oMj~SSKrSSKJr SSKJs Jr SSKJrJr SSK J r SSK J r SSK Jr SSKJr SS KJr SS KJr S S KJrJrJr \ R2"\5rSr"S S\R:5r"SS\R>5r "SS\RB5r""SS\R>5r#"SS\R>5r$"SS\R>5r%"SS\R>5r&"SS\R>5r'"SS\R>5r("SS \R>5r)"S!S"\\\\ 5r*g)#N) ConfigMixinregister_to_config)FromOriginalModelMixin)logging)apply_forward_hook)get_activation)AutoencoderKLOutput) ModelMixin)AutoencoderMixin DecoderOutputDiagonalGaussianDistributionc^\rSrSrSrS S\S\S\\\\\4-S\\\\\4-S\\\\\4-SS 4 U4S jjjrSU4S jjrS r U=r $)QwenImageCausalConv3d(aM A custom 3D causal convolution layer with feature caching support. This layer extends the standard Conv3D layer by ensuring causality in the time dimension and handling feature caching for efficient inference. Args: in_channels (int): Number of channels in the input image out_channels (int): Number of channels produced by the convolution kernel_size (int or tuple): Size of the convolving kernel stride (int or tuple, optional): Stride of the convolution. Default: 1 padding (int or tuple, optional): Zero-padding added to all three sides of the input. Default: 0 in_channels out_channels kernel_sizestridepaddingreturnNc>[TU]UUUUUS9 URSURSURSURSSURS-S4UlSUlg)N)rrrrrr r rrrr)super__init__r_padding)selfrrrrr __class__s C:\Users\johnn\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.13_qbz5n2kfra8p0\LocalCache\local-packages\Python313\site-packages\diffusers/models/autoencoders/autoencoder_kl_qwenimage.pyrQwenImageCausalConv3d.__init__7sx #%#  a$,,q/4<<?DLLYZO]^aeamamnoap]prst   c:>[UR5nUb^URSS:aKURUR5n[R "X!/SS9nUS==UR S-ss'[R"X5n[TU])U5$)Nrr dim) listrtodevicetorchcatshapeFpadrforward)rxcache_xrr s r!r0QwenImageCausalConv3d.forwardKst}}%  4==#3a#7jj*G 7,A.A AJ'--* *J EE! wq!!r#)rr)r rN) __name__ __module__ __qualname____firstlineno____doc__inttuplerr0__static_attributes__ __classcell__r s@r!rr(s &././ !!!5c3// ! eCcM** ! uS#s]++ ! !!(""r#rc N^\rSrSrSrS S\S\S\S\SS4 U4S jjjrS rS r U=r $) QwenImageRMS_normUa A custom RMS normalization layer. Args: dim (int): The number of dimensions to normalize over. channel_first (bool, optional): Whether the input tensor has channels as the first dimension. Default is True. images (bool, optional): Whether the input represents image data. Default is True. bias (bool, optional): Whether to include a learnable bias term. Default is False. r' channel_firstimagesbiasrNcR>[TU]5 U(dSOSnU(aU/UQ7OU4nX lUS-Ul[R "[ R"U55UlU(a0[R "[ R"U55Ul gSUl g)N)r r r )r r g?) rrrBscalenn Parameterr+onesgammazerosrD)rr'rBrCrDbroadcastable_dimsr-r s r!rQwenImageRMS_norm.__init__asw .4Y&.;*)*#*#X \\%**U"34 8# UHo[TR5;v M g7fr4)strdtype).0tr1s r! ,QwenImageRMS_norm.forward..ls!Q '=!QWW '=s$')float4_float8_r r&) rRr+float16bfloat16anyr. normalizefloatrBr)rGrKrD)rr1needs_fp32_normalize normalizeds ` r!r0QwenImageRMS_norm.forwardks ww5==%..*II SQ '=Q N [[.BUYUgUgPQmoqtt GG JJ&3dii??r#)rDrBrKrGTTF) r5r6r7r8r9r:boolrr0r<r=r>s@r!r@r@UsJ FCFFTFX\FimFF@@r#r@c,^\rSrSrSrU4SjrSrU=r$)QwenImageUpsamplevz Perform upsampling while ensuring the output tensor has the same data type as the input. Args: x (torch.Tensor): Input tensor to be upsampled. Returns: torch.Tensor: Upsampled tensor with the same data type as the input. c\>[TU]UR55RU5$r4)rr0r^type_as)rr1r s r!r0QwenImageUpsample.forwards#wqwwy)11!44r#)r5r6r7r8r9r0r<r=r>s@r!rerevs55r#recL^\rSrSrSrS\S\SS4U4SjjrSS/4S jrS r U=r $) QwenImageResampleaP A custom resampling module for 2D and 3D data. Args: dim (int): The number of input/output channels. mode (str): The resampling mode. Must be one of: - 'none': No resampling (identity operation). - 'upsample2d': 2D upsampling with nearest-exact interpolation and convolution. - 'upsample3d': 3D upsampling with nearest-exact interpolation, convolution, and causal 3D convolution. - 'downsample2d': 2D downsampling with zero-padding and convolution. - 'downsample3d': 3D downsampling with zero-padding, convolution, and causal 3D convolution. r'moderNc >[TU]5 XlX lUS:Xa<[R "[ SSS9[R"XS-SSS95UlgUS :XaO[R "[ SSS9[R"XS-SSS95Ul[XS-S S S9Ul gUS :XaE[R "[R"S 5[R"XSSS95UlgUS:XaV[R "[R"S 5[R"XSSS95Ul[XS SSS9Ul g[R"5Ulg)N upsample2d)@rqz nearest-exact) scale_factorrnr rr r upsample3d)rr r )r rr downsample2d)rr rr )r r )r downsample3d)r r r r)rr) rrr'rnrH SequentialreConv2dresampler time_conv ZeroPad2dIdentity)rr'rnr s r!rQwenImageResample.__init__s'   < MM!zP #axA6DM\ !MM!zP #axA6DM33aT]^DN ^ #MM",,|*DbiiPSZ[djFklDM ^ #MM",,|*DbiiPSZ[djFklDM23YybklDNKKMDMr#rc UR5upEpgnURS:XGaUGbUSn X)cSX)'US==S- ss'GOUSS2SS2[[URS5*S2SS2SS24R 5n U RSS:a^X)bYX)S:waQ[ R"X)SS2SS2SSS2SS24RS5RU R5U /SS9n U RSS:aQX)bLX)S:XaD[ R"[ R"U 5RU R5U /SS9n X)S:XaURU5nOURXU 5nXU 'US==S- ss'URUSXVXx5n[ R"USS2SSS2SS2SS2SS24USS2SSS2SS2SS2SS244S5nURXEUS-Xx5nURSnURSSSSS 5RXF-XWU5nUR!U5nUR#XFURS5URS5URS55RSSSSS 5nURS :XaUbUSn X)c!UR 5X)'US==S- ss'U$USS2SS2SS2SS2SS24R 5n UR[ R"X)SS2SS2SS2SS2SS24U/S55nXU 'US==S- ss'U$) NrtrRepr r rYr&rr%rv)sizernminCACHE_Tr-cloner+r, unsqueezer)r* zeros_likerzreshapestackpermuteryview) rr1 feat_cachefeat_idxbcrThwidxr2s r!r0QwenImageResample.forwards aA 99 $%qk?*&+JOQK1$K1s7AGGAJ'?&?&A1a GHNNPG}}Q'!+ 0KPZP_chPh"'))'_Q2q!^<FFqILLW^^\^eflm#}}Q'!+ 0KPZP_chPh"'))U-=-=g-F-I-I'..-Y[b,cij"k!%/ NN1- NN1o>&-sOQK1$K !Qa3A Qq!Q1a'7%8!Aq!Q1s@r!rlrls5 *C*s*t*6%)A3++r#rlc \^\rSrSrSrS S\S\S\S\SS4 U4S jjjrSS /4S jr S r U=r $)QwenImageResidualBlocka) A custom residual block module. Args: in_dim (int): Number of input channels. out_dim (int): Number of output channels. dropout (float, optional): Dropout rate for the dropout layer. Default is 0.0. non_linearity (str, optional): Type of non-linearity to use. Default is "silu". in_dimout_dimdropout non_linearityrNcn>[TU]5 XlX l[ U5Ul[ USS9Ul[XSSS9Ul [ USS9Ul [R"U5Ul [X"SSS9UlX:wa[XS5Ulg[R"5Ulg)NFrCrr rs)rrrrr nonlinearityr@norm1rconv1norm2rHDropoutrconv2r| conv_shortcut)rrrrrr s r!rQwenImageResidualBlock.__init__s   *=9've< *6AqI &wu= zz'* *7QJ JPJ[26AFacalalanr#rc tURU5nURU5nURU5nUbUSnUSS2SS2[[UR S5*S2SS2SS24R 5nUR SS:aVX%bQ[R"X%SS2SS2SSS2SS24RS5RUR5U/SS9nURXU5nXbU'US==S- ss'OURU5nURU5nURU5nURU5nUbUSnUSS2SS2[[UR S5*S2SS2SS24R 5nUR SS:aVX%bQ[R"X%SS2SS2SSS2SS24RS5RUR5U/SS9nURXU5nXbU'US==S- ss'X-$URU5nX-$Nrr rYr&r )rrrrrr-rr+r,rr)r*rrrr)rr1rrrrr2s r!r0QwenImageResidualBlock.forwards   q ! JJqM   a   !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv 1o.A%sO QK1 K 1 A JJqM   a  LLO  !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv 1o.A%sO QK1 K u  1 Au r#) rrrrrrrrr)rFsilu r5r6r7r8r9r:r^rQrr0r<r=r>s@r!rrsd# ooo o  o  oo(%)A3((r#rc2^\rSrSrSrU4SjrSrSrU=r$)QwenImageAttentionBlocki&zm Causal self-attention with a single head. Args: dim (int): The number of channels in the input tensor. c>[TU]5 Xl[U5Ul[ R "XS-S5Ul[ R "XS5Ulg)Nrr ) rrr'r@normrHrxto_qkvproj)rr'r s r!r QwenImageAttentionBlock.__init__.sI &c* ii1Wa0 IIc* r#cpUnUR5up4pVnURSSSSS5RX5-XFU5nURU5nUR U5nURX5-SUS-S5nURSSSS5R 5nUR SSS9upn [R"XU 5nURS5RSSS5RX5-XFU5nURU5nURX5XFU5nURSSSSS5nX-$)Nrr r rr%rYr&) rrrrr contiguouschunkr.scaled_dot_product_attentionsqueezerr) rr1identity batch_sizechannelstimeheightwidthqkvqkvs r!r0QwenImageAttentionBlock.forward7s045FFH1 dE IIaAq! $ , ,Z->RW X IIaLkk!nkk*+Q1 bAkk!Q1%002))A2)&a  * *1 3 IIaL Aq ) 1 1*2CXW\ ] IIaL FF:Xu = IIaAq! $|r#)r'rrr) r5r6r7r8r9rr0r<r=r>s@r!rr&s+r#rc T^\rSrSrSrS S\S\S\S\4U4SjjjrSS /4S jr S r U=r $) QwenImageMidBlockiSz Middle block for QwenImageVAE encoder and decoder. Args: dim (int): Number of input/output channels. dropout (float): Dropout rate. non_linearity (str): Type of non-linearity to use. r'rr num_layersc T>[TU]5 Xl[XX#5/n/n[ U5H8nUR [ U55 UR [XX#55 M: [R"U5Ul [R"U5Ul SUl g)NF) rrr'rrangeappendrrH ModuleList attentionsresnetsgradient_checkpointing) rr'rrrrr_r s r!rQwenImageMidBlock.__init__]s *#GKL z"A   5c: ; NN1#GS T#-- 3}}W- &+#r#NrcURS"XU5n[URURSS5HupEUbU"U5nU"XU5nM U$Nrr )rzipr)rr1rrattnresnets r!r0QwenImageMidBlock.forwardls] LLOA8 4 ab1ABLDGqh/A C r#)rr'rr)rFrr rr>s@r!rrSsB ,C ,% ,c ,`c , ,%)A3  r#rc b^\rSrSrSrSS/SQS//SQSS S 4 S \4U4S jjjrS S/4SjrSrU=r $)QwenImageEncoder3dizaC A 3D encoder module. Args: dim (int): The base number of channels in the first layer. z_dim (int): The dimensionality of the latent space. dim_mult (list of int): Multipliers for the number of channels in each block. num_res_blocks (int): Number of residual blocks in each block. attn_scales (list of float): Scales at which to apply attention mechanisms. temperal_downsample (list of bool): Whether to downsample temporally in each block. dropout (float): Dropout rate for the dropout layers. non_linearity (str): Type of non-linearity to use. r%r r r%r%r rbrFrrrc 4>[TU]5 XlX lX0lX@lXPlX`l[U 5Ul S/U-V s/sHoU -PM n n Sn [XSSSS9Ul [R"/5Ul[[!U SSU SS55Hun up[#U5HSnURR%['XU55 X;a$URR%[)U55 UnMU U [+U5S- :wdM~Xm(aSOSnURR%[-UUS 95 U S -n M [/WXySS 9Ul[3US S 9Ul[XSSS9UlS Ulgs sn f)Nr ?rrrsrYrvrurnrqrFr)rrr'z_dimdim_multnum_res_blocks attn_scalestemperal_downsampler rrconv_inrHr down_blocks enumeraterrrrrlenrlr mid_blockr@norm_outconv_outr)rr'rrrrrrinput_channelsrudimsrGirrrrnr s r!rQwenImageEncoder3d.__init__s   ,&#6 *=9#$x0Aa0-^!WaQRS ==,$-c$s)T!"X.F$G A >*  ''(>vPW(XY'$$++,CG,LM  +CMA%%)<)?~^  ''(9'(MN %H+7GWXY*'%@ -gaK &+#;1s FNrc ZUbUSnUSS2SS2[[URS5*S2SS2SS24R5nURSS:aVX$bQ[R "X$SS2SS2SSS2SS24R S5RUR5U/SS9nURXU5nXRU'US==S- ss'OURU5nURHnUb U"XU5nMU"U5nM URXU5nURU5nURU5nUbUSnUSS2SS2[[URS5*S2SS2SS24R5nURSS:aVX$bQ[R "X$SS2SS2SSS2SS24R S5RUR5U/SS9nURXU5nXRU'US==S- ss'U$URU5nU$r)rrr-rr+r,rr)r*rrrrrr)rr1rrrr2layers r!r0QwenImageEncoder3d.forwards  !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv Q30A%sO QK1 K QA%%E%!2!H & NN1( 3 MM!    a   !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv aC1A%sO QK1 K a Ar#) rrrr'rrrrrrrrr r5r6r7r8r9rQrr0r<r=r>s@r!rrzsM  /#3,3,3,j%)A3%%r#rc h^\rSrSrSrSS\S\S\S\S\S-S \4 U4S jjjrSS /4S jr S r U=r $)QwenImageUpBlockian A block that handles upsampling for the QwenImageVAE decoder. Args: in_dim (int): Input dimension out_dim (int): Output dimension num_res_blocks (int): Number of residual blocks dropout (float): Dropout rate upsample_mode (str, optional): Mode for upsampling ('upsample2d' or 'upsample3d') non_linearity (str): Type of non-linearity to use Nrrrr upsample_moderc D>[T U]5 XlX l/nUn[ US-5H n UR [ XXF55 UnM" [R"U5Ul SUl Ub#[R"[X%S9/5Ul SUl g)Nr rF) rrrrrrrrHrr upsamplersrlr) rrrrrrrr current_dimrr s r!rQwenImageUpBlock.__init__s    ~)*A NN1+_ `!K+}}W-   $ mm->w-[,\]DO&+#r#rcURHnUb U"XU5nMU"U5nM URb0UbURS"XU5nU$URS"U5nU$)a  Forward pass through the upsampling block. Args: x (torch.Tensor): Input tensor feat_cache (list, optional): Feature cache for causal convolutions feat_idx (list, optional): Feature index for cache management Returns: torch.Tensor: Output tensor r)rr)rr1rrrs r!r0QwenImageUpBlock.forwardsullF%1(31I # ?? &%OOA&qh?OOA&q)r#)rrrrr)rFNrrr>s@r!rrsn "$(#,,, ,  , Tz ,,,<%)A3r#rc b^\rSrSrSrSS/SQS//SQSS S 4 S \4U4S jjjrS S/4SjrSrU=r $)QwenImageDecoder3di+a? A 3D decoder module. Args: dim (int): The base number of channels in the first layer. z_dim (int): The dimensionality of the latent space. dim_mult (list of int): Multipliers for the number of channels in each block. num_res_blocks (int): Number of residual blocks in each block. attn_scales (list of float): Scales at which to apply attention mechanisms. temperal_upsample (list of bool): Whether to upsample temporally in each block. dropout (float): Dropout rate for the dropout layers. non_linearity (str): Type of non-linearity to use. rr%rr FTTrFrrrc >[TU]5 XlX lX0lX@lXPlX`l[U 5Ul US/USSS2-V s/sHoU -PM n n SS[U5S- -- n [X+SSSS9Ul [U SXySS9Ul[R "/5Ul[%['U SSU SS55Hgun upU S:aUS-nSnU [U5S- :wa Xm(aS OS n[)UUUUUU S 9nUR"R+U5 UcMbU S -n Mi [-WS S9Ul[XSSS9UlS Ulgs sn f)NrYrr rrr rsrrtrp)rrrrrrrqFr)rrr'rrrrtemperal_upsampler rrrrrrrHr up_blocksrrrrr@rrr)rr'rrrrrrrrrrrGrrrrup_blockr s r!rQwenImageDecoder3d.__init__:s   ,&!2*=9#+2,(4R4.!@A!@Aa!@AaCMA-..-UGQJ +47GWXYr*$-c$s)T!"X.F$G A 1u1!MCMA%%0A0D , (-++ H NN ! !( +( /%H4*'%@ -gqRST &+#QBsE'Nrc @UbUSnUSS2SS2[[URS5*S2SS2SS24R5nURSS:aVX$bQ[R "X$SS2SS2SSS2SS24R S5RUR5U/SS9nURXU5nXRU'US==S- ss'OURU5nURXU5nURH nU"XU5nM URU5nURU5nUbUSnUSS2SS2[[URS5*S2SS2SS24R5nURSS:aVX$bQ[R "X$SS2SS2SSS2SS24R S5RUR5U/SS9nURXU5nXRU'US==S- ss'U$URU5nU$r)rrr-rr+r,rr)r*rrrrrr)rr1rrrr2rs r!r0QwenImageDecoder3d.forward{s  !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv Q30A%sO QK1 K QA NN1( 3H1A' MM!    a   !1+C1s7AGGAJ7791a?@FFHG}}Q!# (C))Z_Q2q!^%D%N%Nq%Q%T%TU\UcUc%dfm$ntuv aC1A%sO QK1 K a Ar#) rrrr'rrrrrrrrrrr>s@r!rr+sM  -#?,?,?,B%)A3##r#rcf^\rSrSrSrSr\SS/SQS//SQS S /S Q/S Q4 S \S\S\\S\S\\ S\\ S\ S\S\\ S\\ SS4U4Sjjj5r S3S\S-S\S-S\ S-S\ S-SS4 Sjjr Sr S \R4S!jr\S4S \RS"\ S\\\-4S#jj5rS4S$\RS"\ 4S%jjr\S4S$\RS"\ S\\R-4S&jj5rS'\RS(\RS)\S\R4S*jrS'\RS(\RS)\S\R4S+jrS \RS\4S,jrS4S$\RS"\ S\\R-4S-jjrS5S.\RS/\ S"\ S0\R:S-S\\R-4 S1jjrS2rU=r $)6AutoencoderKLQwenImageia A VAE model with KL loss for encoding videos into latents and decoding latent representations into videos. This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented for all models (such as downloading or saving). F`rr rrFr)gy):gM Og^)gQ?gtVƿgZӼ?gBfjÿgU0*?gL F%u?gMg&†?gz6>׿g F%uȿg[ AcgMJ?gW2ıҿ)g_L@gNё\C?gQ@g?@g9#J{?g|a2U?gHPs@g0*@gJ{/L& @gJY8@g]C@g(?gK46?gS:?go_Ι@g-?base_dimrrrrrrr latents_mean latents_stdrNc >[T U]5 X lX`lUSSS2Ul[ XS-X4XPRXx5Ul[US-US-S5Ul[X"S5Ul [XX4XPRXx5Ul S[UR5-Ul SUlSUlSUlSUlSUlSUlURb*[)SURR+555OSUR b*[)S UR R+555OSS .Ulg) NrYr r Fc3B# UHn[U[5v M g7fr4 isinstancerrSms r!rU2AutoencoderKLQwenImage.__init__..`I_A:a)>??I_rc3B# UHn[U[5v M g7fr4r rs r!rUrrr)decoderencoder)rrrrrrrr quant_convpost_quant_convrrrspatial_compression_ratio use_slicing use_tilingtile_sample_min_heighttile_sample_min_widthtile_sample_stride_heighttile_sample_stride_widthsummodules_cached_conv_counts) rrrrrrrrrrrr s r!rAutoencoderKLQwenImage.__init__sB  #6 !4TrT!:) ai;H`H`bi  0 519aH4U1E) X{DZDZ\c  *+c$2J2J.K)K&!  '*#%("*-&(+% ||'`I]I]I_``||'`I]I]I_`` $  r#rrrrcSUlU=(d URUlU=(d URUlU=(d URUlU=(d URUlg)a Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow processing larger images. Args: tile_sample_min_height (`int`, *optional*): The minimum height required for a sample to be separated into tiles across the height dimension. tile_sample_min_width (`int`, *optional*): The minimum width required for a sample to be separated into tiles across the width dimension. tile_sample_stride_height (`int`, *optional*): The minimum amount of overlap between two consecutive vertical tiles. This is to ensure that there are no tiling artifacts produced across the height dimension. tile_sample_stride_width (`int`, *optional*): The stride between two consecutive horizontal tiles. This is to ensure that there are no tiling artifacts produced across the width dimension. TN)rrrrr)rrrrrs r! enable_tiling$AutoencoderKLQwenImage.enable_tilingsW0&<&[@[@[#%:%Xd>X>X")B)ddFdFd&(@(aDDaDa%r#cSnU"UR5UlS/UlS/UR-UlU"UR5UlS/UlS/UR -Ulg)NcpSnUR5Hn[U[5(dMUS- nM! U$r)r r r)modelcountrs r! _count_conv3d9AutoencoderKLQwenImage.clear_cache.._count_conv3ds5E]]_a!677QJE%Lr#r)r _conv_num _conv_idx _feat_mapr _enc_conv_num _enc_conv_idx _enc_feat_map)rr*s r! clear_cache"AutoencoderKLQwenImage.clear_cachesd 't||4$..0*4<<8S"Vd&8&88r#r1c URu p#pEUR(a/XPR:dX@R:aUR U5$UR 5 SUS- S--n[ U5HnS/UlUS:Xa9URUSS2SS2SS2SS2SS24URURS9nMJURUSS2SS2SSUS- --SSU--2SS2SS24URURS9n [R"WU /S5nM URW5n UR 5 U $)Nr r%rrrr ) r-rrr tiled_encoder2rr0rr1r+r,r) rr1r num_framerriter_routout_encs r!_encodeAutoencoderKLQwenImage._encodes:)*&1 ??(B(B BfOjOjFj$$Q' ' Y]q((uA"#D Avll1Q2A2q!^#4ASAS^b^p^plq||aAQU Oa!a%i7A=>#11!//$ iid Q/ooc"  r# return_dictc>UR(aYURSS:aFURS5Vs/sHo0RU5PM nn[R "U5nOURU5n[ U5nU(dU4$[US9$s snf)a Encode a batch of images into latents. Args: x (`torch.Tensor`): Input batch of images. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. Returns: The latent representations of the encoded videos. If `return_dict` is True, a [`~models.autoencoder_kl.AutoencoderKLOutput`] is returned, otherwise a plain `tuple` is returned. rr ) latent_dist)rr-splitr<r+r,rr )rr1r>x_sliceencoded_slicesr posteriors r!encodeAutoencoderKLQwenImage.encode.s    QCD771:N:ll73:NN .)A QA03 < "y99OsBzc URu p4pVURUR-nURUR-nUR(aXh:dXW:aUR XS9$UR 5 URU5n [U5Hn S/Ul U S:Xa;URU SS2SS2XS-2SS2SS24URURS9n MLURU SS2SS2XS-2SS2SS24URURS9n [R"W U /S5n M [R"W SSS9n UR 5 U(dU 4$[U S 9$) Nr>rr r5r gr)rmaxsample)r-rrrr tiled_decoder2rrr-rr.r+r,clampr) rrGr>rr7rrtile_latent_min_heighttile_latent_min_widthr1rr9r:s r!_decodeAutoencoderKLQwenImage._decodeIsS)*&1!%!\>\ \ ?? =A`$$Q$@ @    #y!ASDNAvll1Q11u9a%:#;aeaoaolp||AaAAIq!&;$<bfbpbp|qiid Q/ "kk#4S1 6MC((r#cPUR(acURSS:aPURS5Vs/sHo0RU5RPM nn[ R "U5nOURU5RnU(dU4$[US9$s snf)a Decode a batch of images. Args: z (`torch.Tensor`): Input batch of latent vectors. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. Returns: [`~models.vae.DecoderOutput`] or `tuple`: If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is returned. rr rK)rr-rArQrLr+r,r)rrGr>z_slicedecoded_slicesdecodeds r!decodeAutoencoderKLQwenImage.decodebs    QJK''RS*U*wll73::*NUii/Gll1o,,G: G,,Vs$B#ar blend_extentc [URSURSU5n[U5HInUSS2SS2SS2U*U-SS24SXC- - -USS2SS2SS2USS24XC- --USS2SS2SS2USS24'MK U$)Nr rr-r)rrYrrZys r!blend_vAutoencoderKLQwenImage.blend_v{s1772; \B |$A Aq<-!*;Q!>?1qGWCWX[\]^`acdfgij]j[k \ AaAq!m %r#c[URSURSU5n[U5HInUSS2SS2SS2SS2U*U-4SXC- - -USS2SS2SS2SS2U4XC- --USS2SS2SS2SS2U4'MK U$)NrYr r])rrYrrZr1s r!blend_hAutoencoderKLQwenImage.blend_hs1772; \B |$A Aq!l]Q->!>?1qGWCWX[\]^`acdfgij]j[k \ AaAq!m %r#c URu p#pEX@R-nXPR-nURUR-nURUR-n URUR-n UR UR-n X- n X- n /n[ SX@R5GHHn/n[ SXPR 5GHnUR5 /nSUS- S--n[ U5HnS/UlUS:Xa.USS2SS2SS2XUR-2UUUR-24nO >> !%!\>\ \$($B$BdFdFd$d!#'#@#@DDbDb#b -I +F q&"@"@AAC1e%B%BC  ":>a"77 {+A*+D&Av Arr143N3N/N+NPQTUX\XrXrTrPr!rs QU Oa!a%i7D$?$? ??D$>$> >> @  <<9K9KVZVhVh<iD??40DKK%, 599Tq12'D( KK -B.  oFAsJ$S>4q5<<U AlKDq5<<AE D+FD!!$q!Q0J1J0JLeMeLe'e"fg*   uyy< =&ii +Aq!^m^]l],RS r#c URu p4pVXPR-nX`R-nURUR-n URUR-n URUR-n UR UR-n URUR- n URUR - n/n[ SX[5Hn/n[ SXl5HnUR5 /n[ U5HonS/UlUSS2SS2UUS-2UUU -2UUU -24nURU5nURUURURS9nURU5 Mq UR[R"USS95 M URU5 M UR5 /n[U5Hunn/n[U5H~unnUS:aUR!UUS- UUU 5nUS:aUR#UUS- UU5nURUSS2SS2SS2SUR2SUR 245 M UR[R"USS95 M [R"USS9SS2SS2SS2SU2SU24nU(dU4$[%US 9$) a Decode a batch of images using a tiled decoder. Args: z (`torch.Tensor`): Input batch of latent vectors. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. Returns: [`~models.vae.DecoderOutput`] or `tuple`: If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is returned. rNr r5r r&rYrrK)r-rrrrrrr2r-rrr.rr+r,rr_rbr)rrGr>rrerr sample_height sample_widthrOrPrhrirjrkrlrrmrnrrrprVrqrrdecs r!rM#AutoencoderKLQwenImage.tiled_decodes+,'''1&!?!?? === !%!\>\ \$($B$BdFdFd$d!#'#@#@DDbDb#b 22T5S5SS 0043P3PP q&  "z*A&'SDNQ1q1u9a!6L2L.LaRSVkRkNkklD//5D"ll4DNNUYUcUcldGKK( +  599Tq12? KK =  oFAsJ$S>4q5<<QU AlKDq5<<AE D+FD!!$q!Q0P$2P2P0PRqTXTqTqRq'q"rs*   uyy< =&ii +Aq!^m^]l],RS6MC((r#rLsample_posterior generatorcUnURU5RnU(aURUS9nOUR5nUR XsS9nU$)a Args: sample (`torch.Tensor`): Input sample. sample_posterior (`bool`, *optional*, defaults to `False`): Whether to sample from the posterior. return_dict (`bool`, *optional*, defaults to `True`): Whether or not to return a [`DecoderOutput`] instead of a plain tuple. generator (`torch.Generator`, *optional*): A [`torch.Generator`](https://pytorch.org/docs/stable/generated/torch.Generator.html) to make sampling deterministic. Returns: [`~models.vae.DecoderOutput`] or `tuple`: If `return_dict` is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is returned. )rzrI)rEr@rLrnrW) rrLryr>rzr1rDrGrws r!r0AutoencoderKLQwenImage.forward sS. KKN..   9 5A Akk!k5 r#)r!r-r,r0r/r1r.rrrrrrrrrrrrrr)NNNN)T)FTN)!r5r6r7r8r9 _supports_gradient_checkpointingrr:r(r^rcrr$r2r+Tensorr<rr r;rrErQrrWr_rbr6rM Generatorr0r<r=r>s@r!rrs(-$*#%*=%o$d9 9 9 s) 9  9 %[ 9 "$Z9 9 9 5k9 %[9  9 9 z.2,02615 b #d b #Tzb$)4< b #($, b  b<9 037::,0: u%AB B::4))D)2- -4-=SXS_S_C_--0%,,cell%,,cell@ell@/B@D=)ell=)=)Y^YeYeIe=)D"' ,0    ??T)   % r#r)+r+torch.nnrHtorch.nn.functional functionalr.configuration_utilsrrloadersrutilsrutils.accelerate_utilsr activationsr modeling_outputsr modeling_utilsr vaerrr get_loggerr5loggerrConv3drModuler@Upsamplererlrrrrrrrrjr#r!rs* B-8(2'NN   H % *"BII*"Z@ @B 5 5T TnGRYYGT*bii*Z$ $NiiXBryyBJsslIZ)9;H^Ir#