7hSSKrSSKrSSKrSSKrSSKrSSKrSSKrSSKrSSKrSSK J r SSK J r SSKJ r Jr SSKJrJrJrJrJr SSKrSSKJr SSKJrJr SSKJr SS KJr SS KJ r \(aSS K!J"r" SS K#J$r$ SS K%J&r& SSK'J(r( SSK'J)r)J*r*J+r+ SSK,J-r-J.r. SSK/J0r0J1r1J2r2J3r3J4r4J5r5J6r6J7r7J8r8J9r9J:r:J;r;JS?\RRS\P4S@jr[S=S>S?\RRS\\RR4SAjr]S=S>S?\RRS\P4SBjr^S=S>S?\RRS\\R4SCjr`S=S>S?\RRS\P4SDjraS=S>S?\RRS\\R4SEjrbS#\RRSF\\RR/\\RR\P44S\RR4SGjrcSH\V\RRS\V\RR4SIjrd\ SJ5reS#\RRSS4SKjrfS#\RR4SLjrgSSH\V\RRS\\RR4SMjjrhSH\V\RRS\S4SNjriSH\V\RRS\V\RR4SOjrjSP\RRSQ\RRSS4SRjrkS#\RRS\RRSS4SSjrmST\RRS\\RR4SUjrnS\RRS\R4SVjrpSWrqS#\RRSS4SXjrrS#\RRSYSS\RRSZ\K\L\4SS4 S[jrsS\\KS]\PS\K4S^jrtS#\RRS\RRR44S_jrw\ S\RR4S`j5rxSaSbS\P4ScjrySaSbS\P4SdjrzSaSbS\P4Sejr{SaSbS\P4Sfjr|SSgjr}Shr~SSijr\GR"SjSk9S\Sb4Slj5rSm\GRGRS\Sb4SnjrS\Sb4SojrSSpjr\ Sq5rSr\S\RRS\K\L\\R\RR444SsjrS5\Y\RRSS4StjrS5\Y\RRSS4SujrSvrSwr"SxSy\RR5rSzrg)N)Iterable)contextmanager)ismethod Parameter)AnyCallableOptional TYPE_CHECKINGUnion)detect_fake_mode) FakeTensorFakeTensorMode)FunctionalTensor)#first_call_function_nn_module_stack)insert_deferred_runtime_assertsConstantAttrMap) OperatorBase)ExportedProgram)ExportGraphSignature)CustomObjArgument InputKind OutputKind)_deregister_pytree_flatten_specregister_pytree_flatten_spec) _deregister_pytree_node_register_pytree_nodeContext FlattenFuncFromDumpableContextFn GetAttrKeyKeyPathkeystr MappingKey SequenceKeyToDumpableContextFntree_flatten_with_path UnflattenFuncp_b_c_obj_tokenFreturncSSKJn U"5nURVs1sHBnUR[R :XdM#UR (aM6URiMD nnUR5HnupxXv;aM Un URS5GtpU Hn [X5n M U RRU S5 [XU5 URX5 Mp U$s snf)Nrr.)(torch._export.passes.lift_constants_passr input_specskindrBUFFER persistenttargetitemssplitgetattr_bufferspopsetattradd) graph_signature constantsmodrconstant_attrsspecnon_persistent_buffersnamevalue_modatomsattratoms M/var/www/fran/franai/venv/lib/python3.13/site-packages/torch/_export/utils.py_collect_and_set_constant_attrsrL@sI$&N$///D 99 (( ( 15  / !(   ) zz# D4&D $%E"5') #s"C+C+C+rAcrSSKJnJn [5nURR HnUR S:XdM[RRRXR5n[U[R5(dMiURU;dM{URU;dM[RRRXR5 U"XpURURS5 UR!UR5 M UR#5 U$)Nr) _assign_attr _AttrKindget_attrF)torch.export.unflattenrNrOsetgraphnodesoptorchfx graph_module _get_attrr7 isinstanceTensor _del_attrr5r> recompile)rA state_dictrDrNrOtemp_registered_constantsnoder7s rK_register_constants_as_buffersra^s? #  77j XX**44S++FF&%,,//KKz1KK'==HH))33CE dkk9;K;KUS-11$++> MMO $$rc@URH1nURU;dM[RUlSUlM3 UR HLnUR[R:XdM#URU;dM5[SURS35e U$)Nz Constant z< is mutated in the forward method. Pls register it as