Mj;dZddlmZddlZddlmZmZddlZddlmZddl m Z m Z dZ dZ d ZGd d ejZ dd dZGddejZGddejZGddejZGdde ZdZdZGdde ZededdS)!a3YuE2 Oobleck audio decoder with exact-boundary tiled decoding. Oobleck and SnakeBeta derived from stable-audio-tools a6ae0cdf8b2eb1567a4b42ceadddec3712d99d45. Copyright (c) 2023 Stability AI; Copyright (c) 2022 NVIDIA CORPORATION. MIT: see THIRD_PARTY_NOTICES.md shipped with this module/model repository. ) annotationsN)CallableLiteral)nn)PretrainedConfigPreTrainedModelc$tj|i|SN)rConv1dargskwargss ;C:\Users\johnn\OneDrive\Documents\AI\models\TTS\yue2\vae.pyWNConv1drs 9d %f % %%c$tj|i|Sr )rConvTranspose1dr s rWNConvTranspose1drs t .v . ..rcn|d|dzz tjtj||zdzzS)N?& .>)torchpowsin)xalphabetas r snake_betarsA tk)*ei !e)a// rc,eZdZ dfd ZdZxZS) SnakeBetarTc"t||_||_|jr]t jt j||z|_t jt j||z|_ n\t jt j ||z|_t jt j ||z|_ ||j_ ||j _ d|_ dS)Nr) super__init__ in_featuresalpha_logscaler Parameterrzerosrrones requires_gradno_div_by_zero)selfr%ralpha_trainabler& __class__s rr$zSnakeBeta.__init__ s &,   Fek+&>&>&FGGDJ U[%=%=%EFFDIIej&=&=&EFFDJ UZ %<%%       x'!*x2GHHH ' X 5!- ! (: R[]]  mV, rc,||Sr rWrcs rr3zOobleckDecoder.forwardrdr) rrgrhrirlFrmFFFTr4r9s@rrfrf{sp""5-5-5-5-5-5-nrrfcneZdZdZdZehdZfdZegdZ dfd Z xZ S) YuE2VAEConfigzConfiguration shared by YuE2-Vae and YuE2-Vae-legacy. Use ``standard`` for listening and ``legacy`` for the paper metric baseline. yue2_vae>dtypeauto_map use_cache is_decoder model_type return_dict torch_dtype torchscript bos_token_id eos_token_id pad_token_id architecturesoutput_attentionsis_encoder_decoderadd_cross_attentionattn_implementationtie_word_embeddingsoutput_hidden_statestransformers_versiondecoder_start_token_idcfdtDS)Nc@i|]\}}|jvs |jv||S) _hf_fields_inference_fields.0keyvaluer,s r z)YuE2VAEConfig.to_dict..sELLLzsE$/))SD4J-J-JU-J-J-Jr)r#to_dictitems)r,r.s`rrzYuE2VAEConfig.to_dictsPLLLLUWW__->->-D-D-F-FLLL Lr)decoder_config sample_raterwdownsampling_ratioaudio_channelsrelease_variantdecode_core_framesdecode_halo_framesN逻@rstandardc  fd| D} | ddg| ddddtjdi| |pt dd gd gd d d d dd _hd fdjD_t |_t |_t |_ t |_ |_ t |_ t |_ j dks j dkrtdtjjdj krtdjdjkrtddS)Nc.i|]\}}|jv||Sr)rrs rrz*YuE2VAEConfig.__init__..s+XXXeAWAW#uAWAWAWrrYuE2VAErzmodeling_vae.YuE2VAEConfigzmodeling_vae.YuE2VAE) AutoConfig AutoModelrr)rRrrjrkrrh)rrrjrjrJTrmF) rNrErxryrwrzr{r~r> rxryrErzrrwr{r~rNr|r}c$i|] \}}|v || Srr)rrrdecoder_fieldss rrz*YuE2VAEConfig.__init__..s+qqqjc5[^bp[p[psE[p[p[prrRrz#Invalid VAE core/halo configurationryz/Decoder strides do not match downsampling_ratiorwz.Decoder input channels do not match latent_dimr)r setdefaultr#r$dictrintrrwrrrrrrDr_prod) r,rrrwrrrrrrrr.s ` @rr$zYuE2VAEConfig.