biddlmZmZmZmZddlZddlZddlm Z ddl m cm Z ddl mZmZddlmZddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZddl m!Z!m"Z"ejFe$Z%Gdde jLZ'Gdde jPZ)Gdde jPZ*Gdde jPZ+Gdde jPZ,Gdde jPZ-Gdde jPZ.Gdde jPZ/Gd d!e jPZ0Gd"d#eeeZ1y)$)DictOptionalTupleUnionN) ConfigMixinregister_to_config)FromOriginalModelMixin)logging)apply_forward_hook)get_activation)CogVideoXDownsample3D)AutoencoderKLOutput) ModelMixin)CogVideoXUpsample3D) DecoderOutputDiagonalGaussianDistributioncTeZdZdZdej dej ffd ZxZS)CogVideoXSafeConv3dzq A 3D convolution layer that splits the input tensor into smaller parts to avoid OOM in CogVideoX Model. inputreturnc |jd|jdz|jdz|jdz|jdzdzdz }|dkDr|jd}t|dz dz}tj||d}|dkDrZ|dgt dt |Dcgc]4}tj||dz dddd| dzdf||fd6c}z}g}|D] }|jt |)|"tj|d} | St |)|Scc}w)Nrrr ri@dim) shape kernel_sizeinttorchchunkrangelencatappendsuperforward) selfr memory_countrpart_num input_chunksi output_chunks input_chunkoutput __class__s q/home/cdr/jupyterlab/.venv/lib/python3.12/site-packages/diffusers/models/autoencoders/autoencoder_kl_cogvideox.pyr(zCogVideoXSafeConv3d.forward+s_ [[^ekk!n ,u{{1~ = A NQVQ\Q\]^Q_ _cd dgn n  ! **1-KLQ ,Q0"1c,&784II|AE21a+9I9K3KLl[\o^def4  M+ C $$UW_[%AB CYY}!4FM7?5) )4s59E)__name__ __module__ __qualname____doc__r!Tensorr( __classcell__r1s@r2rr&s(*U\\*ell**rc eZdZdZ ddededeeeeeeffdededef fd Z dd e jd e e jd e jfd Z dd e jd e e jd e jfd Z xZS)CogVideoXCausalConv3da=A 3D causal convolution layer that pads the input tensor to ensure causality in CogVideoX Model. Args: in_channels (`int`): Number of channels in the input tensor. out_channels (`int`): Number of output channels produced by the convolution. kernel_size (`int` or `Tuple[int, int, int]`): Kernel size of the convolutional kernel. stride (`int`, defaults to `1`): Stride of the convolution. dilation (`int`, defaults to `1`): Dilation rate of the convolution. pad_mode (`str`, defaults to `"constant"`): Padding mode. in_channels out_channelsrstridedilationpad_modec t |t|tr|fdz}|\}}} |dz } |dz dz} | dz dz} ||_| |_| |_| |_| | | | | df|_d|j |j f|_ d|_ ||_ t|tr|n|ddf}|ddf}t||||||jdk(rdn |jd|_y)Nrrr r replicatezeros)r=r>rr?r@padding padding_mode)r'__init__ isinstancer rA height_pad width_padtime_padtime_causal_paddingconst_padding_conv3d temporal_dimtime_kernel_sizetuplerconv)r)r=r>rr?r@rArOheight_kernel_sizewidth_kernel_sizerKrIrJr1s r2rGzCogVideoXCausalConv3d.__init__Qs  k3 '&.1,KBM?,.?$a'(1,2 &*q0   $"  $-y*jRZ\]#^ %&$H! 0%fe461a.a#'#%#+5A4;T;T   r:inputs conv_cacherc|jdk(r$tj||jd}|S|j}|dkDr5||gn|ddddddfg|dz z}t j ||gzd}|S)NrC)moderr r)rAFpadrLrOr!r%)r)rTrUr cached_inputss r2fake_context_parallel_forwardz3CogVideoXCausalConv3d.fake_context_parallel_forward}s ==K 'UU64#;#;+NF  //KQ0:0F VTUWXZ\[\Z\T\M]L^bmpqbqLr =F8#;C r:c|j||}|jdk(rd}n*|dddd|j dzdfj}|j |}||fS)NrCr)r[rArOclonerQ)r)rTrUr0s r2r(zCogVideoXCausalConv3d.forwardsi33FJG ==K 'J1t'<'<&T) num_channels num_groupsepsaffiner)rr?)r'rGnn GroupNorm norm_layerr<conv_yconv_b)r)rcrdrer1s r2rGzCogVideoXSpatialNorm3D.