buffer) r3r7rCONSTANT_TENSORr4r6 output_specsrBUFFER_MUTATION RuntimeError)sigr_rCs rK7_override_graph_signature_for_temp_registered_constantsrixs ;;3 3!11DI"DO    II33 3 88DKK=(de  ! Jrbnew_sigcZURVs1sHBnUR[R:XdM#UR(aM6UR iMD nnURH<nUR[R:XdM#UR U;dM5SUlM> U$s snf)NF)r3r4rr5r6r7)old_sigrjrCrDs rK/_overwrite_signature_for_non_persistent_buffersrms '''D 99 (( ( 15  ' ## 99 (( (T[[._getattrsD"- D&A= =q  rb call_moduleTF)recurseremove_duplicater1rP call_functioncustom)rVrW GraphModulestrrSrTr7metarUrZnnModulenamed_parameters named_buffers_opsHigherOrderOperator _input_nodesproxy_COPY_META_FIELDS) rAparams_buffers_to_node_metarvr`r7r submodulerE_argentrys rK_collect_param_buffer_metadatars#%!,,!! yy 77m # -I)UXX__55(99 5 : GDHL t0CD  )66 5 7 GDHL t0CD  77j  -Ii)=)=>>6:F3 77o %j KK77/ / ((66Z'!&!A!A H,$ =MQ[7 CEJ "B)7 H '&rbrgmcdUR5H'nURSS5 URSS5 M) URRHnURS:XdMUR UR ;aGUR UR nXP;a)XR5HupgXtRU'M UR UR;dMURUR nX;dMXR5HupgXtRU'M M g)ze Given that we collected param'buffer metadata before, we put them back in newly traced graph module nn_module_stackN stack_trace placeholder) valuesr<rSrTrUr7inputs_to_parametersr8rinputs_to_buffers) rrrjmetadatar` param_namekv buffer_names rK)_populate_param_buffer_metadata_to_new_gmrs0668 &- ]D)9 77m #{{g:::$99$++F < ; G M M O'( ! !P{{g777%77 D = ; H N N P'( ! !QrbcfURRVs/sH1nURRSS5cM"URSPM3 nn[ U5nUb UR $UH:n[ U[R5(dM$URR s $ gs snf)Nval) rSrTrget_detect_fake_mode_from_gm shape_envrZrVSymIntr`)rr`vals fake_moders rK_get_shape_env_from_gmrsHHNN "D 99== %  %"   *"-I"""  a & &66## # s !B.B.name_map orig_namerEis_placeholdercnX R5;a[R"SU5nU(a2U(d+URS5[ URS55pROSnUSUS-3=o`R5;a$US- nUSUS-3=o`R5;aM$X`U'X$X U'X$)z Renames nodes to avoid name collisions, with suffixing. name_map: map from original name to new name orig_name: mapping key name: candidate name (potentially suffixed, e.g. mul_2) is_placeholder: if the node is a placeholder, avoid detecting suffix z (.*)_(\d+)rr)rrematchgroupint)rrrErrndup_names rK_rename_without_collisionsrs   - kk!nc%++a.&9!A"V1QUG,,x1BB FA#V1QUG,,x1BB&  #  rbkey_pathcUSn[U[5(deURS:XaS[USS53$USn[U[5(de[ U5SSnU[USS53$)z}For a given index into the flat_args, return a human readable string describing how to access it, e.g. "*args["foo"][0].bar" rz*argsrNr)rZr%idxr#r$r~)rargs_kwargs_key_path kwarg_keyrEs rK get_keystrrs$A; *K 8 88 81$vhqrl+,--QK )Z00009~a#x|,-..rbsymintrkeypathic jSSKJn [U[R5(a%UR R R(a[X5(agSSKnSSK J n SSK J n [U[R5(Ga2[UR R R5S:XGa[[!UR R R55n X;aVUR R R#U5n