3jʩ^SSKrSSKrSSKrSSKJs Jr SSKJr SSK J r SSK J r J r SSKJr SSKJr SSKJr SS KJr S S KJrJrJrJr S S KJrJrJr S SKJ r S SKJ!r! \RD"\#5r$S1Sjr%"SS\5r&"SS\5r'"SS\5r("SS\5r)"SS\5r*"SS\5r+"SS\5r,"SS \5r-"S!S"\5r."S#S$\.5r/"S%S&\.5r0"S'S(\5r1"S)S*\.5r2"S+S,\.5r3"S-S.\.5r4"S/S0\.5r5g)2N)tqdm) FrozenDict)ClassifierFreeGuidanceClassifierFreeZeroStarGuidance)HeliosTransformer3DModel)HeliosScheduler)logging) randn_tensor) BlockStateLoopSequentialPipelineBlocksModularPipelineBlocks PipelineState) ComponentSpec InputParam OutputParam)calculate_shift)HeliosModularPipelinec Uc[R"US9nO[U[5(aUSnUupn X-n [R"XS9SU--[R "XUS9U-- n U [R"XS9S-- n U R 5n [RRU 5nX-U-X:--XK--n[R"XURUS9RU5nUUR-nURXX#U -XK-X5nURSSSSSS S 5RXX#U5nU$) zGenerate spatially-correlated block noise for pyramid upsampling correction. Uses a multivariate normal distribution with covariance based on `gamma` to produce noise with block structure, matching the upsampling artifacts that need correction. )devicerrg:0yE>r generatorr r)torch Generator isinstancelisteyeonesfloatlinalgcholeskyrandnrtoTviewpermutereshape) batch_sizechannel num_framesheightwidthgamma patch_sizerr_phpw block_sizecovL block_numberznoises d/home/wildlama/miniconda3/lib/python3.13/site-packages/diffusers/modular_pipelines/helios/denoise.pysample_block_noiser>*sA&OO62 It $ $aL IA2J  *,E :UZZ gm=nqv=vv599Z /$ 66C ))+C c"A'*4 EUL LY5E5EQZ[^^_efA GE JJzJ" ekSU ZE MM!Q1aA . 6 6zJ`e fE Lc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )HeliosChunkHistorySliceStepZzSlices history latents into short/mid/long for a T2V chunk. At k==0 with no image_latents, creates a zero prefix. Otherwise uses image_latents (either provided or captured from first chunk by HeliosChunkUpdateStep). heliosreturncg)NzT2V history slice: splits history into long/mid/short. At k==0 with no image_latents, creates a zero prefix; otherwise uses image_latents as prefix for short history.selfs r= description'HeliosChunkHistorySliceStep.descriptioncs  _ r?c [SS[SS9[SS[SS9[SS[RS S9[S S[ S 9/$) Nkeep_first_frameT7Whether to keep the first frame as a prefix in history.default type_hintrI history_sizes=Sizes of long/mid/short history buffers for temporal context.requiredrPrIhistory_latents1Accumulated history latents from previous chunks. latent_shaperTrP)rboolr!rTensortuplerGs r=inputs"HeliosChunkHistorySliceStep.inputsjs` "U   [   !,,O   ~ F'  r?c/$NrFrGs r=intermediate_outputs0HeliosChunkHistorySliceStep.intermediate_outputs r? components