MjHUdZddlmZddlmZddlmZmZddlm Z m Z m Z m Z m Z mZmZmZmZee zeedfzdzZeed <Gd d Zd ed edfdZGddeZee zZeed<eedfezdzZeed<ee zeedfzdzZeed<deeedfzd efdZded efdZded efdZd-deded efdZ d-deded efdZ!deded efdZ"d.deded efd Z#dedzd edzfd!Z$dedzd edzfd"Z%deded efd#Z&deded efd$Z'd%ededed efd&Z(deded efd'Z)deded efd(Z*deded efd)Z+deded efd*Z,d+eedfded eeeffd,Z-y)/z Definition of CuTe Layouts and functions to manipulate them which works with the order of lexicographic instead of co-lexicographic as implemented in the original layout.py )chain) TypeAlias)SelfTypeIs) crd2idxflattenhas_noneIntTupleis_intis_tupleproductslice_suffix_product.NCoordinateTypec eZdZy) LayoutBaseN)__name__ __module__ __qualname___/var/www/html/ia-enhanced/venv/lib/python3.12/site-packages/torch/distributed/_pycute/layout.pyrr=srrxreturnLayoutc"t|tSN) isinstancer)rs r is_layoutr As a $$rceZdZddededzddfdZdedefdZdefdZ d e de ezfd Z d ede fd Z defd ZdefdZdefdZdefdZy)rN_shape_stridercZ||_|t|j|_y||_yr)shaperstride)selfr"r#s r__init__zLayout.__init__Fs% ?(4DK!DKrotherct|tsy|j|jk(xr|j|jk(S)NF)rrr%r&)r'r)s r__eq__z Layout.__eq__Ns4%(zzU[[(HT[[ELL-HHrcXt|jrt|jSyNr)r r%lenr's r__len__zLayout.__len__Ts DJJ tzz? "rargsct|r|t|dk(r:tt|d|jt|d|j Stt||jt||j St|dk(r$t |d|j|j St ||j|j S)a Map a logical coordinate to a linear index (Coord has no Underscore slice operators) OR Slice the layout and return the sublayout (Coord has an Underscore slice op) Follow the same behavior of `Layout::operator(Coord const&)` in cute C++ rr)r r.rrr%r&r)r'r1s r__call__zLayout.__call__[s D>4yA~fT!Wdjj96$q'4;;;WXXfT4::6tT[[8QRR4yA~tAw DKK@@tTZZ==rict|jr&t|j||j|S|dk7rtt|j|jS)Nr)r r%rr&AssertionError)r'r4s r __getitem__zLayout.__getitem__osM DJJ $**Q-Q8 8Av$$$**dkk2 2rc,t|jSr)rr%r/s rsizez Layout.sizexstzz""rc:||jdz dzSr-)r9r/s rcosizez Layout.cosize|sDIIK!O$q((rc8|jd|jS)N:r%r&r/s r__str__zLayout.__str__s**Qt{{m,,rc<d|jd|jdS)NzLayout(,)r>r/s r__repr__zLayout.__repr__s Adkk]!44rr)rrrr r(objectboolr+intr0rrr3r7r9r;strr?rCrrrrrEs"x"(T/"T"IFItI >n>>(3S3T3#c#))--5#5rLayoutOrIntTuple LayoutProfile LayoutInputlayoutsct|dk(rt|ds|d}td|D\}}t||S)Nrrc3LK|]}|j|jfywrr>).0as r zmake_layout..s?!177AHH-?s"$)r.r zipr)rKr%r&s r make_layoutrRsC 7|q71:!6!*?w?@ME6 %  rlayoutcNt|r|jSt|Sr)r r9rrSs rr9r9s {{} 6?rc"|jSr)r;rUs rr;r;s ==?rprofilec  trsttkrttt fdt tDfdt ttDSdg}dg}t ttjttjD]Z\}}|dk(r |ddk(r ||d<||d<|d|dz|k(r |d|z|d<9|j||j|\t|dk(rt|d|dS|j|jtt|t|S)Nc3BK|]}t||ywr)coalescerNr4rSrWs rrPzcoalesce..s NQ&)WQZ0Nc3(K|] }| ywrrrNr4rSs rrPzcoalesce..EqErr)r r.r6rRrrangerQreversedr r%r&appendrreversetuplerSrW result_shape result_strider%r&s`` rrZrZss v;W %  N%G :MNEE#g,F $DE   3LCM&''&--2H)I) v A:  "  "$L  &M"  "  b 1 1V ;+B/%7L     &   (!)$ <Al1o}Q'788eL)5+?