X:wa+[%U5n Ub U SUS3- n ['SU S U S U35eO[UR R UR(5(a[+U5X:'OU "UR-UR R U5U 5n U cC[%U5n Ub U SUS3- n ['S U S US UR R SU S3 5e[+U S5X:'UR R U;aU"X R R 5upUS:a0X:a+[%U5n Ub U SUS3- n ['SU SUS U35eU[.R0:a1X:a+[%U5n Ub U SUS3- n ['SU SUS U35eggg[U[R5(a&UR R R(dgX:wa,[%U5n Ub U SUS3- n ['SU S US US35eg)Nr) _IntWrapper)_convert_range_to_int) try_solverz.shape[]Expected input at  to be equal to , but got zExpected input z = z to be of the form z, where z is an integerrz to be >= z to be <= zt. If you meant for this dimension to be dynamic, please re-export and specify dynamic_shapes (e.g. with Dim.DYNAMIC))torch.export.dynamic_shapesrrZrVrr`expr is_numbersympy@torch._export.passes.add_runtime_assertions_for_constraints_passrtorch.utils._sympy.solverlen free_symbolsnextitersubsrrgSymbolrEqmathinf)rrrange_constraintsunification_maprrrrrrsymbol existing_dimpathsolutionmin_valmax_vals rK _check_symintr/s58 3 %% '' c ' ' 3&%,,''C 0@0@0M0M,NRS,Sd6;;++889:  $!;;++00AL"!'*=gaSN*D"(.>|nJWZV[\ #&++**ELL99 +.c('$UXXfkk.>.>%DfM#%g.D}'!A.&)$s3%8''-{{'7'7&8X /2(1+.>O+ ;;  0 04!++"2"23  G{=%g.D}'!A.&,TF*WIZPSuU!=%g.D}'!A.&,TF*WIZPSuU !" 1* FELL ) )&++2B2B2L2L '" = gaSN "D &6vhjN& &  rbinput_placeholdersc SSKn[U5[U5:wa$[S[U5S[U5S35e0n[X5GHuupVnURR S5n[ U[5(a[ U[R5(d#[S[U5S[U535e[UR5[UR5:wa2[S[U5S URSURS35e[[URUR55Hun up[XX$XY5 M GM[ U[[ ["45(a=[U5[U5:wdXh:wa[S[U5S US U35eGM{[ U[R$5(dGM[XX$US5 GM g) Nrz&Unexpected number of inputs (expected z, got )rrz to be a tensor, but got z,Unexpected number of dimensions in input at z.shape (expected rr)rrrgziprrrZr rVr[rtypeshape enumeraterrfloatr~r) rflat_args_with_pathrrrrrr`node_valjarg_dimnode_dims rK"_check_input_constraints_for_graphrs 3'9#:: /01>WX\]`XaWbc8>>"c#))n4"B:hCWBXY!!) 0syykD +4C 8>>4R*S&&G'88+T 3s"3 4 4CyDN*co"(H)=(>>NxjXbcfbgh/>%,, / / 08T 3"Nrb)serialized_type_nameto_dumpable_contextfrom_dumpable_contextreturn_none_fieldscls flatten_fn unflatten_fnrrrrc ^^^[R"T5(d ST35eS[S[[[[ 44U4SjjnS[ [S[ S[4U4SjjnS[S[[[[ 44U4Sjjn TbTOUmUbUOUnUSLUSL- (a[S TS 35e[TTUUU UUS 9 g) Nz7Only dataclasses can be registered with this function: objr/c>/n/n/n[R"U5HbnUR[XR5peUcT(a$UR U5 UR U5 MQUR U5 Md XU/4$N) dataclassesfieldsrEr:append)r flattened flat_names none_namesfrErrs rKdefault_flatten_fn=register_dataclass_as_pytree_node..default_flatten_fns}   ##C(AVV 4#"4  %!!$'!!$' )z222rbrcontextc j>Uup#T"S0[[X 55D[RU5D6$)N)dictrfromkeys)rrrrrs rKdefault_unflatten_fn?register_dataclass_as_pytree_node..default_unflatten_fns1!