block_statekc  URnURnURnURnURuppn U(aUR SS2SS2[ U5*S24RUSS9upnUc/US:Xa)[R"UU SU U U[RS9nOUn[R"UU/SS9nO3UR SS2SS2[ U5*S24RUSS9upnUUl Xl XlX4$)Nr dimrr)rdtype)rLrQ image_latents_execution_devicerWrUsumsplitrzerosfloat32catlatents_history_shortlatents_history_midlatents_history_long)rHrcrdrerLrQrjrr-num_channels_latentsr4h_latentw_latentrsrrlatents_history_1xlatents_prefixrqs r=__call__$HeliosChunkHistorySliceStep.__call__s4&77#11 #11 --BMBZBZ? !x LWLgLg1s=))++MeMqe) J 7I$a!&(!--""/$)II~?Q.RXY$Z !OZOjOj1s=))++PeMqe) M 7L-B )*='+?(&&r?rFN__name__ __module__ __qualname____firstlineno____doc__ model_namepropertystrrIr!rr\rr`rno_gradrr intry__static_attributes__rFr?r=rArAZs J  S   Z(  .d;&7 ]]_"'#8"'z"'VY"'"'r?rAc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )HeliosI2VChunkHistorySliceStepzSlices history latents into short/mid/long for an I2V chunk. Always uses image_latents as prefix (assumes history pre-seeded with fake_image_latents). rCrDcg)NzyI2V history slice: splits pre-seeded history into long/mid/short, always using image_latents as prefix for short history.rFrGs r=rI*HeliosI2VChunkHistorySliceStep.descriptions  F r?c [SS[SS9[SS[SS9[SS[RS S9[S S[RS S9/$) NrLTrMrNrQrRrSrUrVrjz5First-frame latents used as prefix for short history.)rrYr!rrZrGs r=r\%HeliosI2VChunkHistorySliceStep.inputssl "U   [   !,,O   ,,S  '  r?c/$r_rFrGs r=r`3HeliosI2VChunkHistorySliceStep.intermediate_outputsrbr?rcrdrec|URnURnURnU(aJURSS2SS2[ U5*S24R USS9upxn [ R"Xi/SS9n O3URSS2SS2[ U5*S24R USS9upxn XlXl Xrl X4$)Nr rg) rLrQrjrUrlrmrrprqrrrs) rHrcrdrerLrQrjrsrrrwrqs r=ry'HeliosI2VChunkHistorySliceStep.__call__s&77#11 #11 LWLgLg1s=))++MeMqe) J 7I%*II}.QWX$Y !OZOjOj1s=))++PeMqe) M 7L-B)*='+?(&&r?rFNr{rFr?r=rrs J  S   Z(  8d;&7 ]]_'#8'z'VY''r?rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \ R"5S\S\S \4S j5rS rg ) HeliosChunkNoiseGenStepz7Generates noise latents for a chunk using randn_tensor.rCrDcg)NzEGenerates random noise latents at full resolution for a single chunk.rFrGs r=rI#HeliosChunkNoiseGenStep.descriptionsVr?cL[SS[S9[R"S5/$)NrWTrXr)rr[templaterGs r=r\HeliosChunkNoiseGenStep.inputss( ~ F    ,  r?rcrdrecURn[URURU[R S9UlX4$)Nrrri)rkr rWrrrolatentsrHrcrdrers r=ry HeliosChunkNoiseGenStep.__call__sA--*  $ $ 0E0Ef\a\i\i &&r?rFNr|r}r~rrrrrrIr!rr\rrrr rryrrFr?r=rrswAJ WSWW Z(   ]]_'#8'z'VY''r?rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \ R"5S\S\S \4S j5rS rg ) HeliosPyramidChunkNoiseGenStepi zBGenerates noise latents and downsamples to smallest pyramid level.helios-pyramidrDcg)NzuGenerates random noise at full resolution, then downsamples to the smallest pyramid level via bilinear interpolation.rFrGs r=rI*HeliosPyramidChunkNoiseGenStep.descriptions  8 r?cn[SS[S9[S/SQ[SS9[R"S5/$) NrWTrX pyramid_num_inference_steps_list rr,Number of denoising steps per pyramid stage.rNr)rr[r!rrGs r=r\%HeliosPyramidChunkNoiseGenStep.inputss> ~ F 2$J      ,  r?rcrdrecURnURupVpxn [URURU[R S9n XpU R SSSSS5RXW-XkU 5n [[UR5S- 5H'n U S-n U S-n [R"XU 4SS9S-n M) U RXWXkU 5R SSSSS5Ul X4$) Nrrr rrrbilinearsizemode)rkrWr rrror+r,rangelenrF interpolater)rHrcrdrerr-rtnum_latent_framesrurvrhwr4s r=ry'HeliosPyramidChunkNoiseGenStep.__call__%s--R]RjRjO *;x  $ $ 0E0Ef\a\i\i  1//!Q1a0889WYmrsts;GGH1LMA !GA !GAmmGa&zJQNGN&oojMafghpp q!Q &&r?rFNrrFr?r=rr stL!J  S    Z(     ]]_'#8'z'VY''r?rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )HeliosChunkSchedulerResetStepi<z7Resets the scheduler with timesteps for a single chunk.rCrDcg)NzXResets the scheduler with the correct timesteps and shift parameter (mu) for this chunk.rFrGs r=rI)HeliosChunkSchedulerResetStep.descriptionAsir?c$[S[5/$)N scheduler)rr rGs r=expected_components1HeliosChunkSchedulerResetStep.expected_componentsEs + 7  r?ct[SS[S9[R"SSS9[R"S5/$)NmuTrXsigmas)rTnum_inference_steps)rr$rrGs r=r\$HeliosChunkSchedulerResetStep.inputsKs9 tde <   4 8    5 6  r?rcrdrecURnURRURXBRUR S9 URR UlX4$)N)rrr)rkr set_timestepsrrr timestepsrs r=ry&HeliosChunkSchedulerResetStep.__call__Ss_--**  + +FCUCUZeZhZh + !+ 4 4 > > &&r?rFNr|r}r~rrrrrrIr!rrrr\rrrr rryrrFr?r=rr<sAJ jSjj T-%8   Z(   ]]_'#8'z'VY''r?rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )HeliosChunkDenoiseInnericzLInner timestep loop for denoising a single chunk, using guider for guidance.rCrDcg)NzInner denoising loop that iterates over timesteps for a single chunk. Uses the guider to manage conditional/unconditional forward passes with cache_context, applies guidance, and runs scheduler step.rFrGs r=rI#HeliosChunkDenoiseInner.descriptionhs  9 r?c v[S[5[S[5[S[[ SS05SS9/$)N transformerrguiderguidance_scale@ from_configconfigdefault_creation_methodrrr rrrGs r=r+HeliosChunkDenoiseInner.expected_componentspB -)A B + 7 &!#3S"9:(5    r?c ^[R"S5[R"S5[S[RS9[S[RS9[R"S5[R"S5[R"S5[R"S 5/$) Nrr prompt_embedsrPnegative_prompt_embedsdenoiser_input_fieldsrattention_kwargsr)rrrrZrGs r=r\HeliosChunkDenoiseInner.inputs}s    *    , %,, ? /5<< H    7 8    5 6    2 3    ,  r?rcrdrec \URnURnURnURRn[ U5XaR R-- nSURUR40n [[R"URR5RR55n 0n UR R#5HupX;dM X;dMXU 'M UR$R'U5U S'UR(R'U5U S'UR*R'U5U S'UR,U S'[/US9n[1U5GHunnUR3UR4S5R'[6R85nUR'U5nUR:R=XUS9 UR:R?U 5nUHnUR:RAUR5 U R5Vs0sHo3[CUU5_M nn[CUUR:RD5nURRGU5 UR"S UUS S .UDU D6SUl$SSS5 UR:RKUR5 M UR;U5SnUR RMUUUURNS S 9SnU[ U5S - :Xd/US -U:dGMUS -UR R-S:XdGMURQ5 GM SSS5 XBlX4$s snf!,(df  N=f!,(df  N-=f)Nencoder_hidden_statesrqrrrsrtotalrsteprtimestepF hidden_statesr return_dictrrrrF))rrrrrirrorderrrsetinspect signatureforward parameterskeysritemsrqr(rrrsrr enumerateexpandshaperint64r set_stateprepare_inputsprepare_modelsgetattr_identifier_key cache_context noise_predcleanup_modelsrrupdate)rHrcrdrerrrtransformer_dtypenum_warmup_steps guider_inputstransformer_args shared_kwargs field_name field_value progress_baritrlatent_model_input guider_stateguider_state_batch cond_kwargs context_namers r=ry HeliosChunkDenoiseInner.__call__su%%)) )==&2288y>,?BVBVB\B\,\\ $k&?&?AcAc%d w001G1G1O1OP[[``bc '2'H'H'N'N'P #J-*2Q,7j)(Q 2=1R1R1U1UVg1h -./:/N/N/Q/QRc/d +,0;0P0P0S0STe0f ,-,7,H,H () + , !),188GMM!$4588E%,ZZ0A%B"!!++fg+h)00?? N *6&%%44Z5K5KLN[N`N`Nb"cNbg.@!&D#DNbK"c#*+=z?P?P?`?`#aL#//==lK8B8N8N9*<%-(-9* 9 , 9  9*5L%%44Z5K5KL+7(..|N!NN N N N+rFNrrFr?r=rrcsVJ  S    T-%8      Z(     ]]_B'#8B'zB'VYB'B'r?rc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )HeliosPyramidChunkDenoiseInneria-Nested pyramid stage loop with inner timestep denoising. For each pyramid stage (small -> full resolution): 1. Upsample latents + block noise correction (stages > 0) 2. Compute mu from current resolution, set scheduler timesteps 3. Run timestep denoising loop (same logic as HeliosChunkDenoiseInner) rrDcg)NzPyramid denoising inner block: loops over pyramid stages from smallest to full resolution. Each stage upsamples latents (with block noise correction), recomputes scheduler parameters, and runs the timestep denoising loop.rFrGs r=rI*HeliosPyramidChunkDenoiseInner.descriptions  4 r?c x[S[5[S[5[S[[ SSS.5SS9/$) Nrrrrr )rzero_init_stepsrr)rrr rrrGs r=r2HeliosPyramidChunkDenoiseInner.expected_componentssA -)A B + 7 .!SQ"OP(5    r?c ,[R"S5[S[RS9[S[RS9[R"S5[S/SQ[SS 9[R"S 5[R"S 5/$) NrrrrrrrrrNrr)rrrrZr!rGs r=r\%HeliosPyramidChunkDenoiseInner.inputss{    * %,, ? /5<< H    7 8 2$J      2 3    ,  r?rcrdrec X URnURRnURn[ UR 5nSUR UR40n[[R"URR5RR55n 0n URR5HupX;dM X;dMXU 'M UR R#U5U S'UR$R#U5U S'UR&R#U5U S'UR(U S'[+UR,SS5n [/U5GHnU bUS:aSUR,lURR2R4nUR6SUR6S -UR6S -USUS -US --n[9UUR:R2R=S S5UR:R2R=SS5UR:R2R=SS5UR:R2R=SS55nUR UnUR:R?UUUUS9 UR:R@nUS:GaPUR6unnnnnUS -nUS -nURCSS S SS5REUU-UUU5n[FRH"UUU4SS9nUREUUUUU5RCSS S SS5nS UR:RJU- nUR:R2RLnS [NRP"S S U- -5S U- -U-- nUS U- -[NRP"U5- nUR6unnnnn [SUUUUU UUUURTS9 n!U!R#US9n!UU-UU!--n[ U5UUR:RV-- n"[YUS9n#[[U5GHun$n%U%R]UR6S5R#[^R`5n&UR#U5n'UR,RcU$UU%S9 UR,ReU5n(U(Hn)UR,RgUR5 UR5V*s0sHn*U*[+U)U*5_M n+n*[+U)UR,Rh5n,URRkU,5 UR"S!U'U&SS.U+DU D6SU)l6SSS5 UR,RoUR5 M UR-U(5Sn-UR:RqU-U%UURTSS 9SnU$[ U5S - :Xd/U$S -U":dGMU$S -UR:RV-S:XdGMU#Rs5 GM SSS5 GM U bXR,lXblX4$s sn*f!,(df  N=f!,(df  GM>=f)"Nrrqrrrsrrrrr base_image_seq_lenmax_image_seq_len base_shift? max_shiftffffff?)rrrrnearestrrrirrFrrrF):rkrrirrrrrrrrrrrrrrqr(rrrsrrrrrrr3rrrgetrrr+r,rrori_start_sigmasr2mathsqrtr>rrrrrrrrrrrrrrrr).rHrcrdrerrrpyramid_num_stagesrrrrrorig_zero_init_stepsi_sr3 image_seq_lenrrrr-rtr/ current_h current_wnew_hnew_w ori_sigmar2alphabetarrr<rrr r rr r r kkrrrs. r=ry'HeliosPyramidChunkDenoiseInner.__call__s--&2288%% !M!MN $k&?&?AcAc%d w001G1G1O1OP[[``bc '2'H'H'N'N'P #J-*2Q,7j)(Q 2=1R1R1U1UVg1h -./:/N/N/Q/QRc/d +,0;0P0P0S0STe0f ,-,7,H,H () 'z'8'8:KTR+,C$/C!G45 !!1$//66AAJ$]]2.r1BBW]]SUEVV1 1 - 1 =M!$$++//0DcJ$$++//0CTJ$$++// cB$$++// TB B#."N"Ns"S   . .# /  #,,66IQwU\UbUbR 0*i!A !A !//!Q1a8@@+-A9i--uen9U!//*jBVX]_demmq!Q  4 4 E Ec JJ ",,3399TYYqAI71y=IIUVI .51AAEL]]B 0*a*(!)33 '89'/D5L8 #9~0CjFZFZF`F`0`` /0L%i0DAq xx a(89< noise gen -> scheduler reset -> denoise -> update. history_slice noise_genscheduler_reset denoise_inner update_chunkrDcg)NzT2V chunk denoise step that iterates over temporal chunks. At each chunk: history_slice -> noise_gen -> scheduler_reset -> denoise_inner -> update_chunk.rFrGs r=rI"HeliosChunkDenoiseStep.descriptions  m r?rFN)r|r}r~rrrArrrr: block_classes block_namesrrrIrrFr?r=rXrXs>h $% MeK  S  r?rXcF\rSrSrSr\\\\\ /r /SQr \ S\ 4Sj5rSrg)HeliosI2VChunkDenoiseStepizbI2V chunk-based denoising: I2V history slice -> noise gen -> scheduler reset -> denoise -> update.rYrDcg)NzI2V chunk denoise step that iterates over temporal chunks. At each chunk: history_slice (I2V) -> noise_gen -> scheduler_reset -> denoise_inner -> update_chunk.rFrGs r=rI%HeliosI2VChunkDenoiseStep.descriptions  s r?rFN)r|r}r~rrrrrrr:rarbrrrIrrFr?r=rdrds>l '% MeK  S  r?rdc\rSrSrSrSr\S\4Sj5r\S\ \ 4Sj5r \S\ \ 4Sj5r \R"5S\S \S \4S j5rS rg )'HeliosPyramidDistilledChunkDenoiseInneria|Nested pyramid stage loop with DMD denoising for distilled checkpoints. Same progressive multi-resolution strategy as