Z^(hjUddlmZddlZddlmZmZddlmZmZddl 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..xs-!^,cakktk45dispatchr2output_differentiabilityDefault)r2rQ)r.openyamlloadr,r native_functionsrr)rLr$rrMr1 signatureappendstrrsetpopcreate_differentiability_inforG)derivatives_yaml_pathnative_yaml_pathtags_yaml_pathkeyf definitionsfuncs!native_functions_with_view_groupsrWr<r;functions_by_signaturefunctions_by_schemafunction op_counterr:used_dispatch_keys defn_dict specificationrRr2per_dispatch_diffinfoss rFload_derivativesrnis !"2 3C // ' ( :A))Aj9K :""2NCTT-QQV,W)$ 6 .  7 !56      :<( ?H "8==#:#:#< = D DX Nx}}%-@@ @@6> HMM 2 3 ?S\^ IK'*u$ 1I* ) f 5 +4==.,(&3I@VW +z$create_derivative..sUAa6kx7Usc3fK|])}tj|dj+yw)T)symintN)r return_typeremove_const_ref)r{rs rFr}z$create_derivative..s)?@$'88:s/1z3Out of bounds grads access: derivative formula for z used grads[z ], but the forward only returns z outputs.)formulaoriginal_formula var_names saved_inputs saved_outputsnamed_gradients)runctypertupler return_namesr1returnsziprsaved_variablesresearchr*formatused_gradient_indicesr8 RuntimeErrorr2r )rbrravailable_named_gradientsrarsr return_typesr2type named_returnsrrused_named_gradientsis rFcreate_derivativers -:1-=#()!!##I#UAQAQRSATUULDEFFNNL 25\<1P#-4 4M,GY JG\,WmYOG].  99[''-w 7 #7 + AFFNN# #EchhqvvFVEWXS @QVV^^AT@UU^`  )!#,  =#s!F F 2Fc &|}d}|jjD]%}|j|vs|d}||jfz}'||dk(rIt |jjdk(sJ|jjdjf}ne|D]`}t j d|}t |dk(s(|d}t|d}||jj|jfz}b|Jdt|||ddddS) Nrz)ryrz ^result(\d+)$z4No matching output for forward derivative definitionFrr var_typesrequired_inputs_fw_gradrequired_inputs_primalrequired_original_self_valueis_reusing_outplace_formula) r1rr2rr8rfindallintr ) rbrnamesrrrvar_nameresarg_idxs rFcreate_forward_derivativers-I)-I VV^^. 66Y   !QVVI-I .  #qvv~~&!+ ++*//1I% Ljj!18<s8q= ($& !#a&kG )QVV^^G-D-I-I,K KI  L  X"XX  $#%*$) rPc dfd }g}|D]}|j} || d} | dk(rSjjtjk7sJdjj dt dk(r)t |dkDst |djdkDrtdd t |dk(stdd |dj} dj dfd } tjtjd | | } D]>}|j dfd } tjtj| | } @d| d} t|} | } n)| dk(r#t |dkDst |djdkDrtddDcgc]}|j }}t |dkDsJg}|D]|vrdz|jjj!rdz t"j$j&vrdddj)|d} n>t"j*j&vsJ|ddddj)|ddd} t|} | } || d}|jt-| |j|j.| |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_namerbs rFfind_required_inputsz=postprocess_forward_derivatives..find_required_inputssVV[[%%**55kB %( .CLL"xxHIIk00:GDE".yk9XYaXbcCCK*M!!) *CE IIk00G1CDgNE##H-' .*_%%rP_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 rFreplz-postprocess_forward_derivatives..repl`s(''!