__init__sP ,,J6W[dhi+KQR[\] +KQR[\] r:fzqrUrci}|xsi}|jddkDr|jddzdk(r|ddddddf|ddddddf}}|jdd|jdd}}|ddddddf|ddddddf} } tj| |} tj| |} tj| | gd}n$tj||jdd}|j ||j d\} |d<|j||j d\} |d<|j|} | | z| z}||fS) Nr r)sizerrorUrp) rrX interpolater!r%rogetrprn)r)rqrrrUnew_conv_cachef_firstf_rest f_first_size f_rest_sizez_firstz_restrorpnorm_fnew_fs r2r(zCogVideoXSpatialNorm3D.forwards\%2 771:>aggaj1n11bqbk1Q12X;VG(/ bc(:FLLdropout temb_channelsrerj non_linearity conv_shortcutspatial_norm_dimrAc t ||xs|}||_||_t ||_||_| |_| ;tj||||_ tj||||_ n&t|| ||_ t|| ||_ t||d| |_|dkDrtj|||_tj"||_t||d| |_|j|jk7r7|j rt||d| |_yt+||ddd|_yy) N)rhrirj)rcrdrer)r=r>rrAr) in_features out_featuresr)r=r>rr?rE)r'rGr=r>r nonlinearityuse_conv_shortcutrrlrmnorm1norm2rbr<conv1Linear temb_projDropoutrconv2rr) r)r=r>rrrerjrrrrAr1s r2rGzCogVideoXResnetBlock3D.__init__sV #2{ &(*=9!. 0  #;6WZ[DJrr)r)rTrrrrUry hidden_statess r2r(zCogVideoXResnetBlock3D.forwards%2  >59ZZ r^h^l^lmt^uZ5v 2M>'2 JJ}5M))-8 15MV`VdVdelVm1n. ~g.  )DNN4;L;LT;R,STUWXZ^`dfjTj,kkM >59ZZ r^h^l^lmt^uZ5v 2M>'2 JJ}5M))-8  ]3 15MV`VdVdelVm1n. ~g.   t00 0%%:>:L:Lz~~o'F;M;77++F3%. n,,r:) NrrgswishFNfirstNNN)r3r4r5r6r rfloatr`boolrGr!r7rr(r8r9s@r2rrs:'+ $#*.::sm: :  :  ::::#3-::~(,%)8< )- )-u||$)- U\\ " )- T#u||"345 )-  )-r:rceZdZdZdZ ddedededededed ed ed ed ed edeffd Z dde jde e jde e jde e ee jfde jf dZxZS)CogVideoXDownBlock3Da A downsampling block used in the CogVideoX model. Args: in_channels (`int`): Number of input channels. out_channels (`int`, *optional*): Number of output channels. If None, defaults to `in_channels`. temb_channels (`int`, defaults to `512`): Number of time embedding channels. num_layers (`int`, defaults to `1`): Number of resnet layers. dropout (`float`, defaults to `0.0`): Dropout rate. resnet_eps (`float`, defaults to `1e-6`): Epsilon value for normalization layers. resnet_act_fn (`str`, defaults to `"swish"`): Activation function to use. resnet_groups (`int`, defaults to `32`): Number of groups to separate the channels into for group normalization. add_downsample (`bool`, defaults to `True`): Whether or not to use a downsampling layer. If not used, output dimension would be same as input dimension. compress_time (`bool`, defaults to `False`): Whether or not to downsample across temporal dimension. pad_mode (str, defaults to `"first"`): Padding mode. Tr=r>rr num_layers resnet_eps resnet_act_fn resnet_groupsadd_downsampledownsample_padding compress_timerAc @t|g} t|D]-}|dk(r|n|}| jt |||||||| /t j | |_d|_| r(t j t||| | g|_d|_ y)Nr)r=r>rrrerjrrArErF) r'rGr#r&rrl ModuleListresnets downsamplersrgradient_checkpointing)r)r=r>rrrrrrrrrrArr- in_channelr1s r2rGzCogVideoXDownBlock3D.