o ۷iA.@sddlZddlZddlmmZddlmZmZddl m Z ddddd Z Gd d d ejj Z Gd d d ejj ZGdddejj ZGdddejj ZGddde eZdS)N) ConfigMixinregister_to_config) ModelMixin)r)r)r)rr)g?g?@g@csFeZdZd dededBdeffdd Zdejd ejfd d ZZS) ResBlockNrchannels mid_channelsdimscst|dur |}|dkrtjjntjj}|||ddd|_tjd||_|||ddd|_ tjd||_ tj |_ dS)Nrrr kernel_sizepadding ) super__init__torchnnConv2dConv3dconv1 GroupNormnorm1conv2norm2SiLU activation)selfr r r Conv __class___/home/ubuntu/vllm_env/lib/python3.10/site-packages/diffusers/pipelines/ltx2/latent_upsampler.pyr"s zResBlock.__init__ hidden_statesreturncCsH|}||}||}||}||}||}|||}|SN)rrrrr)rr%residualr#r#r$forward/s     zResBlock.forward)Nr) __name__ __module__ __qualname__intrrTensorr) __classcell__r#r#r!r$r !s  r cs&eZdZdfdd ZddZZS)PixelShuffleNDrrrcs*t||_||_|dvrtddS)N)rrrzdims must be 1, 2, or 3)rrr upscale_factors ValueError)rr r2r!r#r$r<s zPixelShuffleND.__init__c Cs|jdkr*|ddg|jddRdddddddd ddd dddS|jdkrN|ddg|jddRddd dddd dddS|jdkrn|ddg|jddRddddd dddSdS) Nrrrrr)r unflattenr2permuteflattenrxr#r#r$r)Es  < 6zPixelShuffleND.forward)r1)r*r+r,rr)r/r#r#r!r$r0;s r0c sJeZdZdZd dedededdffdd Zd ejdejfd d ZZ S)BlurDownsamplez Anti-aliased spatial downsampling by integer stride using a fixed separable binomial kernel. Applies only on H,W. Works for dims=2 or dims=3 (per-frame). r5r striderr&Nc st|dvrtd|dksddkr!td||_||_|_tfddtD}|dddf|dddf}||  }| d |ddddddfdS) N)rrz$`dims` must be either 2 or 3 but is rrrz0`kernel_size` must be an odd number >= 3 but is csg|] }td|qS)r)mathcomb).0krr#r$ osz+BlurDownsample.__init__..kernel) rrr3r r>rrtensorrangesumfloatregister_buffer)rr r>rrBk2dr!rCr$r_s  $zBlurDownsample.__init__r<c Cs|jdkr|S|jdkr-|jd}|j|d|j|j}tj||d|j|jd|d}|S|j\}}}}}|dd dd}|j|d|j|j}tj||d|j|jd|d}|jdd\}}| d||f |d|||}|S)Nrr)weightbiasr>rgroupsrr4) r>r shaperEexpandrFconv2d transposer:r8reshape) rr<crLbf_h2w2r#r#r$r)ts    zBlurDownsample.forward)r5) r*r+r,__doc__r-rrr.r)r/r#r#r!r$r=Ys r=csBeZdZdZd dedeffdd Zdejd ejfd d Z Z S) SpatialRationalResamplera Scales by the spatial size of the input by a rational number `scale`. For example, `scale = 0.75` will downsample by a factor of 3 / 4, while `scale = 1.5` will upsample by a factor of 3 / 2. This works by first upsampling the input by the (integer) numerator of `scale`, and then performing a blur + stride anti-aliased downsample by the (integer) denominator. r r scalecstt||_t|d}|dur"td|dtt|\|_ |_ t j j ||j d|ddd|_td|j |j fd|_td|j d|_dS) NzThe supplied `scale` z( is not supported; supported scales are rrrr)r2)r r>)rrrIr_ RATIONAL_RESAMPLER_SCALE_MAPPINGgetr3listkeysnumdenrrrconvr0 pixel_shuffler= blur_down)rr r_ num_denomr!r#r$rs     z!SpatialRationalResampler.__init__r<r&cCs"||}||}||}|Sr')rfrgrhr;r#r#r$r)s   z SpatialRationalResampler.forward)r^r ) r*r+r,r\r-rIrrr.r)r/r#r#r!r$r]sr]csleZdZdZe       dd ed ed ed ed edededBffdd Zde j de j fddZ Z S)LTX2LatentUpsamplerModela Model to spatially upsample VAE latents. Args: in_channels (`int`, defaults to `128`): Number of channels in the input latent mid_channels (`int`, defaults to `512`): Number of channels in the middle layers num_blocks_per_stage (`int`, defaults to `4`): Number of ResBlocks to use in each stage (pre/post upsampling) dims (`int`, defaults to `3`): Number of dimensions for convolutions (2 or 3) spatial_upsample (`bool`, defaults to `True`): Whether to spatially upsample the latent temporal_upsample (`bool`, defaults to `False`): Whether to temporally upsample the latent r^rrTFr in_channelsr num_blocks_per_stager spatial_upsampletemporal_upsamplerational_spatial_scaleNc s|t||_|_||_|_||_||_dkrtj j ntj j }||ddd|_ tj d|_tj |_tj fddt|D|_|rd|rdtj tj j ddddtd|_n>|r|durrt|d |_n0tj tj j d dddtd|_n|rtj tj j ddddtd|_ntd tj fd dt|D|_||ddd|_dS) Nrrrrrcg|]}tdqS)r r rArYr r r#r$rDz5LTX2LatentUpsamplerModel.__init__..)r r_rz9Either spatial_upsample or temporal_upsample must be Truecrqrrrsrtrur#r$rDrv)rrrlr rmr rnrorrrr initial_convr initial_normrinitial_activation ModuleListrG res_blocks Sequentialr0 upsamplerr]r3post_upsample_res_blocks final_conv) rrlr rmr rnrorpConvNdr!rur$rsD  "z!LTX2LatentUpsamplerModel.__init__r%r&cCs~|j\}}}}}|jdkrX|ddddddd}||}||}||}|jD]}||}q,||}|j D]}||}q;| |}| d|dfddddd}|S||}||}||}|jD]}||}qj|j r||}|ddddddddddf}n!|ddddddd}||}| d|dfddddd}|j D]}||}q| |}|S)Nrrrrrr4) rPr r9r:rxryrzr|r~rrr8ro)rr% batch_size num_channels num_framesheightwidthblockr#r#r$r)s:               (    z LTX2LatentUpsamplerModel.forward)rkr^rrTFr ) r*r+r,r\rr-boolrIrrr.r)r/r#r#r!r$rjs46rj)r?rtorch.nn.functionalr functionalrRconfiguration_utilsrrmodels.modeling_utilsrr`Moduler r0r=r]rjr#r#r#r$s  0!