SjbjUddlmZddlZddlmZmZddlmZmZddl Z ddl m Z ddl m Z mZmZmZddlmZmZmZmZmZmZmZmZmZmZmZmZmZmZm Z m!Z!m"Z"ddl#m$Z$dd l%m&Z&m'Z'dd l(m)Z)m*Z*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0dd l1m2Z2m3Z3m4Z4dd l5m6Z6erdd l7m8Z8e9e:e*e:e;effede?d<ez"load_derivatives..|s-!^,cakktk45zDuplicate function schema: dispatchr2output_differentiabilityDefault)r2rR)r.openyamlloadr,r native_functionsrr)rMr$rrNr1 signatureappendstrr9rsetpopcreate_differentiability_inforH)derivatives_yaml_pathnative_yaml_pathtags_yaml_pathkeyf definitionsfuncs!native_functions_with_view_groupsrXr=r<functions_by_signaturefunctions_by_schemafunction op_counterr;used_dispatch_keys defn_dict specificationrSr2per_dispatch_diffinfoss rGload_derivativesroms !"2 3C // ' ( :A))Aj9K :""2NCTT-QQV,W)$ 6 .  7 !56      :<( ?H "8==#:#:#< = D DX N8==!%88$'B3x}}CUBV%WXX6> HMM 2 3  ?S\^ IK'*u$ 1I* ) f 5 +4==.,(&3I@VW +z$create_derivative..sUAa6kx7Usc3fK|])}tj|dj+yw)T)symintN)r return_typeremove_const_ref)r|rs rGr~z$create_derivative..s)?@$'88:s/1z3Out of bounds grads access: derivative formula for z used grads[z ], but the forward only returns z outputs.)formulaoriginal_formula var_names saved_inputs saved_outputsnamed_gradients)rvnctypertupler return_namesr1returnsziprsaved_variablesresearchr*formatused_gradient_indicesr8 RuntimeErrorr2r )rcrravailable_named_gradientsrartr return_typesr2type named_returnsrrused_named_gradientsis rGcreate_derivativers -:1-=#()!!##I#UAQAQRSATUULDEFFNNL 25\<1P#-4 4M,GY JG\,WmYOG].  99[''-w 7 #7 + AFFNN# #EchhqvvFVEWXS @QVV^^AT@UU^`  )!#,  =#s!F F 2Fc |}d}|jjD]%}|j|vs|d}||jfz}'||dk(rrt |jjdk7r+t dt |jj|jjdjf}ne|D]`}t jd|}t |dk(s(|d}t|d}||jj|jfz}b| t dt|||dddd S) Nr{)rzz$Expected 1 return for 'result', got rz ^result(\d+)$z4No matching output for forward derivative definitionFrr var_typesrequired_inputs_fw_gradrequired_inputs_primalrequired_original_self_valueis_reusing_outplace_formula) r1rr2rr8r9rfindallintr ) rcrnamesrrrvar_nameresarg_idxs rGcreate_forward_derivativersNI)-I VV^^. 66Y   !QVVI-I .  #166>>"a'$:3qvv~~;N:OP*//1I% Ljj!18<s8q= ($& !#a&kG )QVV^^G-D-I-I,K KI  LSTT  $#%*$) rQc Ddfd }g}|D]}|j} || d} | dk(rWjjtjk(r#t djj dtdk(r)t|dkDst|djdkDrtdd t|dk(stdd |dj} dj dfd } tjtjd | | } D]>}|j dfd } tjtj| | } @d| d} t|} | } nR| dk(rLt|dkDst|djdkDrtddDcgc]}|j }}t|dk(r t dg}|D]|vrdz|j!jj#rdz t$j&j(vrdddj+|d} n^t$j,j(vr"t djj |ddddj+|ddd} t|} | } || d}|j!t/| |j|j0| |dd|Scc}w)Nc  jjjjjd}t }D]}|j dvr|s|j}t jtj||}|rtdd|d|d|d t jtj||z|}|s|j|t|S)N _foreach_)zat::TensorListzconst at::ITensorListRef &zThe forward formula for z is using the base name of the z- argument which is ambiguous. You should use z"_p to access the primal value and z_t to access the tangent.) r1r2base startswithr\rrrr*rraddr) rpostfix is_foreachrequired_inputsargarg_namefoundargs_with_derivatives defn_namercs rGfind_required_inputsz=postprocess_forward_derivatives..find_required_inputs%sVV[[%%**55kB %( .CLL"xxHIIk00:GDE".yk9XYaXbcCCK*M!!) *CE IIk00G1CDgNE##H-' .