__init__js z" A()QLJ NN& *!-#"/(""/%   }}W-    " )$l2 A= ~n=AG!4 ~8VA= ~n=     (#00 ; +M :  ;n,,r:) rrrgrrTrFrr)r3r4r5r6 _supports_gradient_checkpointingr rr`rrGr!r7rrr(r8r9s@r2rrKs8(,$ $#"##-,-,-, -,  -,  -,-,-,-,-, -,-,-,d(,%)8< -|| -u||$ - U\\ " - T#u||"345 -  -r:rceZdZdZdZ ddedededededed ed eed effd Z dd e jdee jdee jdee ee jfde jf dZ xZS)CogVideoXMidBlock3Da A middle block used in the CogVideoX model. Args: in_channels (`int`): Number of input channels. temb_channels (`int`, defaults to `512`): Number of time embedding channels. dropout (`float`, defaults to `0.0`): Dropout rate. num_layers (`int`, defaults to `1`): Number of resnet layers. resnet_eps (`float`, defaults to `1e-6`): Epsilon value for normalization layers. resnet_act_fn (`str`, defaults to `"swish"`): Activation function to use. resnet_groups (`int`, defaults to `32`): Number of groups to separate the channels into for group normalization. spatial_norm_dim (`int`, *optional*): The dimension to use for spatial norm if it is to be used instead of group norm. pad_mode (str, defaults to `"first"`): Padding mode. Tr=rrrrrrrrAc t |g} t|D]%} | jt |||||||||  't j | |_d|_y)N) r=r>rrrerjrrrAF) r'rGr#r&rrlrrr) r)r=rrrrrrrrAr_r1s r2rGzCogVideoXMidBlock3D.__init__sr z" A NN& +!,#"/("%5"/%   }}W- &+#r:rrrrrUrc 6i}|xsi}t|jD]v\}}d|}tjr7|jr+|j |||||j |\}||<V|||||j |\}||<x||fS)z2Forward method of the `CogVideoXMidBlock3D` class.rrv)rrr!rrrrx) r)rrrrrUryr-rrs r2r(zCogVideoXMidBlock3D.forwards%2 "4<<0 IAv&qc]N$$&4+F+F@D@a@aM4Z^^N5SA= ~n=AG!4 ~8VA= ~n= n,,r:)rrrgrrNrr)r3r4r5r6rr rr`rrGr!r7rr(r8r9s@r2rrs0(,$  $*.,,, ,  ,  ,,,#3-,,H(,%)8< -||-u||$- U\\ " - T#u||"345 -  -r:rceZdZdZ ddededededededed ed ed ed ed edeffd Z dde jde e jde e jde e ee jfde jf dZ xZS)CogVideoXUpBlock3Da= An upsampling block used in the CogVideoX model. Args: in_channels (`int`): Number of input channels. out_channels (`int`, *optional*): Number of output channels. If None, defaults to `in_channels`. temb_channels (`int`, defaults to `512`): Number of time embedding channels. dropout (`float`, defaults to `0.0`): Dropout rate. num_layers (`int`, defaults to `1`): Number of resnet layers. resnet_eps (`float`, defaults to `1e-6`): Epsilon value for normalization layers. resnet_act_fn (`str`, defaults to `"swish"`): Activation function to use. resnet_groups (`int`, defaults to `32`): Number of groups to separate the channels into for group normalization. spatial_norm_dim (`int`, defaults to `16`): The dimension to use for spatial norm if it is to be used instead of group norm. add_upsample (`bool`, defaults to `True`): Whether or not to use a upsampling layer. If not used, output dimension would be same as input dimension. compress_time (`bool`, defaults to `False`): Whether or not to downsample across temporal dimension. pad_mode (str, defaults to `"first"`): Padding mode. r=r>rrrrrrr add_upsampleupsample_paddingrrAcBt|g}t|D].