( PT#j12PdmmJ6OPPrbc>T"U5unup#[X!5VVs/sHupE[U5U4PM snnU4$s snnfrrr$)rrr _none_namesrrrs rKdefault_flatten_fn_with_keysGregister_dataclass_as_pytree_node..default_flatten_fn_with_keyssC/9#, ,J/2:/IJ/ItqA"/IJJVVJs<z7Both to_dumpable_context and from_dumpable_context for z must be None or registered.rflatten_with_keys_fnrr) r is_dataclassrtuplelistrr ValueErrorr) rrrrrrrrrr s `` ` rK!register_dataclass_as_pytree_noders   # #   G @FG  3 3d3i.@(A 3QXc]QWQQW#W%S 78J2KW *5;MJ#/#;TU$rr)r` split_maps rK"sequential_split..Us Yt_rbT)keep_original_orderkeep_original_node_name)torch.fx.passes.split_moduler%rSrT_codegenr])rr#r%split_idr`new_gmr's @rKsequential_splitr0@s:IH $   MH" $   $ $ FHH--FLL  MrbrTcNUVs/sHo!"U5(dMUPM sn$s snf)z:Returns the nodes that match the node_call_back as a list.rrTr#r`s rK nodes_filterr3_s!" ;UTnT&:DU ;; ;s""c#8# [nSqSv Uqg!Uqf=f7fNT)_DISABLE_ATEN_TO_ASSERTION_PASS)orig_vals rK$_disable_aten_to_metadata_assertionsr8ds&/H&*#3 *2'('s c SSKJnJn [(ag[R R RR[R R RR[R R RR/nURRGHnURU;dMURR[R R RR :Xa,UR"SURR"S:XaMUR"SR$R'S5=ncMURR)U5 U"U[*R,"UUR$R'S5S95 URR/[R R RR UR"S4URURUR0S.S9 SSS5 SSS5 GM g!,(df  N=f!,(df  GM=f)Nr_node_metadata_hook_set_node_metadata_hookrrr)dtypedevicelayoutargskwargs)(torch._export.passes._node_metadata_hookr;r<r6rVopsatentor?r> dtype_layoutrSrTr7prev_assert_tensor_metadatadefaultrBrrinserting_before functoolspartialr{r@)rr;r<aten_to_variantsr` tensor_vals rK$_insert_aten_to_metadata_assert_passrQos '&      &&  ;;* *   EIINN$J$J$R$RRIIaLDIINN1$55"iil//33E:: GXX..t46M%%+$(IIMM-$@7HH** >>FF"iil_%/%5%5&0&7&7&0&7&7 +7447744s%6I:A8I2I I I I) c SSKJnJn SSKJn [ R RR(deSnU"U[R"X%S95 [U5nU(a"[UUS[UR53SS9 SSS5 [U5 UR!5 U"U5UlX4$!,(df  N9=f) Nrr:)_graph_output_nameszUFile "torch/fx/passes/runtime_assert.py", line 24, in insert_deferred_runtime_assertsr=zexported program: T)export)rDr;r<4torch._functorch._aot_autograd.input_output_analysisrSrV_dynamoconfigdo_not_emit_runtime_assertsrMrNrrrrSrQr] user_outputs)rr?r;r<rSrrs rKapply_runtime_assertion_passrZsY ==   ; ; 1 %  !!"5O /r2I/()LRXX)V(WX    -R0LLN#6r#:O %  s 5B== C cV[X(aUOS5n[U5S:aUS$g)z Returns the first node that matches the node_call_back. If no node matches, returns None. When node_call_back is None, returns the first node in the node list. cgr5rr`s rKr(nodes_first..sQUrbrN)r3r)rTr#rets rK nodes_firstr`s, unDU VC 3x!