*_%%rQ_tauto_element_wisez"Cannot use auto_element_wise with z" because it is an in-place variantrrDerivative definition of z in derivatives.yaml defines the forward definition of gradient as element_wise but this only works for functions with a single differentiable input and a single differentiable output.z in derivatives.yaml defines the forward definition of gradient as element_wise but it does not defines the gradient formula for its argument which is required.cR|jdd|jdS)Nrz _t.conj()group)m input_names rGreplz-postprocess_forward_derivatives..repljs(''!*j\1771:,GGrQgradcR|jdd|jdS)Nr_prr)rrs rGrz-postprocess_forward_derivatives..replss(ggaj\(2aggaj\BBrQ(z).conj() auto_linearz in derivatives.yaml defines the forward definition of gradient as linear but this only works for functions with a single differentiable output.z-Expected at least one differentiable argument_symintzat::z, )z(Expected Variant.method in variants for .rFr)rr[rr[returntuple[str, ...])rrrr[)rr1kindr&inplacer9r2r8rrrrsubr*rrrZ has_symintr(rivariantsjoinrqr r)rcr all_arg_names derivativesforward_derivativesrrupdated_derivativesdefnrrequired_inputs_tangentbackward_formular fw_formulardiff_arg_namesnew_argsrrrs`` ` @@rGpostprocess_forward_derivativesrso&446#@ ,,"6w"E ) )vv{{} 2 22$8 Egh./14*+a/*1-7781<"/ {;44 {#q("/ {;WW" +1~>> .q166J H 2 26 :DBRSJ- T88C VVK$6$6x$@$ S  TZL1J',M&: # G  %'(1,*1-7781<"/ {;II3HH3chhHNH>"a'$%TUUH) *~-'$H)  *vv  "Y& 1::-#I;a (0C/DAF >>3(B166;;-P!) }Ai[$))HQRL:Q9RRST ',N&; # G "6gt!D"" ....(?'=-2,1  m@ D aIs Lc0|D]}||vcStd)Nz Expected `names` to be non-empty)r)rrr2s rG is_forward_derivative_definitionrs*)=(() 9 ::rQc d#d}d$dd%dt d&fd }|jd}t|\}|jddd} r4tdDr"t d k7rt d |d } d g|j |} | s9d jfd|jD} t d|d| | jj} || } t | dk(r.canonical_functions AFF++-((*C 0 011   #:)A,"3"34 4 24& #((9UV.split_names..s=1QWWY=,)rsplit) raw_namess rG split_namesz2create_differentiability_info..split_namess= (<===rQc d}d}d}g}|D]}|j}|xs2ttjt j d|}|t tjt j d|z }|xst|j}|jt||t |krtd|dt |d|t |k(}|r|dkDrtd|d |rt|dhk(rtd|d |r|s|dkDrtd|d y y ) z Check for some subtle mistakes one might make when writing derivatives. These mistakes will compile, but will be latent until a function is used with double backwards. Frrgradsznum_grads_uses (z) < len(used_grads_indices) (rrzx in derivatives.yaml illegally mixes use of 'grad' and 'grads'. Consider replacing occurrences of 'grad' with 'grads[0]'z in derivatives.yaml solely refers to 'grads[0]'. If the first output is indeed the only differentiable output, replace 'grads[0]' with 'grad'; otherwise, there is a likely error in your derivatives declaration.z in derivatives.yaml illegally mixes use of "grad_RETURN_NAME" and "grad" or "grads[x]". Use only one method for identifying gradients.N) rboolrrr*rr8rextendrr9rr\) rr uses_gradnum_grads_usesuses_named_gradsused_grads_indicesdronly_used_grads_indicess rGcheck_grad_usagez7create_differentiability_info..check_grad_usages  (* FAiiG!T ;--f5w?&I c"**[-?-?-H'"RS SN/J48I8I3J   % %&;G&D E F C 23 3 ">"22OPSTfPgOhhij  #1C8J4K"K !++I;788  #s+='>1#'E+I;7  nq.@+I;7== /A rQc g}g}g}t}t|Dcgc]}|j}}|jjDcgc]}|j}}xs#dgt |jjz} t |jj| D cgc]<\} } | r5| j)| jjrd| j>} } } tjD]} | }| }|D]4}||vs||vs td|jjd|d|dt||r|jt|||r|jj!dk(r||z }t#|||| }|j||t|z}|j%|}|rt'dd |t|Dcgc]}|j|vs|}}t)|||||}|||||| fScc}wcc}wcc} } wcc}w) NTgrad_z-While processing the derivative formula for 'z' wrt 'z ', expected 'z/' to not be both an input arg and named return.non_differentiablezderivatives definition for zB have overlapped non_differentiable and differentiable variables: )r\rvr2r1rr8rris_tensor_likesortedkeysr9rrZrlowerrr intersectionrr)rcrrnon_differentiable_arg_namesargs_with_derivatives_setrrr all_ret_namesdifferentiabilityretdifferentiablerrrrr2 derivativeoverlaprrrrrSrs rGset_up_derivativesz9create_differentiability_info..set_up_derivatives"s)+ 7924$.1e!)6q)9:A: :FFNN AFF  5TQVV^^AT8T(+166>>;L'M % #^$'')CHH:  % ! %   ,     !   K0   ! ( %   Y;  % X! sH; I+AII I rSNc3<K|]}t|tywr)rMr[)r|diffs rGr~z0create_differentiability_info..s("& 4(rrzNot supported: for z,output_differentiability must either be list[bool] or a list[str] where each str is a condition. In the case where it is a condition, we only support single-output functions. Please file us an issue. T c3nK|],\}}tj|jk(s)|.ywr)rr2r1r|kvrs rGr~z0create_differentiability_info..s. !Q!&&9IY9VA s*55z)could not find ATen function for schema: z . Available signatures: rc3jK|]*\}}tj|k(r t|,ywr)rr2r[rs rGr~z0create_differentiability_info..s1 1xx{i' F s03z3could not find ATen function for legacy signature: z corresponding to schema z9. Please report a bug to PyTorch. Available signatures: grad_input_maskc34K|]}|jywrr2r|rs rGr~z0create_differentiability_info..sFQVVFz Schema for z has an argument named grad_input_mask, but this name would be shadowed by our codegen. Please use a different name in native_functions.yaml.rzc34K|]}|jywrrrs rGr~z0create_differentiability_info..s=qAFF=rzy has an argument named result, but this is only allowed for outputs.Please use a different name in native_functions.yaml.rRrTzInvalid dispatch key z in derivatives.yaml for z, expected key to be one of ) r2r1oprrall_saved_inputsall_saved_outputsrrrrrS#output_differentiability_conditions)rOzSequence[NativeFunction]r2r[rr#)rr[rr)rr[rzSequence[Derivative]rNone)rcr#rzituple[Sequence[Derivative], Sequence[ForwardDerivative], Sequence[Binding], Sequence[str], Sequence[str]])rr]r+anyr8rr5rr4r1rYrv_VALID_AUTOGRAD_KEYSrr\r_create_op_prefixr dedup_varsrr) rlrgrhrjrkrrrmrr schema_functionavailrYrO canonical diffinfo_dictrbrrrrrrrr op_prefixrrrrrSrs @@@@@rGr^r^s+36 ">4la a  a a a HMM&)M$]3LIq )}}-GN*.'C(*B(% ' (A -%m_5,- /G+$(6 )--m>= ((ABs #KKz(?:^|\W)grads\[(\d+)\]cntjt|Dcgc] }t|c}Scc}w)zDetermine a list of gradient indices (the i in grads[i]) that are used by the formula. >>> used_gradient_indices("foo(grads[0], grads[1])") [0, 1] )rrGRAD_INDEX_REGEXr)rrs rGrrs(JJ'7A BqCF BB Bs2c* dMfd }ddddfdddd d fd d d dd fddddfddddddfddddfddddfddd dfd!d"d#dfd$d%d&dfd'd(d)dfd*d+d,dfd-d.d/dfd0d1d2dfd3d4d5|d fd6d7d8dfd9d:d;dfg}g d<|vsd=|vrtd>d?|ztjd@|stjdA|rtdBdC|zdD|vsdE|vrtdFdG|z|D]}t|jt r|jjn |j |D]3\}dN fdH }tj |j ||}5|jtttk(r!tj dI dI dJ dK|}tjtj |s jt| L|t fS)ONc8|fk7r tdd|d|dS)NzReplacement for ".strides()" is currently only supported for single derivatives of the same tensor that ".strides()" is being called on.zstrides_or_error(z, "z"))r9)r2rs rG stride_exprz$saved_variables..stride_exprs6   8 #4&D644rQz{}.sym_sizes\(\) _sym_sizesc4t|ttSrrr rrs rGrPz!saved_variables..z$ /8R'SrQ)suffixrz{}->sym_sizes\(\)_sym_sizes_optcFt|tttSrrrr rrs rGrPz!saved_variables..$z- /(BC(rQc|d|dS)Nz2.has_value() ? std::optional(z->sym_sizes()) : std::nulloptr{rs rGrPz!saved_variables..'sv-_`d_efC&DrQ)r4rexprz{}.sym_blocksize\(\)_self_sym_blocksize_optcFt|tttSrr7rs rGrPz!saved_variables../r8rQcd|dS)Nz#at::sparse_csr::getSymIntBlockSize(rr{rs rGrPz!saved_variables..2s'J4&PQ%RrQz{}.options\(\)_optionsc4t|ttSr)rr rrs rGrPz!saved_variables..:sz$ .8Q'RrQzzeros_like\({}\)_infoc4t|ttSr)rr rrs rGrPz!saved_variables..Bsz$ ,8O'PrQc|Srr{rs rGrPz!saved_variables..CsTrQc |dzS)Nz _info.zeros()r{rs rGrPz!saved_variables..Ds D?$:rQ)r4rr:rz{}.sym_size\((-?\w+)\)cNd|jdjddS)N _sym_argsize_r-minus_)groupsreplace)rs rGrPz!saved_variables..Ks%mAHHJqM4I4I#x4X3Y$ZrQc4t|ttSrrr rrs rGrPz!saved_variables..Lz$ '8J'KrQz {}.numel\(\)_numelc4t|ttSrrr rrs rGrPz!saved_variables..Tz$ %8H'IrQz{}.sym_numel\(\) _sym_numelc4t|ttSrrKrs rGrPz!saved_variables..\rLrQzto_args_sizes\({}\) _args_sizesc Xt|ttttSr)rrr rrs rGrPz!saved_variables..dsz+k)E2B&CD(rQzto_args_sizes_symint\({}\)_args_sizes_symintc Xt|ttttSr)rrr rrs rGrPz!saved_variables..nsz+k)G2D&EF(rQzto_args_scalartypes\({}\)_args_scalartypescFt|tttSr)rrr rrs rGrPz!saved_variables..xsz+i &<=(rQzTensorGeometry\({}\) _geometryc4t|ttSr)rr rrs rGrPz!saved_variables..r3rQz{}.scalar_type\(\) _scalar_typec4t|ttSr)rr rrs rGrPz!saved_variables..sz$ +8N'OrQz {}.dim\(\)_dimc4t|ttSrrOrs rGrPz!saved_variables..rPrQz{}.sym_strides\(\) _sym_stridesc4t|ttSrr2rs rGrPz!saved_variables..r3rQz {}.layout\(\)_layoutc4t|ttSr)rr rrs rGrPz!saved_variables..rLrQz{}.is_conj\(\) _conjugatec4t|ttSr)rr rrs rGrPz!saved_variables..rPrQz.sizes()z ->sizes()zY.sizes() is not supported in derivative formulas. Instead, please use the SymInt version,z<.sym_sizes(), which returned a c10::SymIntArrayRef. formula=z\.size\([-]?