}|dk(r|n|}|jt |||||||| |  0t j ||_d|_| r(t j t||| | g|_d|_ y)Nr) r=r>rrrerjrrrArF) r'rGr#r&rrlrr upsamplersrr)r)r=r>rrrrrrrrrrrArr-rr1s r2rGzCogVideoXUpBlock3D.__init__2s z" A()QLJ NN& *!-#"/(""/%5%    }}W-   mm'$l2 A= ~n=AG!4 ~8VA= ~n=  ?? &!__ 9 )- 8  9n,,r:) rrrgrrTrFrr)r3r4r5r6r rr`rrGr!r7rrr(r8r9s@r2rrsF $ "! !#/,/,/, /,  /,  /,/,/,/,/,/,/,/,/,h(,%)8< -|| -u||$ - U\\ " - T#u||"345 -  -r:rceZdZdZdZ ddededeedfdeedfded ed ed ed ed edeffd Z dde jde e jde e ee jfde jfdZxZS)CogVideoXEncoder3DaG The `CogVideoXEncoder3D` layer of a variational autoencoder that encodes its input into a latent representation. Args: in_channels (`int`, *optional*, defaults to 3): The number of input channels. out_channels (`int`, *optional*, defaults to 3): The number of output channels. down_block_types (`Tuple[str, ...]`, *optional*, defaults to `("DownEncoderBlock2D",)`): The types of down blocks to use. See `~diffusers.models.unet_2d_blocks.get_down_block` for available options. block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`): The number of output channels for each block. act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. See `~diffusers.models.activations.get_activation` for available options. layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. norm_num_groups (`int`, *optional*, defaults to 32): The number of groups for normalization. Tr=r>down_block_types.block_out_channelslayers_per_blockact_fnnorm_epsnorm_num_groupsrrAtemporal_compression_ratioc t|ttj| } t ||dd| |_tjg|_ |d} t|D]d\}}| }||} |t|dz k(}|| k}|dk(rt|| d| ||||| | }n td|jj|ft|dd| d |||| |_tj"||dd |_tj&|_t |dd |zd| |_d |_y)NrrrrArr) r=r>rrrrrrrrzEInvalid `down_block_type` encountered. Must be `CogVideoXDownBlock3D`r )r=rrrrrrrArg)rjF)r'rGr nplog2r<conv_inrlr down_blocksrr$r ValueErrorr&r mid_blockrmnorm_outSiLUconv_actconv_outr)r)r=r>rrrrrrrrArtemporal_compress_leveloutput_channelr-down_block_type input_channelis_final_blockr down_blockr1s r2rGzCogVideoXEncoder3D.__init__s{& #&bgg.H&I"J,[:LQ:O]^iqr ==,,A."+,<"= 0 A*M/2N#&8"9A"==N 77M"881 -!/"##/'"("1'5#5"/  !!hii    # #J /- 02-*2. )   _6H6LRVW   - r "A $4!h  ',#r:samplerrUrc i}|xsi}|j||jd\}|d<tjr|jrt |j D]4\}}d|}|j|||d|j|\}||<6|j|j||d|jd\}|d<nlt |j D]*\}}d|}|||d|j|\}||<,|j||d|jd\}|d<|j|}|j|}|j||jd\}|d<||fS)z5The forward method of the `CogVideoXEncoder3D` class.rrv down_block_Nrr) rrxr!rrrrrrrrr) r)rrrUryrr-rrs r2r(zCogVideoXEncoder3D.forwards%2 37<<S]SaSabkSl<3m0 ~i0  "t'B'B!*4+;+;!<  :#.qc!2@D@a@a!NN>2 A= ~n= :>9Z9Z{+ : 6M>+6"+4+;+;!