HeliosPyramidChunkDenoiseInner, but: - Guidance is disabled (guidance_scale=1.0, no unconditional pass) - Supports is_amplify_first_chunk (doubles first chunk's timesteps via scheduler) - Tracks start_point_list and passes DMD-specific args to scheduler.step() rrDcg)NzDistilled pyramid denoising inner block for DMD checkpoints. Loops over pyramid stages from smallest to full resolution with guidance disabled and DMD scheduler support.rFrGs r=rI3HeliosPyramidDistilledChunkDenoiseInner.description's  a r?c v[S[5[S[5[S[[ SS05SS9/$)Nrrrrg?rrrrGs r=r;HeliosPyramidDistilledChunkDenoiseInner.expected_components.rr?c J[R"S5[S[RS9[S[RS9[R"S5[S/SQ[SS 9[S S [ S S 9[R"S 5[R"S5/$)Nrrrrrr)r r r rrNis_amplify_first_chunkTzYWhether to double the first chunk's timesteps via the scheduler for amplified generation.rr)rrrrZr!rYrGs r=r\.HeliosPyramidDistilledChunkDenoiseInner.inputs;s    * %,, ? /5<< H    7 8 2!J   (w      2 3    ,%  r?rcrdrec URnURRnURn[ UR 5nUS:HnU/n SUR UR40n [[R"URR5RR55n 0n URR5HupX;dM X;dMXU 'M UR R#U5U S'UR$R#U5U S'UR&R#U5U S'UR(U S'[+U5GH0nURR,R.nUR0SUR0S-UR0S -USUS -US --n[3UUR4R,R7S S 5UR4R,R7SS5UR4R,R7SS5UR4R,R7SS55nUR UnUR4R9UUUUUR:=(a US9 UR4R<nUS:GaaUR0unnnnnUS -nUS -nUR?SS S SS5RAUU-UUU5n[BRD"UUU4SS9nURAUUUUU5R?SS S SS5nS UR4RFU- nUR4R,RHnS [JRL"S S U- -5S U- -U-- nUS U- -[JRL"U5- nUR0unnnn n![OUUUU U!UUUURPS9 n"U"R#US9n"UU-UU"--nU RSU5 [ U5UUR4RT-- n#[WUS9n$[YU5GH(un%n&U&R[UR0S5R#[\R^5n'UR#U5n(UR`RcU%UU&S9 UR`ReU 5n)U)Hn*UR`RgUR5 U R5Vs0sHo3[iU*U5_M n+n[iU*UR`Rj5n,URRmU,5 UR"S U(U'SS.U+DU D6SU*l7SSS5 UR`RqUR5 M URaU)5Sn-UR4RsU-U&UURPSU%XUR4RtUR4R<US9 SnU%[ U5S - :Xd/U%S -U#:dGMU%S -UR4RT-S:XdGMU$Rw5 GM+ SSS5 GM3 XblX4$s snf!,(df  GN=f!,(df  GMe=f)!Nrrrqrrrsrrrrrr rrr r!r"r#r$r%)rrrnrrr&rrr'rrFr)rrcur_sampling_stepdmd_noisy_tensor dmd_sigmas dmd_timesteps all_timestepsrF)rrCrrrrrrrrrrrrrrrrrr).rHrcrdrerrrr,is_first_chunkstart_point_listrrrrrr.r3r/rrrr-rtr/r0r1r2r3r4r2r5r6rrr<rrr r rr r r rrrs. r=ry0HeliosPyramidDistilledChunkDenoiseInner.__call__Rs--&2288%% !M!MNa$9 $k&?&?AcAc%d w001G1G1O1OP[[``bc '2'H'H'N'N'P #J-*2Q,7j)(Q 2=1R1R1U1UVg1h -./:/N/N/Q/QRc/d +,0;0P0P0S0STe0f ,-,7,H,H ()+,C#//66AAJ%]]2.r1BBW]]SUEVV1 1 - 1 =M!$$++//0DcJ$$++//0CTJ$$++// cB$$++// TB B#."N"Ns"S   . .#'2'I'I'\n / #,,66IQwU\UbUbR 0*i!A !A !//!Q1a8@@+-A9i--uen9U!