS[PRS[P/#r/r0)r>r?s"r5r@rGs#  U\\ ell r7cVPV4pVPV4pTpVPV4pVPV4pVP V4pW,pVP V4pV#rC)r)rr,rrrr4s&& r5rXcodec2SemanticAdapter.forwardGsi =1  -0   =1  -0  =1 %0  =1 r7)rrr)r,rrrrss@@r5rr%s  B   r7rc>a]tRtRtoRtRtR tR]R]/tRt Rt Vt R#) Xcodec2PreTrainedModeliSr' input_valuesr attentionsc\P!V4\V\4'dC\P !VP 4\P !VP4R#\V\4'dD\P!VP4p\P!VPV4R#\V\4'dVPVP P"R7wr4p\P!VP V4\P!VP$V4\P!VP&V4R#\V\(4'd]\+RVP,, RVP,, VP.4p\P!VP0V4R#\V\24'dO\+VP4VP6VP.4p\P!VP0V4R#R#)rFrYg333333?N)r _init_weightsr-rinitzeros_alphabetartr{ryrMcopy_rzr8r>r@rErArBrr"rr"rrcutoffrO)r2modulerzr@rArB filter_tensors&& r5r$Xcodec2PreTrainedModel._init_weights\sc%%f- f. / / KK % KK $  0 1 1&&v||4F JJv}}f -  ? @ @&,&=&=V]]EYEY&=&Z #F8 JJv}}f - JJv||U + JJv 1  1 2 20v||1CS6<r?s"r5r@Xcodec2Model.__annotate__us}r7c<\SV`V4VPVn\P!VP 4Vn\V4Vn\V4Vn \P!VPVP P,VPVP P,4Vn\V4Vn\#V4VnVP'4R#rC)r0rrHr from_configr(semantic_encoderrsemantic_adapterracoustic_encoderrrrU fc_encoderrrracoustic_decoder post_initrs&&r5rXcodec2Model.__init__us   ++ ) 5 5f6R6R S 6v > .v 6))   !=!=!I!I I   !=!=!I!I I *&1 .v 6 r7c<V^8dQhRS[PRS[PRS[PR,RS[PR,RS[RS[S[,RS[S[,/# r:rinput_features padding_maskNinput_features_maskoutput_latentsr3r;)r{r~boolr r r\r)r>r?s"r5r@rsw2 2 ll2  2 llT) 2 #\\D0 2  2 +,2  % %2 r7c L\P!4;_uu_4VPW$R7pRRR4XPP ^^4pVP V4pVP V4p \P!W.^R7p VPV P ^^44p VPV 4wrV P ^^4p V P ^^4p Rp VeVPRRR7pWP,p\P!V PR,VPR7P^R4pVV8P!VP"4p \%T V'd V V R7#RV R7# +'giELj;i) a input_values (`torch.Tensor` of shape `(batch_size, 1, sequence_length)`): Input audio waveform. input_features (`torch.Tensor` of shape `(batch_size, mel_bins, time_steps)`): Input audio mel spectrogram for semantic encoding. padding_mask (`torch.Tensor` of shape `(batch_size, 1, sequence_length)`): Padding mask used to pad `input_values`. input_features_mask (`torch.Tensor` of shape `(batch_size, time_steps)`, *optional*): Attention mask for the spectrogram input to the semantic encoder. `1` for valid frames, `0` for padding. output_latents (`bool`, *optional*, defaults to `False`): Whether to return the continuous latent representation from the quantizer. )rNriT)rGkeepdimrF)rxryrzr)r{no_gradrlast_hidden_staterrrcatrrrnrHrKrrErrrr)r2rrrrrr3semantic_outputsemantic_hidden_statesacoustic_hidden_statesrryrxrz audio_length token_lengthidxs&&&&&&, r5encodeXcodec2Model.encodesz2]]__"33N3gO!0!B!B!L!LQPQ!R!%!6!67M!N"&!6!6|!D #9"RXYZ  (?(?1(EF  $~~m<##Aq)!++Aq1    #'++D+AL'??:L,,{004\=P=PQVVWXZ\]C #l 266|7I7IJ ##-G-  37-  /__s F F# c <V^8dQhRS[PR,RS[PR,RS[S[,RS[S[,/#)r:rxNryr3r;)r{r~r r r\r)r>r?s"r5r@rsQ>>\\D(>$>+, > % % >r7c VfVf \R4hVe-VPPVP^^44pMVP^^4pVP!V3/VBp\ VR7#)a  audio_codes (`torch.LongTensor` of shape `(batch_size, 1, codes_length)`): Discrete code indices computed using `model.encode`. latents (torch.Tensor of shape `(batch_size, dimension, time_steps)`, *optional*): Quantized continuous representation of input. z3Either `latents` or `audio_codes` must be provided.)rw) ValueErrorrrrrr)r2rxryr3 recon_audios&&&, r5decodeXcodec2Model.decodess ?{2RS S  "nn// 0E0Ea0KLG''1-G++G>v> #==r7c<V^8dQhRS[PRS[PRS[PR,RS[PR,RS[RS[S[,RS[S[,/#r)r{r~rr r r\ru)r>r?s"r5r@rsv: : ll:  : llT) : #\\D0 :  : +,:   : r7c (VPR,pVPVVVVRRR7pVP!R RVPRR/VB^,RRV13,p \ T VP V'd VPMRVP R7#) aG input_values (`torch.Tensor` of shape `(batch_size, 1, sequence_length)`): Input audio waveform. input_features (`torch.Tensor` of shape `(batch_size, mel_bins, time_steps)`): Input audio mel spectrogram for semantic encoding. padding_mask (`torch.Tensor` of shape `(batch_size, 1, sequence_length)`): Padding mask used to pad `input_values`. input_features_mask (`torch.Tensor` of shape `(batch_size, time_steps)`, *optional*): Attention mask for the spectrogram input to the semantic encoder. `1` for valid frames, `0` for padding. output_latents (`bool`, *optional*, defaults to `False`): Whether to return the continuous latent representation from the quantizer. Examples: ```python >>> from datasets import load_dataset >>> from transformers import AutoFeatureExtractor, Xcodec2Model >>> dataset = load_dataset("hf-internal-testing/librispeech_asr_dummy", "clean", split="validation") >>> audio = dataset["train"]["audio"][0]["array"] >>> model_id = "HKUSTAudio/xcodec2-hf" >>> model = Xcodec2Model.from_pretrained(model_id) >>> feature_extractor = AutoFeatureExtractor.from_pretrained(model_id) >>> inputs = feature_extractor(audio=audio, sampling_rate=feature_extractor.sampling_rate, return_tensors="pt") >>> outputs = model(**inputs) >>> audio_codes = outputs.audio_codes >>> audio_values = outputs.audio_values ```T)rrrr return_dictryr .N)rwrxryrzrr,)rrr ryrurxrz) r2rrrrrr3lengthencoder_outputsrws &&&&&&, r5rXcodec2Model.forwardsV##B'++ )% 3 & {{_?+B+B_PT_X^_`abcfhoiohocop %'33/=O++4,==   r7)rrrrHrrr)NNF)NN)r_r`rarbr& config_classrrrrr rrprqrrrss@@r5rrqs} L 2 2 2 h>>>.: : : : r7r)r&rr)Tcollections.abcr dataclassesrnumpyrDr{torch.nnrtorch.nn.functionalr3rhuggingface_hub.dataclassesrrr cache_utilsrconfiguration_utilsrmodeling_utilsr r processing_utilsr utilsr r rr utils.genericrautorrrclip.modeling_cliprdac.modeling_dacrrrllama.configuration_llamarllama.modeling_llamarrrrr"qwen2_5_omni.modeling_qwen2_5_omnirrr r!r"voxtral.modeling_voxtralr#r&rurrrrrrrrrrrrrModulerr8rtrrrrr__all__r,r7r5r(s5%! .& 3F& ,88(KK3>23E#KE#4E#P 1K1 1( 1;1 1" 2;2 2 1 WW.)~.)b +  +  1  - L%GL(]/]b/b[Z[::0P1biiP1f3"ryy3"l&ryy&,)3RYY)3X+RYY+\535<@A^ )^ B^ B Fr7