<  :#.qc!2@J!4z~~n/MA= ~n= :>tTjnn[6Q:H: 6M>+6  m4  m4 48MM-\f\j\jku\vM4w1 ~j1n,,r:) rrrrrrrrrsilurgrrrrNNr3r4r5r6rr rr`rrGr!r7rrr(r8r9s@r2rrs*(,$- /C !!,-#G,G,G, S/ G,"#s(OG,G,G,G,G,G, !G,"%*#G,X(,8< 4- 4-u||$4-T#u||"345 4-  4-r:rceZdZdZdZ ddededeedfdeedfded ed ed ed ed edeffd Z dde jde e jde e ee jfde jfdZxZS)CogVideoXDecoder3Da@ The `CogVideoXDecoder3D` layer of a variational autoencoder that decodes its latent representation into an output sample. Args: in_channels (`int`, *optional*, defaults to 3): The number of input channels. out_channels (`int`, *optional*, defaults to 3): The number of output channels. up_block_types (`Tuple[str, ...]`, *optional*, defaults to `("UpDecoderBlock2D",)`): The types of up blocks to use. See `~diffusers.models.unet_2d_blocks.get_up_block` for available options. block_out_channels (`Tuple[int, ...]`, *optional*, defaults to `(64,)`): The number of output channels for each block. act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use. See `~diffusers.models.activations.get_activation` for available options. layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block. norm_num_groups (`int`, *optional*, defaults to 32): The number of groups for normalization. Tr=r>up_block_types.rrrrrrrArc t|tt|} t || dd| |_t | ddd||||| |_tjg|_ | d} ttj| }t|D]k\}}| }| |} |t|dz k(}||k}|dk(rt!|| d| |dz||||| ||  }| }n t#d |jj%|mt'| d || |_tj*|_t | d |d| |_d |_y) Nrrrr )r=rrrrrrrArr) r=r>rrrrrrrrrrAzAInvalid `up_block_type` encountered. Must be `CogVideoXUpBlock3D`r)reF)r'rGlistreversedr<rrrrlr up_blocksr rrrr$rrr&rbrrrrr)r)r=r>rrrrrrrrArreversed_block_out_channelsrrr- up_block_typeprev_output_channelrrup_blockr1s r2rGzCogVideoXDecoder3D.__init__6s& &*84F+G&H#, 4Q7QQY  -3A6 )(  r*4Q7"%bgg.H&I"J ). 9 , A}"0 8;N#&8"9A"==N 77M 44- 3!/"##/!3'"("1%0%3!3"/% '5# !dee NN ! !( +3 ,6//J2/NP[dst   - ' +\qS[  ',#r:rrrUrc Ji}|xsi}|j||jd\}|d<tjr|jr|j |j ||||jd\}|d<t|jD]4\}}d|}|j |||||j|\}||<6nm|j ||||jd\}|d<t|jD]+\}}d|}|||||j|\}||<-|j|||jd\}|d<|j|}|j||jd\}|d<||fS)z5The forward method of the `CogVideoXDecoder3D` class.rrvr up_block_rr) rrxr!rrrrrrrrr) r)rrrUryrr-rrs r2r(zCogVideoXDecoder3D.forwards%2 37<<S]SaSabkSl<3m0 ~i0  "t'B'B9=9Z9Z{+ : 6M>+6 )8  8#,QC@D@a@a!NN>2 A= ~n= :>tV {8S:H: 6M>+6 )8  8#,QC@H!4JNN>rrrlatent_channelsrrrrr sample_height sample_widthscaling_factor shift_factor latents_mean latents_std force_upcastuse_quant_convuse_post_quant_convinvert_scale_latentsc ft|t||||||| | |  |_t ||||||| | |  |_|rt d|zd|zdnd|_|r t ||dnd|_d|_ d|_ d|_ d|_ | dz|_ | dz|_t|jdt!|j"j$dz zz |_t|jdt!|j"j$dz zz |_d|_d|_y) N) r=r>rrrrrrr) r=r>rrrrrrrr rFgUUUUUU?g?)r'rGrencoderrdecoderr quant_convpost_quant_conv use_slicing use_tilingnum_latent_frames_batch_sizenum_sample_frames_batch_sizetile_sample_min_heighttile_sample_min_widthr r$configrtile_latent_min_heighttile_latent_min_widthtile_overlap_factor_heighttile_overlap_factor_width)r)r=r>rrrrrrrrrrrrrrrrrrrr1s r2rGzAutoencoderKLCogVideoX.