*j\1771:,GGrPgradcR|jdd|jdS)Nr_prr)rrs rFrz-postprocess_forward_derivatives..replis(ggaj\(2aggaj\BBrP(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._symintzat::z, ).rFr)rrZrrZreturntuple[str, ...])rrrrZ)rr1kindr&inplacer2r8rrrrsubr*rrrY has_symintr(rhvariantsjoinrpr r)rbr all_arg_names derivativesforward_derivativesrrupdated_derivativesdefnrrequired_inputs_tangentbackward_formular fw_formulardiff_arg_namesnew_argsrrrs`` ` @@rFpostprocess_forward_derivativesrsN&446#z ,,"6w"E ) )66;;=J$6$66 4QVV[[MAcd 6./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~&* **H) *~-'$H) * vv  "Y& 1::-#I;a (0C/DAF ~~333 ( }Ai[$))HQRL:Q9RRST ',N&; # G "6gt!D"" ....(?'=-2,1  az x WIsK0c0|D]}||vcStd)Nz Expected `names` to be non-empty)r)rrr2s rF is_forward_derivative_definitionrs*)=(() 9 ::rPc 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  czSXXil&7&78888|rPcDtd|jdDS)z(Given "foo, bar", return ["foo", "bar"].c3<K|]}|jywN)strip)r{xs rFr}zEcreate_differentiability_info..split_names..s=1QWWY=,)rsplit) raw_namess rF split_namesz2create_differentiability_info..split_namess= (<===rPc Pd}d}d}g}|D]}|j}|xs2ttjt j d|}|t tjt j d|z }|xst|j}|jt||t |k\sJ|t |k(}|r|dkDrtd|d|rt|dhk(rtd|d|r|s|dkDrtd|dy 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. Frrgradsrzx 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*rr8rextendrrr[) rr uses_gradnum_grads_usesuses_named_gradsused_grads_indicesdronly_used_grads_indicess rFcheck_grad_usagez7create_differentiability_info..check_grad_usagesg  (* FAiiG!T ;--f5w?&I c"**[-?-?-H'"RS SN/J48I8I3J   % %&;G&D E F%7!8888#1C8J4K"K !++I;788  #s+='>1#'E+I;7  nq.@+I;7== /A rPc g}g}g}t}t|Dcgc]}|j}}|jjDcgc]}|j}}xs#dgt |jjz} t |jj| D cgc]<\} } | r5| j)| jjrd| j>} } } tjD]} | }| }|D]0}||vs||vs Jd|jjd|d|dt||r|jt|||n|jjdk(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 'z0' to not be both an input arg and named return. non_differentiablezderivatives definition for zB have overlapped non_differentiable and differentiable variables: )r[rur2r1rr8rris_tensor_likesortedkeysrrYrlowerrr intersectionrr)rbrrnon_differentiable_arg_namesargs_with_derivatives_setrrr all_ret_namesdifferentiabilityretdifferentiablerrrrr2 derivativeoverlaprrrrrRrs rFset_up_derivativesz9create_differentiability_info..set_up_derivatives s)+ 7924$.1e!)6q)9:A: :FFNN AFF  5TQVV^^AT8T(+166>>;L'M % #^$'')CHH:  % ! %   ,     !   K0   ! ( %   W;  % V! sH7 H<+AI=IIrRNc3<K|]}t|tywr)rLrZ)r{diffs rFr}z0create_differentiability_info..ws("& 4(rrzNot supported: for z,output_differentiability must either be list[bool] or a list[str] where each str is a condition. In the case where it is a condition, we only support single-output functions. Please file us an issue. T c3nK|],\}}tj|jk(s)|.ywr)rr2r1r{kvrs rFr}z0create_differentiability_info..s. !Q!