|1v rbc*[[X55$)z:Returns the number of nodes that match the node_call_back.)rr3)rTr#s rK nodes_countrbs |E2 33rbc(UH nU"U5 M U$)z Sequentially visit the nodes list and invoke node_call_back on each element. Returns the nodes list after the node_call_back is invoked on each element. rr2s rK nodes_maprds t Lrbold_nodenew_nodecURU5 URR5 URR U5 g)z- Replace all uses of old_node with new_node. N)replace_all_uses_withusersclearrS erase_node)rerfs rK node_replace_rls4 ""8, NN NNh'rbc[U[RR5(aN[ US5(a<SUR ;a+UR R SUR S05 gggg)Nrr|)rZrVrWr}hasattrrupdate)rrAs rK_update_gm_meta_if_possiblerps_3,,-- C   #((8"456 ! ! .rb call_mod_nodec b^ URS:XdeURRnUce[UR[ 5(de[ XR5nSURR5nSURR5nURRVs/sHoURS:XdMUPM nn[X0R5H;upx[U[RR5(de[Xx5 M= URRU5 UHnURRU5n URS:Xa~U Rn [!X5(a3Sn SU 3n [!X5(aU S- n SU 3n [!X5(aMXl[#XR[ X%R55 [XY5 M [%U5S :GaC[%U5S:Xa[%US R5S:XdeUS RS m [T [RR5(a'T R&R)5 [UT 5 O[T [*[,45(aT H!nUR&R/US 5 M# [1[+UR&R355S 5n [5U U 4S j5 URR7U5 O3[9S [;T 5S 35eURR7U5 SSS5 UR=5 UR?5 U$s snf!,(df  N5=f)z Inline the submodule of the given node into the parent module. Note: we only support the case where submodule takes tensors inputs. rxNc3H# UHoRS:XdMUv M g7f)rNrU.0r`s rK node_inline_..s K.D'']2J44." "c3H# UHoRS;dMUv M g7f))routputNrtrus rKrwrxs  +ww>W/W+ryr{rPrsubmod_rchURS:H=(a UR[R:H$)Nr{)rUr7operatorgetitemr]s rKr(node_inline_..s)O!;"8 x'7'77"8rbc<>[UTURS5$Nr)rlrB) get_item_node new_outputs rKr(r%s -%"=#5#5a#89+rbzUnsupported output type z2. Expect it to be a Node or a list/tuple of Nodes.) rUrS owning_modulerZr7r~r:rTrrBrVrWNoderlrL node_copyrnr=rrirjrrr<r3keysrdrkNotImplementedErrorrdelete_all_unused_submodulesr])rqrsub_gmphsbodyr`r{phrrfnew_target_namerget_item_usersrs @rK node_inline_rs   } ,, ,    * *B >> m**C 0 00 0 R-- .F KFLL.. KC ++ D &||11 I1tWW5Hd1F Is../#uxx}}----b0  " "= 1Dxx))$/Hww*$"*//2//A(/smO!"66Q,3A3-""66#2OOWV[[-IJ $ ) v;?v;!#F1INN(;q(@ @@*J*ehhmm44  &&(mZ8Ju 66&DJJNN6!9-'".,,11348" "##..}=).tJ/?.@@rs    * *= 9e 2h##%LLN IyJ 2 1s 7NNA>N F,N  N.c[R"UR5nURSnS[R 0n/nUR 5HjupV[URU5Vs/sHowRPM nnUH0n U S:XaM UR[R"X55 M2 Ml [R"US9$s snf)z Get source code and parse argument names using AST. The function returns a signature of the forward() function. # TODO: Directly provide inspect.signature compatible TS-d module. rrBself) parameters) astparsecoderrPOSITIONAL_OR_KEYWORDr8r:rBrrinspect Signature) rAast_mod ast_func_def arg_type_map param_listarg_type param_typea arg_name_listarg_names rK&_get_torch_jit_trace_forward_signaturer7sii!G$+LLOLI;;>"c)ns;7G&GG  J G  , Ec *44 D DDrbc/nURRGHvnURS:XdM[UR[ R R5(dMKURRS:Xa]URup4pVUR[XR5U45 