__init__sXF )#(-1-+'A   *'%)1-+'A   Yg-a,.>L@PRSTlpUh2<qQnr $-.),-)'4q&8#%1Q%6"&)  ' '1T[[5S5S1TWX1X+Y Z' #&))C)CqSQUQ\Q\QoQoMpstMtGu)v%w" +0').&r:rrrrrcd|_|xs |j|_|xs |j|_t|jdt |j j dz zz |_t|jdt |j j dz zz |_|xs |j|_ |xs |j|_ y)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_overlap_factor_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. Must be between 0 and 1. Setting a higher value might cause more tiles to be processed leading to slow down of the decoding process. tile_overlap_factor_width (`int`, *optional*): The minimum amount of overlap between two consecutive horizontal tiles. This is to ensure that there are no tiling artifacts produced across the width dimension. Must be between 0 and 1. Setting a higher value might cause more tiles to be processed leading to slow down of the decoding process. Tr rN) rrrr r$rrrrrr)r)rrrrs r2 enable_tilingz$AutoencoderKLCogVideoX.enable_tilingCs4&<&[@[@[#%:%Xd>X>X"&)  ' '1T[[5S5S1TWX1X+Y Z' #&))C)CqSQUQ\Q\QoQoMpstMtGu)v%w"*D*gHgHg')B)ddFdFd&r:cd|_y)z Disable tiled VAE decoding. If `enable_tiling` was previously enabled, this method will go back to computing decoding in one step. FN)rr)s r2disable_tilingz%AutoencoderKLCogVideoX.disable_tilinggs  r:cd|_y)z Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to compute decoding in several steps. This is useful to save some memory and allow larger batch sizes. TNrrs r2enable_slicingz%AutoencoderKLCogVideoX.enable_slicingns  r:cd|_y)z Disable sliced VAE decoding. If `enable_slicing` was previously enabled, this method will go back to computing decoding in one step. FNrrs r2disable_slicingz&AutoencoderKLCogVideoX.disable_slicingus !r:xc|j\}}}}}|jr/||jkDs||jkDr|j |S|j }t ||zd}d} g} t|D]s} ||z} || z| dk(rdn| z} || dzz| z}|dddd| |f}|j|| \}} |j|j|}| j|utj| d} | S)Nrrrvr r) rrrr tiled_encodermaxr#r r r&r!r%)r)r# batch_sizerh num_framesheightwidthframe_batch_size num_batchesrUencr-remaining_frames start_frame end_framex_intermediates r2_encodezAutoencoderKLCogVideoX._encode|s'>?gg; L*fe ??(B(B BftOjOjFj$$Q' '<<*(88!<  {# 'A),<< *Q.qAv!CSTK(AE25EEIq![%::;N)-nQ[)\ &NJ*!%!@ JJ~ & 'ii# r: return_dictc(|jrU|jddkDrC|jdDcgc]}|j|}}t j |}n|j|}t |}|s|fSt|Scc}w)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)rrsplitr2r!r%rr)r)r#r3x_sliceencoded_slicesh posteriors r2encodezAutoencoderKLCogVideoX.encodes    QCD771:Ndll73NNN .)A QA03 < "y99OsBzc0|j\}}}}}|jr1||jkDs||jkDr|j ||S|j }t ||zd} d} g} t| D]s} ||z} || z| dk(rdn| z}|| dzz| z}|dddd||f}|j|j|}|j|| \}} | j|utj| d} |s| fSt| S)N)r3rrrvr rr)rrrr tiled_decoderr&r#r r r&r!r%r)r)r<r3r'rhr(r)r*r+r,rUdecr-r.r/r0z_intermediates r2_decodezAutoencoderKLCogVideoX._decodes>>?gg; L*fe ??(B(B BftOjOjFj$$QK$@ @<<*(88!