&&9IY9VA s*55z)could not find ATen function for schema: z . Available signatures: rc3jK|]*\}}tj|k(r t|,ywr)rr2rZrs rFr}z0create_differentiability_info..s1 1xx{i' F s03z3could not find ATen function for legacy signature: z corresponding to schema z9. Please report a bug to PyTorch. Available signatures: grad_input_maskc34K|]}|jywrr2r{rs rFr}z0create_differentiability_info..sFQVVFz Schema for z has an argument named grad_input_mask, but this name would be shadowed by our codegen. Please use a different name in native_functions.yaml.ryc34K|]}|jywrrrs rFr}z0create_differentiability_info..s=qAFF=rzy has an argument named result, but this is only allowed for outputs.Please use a different name in native_functions.yaml.rQrSzInvalid dispatch key z in derivatives.yaml for z, expected key to be one of ) r2r1oprrall_saved_inputsall_saved_outputsrrrrrR#output_differentiability_conditions)rNzSequence[NativeFunction]r2rZrr#)rrZrr)rrZrzSequence[Derivative]rNone)rbr#rzituple[Sequence[Derivative], Sequence[ForwardDerivative], Sequence[Binding], Sequence[str], Sequence[str]])rr\r+anyr8rr5rr4r1rXru_VALID_AUTOGRAD_KEYSrr[r_create_op_prefixr dedup_varsrr) rkrfrgrirjrr rlrrschema_functionavailrXrN canonical diffinfo_dictrarrrrrrrr op_prefixrrrrrRrs @@@@@rFr]r]s + 36   >1f` `  ` ` ` FMM&)M$]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 rFrrs(JJ'7A BqCF BB Bs2c* dMfd }ddddfdddd d fd d d dd fddddfddddddfddddfddddfddd dfd!d"d#dfd$d%d&dfd'd(d)dfd*d+d,dfd-d.d/dfd0d1d2dfd3d4d5|d fd6d7d8dfd9d:d;dfg}g d<|vsd=|vrtd>d?|ztjd@|stjdA|rtdBdC|zdD|vsdE|vrtdFdG|z|D]}t|jt r|jjn |j |D]3\}dN fdH }tj |j ||}5|jtttk(r!tj dI dI dJ dK|}tjtj |s jt| L|t fS)ONc0|fk(sJdd|d|dS)NzReplacement for ".strides()" is currently only supported for single derivatives of the same tensor that ".strides()" is being called on.zstrides_or_error(z, "z")rz)r2rs rF stride_exprz$saved_variables..stride_exprs6TG#  4 ##4&D644rPz{}.sym_sizes\(\) _sym_sizesc4t|ttSrrr rrs rFrOz!saved_variables..z$ /8R'SrP)suffixrz{}->sym_sizes\(\)_sym_sizes_optcFt|tttSrrrr rrs rFrOz!saved_variables.. z- /(BC(rPc|d|dS)Nz2.has_value() ? std::optional(z->sym_sizes()) : std::nulloptrzrs rFrOz!saved_variables..sv-_`d_efC&DrP)r2rexprz{}.sym_blocksize\(\)_self_sym_blocksize_optcFt|tttSrr5rs rFrOz!saved_variables..r6rPcd|dS)Nz#at::sparse_csr::getSymIntBlockSize(rrzrs rFrOz!saved_variables..s'J4&PQ%RrPz{}.options\(\)_optionsc4t|ttSr)rr rrs rFrOz!saved_variables..#sz$ .8Q'RrPzzeros_like\({}\)_infoc4t|ttSr)rr rrs rFrOz!saved_variables..+sz$ ,8O'PrPc|Srrzrs rFrOz!saved_variables..,sTrPc |dzS)Nz _info.zeros()rzrs rFrOz!saved_variables..-s D?$:rP)r2rr8rz{}.sym_size\((-?\w+)\)cNd|jdjddS)N _sym_argsize_r-minus_)groupsreplace)rs rFrOz!saved_variables..4s%mAHHJqM4I4I#x4X3Y$ZrPc4t|ttSrrr rrs rFrOz!saved_variables..5z$ '8J'KrPz {}.numel\(\)_numelc4t|ttSrrr rrs rFrOz!saved_variables..