UR[XR5U45 MURRS:XaFURSURSSpUR[XR5U45 GM"URRS:XdGM?URupn UR[X R5X-45 GMy UHup|0n [URR5HoupU[U 5:a6XRXR'XR=UlUlMJ[XRUR5UlMq [!U5 UR#5 M g)a Propagate placeholder names from the top-level graph into HigherOrderOp subgraphs, and handle collisions with non-placeholders by count suffixing. Different HOO subgraph types have different input schemas, so we first enumerate them and gather the top-level named placeholder nodes. r{condwrap_with_set_grad_enabledrrNmap_impl)rSrTrUrZr7rVrr_name_argsrr:rrrEr_name_hoo_subgraph_placeholdersr])rsubgraph_ph_tuplesr`r true_graph false_graph cond_argssubgraphr body_grapharrayrBhoo_phsrrs rKrrcsRT 77o %* KK77+ + {{  F*8< 5{"))727H7H+I9*UV"))727I7I+JI*VW""&BB $ 1 tzz!"~#"))72+G*MN""j0*.**' 4"))R!2!23U\B&0#% !5!56GA3w<&-joo#*1*//9 DK6xDIIV 7 (10rbexport_graph_signaturer@c J ^0n[U[RR5(aJURR H0nSUR ;dMUR SXxR'M2 Sn Sm0n [X#U5n [U 5upURVs/sH9nUR[R:XdM#URRPM; nn[X5HOuunnnU(dM[!U U["[RSR%U4SjU55-SS9 MQ URHnUR[R:XaM#UR[R&:XaSnO U "UR(5R+5n[,R."S SU5n[!U URR["URU-SS9 UU;dMUUXzURR'UU M URR H5nUR0S :XaM[!XRUR5 M7 URR GHnUR0S :XaURU ;deXR=UlUlURU;aYUR R3S5cXxRUR S'O!UR SXxR:Xde[UR S [45(a URUR S lMMURU ;dGMXRUlGM [7U5 UR95 URHnURRU ;deXRRURlUR[R::XdMfUR(U ;dMxXR(S S UlM UR<HnURRU ;a'XRRURlUR[>R@:XdMdUR(U ;dMvXR(UlM [CURE55HnUUnUU ;dM[U[RF5(aM1U UnUU:wdM>[,RH"SU5(dM\U["[R:U-:wdM|UUU'UU M g s snf)a This pass is run at the end of _export_non_strict() to assign better placeholder node names: - User inputs: These follow the signature of mod.forward(), e.g. forward(x, y) produces nodes x, y. For nested inputs from dictionaries, lists, tuples, or dataclasses, the names are a concatenation of the path to the tensor. e.g. x = { 'a': torch.randn(), 'b': [torch.randn(), torch.randn()] } produces nodes x_a, x_b_0, x_b_1. - Parameters/buffers/constants/custom objects: These follow the FQN of the object, prefixed by "p", "b", "c", "obj" respectively. e.g. self.bar.l0.weight produces "p_bar_l0_weight". - Effect tokens: These are named token, token_1, ... r|cURS5(aU[S5SnO$URS5(aU[S5Sn[R"SSU5nU$)N L__self___self_ [^a-zA-Z0-9]r) startswithrrsubxs rK _strip_name,placeholder_naming_pass.._strip_namesY << % %#l#%&A \\' " "#g,.!A FF?C +rbcP[U[5(a-[R"SS[ UR 55nU$[U[ 5(a[ UR5$[U[5(a UR$[S[U5SU35e)NrrzPytree key of type z not handled for ) rZr$rrr~keyr%rr!rErgrrs rK_extract_pytree_key4placeholder_naming_pass.._extract_pytree_keys} a $ $SZ8AH ; ' 'quu:  : & &66M!4T!WI=NqcRS Srbrc3P># UHnT"U5R5v M g7fr)lower)rvrrs