oMjSSKrSSKrSSKJr SSKJs Jr SSKJrJ r SSK J r SSK J r SSKJrJrJr SSKJr SS KJr SS KJrJr S S KJrJrJr \ R:"\5r"S S\R@5r!"SS\RD5r#"SS\RH5r%"SS\RH5r&"SS\RH5r'"SS\RH5r("SS\RH5r)"SS5r*"SS\RH\5r+"SS \RH5r,"S!S"\RH5r-"S#S$\\\\5r.g)%N) ConfigMixinregister_to_config)logging)apply_forward_hook)AttentionMixinAttentionModuleMixin FeedForward)dispatch_attention_fn)AutoencoderKLOutput) ModelMixinget_parameter_dtype)AutoencoderMixin DecoderOutputDiagonalGaussianDistributionc^\rSrSrSrSS\S\S\\\\\4-S\\\\\4-S\S\S \S S 4U4S jjjrS \ RS \ R4Sjr Sr U=r $)MiniMaxH3VideoCausalConv3d"a 3D convolution used throughout the MiniMax-H3 video encoder. Spatial padding is symmetric and uses `spatial_padding_mode` (`"reflect"` in the released checkpoint); temporal padding is causal, i.e. `kernel_size_t - 1` zero frames are prepended and nothing is appended. in_channels out_channels kernel_sizestridespatial_paddingtemporal_paddingspatial_padding_modereturnNcH>[TU]XX4SS9 XPlX`lXplg)Nr)rrpadding)super__init__rrr) selfrrrrrrr __class__s C:\Users\johnn\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.13_qbz5n2kfra8p0\LocalCache\local-packages\Python313\site-packages\diffusers/models/autoencoders/autoencoder_kl_minimax_h3.pyr"#MiniMaxH3VideoCausalConv3d.__init__*s,  def. 0$8! hidden_statesc nURS:a0URn[R"XX"USS4URS9nURS:a&[R"USSSSURS4SS9n[R "XR URURSURS9$)Nrmodeconstant)rr dilation) rFpadrrconv3dweightbiasrr-)r#r(r s r%forward"MiniMaxH3VideoCausalConv3d.forward9s   ! #**GEE7AqIPTPiPiM  1 $EE-!Q1d>S>SUV1W^hiMxx {{DIIdkk[\gkgtgtuur')rrr)rrrreflect)__name__ __module__ __qualname____firstlineno____doc__inttuplestrr"torchTensorr3__static_attributes__ __classcell__r$s@r%rr"s./ !$- 9 9 95c3// 9 eCcM** 9  9 9" 9  9 9vU\\vellvvr'rc`^\rSrSrSrS\R S\R 4U4SjjrSrU=r $)MiniMaxH3VideoGroupNormDz Group normalization applied to each latent frame in isolation (`use_t_isolated_gn` in the original config): the temporal axis is folded into the batch axis so statistics never mix across frames. r(rc >URup#pEnURSSSSS5R5nURX$-USXV5n[TU]U5nURX$X5U5nURSSSSS5R5$)Nrrrr)shapepermute contiguousviewr!r3)r#r( batch_size num_channels num_framesheightwidthr$s r%r3MiniMaxH3VideoGroupNorm.forwardJs>K>Q>Q; *e%--aAq!<GGI %**:+BLRSU[c  6 %**:>@@r') r6r7r8r9r:r>r?r3r@rArBs@r%rDrDDs, AU\\AellAAr'rDc ^\rSrSrS S\S\S\S\S\SS4 U4S jjjrS \RS\R4S jr S r U=r $)MiniMaxH3VideoResnetBlock3dSrrnorm_num_groupsnorm_epsrrNc >[TU]5 XlX l[ X1USS9Ul[ UUSSSUS9Ul[ X2USS9Ul[ UUSSSUS9Ul SUl X:wa[ XSS9Ul gg)NTepsaffinerrrrrrrr) r!r"rrrDnorm1rconv1norm2conv2 conv_shortcut)r#rrrVrWrr$s r%r"$MiniMaxH3VideoResnetBlock3d.__init__Ts &(,_x`de /  !5   -_PXaef /  !5   "  &!;Kcd!eD  'r'r(c"Un[R"URU55nURU5n[R"UR U55nUR U5nUR bUR U5nX!-$N)r.silur^r_r`rarb)r#r(residuals r%r3#MiniMaxH3VideoResnetBlock3d.forwardvsv tzz-89  =1 tzz-89  =1    )))(3H''r')r_rarbrr^r`r ư>r5 r6r7r8r9r;floatr=r"r>r?r3r@rArBs@r%rTrTSs{ "$- f f f f  f " f  f fD(U\\(ell((r'rTc ^\rSrSrSrSS\S\S\S\S\SS 4 U4S jjjrS \RS\R4S jr S r U=r $)MiniMaxH3VideoDownsample3dz Strided 3x3x3 downsampling convolution. A spatial stride of 2 is preceded by an asymmetric bottom/right pad of 1 (the convolution itself carries no spatial padding), so the output is exactly `ceil(size / 2)`. rrtemporal_stridespatial_striderrNc f>[TU]5 X@lXPl[ UUSX4U4SSUS9Ulg)Nrrr)rrrrr)r!r"rrrrconv)r#rrrqrrrr$s r%r"#MiniMaxH3VideoDownsample3d.__init__sB ,$8!.  #^D!5  r'r(cURS:Xa [R"USURS9nUR U5$)Nr)rrrrrrr*)rrr.r/rrt)r#r(s r%r3"MiniMaxH3VideoDownsample3d.forwards8   ! #EE-1C$JcJcdMyy''r')rtrrr)rrr5) r6r7r8r9r:r;r=r"r>r?r3r@rArBs@r%rorosw !$-        "     *(U\\(ell((r'roc^\rSrSrSS\S\S\S\S\S\S\S \S S 4U4S jjjrS \RS \R4Sjr Sr U=r $)MiniMaxH3VideoDownBlock3drr num_layerstemporal_downsample_factorspatial_downsample_factorrVrWrrNc 6>[T U]5 [R"[ U5V s/sHn [ U S:XaUOUUUUUS9PM sn 5UlSUlXE-S:a'[R"[UUUUUS9/5UlSUl gs sn f)Nr)rrrVrWrr)rqrrrF) r!r"nn ModuleListrangerTresnets downsamplersrogradient_checkpointing) r#rrr{r|r}rVrWrir$s r%r""MiniMaxH3VideoDownBlock3d.__init__s }}z* +A,/0Av is_grad_enabledr_gradient_checkpointing_funcr)r#r(resnet downsamplers r%r3!MiniMaxH3VideoDownBlock3d.forwardsollF$$&&4+F+F $ A A& X &} 5 #    (#00 +M :  1r')rrrrirlrBs@r%ryrys "$-&,&,&, &, %( &, $' &,&,&,"&, &,&,P U\\ ell  r'ryc^\rSrSrSrSS\S\S\\S4S\S\\S4S \\S4S \S \S \S S4U4Sjjjr S\ RS \ R4Sjr Sr U=r$)MiniMaxH3VideoEncoder3dz Causal 3D CNN encoder. `block_out_channels` gives the channel count of every level; the per-level `spatial_downsample_factors` / `temporal_downsample_factors` multiply out to the total compression ratios. rrblock_out_channels.layers_per_blockspatial_downsample_factorstemporal_downsample_factorsrVrWrrNc |>[T U]5 [UUSSSSU S9UlUS4[ USS5-n [ R "[[U55V s/sHn [XX;UXkX[UUU S9PM sn 5Ul [XsSUSS 9Ul [USUSSSU S9Ul gs sn f) Nrrrrr\)rrr{r|r}rVrWrTrY)r!r"rconv_inr<rrrlenry down_blocksrDnorm_outconv_out) r#rrrrrrrVrWrblock_in_channelsrr$s r%r" MiniMaxH3VideoEncoder3d.__init__s 1  q !!5   024u=OPSQS=T7UU==s#567 8A* 1 4!3!6//J/M.H.K$3%)= 8   0TVAW]enrs 2 r " !5   ! s!"B9r(cURU5nURH nU"U5nM [R"UR U55nUR U5$re)rrr.rfrr)r#r( down_blocks r%r3MiniMaxH3VideoEncoder3d.forwardsP ]3 **J&}5M+t}}];< }}]++r')rrrr) r0rrirrrrrrrrrrrrrrjrkr5)r6r7r8r9r:r;r<rmr=r"r>r?r3r@rArBs@r%rrs.M !6H7I!$-0 0 0 "#s(O 0  0 %*#s(O 0 &+38_0 0 0 "0  0 0 d,U\\,ell,,r'rc ^\rSrSrSrS S\S\S\SS4U4SjjjrS \RS\ \R\R44S jr S r U=r $) MiniMaxH3VideoRotaryPosEmbedia  3-axis rotary embedding for the ViT decoder. Coordinates are length-normalized to `[-1, 1)` per axis and scaled by `2 * pi`, and the resulting `(t, h, w)` angles are concatenated and then duplicated, so the first `rope_dim_ratio * attention_head_dim` channels of every head are rotated. dimthetanum_axesrNc >[TU]5 USU--S:wa[SUSSU-S35eSU[R"SSSU-U- [R S9-- nUR S US S 9 g) Nrrz`dim` z% must be divisible by `2 * num_axes` .?r)dtypeinv_freqF) persistent)r!r" ValueErrorr>arangefloat32register_buffer)r#rrrrr$s r%r"%MiniMaxH3VideoRotaryPosEmbed.__init__s  !h, 1 $vcU*OPQT\P\~]^_` `%,,q!Q\C5Gu}}"]]] ZeDr' position_idscS[R-USS2SS2SS2S4-URSSSSS24-nURSS5R S5R S5nUR 5UR54$)N@rr)mathpirflattentile unsqueezecossin)r#rangless r%r3$MiniMaxH3VideoRotaryPosEmbed.forward%sxtwwaAtm!<r?r<r3r@rArBs@r%rrsb ECEEETEE*ELL*U5<<;U5V**r'rc \rSrSrSrSrS SSS\RS\\R\R4S-S\R4Sjjr S r g) MiniMaxH3VideoAttnProcessori+NattnMiniMaxH3VideoAttentionr( rotary_embrc URU5RSURS45nURU5RSURS45nUR U5RSURS45nUR UR 55RUR5nURUR 55RUR5nUbUupxURUR5nURUR5nURSn USSU 24USU S24pUSSU 24USU S24pU RSSS9upU RSSS9unn[R"U*U/SS9n[R"U*U/SS9n[R"X-UU--U /SS9n[R"X-UU--U /SS9n[UUUSURUR S9nUR#SS5nUR$S"U5$)Nrr.r) attn_maskbackendparallel_configrr)to_q unflattenheadsto_kto_vnorm_qrmtornorm_krHchunkr>catr _attention_backend_parallel_configrto_out)r#rr(rquerykeyvaluerr rotary_dim query_rotary query_pass key_rotarykey_pass query_first query_second key_first key_second query_rotated key_rotateds r%__call__$MiniMaxH3VideoAttnProcessor.__call__/s-  -(221tzz26FGii &00TZZ4DE -(221tzz26FG EKKM*--ekk:kk#))+&))#))4  !!HC&&%C&&%C2J',S+:+-='>c:;FV@W*#&sKZK'7#8#c:;>N:O(4(:(:1"(:(E %K$.$4$4QB$4$? !Iz!II }k&BKM))j[)$<"EKII|1MC4GGTZ\]E))Z- c0AA8LRTUC-   ++ 11  &--a3 {{1~m,,r'rRre) r6r7r8r9rrr>r?r<rr@rRr'r%rr+sd @D %-'%-||%-%,, 45< %-  %-%-r'rc ^\rSrSr\r\/rSS\S\S\S\S\ SS4 U4S jjjr SS \ RS \ \ R\ R4S-S\ R4S jjrS rU=r$)riWrrdim_headrZr2rNc >[TU]5 X lX0lXPlX#-n[ R "X4SS9Ul[ R "X4SS9Ul[ R"XUS9Ul [ R"XUS9Ul [ R"XUS9Ul [ R"[ R"XaUS9[ R"S5/5UlUR![#55 g)NFrZelementwise_affine)r2)r!r"rruse_biasrRMSNormrrLinearrrrrDropoutr set_processorr)r#rrrrZr2 inner_dimr$s r%r" MiniMaxH3VideoAttention.__init__[s    $ jjuM jjuM IIc48 IIc48 IIc48 mmRYYyD%I2::VY?$[\  689r'r(rc&URXU5$re) processor)r#r(rs r%r3MiniMaxH3VideoAttention.forwardks~~d:>>r') rrrrrrrrr)h㈵>Tre)r6r7r8r9r_default_processor_cls_available_processorsr;rmboolr"r>r?r<r3r@rArBs@r%rrWs889:C::s::UY:ei::"cg?"\\?7[TU]5 [R"XSS9Ul[ XX5US9Ul[R"[R"U55Ul [R"XSS9Ul [XSUS9Ul [R"[R"U55Ulg)NTr)rrrrZr2swiglu)mult activation_fnr2)r!r"rrr^rr Parameterr>zerosscale1r`r ffscale2)r#rrrrrZr2r$s r%r"'MiniMaxH3VideoTransformerBlock.__init__rs ZZF +8cgh ll5;;s#34 ZZF ctTll5;;s#34 r'r(rcfURUR55RUR5nXR X25UR --nUR UR55RUR5nXRU5UR--nU$re) r^rmrrrrr`r r )r#r(rnorm_hidden_statess r%r3&MiniMaxH3VideoTransformerBlock.forwards"ZZ (;(;(=>AA-BUBUV% 2D(QTXT_T_(__ !ZZ (;(;(=>AA-BUBUV%0B(Cdkk(QQ r')rr r^r`rr )rGrTre)r6r7r8r9r;rmrr"r>r?r<r3r@rArBs@r%rrqs 5 55 5  5  55 55$cg"\\7[TU]5 Xg-n X0lX@lX lXl[ [X{-5U S9Ul[R"X5Ul [R"[R"SX55Ul[R "[#U5Vs/sHn[%U UUU U S9PM sn5Ul[R("U SU S9Ul[R"XU-U-U-5UlSUlgs snf)N)rr)rrrrrZT)rrZF)r!r"rrrrrr;roperrproj_inrr>rregister_tokensrrrtransformer_blocks LayerNormrproj_outr)r#rrrrr{rrrrrrrWr_r$s r%r"#MiniMaxH3VideoViTDecoder3d.__init__s !6$((#6 05G5X1Yakl yy2 !||EKK;N,TU"$--z* +A/-/% +  #  STxP  #l'BZ'OR\'\] &+# s.Dr(c jURup#pEnURSSSSS5RX$U-U-U5nURU5nURSnURR USS5n[ R"USS2SS2SS245n [ R"XU /SS9nXEU4V s/sH;n S[ R"S U [ RURS 9U - -S - PM= n n [ R"[ R"U S S 06SS9RSS5n U RS5R USS5n U R!X R"S-S45n [ R"X/SS9n UR%U 5nUR&HJn[ R("5(a%UR*(aUR-XU5nMBU"X5nML UR/U5nUR1U5nUSS2SU2SS24nUR2UR4nnUR7UUUUUR8UUU5nURSSSSSSSS5R;5nURUUR8UU-UU-UU-5$s sn f)NrrrrGrrrrg?)rdevicerindexingij)rHrIreshaperrexpandr> zeros_likerrrr"stackmeshgridrr new_zerosrrrrrrrrrrrKrrJ)r#r(rLrMrNrOrP num_patchesr cls_tokensizegridsr suffix_idsrblockrrs r%r3"MiniMaxH3VideoViTDecoder3d.forwards>K>Q>Q; *e%--aAq!<DD V+e3\  ]3 #))!, ..55j"bI$$]1bqb!8%<=  =9"MSTU $U3 3 5<<T}G[G[\_cc dgj j3  {{5>>5#H4#HbQYYZ[]^_ #--a077 BK !++Z9Q9QTU9UWX,YZ yy,!;C YY|, ,,E$$&&4+F+F $ A A%Xb c %m @ -  m4  m4 %a+q&89 #'??D4E4EL %**            &--aAq!Q1EPPR $$      % Z  J    A s8AJ0) rrrrrrrrrrr) rrG$rj@rGrGr?r) r6r7r8r9r:r;rmr"r>r?r3r@rArBs@r%rrs#%"$#$! $(,(,(, (,  (,  (,!(, (,!(,(,(,(,(, (,(,T3 U\\3 ell3 3 r'rc0T^\rSrSrSrSrSS/rS/rS/r/SQr \ SBS\ S \ S \ S \ \ S 4S \ S\ \ S 4S\ \ S 4S\ S\ S\S\ S\ S\ S\ S\ S\ S\ S\ S\ S\ S\ \ S 4S\ \ S 4SS 4.U4S!jjj5rSCS"\ S -S#\ S -S$\ S -S%\ S -SS 4 S&jjrS'\ S(\ S)\ S\ \\ \\ \\ 44S*jrS+\R(S,\R(S-\ S.\ S\R(4 S/jrS0\\\R(S1\\ S2\\ S\R(4S3jrS4\R(S\R(4S5jrS6\R(S\R(4S7jrS4\R(S\R(4S8jrS6\R(S\R(4S9jr\SDS4\R(S:\S\\ \R(-4S;jj5r\SDS6\R(S:\S\\ \R(-4S<jj5r SES=\R(S>\S?\RBS -S:\S\\ \R(-4 S@jjr"SAr#U=r$$)FAutoencoderKLMiniMaxH3ia A VAE model with a causal 3D CNN encoder and a non-causal ViT decoder, used in [MiniMax-H3](https://huggingface.co/MiniMaxAI). This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented for all models (such as downloading or saving). Latents are normalized with per-channel `latents_mean` / `latents_std` rather than a `scaling_factor`; a pipeline encodes with `(latent - latents_mean) / latents_std` and decodes with `latent * latents_std + latents_mean`. The pixel convention is ImageNet-normalized RGB over a `[0, 1]` base range, not the usual `[-1, 1]`: `encode` expects `(pixel - imagenet_mean) / imagenet_std` and `decode` returns values in that same space, so a pipeline has to apply `sample * imagenet_std + imagenet_mean` (mean `(0.485, 0.456, 0.406)`, std `(0.229, 0.224, 0.225)`) and clamp to `[0, 1]` before postprocessing. The temporal geometry is fixed by `clip_length` (17 pixel frames per encoder chunk) and `token_drop` (3 trailing latent frames dropped per encode): `17 * n + 5` pixel frames map to `5 * n + 2` latent frames. Unlike most autoencoders in the library, spatial tiling is **on by default**: MiniMax-H3 was released with tiling enabled for both encoding and decoding, and the released frames are the blended-tile ones, so disabling tiling changes the output. Use `enable_tiling` to change the tile geometry, `disable_tiling` to turn it off. TrTrnorm)encoderdecoder quant_convpost_quant_convrrlatent_channelsr.rrrrVrWrdecoder_num_layersdecoder_num_attention_headsdecoder_attention_head_dimdecoder_num_register_tokensdecoder_ffn_multdecoder_rope_thetadecoder_rope_dim_ratiodecoder_norm_eps clip_length token_drop latents_mean latents_stdrNc>[TU]5 [R"U5Ul[R"U5Ul[ USU-UUUUUU U S9 Ul[R"SU-SU-SS9Ul [R"X3SS9Ul [UUURUR U U U UUUUUS9 Ul U*UR -Ul[R"UUR - 5UlU*UR -Ul[%UR"UR -UR- S5UlSUlSUlS UlS UlS UlS Ulg) Nr) rrrrrrrVrWrrr]) rrrrr{rrrrrrrWrFTrr8)r!r"rprodspatial_compression_ratiotemporal_compression_ratiorr=rConv3dr?r@rr>frame_pre_paddingceiltokens_chunk_size token_overlapmax frame_overlap use_slicing use_tilingtile_sample_min_heighttile_sample_min_widthtile_sample_min_overlap_heighttile_sample_min_overlap_width)r#rrrArrrrrVrWrrBrCrDrErFrGrHrIrJrKrLrMr$s r%r"AutoencoderKLMiniMaxH3.__init__s{4 )-3M)N&*.))4O*P'.#_,1-'A(C+!5   ))A$7_9LZ[\!yyWXY1'%5588) ;9 ;%)1%   $$/,$2Q2Q!Q!%;9X9X+X!Y)kT-C-CC !3!3d6U6U!UX\XnXn!npqr! '*#%(".0+-/*r'r[r\r]r^cSUlU=(d URUlU=(d URUlU=(d URUlU=(d URUlg)aH Enable tiled VAE encoding/decoding. When this option is enabled, the VAE splits the frames into tiles, encodes or decodes each tile separately and linearly blends the overlaps back together. This lowers the memory requirement and allows processing larger frames. Args: tile_sample_min_height (`int`, *optional*): The tile height in pixel space. Frames taller than this are split along the height dimension. tile_sample_min_width (`int`, *optional*): The tile width in pixel space. Frames wider than this are split along the width dimension. tile_sample_min_overlap_height (`int`, *optional*): The minimum overlap, in pixels, between two consecutive vertical tiles. tile_sample_min_overlap_width (`int`, *optional*): The minimum overlap, in pixels, between two consecutive horizontal tiles. TN)rZr[r\r]r^)r#r[r\r]r^s r% enable_tiling$AutoencoderKLMiniMaxH3.enable_tilingisX,&<&[@[@[#%:%Xd>X>X".L.sPTPsPs+-J-pdNpNp*r'length tile_size min_overlapcX!:aS/U//4$[R"X- 5nX$-X4S- -- U- S:aUS- nX$-X4S- -- U- S:aMU/US- -nX$-[U5- U- n[X`R-5HnXWUS- -==UR- ss'M! S/n[US- 5HnUR USU-XW- 5 M! X/U-U4$)aF Lay `tile_size`-wide tiles over `length` pixels. The number of tiles is the smallest one whose union can cover `length` while keeping every overlap at least `min_overlap`; the slack is then distributed round-robin over the overlaps in whole `spatial_compression_ratio` steps so that every tile boundary stays latent-aligned. rrr)rrTsumrrPappend) r#rcrdre num_tilesoverlaps remainingrtile_start_indicess r% _split_tiles#AutoencoderKLMiniMaxH3._split_tiless  3"$ $IIf01 #k]&CCfLqP NI#k]&CCfLqP =IM2)CM9FB y$B$BBCA )a-( )T-K-K K )D Sy1}%A  % %&8&rr"rndimrKslicer<r) r#rorprqr positionsrHweight_aweight_bslice_aslice_bblended slice_rests r%_blendAutoencoderKLMiniMaxH3._blends81773<|D LLhhaggN aff !c  0066u=,2259;-!&&(l]D1 ;-!&&(Q - E'N#h.5>1BX1MM 773< 'NDk]QVV+  d3 yy'U:%6#78cBBr'tilesheight_overlapswidth_overlapsc /n[U5HupV/n[U5HupUS:aURXS- UXUS- SS9n US:aURXhS- XUS- SS9n U[U5S- :aU SSX%*2SS24n U[U5S- :aU SSS2SX8*24n URU 5 M UR[R "USS95 M [R "USS9$)Nrrrr.) enumerater}rrhr>r) r#rrr result_rowsrrow result_rowjrs r% _stitch_tiles$AutoencoderKLMiniMaxH3._stitch_tiless   &FAJ$S>q5;;uU|AaRSe>TZ\;]Dq5;;sq5z4A9NTV;WDs5zA~%%:(:':%:A =>Ds3x!|#Q(<>+<*<(< <=D!!$'*   uyy< ='yy"--r'xc UR(d URURU55$URSURSp2UR X R UR 5upEnUR X0RUR5upxn /n [XE5Hkup/n [Xx5HDupUSXU -2XU-24nU RURURU555 MF U RU 5 Mm UVs/sHnUUR-PM nnU Vs/sHnUUR-PM nnURU UU5$s snfs snf)zAEncode one temporal clip, spatially tiled when tiling is enabled.rr.) rZr?r=rHrmr[r]r\r^ziprhrPr)r#rrOrP y_indices y_lengths y_overlaps x_indices x_lengths x_overlapsrowsi_posi_lenrj_posj_lenroverlaplatent_y_overlapslatent_x_overlapss r% _encode_clip#AutoencoderKLMiniMaxH3._encode_clipsf??4<<?3 3 QWWR[+/+<+< //1T1T, ( j,0+<+< --t/Q/Q, ( j 5LEC #I 9 eem3UU]5JJK 4??4<<+=>?!: KK  6WaaV`7W(F(FFV`aV`aV`7W(F(FFV`a!!$(9;LMMbas !E.E3zc UR(d URURU55$URSUR-nURSUR-nUR X R UR5upEnUR X0RUR5upxn URn /n [XE5H{up/n[Xx5HTunnUSX-X-X--2X-X-UU --24nURURURU555 MV U RU5 M} URXU 5$)zADecode one temporal clip, spatially tiled when tiling is enabled.rr.) rZr>r@rHrPrmr[r]r\r^rrhr)r#rrOrPrrrrrrratiorrrrrrrs r% _decode_clip#AutoencoderKLMiniMaxH3._decode_clips^<< 4 4Q 78 8t=== d<<<+/+<+< //1T1T, ( j,0+<+< --t/Q/Q, ( j.. 5LEC #I 9 uNU^en%DDNU^eun%DDF  4<<(<(rrrK)r#rrJrN pad_framesrmomentss r%_encodeAutoencoderKLMiniMaxH3._encodes-kk-- WWQZ ?$$Q' '  #q (1a9,,QZK;3NPQSTUJ 1/q1A))qwwqz[89 9A!!!Aq!/QUkG s-/D!c ^^^^URmURRnURmTT-nURSU-nU*T-nXE-T-[ US:5- nUS:a5[ R"XSS2SS2SS24RSSUSS5/SS9n/nSn[U5Hn U T-n URUSS2SS2XT-UR-245n [[ US:5S-5Hhn X-n U SS2SS2XU-24nUSS2SS2URS24nU S:Xa1UbURXURSS9nURU5 MfUnMj M UbURU5 [ R"USS9nUS:a[URR T-mURSU- m[#UUUU4Sj[U555nUSS2SS2SU*24nU$) an Decode a latent video, mirroring the chunking that `_encode` applied. `token_drop` removed the tail of every encoded chunk, so consecutive decoded chunks overlap by `frame_overlap` pixel frames and are linearly cross-faded. Latent frames are repeated at the end when the length is not a whole number of chunks; the extra pixel frames are cut off again at the end. rrNrrrc3R># UHnT(aTU-T-S:XaTOTv M g7f)rNrR).0k intra_tailnum_tokens_before_padtemporal_ratiorUs r% 1AutoencoderKLMiniMaxH3._decode..Es4*A).Ca.GK\-\`a-a guu*s$')rUrrKrQrHr;r>rrrrrVrSr}rXrhrJrg)r#rrKchunk_num_frames num_tokens pad_tokens num_chunksdecoded_chunksrrstartclipr frame_startrdecrrrrrUs @@@@r%_decodeAutoencoderKLMiniMaxH3._decodes%!22[[++ 88,~=WWQZ*, !k%66  -2CCc*WX.FYY > 11bc l11!Q AqIJPQRAz"A))E$$Qq!U=N5NQUQcQc5c-c'c%deD3zA~.232 Q;?O1O#OOPaD$:$:$<<=6* $ GD00>AJ$%GGAJ$; !z*JaLj[L()C r' return_dictcUR[UR55nUR(aXURSS:aE[ R "URS5Vs/sHo0RU5PM sn5nOURU5n[U5nU(dU4$[US9$s snf)a Encode a batch of videos into latents. Args: x (`torch.Tensor`): Input batch of videos, shape `(batch_size, in_channels, num_frames, height, width)`. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.autoencoders.autoencoder_kl.AutoencoderKLOutput`] instead of a plain tuple. Returns: The latent distribution of the encoded videos. Note that MiniMax-H3 normalizes the sampled latents with `latents_mean` / `latents_std` afterwards. rr) latent_dist) rrr=rYrHr>rsplitrrr )r#rrx_slicer posteriors r%encodeAutoencoderKLMiniMaxH3.encodeLs$ DD$T\\2 3    Qiiaggaj Qj7g!6j QRGll1oG09 < "y99 !Rs,B=cnUR[UR55nUR(aXURSS:aE[ R "URS5Vs/sHo0RU5PM sn5nOURU5nU(dU4$[US9$s snf)a Decode a batch of latent videos. Args: z (`torch.Tensor`): Input batch of latent videos, shape `(batch_size, latent_channels, num_latent_frames, height, width)`. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.autoencoders.vae.DecoderOutput`] instead of a plain tuple. Returns: [`~models.autoencoders.vae.DecoderOutput`] or `tuple`: The decoded videos, shape `(batch_size, out_channels, num_frames, height, width)`. rr)sample) rrr>rYrHr>rrrr)r#rrz_slicedecodeds r%decodeAutoencoderKLMiniMaxH3.decodehs DD$T\\2 3    Qiiaggaj Qj7g!6j QRGll1oG: G,, !Rs,B2rsample_posterior generatorcURU5RnU(aURUS9OUR5nUR XdS9$)a Encode then decode a batch of videos. Args: sample (`torch.Tensor`): Input batch of videos, shape `(batch_size, in_channels, num_frames, height, width)`. sample_posterior (`bool`, *optional*, defaults to `False`): Whether to sample the posterior instead of taking its mode. generator (`torch.Generator`, *optional*): Generator used when `sample_posterior=True`. return_dict (`bool`, *optional*, defaults to `True`): Whether to return a [`~models.autoencoders.vae.DecoderOutput`] instead of a plain tuple. Returns: [`~models.autoencoders.vae.DecoderOutput`] or `tuple`: The round-tripped videos, shape `(batch_size, out_channels, num_frames, height, width)`. )r)r)rrrr+r)r#rrrrrrs r%r3AutoencoderKLMiniMaxH3.forwardsG0KK'33 5EI  y  19>>K[{{1{66r')r>r=rXrSr@r?rPrQr[r]r^r\rVrUrYrZ)rrr5rrrrrjrkr5r7rjr8rGrGrr9rr)rrrrrrrrrrrrrrrrrrrrrrrr)rrrrrrrrrrrrrrrrrrrrrrrr)NNNN)T)FNT)%r6r7r8r9r: _supports_gradient_checkpointing_no_split_modules_repeated_blocks _skip_layerwise_casting_patterns_keep_in_fp32_modulesrr;r<rmr=r"ralistrmr>r?r}rrrrrrrr rrr Generatorr3r@rArBs@r%r;r;s(.(,$68XY89(.x$T!.M !6H7I!$-"$+-*,+, !$)(,"&*5)4/P0P0P0 P0 "#s(O P0  P0%*#s(OP0&+38_P0P0P0"P0 P0&)P0%(P0&)P0 !P0""#P0$!&%P0& 'P0()P0*+P0,E3J'-P0.5#:&/P00 1P0P0h.2,05948 q #d q #Tzq),d q (+Tz q  q8E3E3ESEUSWX[S\^bcf^gimnqirSrMsE.C CCSCsCW\WcWcC(.D&'.c.S .  .,NellNu||N2@ell@u||@<%,,811%,,1f: :4:CVY^_d_k_kYlCl::6- -4-=SXY^YeYeSfCf--4"',0 7 77??T) 7  7 u||, , 77r'r;)/rr>torch.nnrtorch.nn.functional functionalr.configuration_utilsrrutilsrutils.accelerate_utilsr attentionr r r attention_dispatchr modeling_outputsr modeling_utilsrrvaerrr get_loggerr6loggerrRr GroupNormrDModulerTroryrrrrrrr;rRr'r%rs B8II62<NN   H %vvD Abll A+("))+(\((B2 2j=,bii=,@*299*()-)-X?bii)=?4RYY:d d Ne7ZnFVe7r'