o ۷i@srddlZddlZddlmZddlmmZddlmZm Z ddl m Z Gdddej Z Gddde eZdS) N) ConfigMixinregister_to_config) ModelMixinc s`eZdZ     ddededed eed fd ed ef fd d Zdej dej fddZ Z S)ResBlockrrr皙?samechannels kernel_sizestride dilations.leaky_relu_negative_slope padding_modecsbt||_||_tfdd|D|_tfddtt|D|_ dS)Nc s"g|] }tj|dqS)rdilationpaddingnnConv1d).0rr r rrV/home/ubuntu/vllm_env/lib/python3.10/site-packages/diffusers/pipelines/ltx2/vocoder.py z%ResBlock.__init__..c s"g|] }tjddqS)rrr)r_rrrr!r) super__init__rnegative_sloper ModuleListconvs1rangelenconvs2)selfr r rrrr __class__rrr s   zResBlock.__init__xreturncCsTt|j|jD] \}}tj||jd}||}tj||jd}||}||}q|S)Nr!)zipr#r&F leaky_relur!)r'r*conv1conv2xtrrrforward's zResBlock.forward)rrrr r ) __name__ __module__ __qualname__inttuplefloatstrr torchTensorr3 __classcell__rrr(rr s( rcseZdZdZedddgdgdgdgdgdgdgd d f d ed ed edeedeedeedeeededeffdd Zdde j de de j fddZ Z S) LTX2Vocoderz\ LTX 2.0 vocoder for converting generated mel spectrograms back to audio waveforms. i)rD)r r@r@r@)r rr i] in_channelshidden_channels out_channelsupsample_kernel_sizesupsample_factorsresnet_kernel_sizesresnet_dilationsroutput_sampling_ratec  sNtt||_t||_||_t||_||_ |jt|kr0t d|jdt|d|jt|krGt dt|jdt|dt j ||dddd|_ t |_t |_|} tt||D]4\} \} } | d }|jt j| || | | | d d t||D]\} }|jt|| ||d q|} qet j ||dddd |_dS) Nza`upsample_kernel_sizes` and `upsample_factors` should be lists of the same length but are length z and z, respectively.z_`resnet_kernel_sizes` and `resnet_dilations` should be lists of the same length but are length rFrr)r rrr@)rr)rr)rr r%num_upsample_layersresnets_per_upsamplerJmathprodtotal_upsample_factorr! ValueErrorrrconv_inr" upsamplersresnets enumerater-appendConvTranspose1drconv_out)r'rHrIrJrKrLrMrNrrOinput_channelsirr output_channelsrr(rrr 6s`       zLTX2Vocoder.__init__F hidden_states time_lastr+cs|sddddtjD]6}tjjdj||j }|dj }t j fddt||Ddd}t j |ddqtjd d t S) a Forward pass of the vocoder. Args: hidden_states (`torch.Tensor`): Input Mel spectrogram tensor of shape `(batch_size, num_channels, time, num_mel_bins)` if `time_last` is `False` (the default) or shape `(batch_size, num_channels, num_mel_bins, time)` if `time_last` is `True`. time_last (`bool`, *optional*, defaults to `False`): Whether the last dimension of the input is the time/frame dimension or the Mel bins dimension. Returns: `torch.Tensor`: Audio waveform tensor of shape (batch_size, out_channels, audio_length) r@rrr,csg|] }j|qSr)rX)rjr`r'rrrsz'LTX2Vocoder.forward..r)dimg{Gz?) transposeflattenrVr$rPr.r/r!rWrQr;stackmeanr\tanh)r'r`rar^startendresnet_outputsrrcrr3ts    $  zLTX2Vocoder.forward)F)r4r5r6__doc__rr7listr9r r;r<boolr3r=rrr(rr>1s@   $=r>)rRr;torch.nnrtorch.nn.functional functionalr.configuration_utilsrrmodels.modeling_utilsrModulerr>rrrrs  &