rKrw*placeholder_naming_pass..s#L8a.q177998s#&T)rr)rrNrz arg(\d+)_1)%rZrVrWr}rSrTrrErr'r3r4r USER_INPUTrrrplaceholder_prefixesjoinTOKENr7rrrrUrrrr] CUSTOM_OBJrerUSER_INPUT_MUTATIONrrr[r)rrrArrfake_params_buffersr@ custom_metar`rr combined_argsrrrCuser_input_namesarg_path_arguser_input_name base_namerEconstantnew_namers @rKplaceholder_naming_passrsq6#%K#uxx++,,IIOOD499$)-8)< II&$ T "H.ckJM3MB+666D 99 ,, ,  6.11D-W)4/ ? &$Y%9%9:((L8LLM#  .X'22 99 ,, ,  99  'I#DKK0668IFF?C; "  HHMM  +i 7   #4?y3IK/ 0I&)34 77m # "8YY B  77m #99( ((&.yy&9 9DI yyK'99==*2*5ii*@DIIh'99X.+ii2HHHH$))E*,=>>(,  % %? YY( " +DI"$B'LLN'22xx}}((( / II-- -$++2I";;/3DK 3'33 88==H $$XX]]3DHHM 99 66 64;;(;R";;/DK 4Y^^%&T? 8 J ell% %  ~HD HH]D11 4Y5I5I JT QQ&. (#dO'ss &"V  V r^in_placec0U(a<UR5H&up#[US5(dM[XS5 M( U$0nUR5H<up#[US5(a"UR5R 5XB'M8X4U'M> U$)z If `in_place` is false, return a new copy of `state_dict` with "proxy" removed from `v.__dict__`. `v` is the values in the dictionary. If `in_place` is true, modify `state_dict` in place. r)r8rndelattrdetachclone)r^rrrnew_state_dicts rKremove_proxy_from_state_dictr/s $$&DAq'"" w/'$$&DAq'""$%HHJ$4$4$6!$%q! ' rbc/n/nURRGHnURS:XaXSUR;aHURSnUb4[ U[ R 5(aURU5 MhMjMl[U5S:XdM}SUR;dSUR;dMSnSUR;aURSnOSUR;aURSnUcM[ U[ R 5(dGMURU5 GM [X-5$)a For a given graph module, we look at the "val" of placeholder nodes to find the fake inputs. Additionally, if gm doesn't have placeholders, we further look at the "example_value" or "val" of other nodes. If no fake mode is found, we return None for fake_mode. rrNr example_value) rSrTrUrrZrVr[rrr )r fake_inps fake_valsr`fake_vals rKrrDs%'I$&I 77m #(:yy'H# 8U\\(J(J  *)K# ^q tyy (ETYY,>H$))+99_5$))#99U+# 8U\\(J(J  * I1 22rbc# # [UR5n[UR5nURR5 URR5 Sv XlX lg!XlX lf=f7fr)r_state_dict_hooks_state_dict_pre_hooksrj)rAstate_dict_hooksstate_dict_pre_hookss rK_disable_load_state_dict_hooksrbsn,01F1F,G04S5N5N0O!##%9 0$8!!1$8!sAB"A3& B3BBrUrc[RRUR5[RRR 5=(d1 [RRR UR ;$r)rV_C%_dispatch_has_kernel_for_dispatch_keyrE DispatchKeyCompositeImplicitAutograd py_kernelsrts rK _is_cia_oprosX 66 GGIuxx++EE  K 88   9 9R]] J rbc<[U5=(a [U5$r)_check_valid_to_preserverrts rK_is_preservable_cia_oprxs #B ' :JrN:rbcLUR5RS5SS:H$)N::rrF)rEr9rts rK _is_aten_opr|s! 779??4  #v --rbc"[U5(+$r)rrts rK _is_custom_oprs2 rbc [RRS5nUHn[UR S55up#UR S5n[ U5S:Xd[ U5S:XdeUSnSn[ U5S:XaUSn[ [ [ [RU5U5U5nM g) zu Utility function to query C++ dispatcher to get the all possible CIA ops and populate them into torch.ops namespace rrr1rrrrKN)rVr,_dispatch_get_registrations_for_dispatch_keyrr9rr:rE)cia_opsrU namespaceop_name split_listop_overload_namers rK_materialize_cpp_cia_opsr s hhCC#G "288D>2 ]]3' :!