//*jBVX]_demmq!Q  4 4 E Ec JJ ",,3399TYYqAI71y=IIUVI .51AAEL]]B 0*a*(!)33 '89'/D5L8 ''0 #9~0CjFZFZF`F`0`` /0L%i0DAq xx a(89<#-#7#7#>#>&0&:&:&D&D&/8  GC NQ..Q"22AAUAUA[A[7[_`7`$++-S110G-^&&&G'hPO10sD CZ1Z 2=Z1/!Z B7Z1 Z1/Z1Z1 Z.)Z11 [ rFNrrFr?r=rhrhs"J  S    T-%8     Z(  , ]]_L'#8L'zL'VYL'L'r?rhcD\rSrSrSr\\\\/r /SQr \ S\ 4Sj5r Srg)HeliosPyramidChunkDenoiseStepizcT2V pyramid chunk denoising: history slice -> pyramid noise gen -> pyramid denoise inner -> update.rZr[r]r^rDcg)NzT2V pyramid chunk denoise step that iterates over temporal chunks. At each chunk: history_slice -> noise_gen (pyramid) -> denoise_inner (pyramid stages) -> update_chunk. Denoising starts at the smallest resolution and progressively upsamples.rFrGs r=rI)HeliosPyramidChunkDenoiseStep.description  W r?rFN)r|r}r~rrrArrr:rarbrrrIrrFr?r=rzrzs;m $&& M RK  S  r?rzcD\rSrSrSr\\\\/r /SQr \ S\ 4Sj5r Srg) HeliosPyramidI2VChunkDenoiseStepizgI2V pyramid chunk denoising: I2V history slice -> pyramid noise gen -> pyramid denoise inner -> update.r{rDcg)NzI2V pyramid chunk denoise step that iterates over temporal chunks. At each chunk: history_slice (I2V) -> noise_gen (pyramid) -> denoise_inner (pyramid stages) -> update_chunk. Denoising starts at the smallest resolution and progressively upsamples.rFrGs r=rI,HeliosPyramidI2VChunkDenoiseStep.descriptionr~r?rFN)r|r}r~rrrrrr:rarbrrrIrrFr?r=rrs;q '&& M RK  S  r?rcD\rSrSrSr\\\\/r /SQr \ S\ 4Sj5r Srg)&HeliosPyramidDistilledChunkDenoiseStepi zDT2V distilled pyramid chunk denoising with DMD scheduler and no CFG.r{rDcg)NzT2V distilled pyramid chunk denoise step with DMD scheduler. At each chunk: history_slice -> noise_gen (pyramid) -> denoise_inner (distilled/DMD) -> update_chunk.rFrGs r=rI2HeliosPyramidDistilledChunkDenoiseStep.descriptions  t r?rFN)r|r}r~rrrArrhr:rarbrrrIrrFr?r=rr s;N $&/ M RK  S  r?rcD\rSrSrSr\\\\/r /SQr \ S\ 4Sj5r Srg))HeliosPyramidDistilledI2VChunkDenoiseStepizDI2V distilled pyramid chunk denoising with DMD scheduler and no CFG.r{rDcg)NzI2V distilled pyramid chunk denoise step with DMD scheduler. At each chunk: history_slice (I2V) -> noise_gen (pyramid) -> denoise_inner (distilled/DMD) -> update_chunk.rFrGs r=rI5HeliosPyramidDistilledI2VChunkDenoiseStep.description(s  z r?rFN)r|r}r~rrrrrhr:rarbrrrIrrFr?r=rrs;N '&/ M RK  S  r?r))rr r NN)6rr*rtorch.nn.functionalnn functionalr tqdm.autorconfiguration_utilsrguidersrrmodelsr schedulersr utilsr utils.torch_utilsr modular_pipeliner rrrmodular_pipeline_utilsrrrbefore_denoiserr get_loggerr|loggerr>rArrrrrrr:rFrXrdrhrzrrrrFr?r=rsJ -M.)- LK+3   H % (`O'"7O'dD'%:D'N'3'2-'%:-'`'$9'Nj'3j'Z'%:'N#'1#'V*!9*!d 3 ( 6 (C'.CC'L $: ( '= ( -C & 0F r?