__init__s YXXXv||~~XXX/I;777*6/' '     ""6""",FR1E1E1E&&&2 Uu1F1F1FNNNrqqqD>ICDy&,'')#FM3==11c{{C4LL((&",--AmA&q(9oa(&*)BCDFGH=#$&'( )&<BFBKQRR I$v,,2GII J JJrceZdZdZeZdZdZfdZe dZ e j dZ dZdd Ze j ddd ZxZS)ra2Strict EMA model; only the decoder needs to reside on an accelerator. ``decode_tiled`` preserves finite receptive-field context and writes cropped cores to CPU. The mathematical waveform is the full decoder's waveform; convolution kernel choices may cause small FP32 rounding differences. audioct|tdi|j|_|ddS)NFr)r#r$rfrdecoderevalrequires_grad_)r,configr.s rr$zYuE2VAE.__init__'sU    %>>(=>>  ""5)))))rcXt|jjSr )nextr parametersdevice)r,s rdecoder_devicezYuE2VAE.decoder_device.s!DL++--..55rc &tj|jjd5|||jt |jjcdddS#1swxYwYdS)z0Full waveform [B,2,1920*T-64], without clipping.F) device_typeenabled)rrN) rautocastrrrtorrr)r,latents rdecodezYuE2VAE.decode2s^(;(@% P P P t t<< 1DDQUQ]QhQhQjQjLkLkLq r rss t t t t t t t t t t t t t t t t t tsABB  B ct|dkrtdt|jt|S)NrRzframes must be positive)rrDrr)r,framess rnatural_output_lengthzYuE2VAE.natural_output_length8s8 v;;??677 7dlCKK888rNc| |jjn|}|jj}t|jd||zdz \}}t d| ||z dzS)NrrR)rrrrrr)r, core_framesratiorrs r required_halozYuE2VAE.required_halo=sa8C8Kdk44Q\  .(q+:MPQ:QRR T1sdD;.2333rcpu on_progress!Callable[[int, int], None] | Nonec0| |jjn|}| |jjn|}t|tr|dkrt d||}t|tr||krt d|d|jd}|jj}| |} tj |jd|jj | ftj |} ||zdz |z} ttd||D]\} } t!|| |z}t#d| |z }t!|||z}||d ||f}| |zt!||z| }}| |z |z}|d |||z|z f}|jd||z krt'd | d ||f||~~||| dz| | S) aDecode bounded tiles, retaining exact cores with natural end length. Each crop has enough left/right context for every dependency. There is no crossfade or boundary smoothing, and no zero padding of final audio. CPU output prevents an entire song from accumulating on the GPU. ``on_progress(completed, total)`` runs after each existing crop copy; no extra synchronization is added. With a CUDA output device, queued work may still be executing. Callback exceptions propagate. NrRz&core_frames must be a positive integerzhalo_frames must be at least z for this decoderr0r)rr.z0VAE tile did not cover its requested output core)rrrrrrDrshaperrremptyrfloat32 enumeratersrrr RuntimeErrorcopy_r)r,rr halo_frames output_devicerrequiredrrtotalrtiles tile_indexstartendleftrighttile out_startout_end crop_startcrops r decode_tiledzYuE2VAE.decode_tiledCsX9D8Kdk44Q\ 8C8Kdk44Q\ +s++ G{QEFF F%%k22+s++ Z{X/E/EXXXXXYY Yb! .**622 V\!_dk.H%P"'- GGG+%)k9!*5FK+H+H!I!I 3 3 Jfek122Cq%+-..Dk 122E;;vc4:o677D!&C%K0G0GwI$,%/JZ W(rs. #""""" $$$$$$$$ ::::::::&&&/// ***** ***>;? 9 9 9 9 929829,99999RY999x7O7O7O7O7O$7O7O7OtLLL.KKK&PPPPPoPPPf%%l333 ,,,,,r