\d+\)z->size\([-]?\d+\)z[.size(int) is not supported in derivative formulas. Instead, please use the SymInt version,z>.sym_size(int), which returned a c10::SymIntArrayRef. formula=z .strides()z ->strides()z[.strides() is not supported in derivative formulas. Instead, please use the SymInt version,z>.sym_strides(), which returned a c10::SymIntArrayRef. formula=ctdr d|nd}dvr dn|jd}jtd|z|dvr d}|S|zS)Nr4r:rrrr:r)callablerrZr )rr4r: replacementrDr2saveds rGrzsaved_variables..repls*2$x.)ANDN1%tH~39D.LDL.aggaj "-tH~dVm<! D='2tE{4'8K&&f}$rQz\bz/.has_value() ? std::optional(z.value()) : std::nulloptrfr2r[rr[)rz re.Match[str]rr[)rrrrMr2rrrrrr rr*rZr r) rnctypesrr/ REPLACEMENTSrregexrrDr2ris ` @@@rGrr s? 5 &S   !*D    $3S    $R   !P):    &ZK   "I   &K   #'   *.   )-   $%S   "(O    I   "(S#   #K   &I  aW6Lt#%EW w 6 gLWIV W   yy%w/299g4 iNwiX Y  w-7": iNwiX Y  -!+6;; GFKK  V[[ ( @KE4 %"ffU\\$/w?G' @0 ;;- '(:; ;ffdV2&GvMefG 99[''-w 7 LL! Q-^ E%L  rQcdj|jdDcgc]}|jc}}|dzjddScc}w)aaTakes a native function name converts to an op prefix name. Note that the "name" parameter must be the native function name without the optional variant suffix, so "add" instead of "add.out". OP names correspond to classes, hence the change to title case. Example:: >>> _create_op_prefix("add") 'AddBackward' rBackwardForwardBackward)rrtitlerI)r2p camel_cases rGr$r$sITZZ_=!'')=>J  # , ,-> KK>sAc.t}g}|D]}t|jjtr |jjjn|jj}||vrb|j ||j ||Sr)r\rMrr2rrrZ)varsseenrivarr2s rGr%r%s~UD"$E #**//>: JJOO  4<   S  LrQ)r;z6dict[FunctionSchema, dict[str, DifferentiabilityInfo]]r<zlist[NativeFunctionsViewGroup]rr!)r_r[r`r[rar[r DerivativeRet)rcr#rSequence[Binding]) rcr#rr[rrrz Sequence[str]rr )rcr#rr[rrrr )rcr#rr[r list[str]rzlist[Derivative]rlist[ForwardDerivative]rrzrr|)rr{rrrr) rlzdict[Any, Any]rgz*dict[FunctionSchema, list[NativeFunction]]rhzdict[str, NativeFunction]rjz Counter[str]rkzset[str]rz7tuple[FunctionSchema, dict[str, DifferentiabilityInfo]])rr[rz list[int])rr[rkzlist[NamedCType]rrrz&tuple[str, tuple[SavedAttribute, ...]]rj)rvSequence[SavedAttribute]rr})N __future__rr collectionsrrtypingrrrV torchgen.apirtorchgen.api.autogradr r r r torchgen.api.typesr rrrrrrrrrrrrrrrrtorchgen.contextr torchgen.genrr torchgen.modelr!r"r#r$r%r&r'r(torchgen.utilsr)r*r+torchgen.yaml_utilsr,collections.abcr-rdictr[r\ryr.__annotations__r#rHrorvrrrrr^r,rrr$r%r{rQrGrs\ # ,%  &2P   ED*(d>45J0J+KKLcRUhVW FHCH=)# A#/# #LK.K.25K.GJK.K.^**,, ,, - ,  ,^&& #&,;&&Rffff" f 1 f - ffR;;%4; ;k)k)Fk)3k) k) ! k) = k)\ -Ci! i! i!i!, i!XL$ rQ