(ti= ddlZddlmZddlmcmZddlmZddlm Z ddl m Z GddejZ Gd d ejZGd d ejZGd dejZGddejZ ddej$dej$dzdededej$f dZy)N) deprecate)RMSNorm)upfirdn2d_nativec eZdZdZ d dedededzdedef fd Zd ejd ejfd Z xZ S) Downsample1DaA 1D downsampling layer with an optional convolution. Parameters: channels (`int`): number of channels in the inputs and outputs. use_conv (`bool`, default `False`): option to use a convolution. out_channels (`int`, optional): number of output channels. Defaults to `channels`. padding (`int`, default `1`): padding for the convolution. name (`str`, default `conv`): name of the downsampling 1D layer. Nchannelsuse_conv out_channelspaddingnamecNt|||_|xs||_||_||_d}||_|r4tj|j|jd|||_ y|j|jk(sJtj|||_ y)Nrstrider  kernel_sizer) super__init__r r r r rnnConv1dconv AvgPool1d)selfr r r r rr __class__s l/Volumes/fast/ai/experiments/MLX_z-image/.venv/lib/python3.12/site-packages/diffusers/models/downsampling.pyrzDownsample1D.__init__(s   (4H     $--1B1BAf^efDI==D$5$55 55 GDIinputsreturnc`|jd|jk(sJ|j|S)Nr)shaper r)rrs rforwardzDownsample1D.forward>s+||A$--///yy  r)FNrr __name__ __module__ __qualname____doc__intboolstrrtorchTensorr# __classcell__rs@rr r sp $#' HHHDj H  H  H,!ell!u||!rr c eZdZdZ d dedededzdedef fd Zd ejd ejfd Z xZ S) Downsample2DaA 2D downsampling layer with an optional convolution. Parameters: channels (`int`): number of channels in the inputs and outputs. use_conv (`bool`, default `False`): option to use a convolution. out_channels (`int`, optional): number of output channels. Defaults to `channels`. padding (`int`, default `1`): padding for the convolution. name (`str`, default `conv`): name of the downsampling 2D layer. Nr r r r rc 0t |||_|xs||_||_||_d} ||_|dk(rtj||| |_ n0|dk(rt||| |_ n|d|_ ntd||r0tj|j|j|| || } n2|j|jk(sJtj| | } |dk(r| |_| |_y|dk(r| |_y| |_y) Nrln_normrms_normzunknown norm_type: )rrr biasrrConv2d_0)rrr r r r rr LayerNormnormr ValueErrorConv2d AvgPool2dr6r)rr r r r rr norm_typeepselementwise_affiner5rrrs rrzDownsample2D.__init__Ss   (4H     ! Xs4FGDI * $#/ABDI  DI29+>? ? 99 t00kRXbiptD==D$5$55 55< DMDI Z DIDIr hidden_statesr ct|dkDs|jddd}tdd||jd|jk(sJ|j 5|j |j ddddj dddd}|jr*|jdk(rd}tj||d d }|jd|jk(sJ|j|}|S) NrscalezThe `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`.z1.0.0rrrrrrrconstantmodevalue) lengetrr"r r8permuter r Fpadr)rr?argskwargsdeprecation_messagerKs rr#zDownsample2D.forwards t9q=FJJw5A#U  gw(; <""1%666 99  IIm&;&;Aq!Q&GHPPQRTUWXZ[\M ==T\\Q.CEE-:QOM""1%666 -0 r) FNrrrNNNTr$r/s@rr1r1Csz $#'  ---Dj -  -  -^U\\u||rr1c eZdZdZ ddedzdedzdedeeeeefffd Z ddejd ejdzd ejdzd ed e d ejf dZ dejd ejfdZ xZ S)FirDownsample2DaA 2D FIR downsampling layer with an optional convolution. Parameters: channels (`int`): number of channels in the inputs and outputs. use_conv (`bool`, default `False`): option to use a convolution. out_channels (`int`, optional): number of output channels. Defaults to `channels`. fir_kernel (`tuple`, default `(1, 3, 3, 1)`): kernel for the FIR filter. Nr r r fir_kernelct||r|n|}|rtj||ddd|_||_||_||_y)Nrrrrr )rrrr:r6rQr r )rr r r rQrs rrzFirDownsample2D.__init__sL '3| IIh !TU_`aDM$  (rr?weightkernelfactorgainr ct|tr|dk\sJ|dg|z}tj|tj}|j dk(rtj ||}|tj|z}||z}|jr|j\}}}}|jd|z |dz z} ||g} t|tj||j| dzdz| dzf} tj| || d} | S|jd|z } t|tj||j|| dzdz| dzf} | S) a"Fused `Conv2d()` followed by `downsample_2d()`. Padding is performed only once at the beginning, not between the operations. The fused op is considerably more efficient than performing the same calculation using standard TensorFlow ops. It supports gradients of arbitrary order. Args: hidden_states (`torch.Tensor`): Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`. weight (`torch.Tensor`, *optional*): Weight tensor of the shape `[filterH, filterW, inChannels, outChannels]`. Grouped convolution can be performed by `inChannels = x.shape[0] // numGroups`. kernel (`torch.Tensor`, *optional*): FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which corresponds to average pooling. factor (`int`, *optional*, default to `2`): Integer downsampling factor. gain (`float`, *optional*, default to `1.0`): Scaling factor for signal magnitude. Returns: output (`torch.Tensor`): Tensor of the shape `[N, C, H // factor, W // factor]` or `[N, H // factor, W // factor, C]`, and same datatype as `x`. rdtyperdevicer)rKrdownrK) isinstancer)r,tensorfloat32ndimoutersumr r"rr\rJconv2d) rr?rTrUrVrW_convHconvW pad_value stride_value upfirdn_inputoutputs r_downsample_2dzFirDownsample2D._downsample_2ds_B&#&6Q;66 >S6\FfEMM: ;;! [[0F%))F##$ ==!' Aq%a61eai@I"F+L, VM,@,@A!m)9>:M XXmVLRSTF  Q&0I% VM,@,@A!m)9>: F rc|jr_|j||jj|j}||jj j ddddz}|S|j||jd}|S)N)rTrUrr)rUrV)r rmr6rTrQr5reshape)rr?downsample_inputs rr#zFirDownsample2D.forwards ==#22=I]I]fjfufu2v ,t}}/A/A/I/I!RQRTU/VVM!// doo^_/`Mr)NNF)rrrr)NNrr)r%r&r'r(r)r*tuplerr,r-floatrmr#r.r/s@rrPrPs  $#'0< )* )Dj ) ) #sC,- )$'+&* @||@ t#@ t# @  @  @ @DU\\ellrrPcbeZdZdZddeffd ZdejdejfdZxZ S) KDownsample2DzA 2D K-downsampling layer. Parameters: pad_mode (`str`, *optional*, default to `"reflect"`): the padding mode to use. pad_modect|||_tjgdg}|j ddzdz |_|jd|j|zdy)N)??ryrxrrrUF) persistent) rrrvr,r`r"rKregister_bufferT)rrv kernel_1drs rrzKDownsample2D.__init__s_   LL">!?@ ??1%*Q. Xy{{Y'>5Qrrr c4tj||jfdz|j}|j|jd|jd|j jd|j jdg}t j|jd|j}|j j|dddfj|jddd}||||f<tj||dS)Nrrr[ror)r) rJrKrv new_zerosr"rUr,aranger\toexpandre)rrrTindicesrUs rr#zKDownsample2D.forwardsv{Q >!! Q Q !!!$ !!!$   ,,v||Av}}E'a077 QRP#)w xxq11r)reflect) r%r&r'r(r+rr,r-r#r.r/s@rrurus1 RR 2ell 2u|| 2rruc ~eZdZdZ d dedededededef fd Zd ejd ejfd Z xZ S) CogVideoXDownsample3Da A 3D Downsampling layer using in [CogVideoX]() by Tsinghua University & ZhipuAI Args: in_channels (`int`): Number of channels in the input image. out_channels (`int`): Number of channels produced by the convolution. kernel_size (`int`, defaults to `3`): Size of the convolving kernel. stride (`int`, defaults to `2`): Stride of the convolution. padding (`int`, defaults to `0`): Padding added to all four sides of the input. compress_time (`bool`, defaults to `False`): Whether or not to compress the time dimension. in_channelsr rrr compress_timecnt|tj||||||_||_y)NrS)rrrr:rr)rrr rrr rrs rrzCogVideoXDownsample3D.__init__2s2 IIk<[Y_ipq *rxr c|jr*|j\}}}}}|jdddddj||z|z||}|jddzdk(r|d|dddf}}|jddkDrt j |dd }t j|d |gd }|j|||||jdjddddd}nMt j |dd }|j|||||jdjddddd}d } t j|| d d}|j\}}}}}|jdddddj||z|||}|j|}|j|||jd|jd|jdjddddd}|S)Nrrrrrro).r.r).N)dimrBrCrD) rr"rIrprJ avg_pool1dr,catrKr) rr batch_sizer framesheightwidthx_firstx_restrKs rr#zCogVideoXDownsample3D.forward@s   :;'' 7J&&% !Q1a(00f1Du1LhX^_Awwr{Q!#"#F)QsABwZ<<#a'\\&aJFIIwy16:CIIj&%1772;OWWXY[\^_abdefLL!<IIj&%1772;OWWXY[\^_abdef EE!Sz 367gg3 Hffe IIaAq! $ , ,Z&-@(FTY Z IIaL IIj&!''!*aggaj!''!* M U UVWYZ\]_`bc dr)rrrF) r%r&r'r(r)r*rr,r-r#r.r/s@rrrsp,# + + + +  +  + +%,,rrr?rUrVrWr ct|tr|dk\sJ|dg|z}tj|tj}|j dk(rtj ||}|tj|z}||z}|jd|z }t||j|j||dzdz|dzf}|S)aEDownsample2D a batch of 2D images with the given filter. Accepts a batch of 2D images of the shape `[N, C, H, W]` or `[N, H, W, C]` and downsamples each image with the given filter. The filter is normalized so that if the input pixels are constant, they will be scaled by the specified `gain`. Pixels outside the image are assumed to be zero, and the filter is padded with zeros so that its shape is a multiple of the downsampling factor. Args: hidden_states (`torch.Tensor`) Input tensor of the shape `[N, C, H, W]` or `[N, H, W, C]`. kernel (`torch.Tensor`, *optional*): FIR filter of the shape `[firH, firW]` or `[firN]` (separable). The default is `[1] * factor`, which corresponds to average pooling. factor (`int`, *optional*, default to `2`): Integer downsampling factor. gain (`float`, *optional*, default to `1.0`): Scaling factor for signal magnitude. Returns: output (`torch.Tensor`): Tensor of the shape `[N, C, H // factor, W // factor]` rrYrr[rr]) r_r)r,r`rarbrcrdr"rrr\)r?rUrVrWrirls r downsample_2drbs8 fc "v{2 2 ~v \\& 6F {{aVV, eiiF d]F Q&(I  -- . !m !9> 2 F Mr)Nrr)r,torch.nnrtorch.nn.functional functionalrJutilsr normalizationr upsamplingrModuler r1rPrurr-r)rsrrrrs "((!299(!VP299PffbiifT2BII2