<  {# 'A),<< *Q.qAv!CSTK(AE25EEIq![%::;N##/!%!5!5n!E)-nQ[)\ &NJ JJ~ & 'ii#6MC((r:c:|jr_|jddkDrM|jdDcgc]}|j|j}}t j |}n|j|j}|s|fSt|Scc}w)a Decode a batch of images. Args: z (`torch.Tensor`): Input batch of latent vectors. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.vae.DecoderOutput`] instead of a plain tuple. Returns: [`~models.vae.DecoderOutput`] or `tuple`: If return_dict is True, a [`~models.vae.DecoderOutput`] is returned, otherwise a plain `tuple` is returned. rrr>)rrr6rBrr!r%r)r)r<r3z_slicedecoded_slicesdecodeds r2decodezAutoencoderKLCogVideoX.decodes    QJK''RS*Uwdll73::UNUii/Gll1o,,G: G,,Vs"Bab blend_extentc t|jd|jd|}t|D]J}|dddddd| |zddfd||z z z|dddddd|ddf||z zz|dddddd|ddf<L|S)Nrrminrr#)r)rHrIrJys r2blend_vzAutoencoderKLCogVideoX.blend_vs1771:qwwqz<@ |$ A Aq<-!*;Q!>?1q!>?1q\>\:\]^ !$"="=@_@_"_` !;!;d>\>\!\]669LL447II<<q&.1 AC1e]3 3"*0@"@!D ! {+&A'14D'D$"2Q"6qAv!K["\K 0AE :=M MI#I-A ; ;;;A : ::: <D(,||DZ|'P$D*2#t4KK%&  599Tq12/ 30 KK 5 8 o =FAsJ$S> V4q5<<QU A>QRDq5<<AE D:LMD!!$q!Q0A1A0ACSOCS'S"TU V   uyy; < =ii + r:c t|j\}}}}}t|jd|jz z}t|jd|j z z} t|j |jz} t|j|j z} |j | z } |j| z } |j}g}td||D]}g}td|| D]}t||zd}d}g}t|D]}||z}||z|dk(rdn|z}||dzz|z}|dddd|||||jz|||jzf}|j|j|}|j||\}}|j||jtj|d|j|g}t!|D]\}}g}t!|D]g\}}|dkDr|j#||dz ||| }|dkDr|j%||dz || }|j|ddddddd| d| fi|jtj|dtj|d}|s|fSt'| S) 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. rrNrvr rrrr>)rr rrrrrrrr#r&r r r&r!r%rrOrQr)r)r<r3r'rhr(r)r*rSrTrUrVrWrXr+rYr-rZr[r,rUr\r]r.r/r0r^r_r`r@s r2r?z#AutoencoderKLCogVideoX.tiled_decodeAs0?@gg; L*feT88A@_@_<_`aD66!d>\>\:\]^ !$"="=@_@_"_` !;!;d>\>\!\]669LL447II<<q&.1 AC1e]3 3!*0@"@!D ! {+&A'14D'D$"2Q"6qAv!K["\K 0AE :=M MI#I-A ; ;;;A : ::: <D++7#33D9'+||DZ|'P$D*KK%&  599Tq12+ 3, KK 1 4 o =FAsJ$S> V4q5<<QU A>QRDq5<<AE D:LMD!!$q!Q0A1A0ACSOCS'S"TU V   uyy; < =ii +6MC((r:rsample_posterior generatorc|}|j|j}|r|j|}n|j}|j |j}|s|fSt |S)N)rcr>)r;r5rrWrGr) r)rrbr3rcr#r:r<r@s r2r(zAutoencoderKLCogVideoX.forwardsf KKN..   9 5A Akk!n##6MC((r:)rrrrrrrrrgrriigYi1p?NNNTFFF)NNNN)rN)T)FTN)$r3r4r5r6r_no_split_modulesr r rr`rrrrGrrr r"r!r7r2r rrrr;rrBrGrOrQr%r? Generatorr(r8r9s@r2rrs>(,$12( & *>! !!,-  *(,/3.2"$$)%*A_/_/_/ * _/c _/ "#J!_/"#_/$%_/&'_/()_/*+_/,%*-_/./_/01_/23_/4uo5_/6uU|,7_/8eEl+9_/:;_/<=_/>"?_/@#A_/_/F15/36:59 "e ( "e (}"e%-UO "e $,E? "e  "eH  !%,,437::,0: "E*F$GG H::6))D)E-Y^YeYeJeDf)8- -4-5X]XdXdIdCe--0%,,cell%,,cellHellHu||HTQ)ellQ)Q)}^c^j^jOjIkQ)l"' /3 ) )) ) EOO, ) u||U\\) * )r:r)2typingrrrrnumpyrr!torch.nnrltorch.nn.functional functionalrXconfiguration_utilsrr loaders.single_file_modelr utilsr utils.accelerate_utilsr activationsr downsamplingrmodeling_outputsrmodeling_utilsr upsamplingrvaerr get_loggerr3loggerConv3drModuler<rbrrrrrrrr:r2r{s 0/ B?8(02',<   H %*"))*>M"BIIM"`0%RYY0%f@-RYY@-Fn-299n-bT-"))T-np-p-fU-U-p]-]-@f)Z6Lf)r: