ϑieV SSKrSSKrSSKJr SSK7 SSKJrJr Sr\"S5S5r \"S 5S 5r \"S 5S 5r \"S 5S5r \"S5S5r \"S5S5r\"S5S5r\"S5S5r\"S5S5r\"S5S5r\"S5S5rS:Sjr\"S5S5r\"S 5S!5r\"S"5S#5r\"S$5S%5r\"S&5S'5r\"S(5S;S)j5r\"S*5S+5r\"S,5S-5r\"S.5S/5r\"S05S15r\"S25S35r\"S45S55r \"S65S75r!\"S85S95r"g)<N)core)*)REGISTER_COMPOSITElookup_compositec>[UR5nU"U/UQ76$N)rtype)opargs _lowerrules h/var/www/html/banglarbhumi/venv/lib/python3.13/site-packages/paddle/incubate/autograd/composite_rules.py _compositers !"''*J b 4  softmaxcRSnSSKJn U"UR5nUS;aSn[US5nUR(d"[ X- 5nU(a [US5nU$[ XSS9nSUl[ X- 5n[XqSS 9n[Xx5nU(a [XT5nU$) z#define composite rule of op softmaxFr convert_dtypefloat16uint16Tfloat32r)keepdimaxisr) paddle.base.data_feederrdtypecastshapeexpmax stop_gradientsumdivide) xris_amprrresmax_temp molecular denominators rsoftmax_compositer+"sF5 !'' "E %% I  77!%j sI&C 1D)H!HAL!IiD9K  (C 3 Jr batch_normc (^^Sn SSKJn U "UR5n U S;a8Sn [US5nU(a [US5OUnU(a [US5OUnUS;aSO[ UR 5S- mU=(a U (+=(d U n[ U4S j[[ UR 5555m[ U4S j[UR 555n[S/S UR5nU(dw[UT5n[X-T5nUUU-- n[UU-U5nUS :Xa UU- U-nOU[UU5- [UU5-nXa-SU- U--nXb-SU- U--nO[UR UR5n[UR UR5n[UU-U5nUS :XaX- [X'-U5-nO'U[X5- [[X/5U-U5-nUS :Xa UU-U-nO[X?5U-[XO5-nU (a [UU 5n[U5n[U5n[U5n[U5nS nU(dUUUUUU4$UUUS S U4$)z~ define composite rule of op batch_norm As the same with op kernel, the position of saved variance indeed return inverse std. FrrrTr)NCNCLNCHWNCHWDrc36># UHoT:wdM Uv M g7fr ).0i feature_axiss r &composite_batchnorm.._sL#6a|:K#6s  c3<># UHupUT;aSOUv M g7f)rNr3)r4r5s reduce_axess rr7r8`s#1CQ+ 1$1CsgNHWCN)rrrrlenrtuplerange enumeratefullmeanpowreshapezerosassign)r%run_meanrun_varscalebiasis_testmomentumepsilon data_layoutuse_global_statstrainable_statisticsr&rr use_run_stat stats_shapehalf batch_meantemp batch_varinv_stdx_haty batch_mean_inv_std_ run_mean_run_var_ reserve_spacer6r;s @@rcomposite_batchnormr_=s&F5 !'' "E %% I */UI&U(,tD)$$: :AGG q@P:%9!9O?OLL5QWW#6LLK1:1771CK T177 #D ![) AE;': 22 y7*T2 & ^w.E[99W>E&!h,*)DD$H 'AA8>>8>>: '--7 y7*T2 & \S'*;d%CCE773.84<Ef EMD  E '% /'$2L L EN$KgHx IgHM )X{HmKK)XtT=@@r layer_normcSnSSKJn U"UR5nUS;a8Sn[US5nU(a [US5OUnU(a [US5OUn[ [ U[ UR555n[XSS9n X - n X-n [XSS9n X-n [U 5nX-nUb9URUSUR:wa[XRUS5nX-nUb9URUSUR:wa[X RUS5nX-n[XRSU5n [XRSU5n U(a [X5nXU 4$) zm define composite rule of op layer_norm out = (x - mean(x)) / sqrt(var + epsilon)) var = mean((x-mean(x))^2) FrrrTrrN) rrrrr>r?r=rrBrsqrtrD)r%rIrJrMbegin_norm_axisr&rrrmean_ differencevar_tmp1variancevar_tmp3 rsqrt_varouts rlayernorm_compositerksiF5 !'' "E %% I */UI&U(,tD)$$ AGG 5 6D t ,EJ&HH6H!HhI  C  77?# $ 3E77?+;#<=Ek  77?# $ 24)9!:;Dj E77#3O4 5Ex)9/!:;H 3 x r instance_normc xSnSSKJn U"UR5nUS;a8Sn[US5nU(a [US5OUnU(a [US5OUnURupxp[ [ S[UR555n [X SS9n X - n X-n[XSS9nX-n[U[S /S URS 95nU U- nUb[US US S /5nUU-nUb[US US S /5nUU-n[U S /5n S U- n[US /5nU(a [UU5nUU U4$) zp define composite rule of op instance_norm out = (x - mean(x)) / sqrt(var + epsilon)) var = mean((x-mean(x))^2) FrrrTrrr?r) rrrrrr>r?r=rBrCrArD)r%rIrJrMr&rrnchwrrdrerfrgrhsqrt_varrj scale_tile bias_tilesaved_variances rinstancenorm_compositerzsVF5 !'' "E %% I */UI&U(,tD)$$JA! q#agg,' (D t ,EJ&HH6H!H8T1#s(..ABH x C UQ1aL1 J D1aA,/ Io EB4 E\N^bT2N 3 ~ %%rgeluc SnSn[UR5S:Xa UROS/n[X@R5n[ USUR5nU(aS[ XCU-UR5n[ USUR5n[ XpX-U-U---5n X-XY--n U $Xe[ U[ URX R5-5--n X -n U $)z define composite rule of op gelug;f?gmBP ?rrrogHm?)r=ronesrrAtanherf) r% approximate M_SQRT1_2 M_2_SQRTPI full_shapeonerSkAlpha GELU_CONSTANTtanh_outrjcdfs rgelu_compositers (JLA-A3J z77 #C  C )Djy"8!''BZ177; m&7!&;a&?"?@Ah#.) CD)WW$E EFFGg r reduce_meancSnSSKJn U"UR5nUS;aSn[US5nUS/4;a([ [ S[ UR555n[U[5(aU4OUn[XUS9nUVs/sHoRUPM nnU/:XaS n O%[R"[RU5n [/U URS 9n [!Xz5n U(a [X5n U $s snf) z define composite rule of op meanFrrrTrNrr)rvaluer)rrrrr>r?r=r isinstanceintr# functoolsreduceoperatormul fill_constantr$) r%rrr&rraxessum_x ele_nums_list value_to_fillnormr's rmean_compositersF5 !'' "E %% I  bzU1c!''l+, s++D7D g .E/34ttWWT]tM4 !((}E kk D  C 3 J5sD expand_v2cURn[U5n[U5nXC::aUS:de/n[U5HZnX6- nXG- nUS:aX(OSn Xn U S:XdU S:Xa US:deSn OX-S:Xde[X- 5n UR U 5 M\ XC:aL/n [X4- 5HnU R S5 M U R U5 [ X 5n [XS9$[XS9$)z define composite rule of op expand_v2, expand_v2->expand repeat_times = shape / x.shape out = tile(x, repeat_times = repeat_times) rrrq repeat_timesrr=r?rappendextendrDtile)r%rshape_indim_outdim_inrr5offsetdimsize_insize_outrepeatshape_in_expand x_reshapes rexpand_v2_compositer"s wwH%jG ]F  A- -L 7^o#&!8(-8 r>W\!8O8F%* **+,FF#w'(A  " "1 %)x(A/ I99  --r expand_as_v2cURnUb URnUce[U5n[U5nXT::aUS:de/n[U5HTnXG- nXX- n U S:aX9OSn X'n U S:Xa U S:deSn OX-S:Xde[X- 5n UR U 5 MV XT:aL/n [XE- 5HnU R S5 M U R U5 [ X 5n[XS9$[XS9$)z define composite rule of op expand_as_v2, expand_as_v2->expand_as repeat_times = target_shape / x.shape out = tile(x, repeat_times = repeat_times) rrrqrr)r%rY target_shaperrrrr5rrrrrrrs rexpand_as_v2_compositerDs!wwH}ww  ## #,G ]F  A- -L 7^o#&!8(-? r>!8O8F%* **+,FF#w'(A  " "1 %)x(A/ I99  --rstackc USRnUS:aU[U5S-- n/USUQSPX!SQ7n[UVs/sHn[XC5PM snU5nU$s snf)zh define composite rule of op stack unsqueeze each dimension of the input (use reshape), and then concat rrN)rr=concatrD)r%rx_shape out_shapeitemrjs rstack_compositeriss djjG ax G q  5'%4.5!5gen5I q9qt'$*q94 @C J:sAflatten_contiguous_rangecURn[U5S:waUOSn[U5S:waUOSnXE::de[U5S:Xa [US/S9S4$XE:Xa [XS9S4$Sn[XES-5H nXcU-nM /n[U5HnUR X75 M UR U5 [US-[U55HnUR X75 M [XS9S4$)at define composite rule of op flatten, flatten_contiguous_range -> flatten. xshape is the dim with 0 added to the front of x, keep the shape information of x to calculate the grad. CINN doesn't need xshape for backward pass, return none instead of xshape. shape_out is the parameter of reshape, get from start_axis and stop_axis. out = reshape(x, shape=shape_out), xshape rr)rN)rr=rDr?r) r% start_axis stop_axisr start_dimend_dim slice_numelr5 shape_outs r"flatten_contiguous_range_compositerwswwH!(mq0 aIx=A-i1G    8}q$d**q)4//K 9k *{" +I 9 % [! 7Q;H .%/ 1 & ,,rdropoutcUcSOUnU(aUOSnUS:Hn[URURX%S9nU(aU(d{US:Xa=SU-[UR[R R R54$X-SU- - [U[R R R54$[U5[U[R R R54$U(d2X-[U[R R R54$USU- -[U[R R R54$)zdefine composite rule of op dropout. upscale_in_train: train: out = input * mask / ( 1.0 - p ) inference: out = input downscale_in_infer train: out = input * mask inference: out = input * (1.0 - p) Trupscale_in_train)rrpseed?) bernoullirrrErVarDescVarTypeUINT8rrF) r% seed_tensorrrKmoderfix_seedrmasks rdropout_compositers 'tXH4D11 177!''Q BDCxQwaggt||/C/C/I/I JJJx37+T$,,..44.!9d4)=)=)C)CDD D8T$ (<(<(B(BCC Ca=$tT\\-A-A-G-G"HH Hrc SSKJn U"U5S;aSOUn[[[ XSSUS9[ [ U5S:XaUOS/XR55U5$) Nrrrrrr)minr!rr)rrr greater_equaluniformrr=)rrrrr new_dtypes rrrs`5#5)-BB   E#3T B 3u:?%Y J    r hard_swishc 2SnSnSn[UR5S:Xa UROS/n[[U[ XCUR S9-[ USUR S95[ XAUR S95U-[ XBUR S9- nU$)zdefine composite rule of op hard_swish. offset=3, threshold=6, scale=6 out = minimum( maximum(x + offset, 0), threshold ) * x / scale g@g@rrrpr)r=rminimummaximumrAr)r% thresholdrIrrr's rhard_swish_compositersI E FLA-A3J D177;;ZAGG4  agg 6     z 0 1 Jr index_selectcVUS:a[UR5U-n[XUS9nU$)z)define composite rule of op index_select.r)r)r=rgather)r%indexrr's rindex_select_compositers. ax177|d"  %C JrsigmoidcSnSSKJn U"UR5nUS;aSn[US5nS[ U*5-nSU- nU(dU$[XS5$)z= define composite rule of op sigmoid res = 1 / (1 + exp(-x)) FrrrTrrrrrrr r%r&rrsum_tempr's rsigmoid_compositers^ F5 !'' "E %% I 3r7{H h,C32$s"22rsilucSnSSKJn U"UR5nUS;aSn[US5nS[ U*5-nX- nU(dU$[XS5$)z: define composite rule of op silu res = x / (1 + exp(-x)) FrrrTrrrrs rsilu_compositers\ F5 !'' "E %% I 3r7{H ,C32$s"22rmeshgridcd[U5nS/U-n[U5H>nXR5nUS:XdUS:XdeUS:XdM+XRSX#'M@ /n[U5H@nS/U-nX#Xc'UR XR U5R U55 MB U$)a^ define composite rule of op meshgrid If the input has N tensors of size S_0, ... S_n-1, then the output will also have N tensors, where each tensor is of shape (S_0, ..., S_n-1). E.g. a1 is Tensor [1,2,3] b1 is Tensor [4,5] r1, r2 = paddle.meshgrid([a1, b1]) r1 is Tensor [[1,1], [2,2], [3,3]] r2 is Tensor [[4,5], [4,5], [4,5]] rr)r=r?rrrrD broadcast_to)inputssizerr5routputs view_shapes rmeshgrid_compositers v;D C$JE 4[immoax3!8## !8yq)EH  G 4[S4Z  vy((4AA%HI Nr fill_any_likec6[URX5nU$)z&define composite rule of op full_like.)rAr)r% fill_valuerplacevals rrr+s;H qww *C Jrsqueeze2ch[UR5nUS:Xa [U5S/$[U5S:Xa[[ U55nOUVs1sHoDU-iM nn/n[ UR5H#upgUS:XaXc;aMUR U5 M% [X5nUS/$s snf)z define composite rule of squeezerNr)r=rrFsetr?r@rrD) r%rrankdimsax new_shapedr:rjs rsqueeze2_compositer4s qwwr&s(9! I4L!RA"RAjL!' "' TO$(&%(&VF0M"#<K .!.BN#!.$!.HG  ./-0-:IIIB  L!"2N#$I33$F33$J 4O$%J 0F33$E*F33L! " L!%$"%$PEL!.".r