@@rc \trsttkrttt fdt tDfdt ttDSg}g}t tjtjD]3\}}|dk(r |dk(r|j||j|5t|dk(r tddSttt|t|S)Nc3BK|]}t||ywr)filterr[s rrPzfilter..s L1q 71:.Lr\c3(K|] }| ywrrr^s rrPzfilter..r_r`rr)r r.r6rRrrbrQr r%r&rdrrZrfrgs`` rrlrls v;W %  Lc'l8KLEE#g,F $DE   LMWV\\2GFMM4JK) v fk    &   ( )  <Aa|u\2E-4HIJJrlayoutAlayoutBc StrttStrst t krt t tfdtt Dfdtt t DStjrt fdDSjdk(rtjdSg}g}j}j}t}ttt|jddtt|jddD]j\}}||zdk(s||zdk(st tt!d||z|} | dk7r%|j#| |j#||z|| z}| |z }l|dk7st |dk(r;|j#||j#|t|jdz|j%|j%t |dk(rt|d|dStt'|t'|S)Nc3BK|]}t||ywr compositionrNr4rnros rrPzcomposition..s RWQZ4Rr\c3(K|] }| ywrrrNr4rns rrPzcomposition..GGr`c36K|]}t|ywrrr)rN layoutB_irns rrPzcomposition..sTy;w :Tsrr)r rsrr r.r6rRrrbr%r&rZrQrcr minmaxrdrerf) rnrorhri rest_shape rest_strideflat_A curr_shape curr_stride new_shapes `` rrsrssC 7F7O44 '  w<#g, &  ReCL>QRGU3w<W%FG    '-- TGTTT~~gmmQ''  ]] nn '"(+ WV\\*12. /'&--:PQRQS:T1U(   #J ,1[:5MQR5R$$C:#<=zJIA~##I.$$[;%>?#y0J *K & ?c,/14    +  wv}}/Ea/H!H I  |  !,q/=+;< <% -u]/CD Drmax_idxcft|rtt|Sg}g}d}tt t |j t |j}|D][\}}|dk(s|dk(r|||zk}t|tur|st|j||z|j|||z}]|j||zdz |z|j||j|jttt|t|S)Nrr)r complementrsortedrQr r&r%typerEr6rdrerZrf) rSrrhri current_idx sorted_DSr&r%in_bounds rrr/s f~&.))LMKs76==176<<3HIJI" %  Q;%1* %&.0 Nd "H Fk12[)fn  %;.2{BC% F5.m0DE FFrc |yt|r t|Sg}g}d}t|j}t|j}t t ||t|}|D]:\}}} |dk(r ||k7rn)|j||j| ||z}<|j|jttt|t|Sr-) r rr r%r&rrQrrdrerZrf) rSrhrir flat_shape flat_stride sorted_DSAr&r%rstrides r right_inverserPs ~ f~LMK&J&--(KK^J5OPQJ",%w A:  & E"W%fn % F5.m0DE FFrcp|yt|r t|Sttt ||Sr)r rrrRrrUs r left_inverserms2 ~ f~ Z%7@ AArc StrttStrst t krt t tfdtt Dfdtt t DStt ttS)Nc3BK|]}t||ywr)logical_dividerts rrPz!logical_divide..s(#71:wqz:r\c3(K|] }| ywrrrvs rrPz!logical_divide..rwr`) r rrr r.r6rRrrbrsrr9rnros``rrrws gvg77 '  w<#g, &  "3w<0HU3w<W%FG     GZg?@ rc StrttStrst t krt t tfdtt Dfdtt t DSt ttttzS)Nc3BK|]}t||ywr)logical_productrts rrPz"logical_product..s($GAJ ;r\c3(K|] }| ywrrrvs rrPz"logical_product..rwr`) r rrr r.r6rRrrbrsrr9r;rs``rrrs gvg77 '  w<#g, &  "3w<0HU3w<W%FG     JwW w(GH'R rsplittercttddStrttkrttfdt tDttfdt tDtt fdt tDfdt ttDSS)Nrrc3DK|]}t||ywr) hier_unzip)rNr4rnrors rrPzhier_unzip..s*  xWQZ 8 s c3.K|] }|dyw)rNrrNr4splits rrPzhier_unzip..sAa Ac3.K|] }|dyw)rNrrs rrPzhier_unzip..s>QU1Xa[>rc3(K|] }| ywrrrvs rrPzhier_unzip..sKAWQZKr`)rRrr r.r6rbr)rrnrors```@rrrs 6!Q<11 '  w<#g, &  3w<(    AU3w<-@A A >%G *=>Ks7|S\)JK    GW %%rc$tt||Sr)rrrs r zipped_dividers ngw 77rc t||}t|dgtt|dDcgc] }|d| c}zScc}wNrr)rrRrbr.rnroresultr4s r tiled_dividersG 7G ,F q {E#fQi.rsB *   (NU63;''$.    %%F8,%@5Z@5H%x/)/ -6= y=(*U63;-??$F YF!&5#55!&!!c6c 'AV'Am'Av'AVK6KMKVK8@E@E+@E&@EHG'G#GfGBG*T1GftmG:B)D0BVd]BF[V6Vkf4&& && &>868K8F8 T&T;T6T 9F9[9V9 T6TKTFT % ,fvs{ASr