=z$ %8H'IrPz{}.sym_numel\(\) _sym_numelc4t|ttSrrIrs rFrOz!saved_variables..ErJrPzto_args_sizes\({}\) _args_sizesc Xt|ttttSr)rrr rrs rFrOz!saved_variables..Msz+k)E2B&CD(rPzto_args_sizes_symint\({}\)_args_sizes_symintc Xt|ttttSr)rrr rrs rFrOz!saved_variables..Wsz+k)G2D&EF(rPzto_args_scalartypes\({}\)_args_scalartypescFt|tttSr)rrr rrs rFrOz!saved_variables..asz+i &<=(rPzTensorGeometry\({}\) _geometryc4t|ttSr)rr rrs rFrOz!saved_variables..kr1rPz{}.scalar_type\(\) _scalar_typec4t|ttSr)rr rrs rFrOz!saved_variables..rsz$ +8N'OrPz {}.dim\(\)_dimc4t|ttSrrMrs rFrOz!saved_variables..zrNrPz{}.sym_strides\(\) _sym_stridesc4t|ttSrr0rs rFrOz!saved_variables..r1rPz {}.layout\(\)_layoutc4t|ttSr)rr rrs rFrOz!saved_variables..rJrPz{}.is_conj\(\) _conjugatec4t|ttSr)rr rrs rFrOz!saved_variables..rNrPz.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)Nr2r8rrrr8r)callablerrYr )rr2r8 replacementrCr2saveds rFrzsaved_variables..repls)1$x.)ANDN1%tH~39D.LDL.aggaj "-tH~dVm<! D='2tE{4'8K&&f}$rPz\bz/.has_value() ? std::optional(z.value()) : std::nulloptrdr2rZrrZ)rz re.Match[str]rrZ)rrrrLr2rrrrrr rr*rYr r) rnctypesrr- REPLACEMENTSrregexrrCr2rgs ` @@@rFrrs= 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% @. ;;- '(:; ;ffdV2&GvMefG 99[''-w 7 LL! M+Z E%L  rPcdj|jdDcgc]}|jc}}|dzjddScc}w)aWTakes a native function name converts to a 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)rrtitlerG)r2p camel_cases rFr"r"sITZZ_=!'')=>J  # , ,-> KK>sAc.t}g}|D]}t|jjtr |jjjn|jj}||vrb|j ||j ||Sr)r[rLrr2rrrY)varsseenrgvarr2s rFr#r#s~UD"$E #**//>: JJOO  4<   S  LrP)r:z6dict[FunctionSchema, dict[str, DifferentiabilityInfo]]r;zlist[NativeFunctionsViewGroup]rr)r^rZr_rZr`rZr DerivativeRet)rbr#rSequence[Binding]) rbr#rrZrrrz Sequence[str]rr )rbr#rrZrrrr )rbr#rrZr list[str]rzlist[Derivative]rlist[ForwardDerivative]rrxrrz)rryrrrr) rkzdict[Any, Any]rfz*dict[FunctionSchema, list[NativeFunction]]rgzdict[str, NativeFunction]riz Counter[str]rjzset[str]rz7tuple[FunctionSchema, dict[str, DifferentiabilityInfo]])rrZrz list[int])rrZrizlist[NamedCType]rrrz&tuple[str, tuple[SavedAttribute, ...]]rh)rtSequence[SavedAttribute]rr{)N __future__rr collectionsrrtypingrrrU torchgen.apirtorchgen.api.autogradr r r r torchgen.api.typesr rrrrrrrrrrrrrrrrtorchgen.contextr torchgen.genrr torchgen.modelr!r"r#r$r%r&r'r(torchgen.utilsr)r*r+torchgen.yaml_utilsr,collections.abcr-rdictrZr[rwr.__annotations__r!rGrnrurrrrr]r*rrr"r#rzrPrFrs\ # ,%  &2P   ED*(d>45J0J+KKLcRUhVW FHCH=) A/ DJ.J.25J.GJJ.J.\**,, ,, - ,  ,^"" #",;""J````" ` 1 ` - ``F;;%4; ;d)d)Fd)3d) d) ! d) = d)N -Cf! f! f!f!, f!RL" rP