#s:!';;;Q-$ z?a )!}  GGEIIy97CEU Vrbc[$)zU This is an special marker that tells our infra that we shouldn't decompose this op. )NotImplementedrAs rK_special_op_to_preserve_ciars  rbcSSKJn U"U5(agU[R;ag[ US5(dg[ UR RVs/sHo"RcMUPM sn5nUS:g=(d UR RnU(ag[RRUR55(dggs snf)Nr)#_should_decompose_because_unsafe_opF_schemaT) torch._decomprr metadata_fnsrnrrr alias_info is_mutablerVr_dispatch_has_kernelrE) op_overloadrrris_mutating_or_aliasings rKrrsA*;77&333 ; * *''11N1q\\1NJ)AoO1D1D1O1O 88 ( ()9)9); < <  Os C-Cr)maxsizecH[[RR5$r)(_collect_all_valid_cia_ops_for_namespacerVrErFrrbrK-_collect_all_valid_cia_ops_for_aten_namespacers 3EIINN CCrb op_namespacec[5 [5nUHSn[X5nUR5H1n[X45n[ U5(dM UR U5 M3 MU U$r)r rRr: overloadsrr>)rrrU op_packetoverloadrs rKrrsbeGL- !++-H!)6K%k22 K(. Nrbcn[5n[RRHnUS:wav[ [RU5(de[ [RU5n[ U[RR5(aU[U5-nM}MU[5-nM U$)a This is an util function that gets the all CIA functional ops. The algorithm is in 2 steps: 1. We first query C++ dispatcher to get the list of CIA ops and then we call getattr on torch.ops.aten to lazily populate them. 2. Sometimes, handful of ops have CIA registered in python dispatcher but not on the C++ side, these can't be caught at the first step. So we walk again to get the final list. Note that the output of this function should never be modified rF) rRrVrE_dirrnr:rZr _OpNamespacerr)rop_namespace_namers rK_collect_all_valid_cia_opsr'seG"YY^^  &599&788 88"599.?@L, (?(?@@CLQQA DF FG, Nrbc [RRRnXR;aE[ URU[RR5(dURU$Sn[ R"X S9$)NcNUSnUS [RRRn[RR UR 5[RRR5(aUR "U/UQ70UD6$[SUS35e)Nkernel Expected z) to have CompositeImplicitAutograd kernel)rVrrrrrE_op_dkAssertionError)rBrCr*dks rK_special_op_to_decompose_cia9_get_decomp_for_cia.._special_op_to_decompose_cias! 8 XX ! ! ; ; 88 9 9 KKM588//II  ==5d5f5 5 F8#LM rb)r*)rVrrrrrZrMrN)rUr.r/s rK_get_decomp_for_ciar1sj    7 7B ]]:bmmB.?AUAU#V#V}}R      9 EErbc## [RRn[RRnS[RlS[RlSv U[RlU[Rlg!U[RlU[Rlf=f7fr5)rVcompiler_is_compiling_flag_is_exporting_flag)old_compiling_flagold_exporting_flags rK_compiling_state_contextr8sx::::?,0),0) ,>),>)-?),>)s5C.B&+C,B==Crc"0[URSS95E[URSS95En0n0nUR5HDupV[ U5U;aU[ U5nOUR USS9nXt[ U5'XsU'MF U$)NF)rzT) static_shapes)rrrr8id from_tensor)rrAparams_buffersfaked_params_buffersmemorrF fake_tensors rK_fakify_params_buffersrAs s##U#; < s  % 8 9N "$D$**,  e9 r%y/K#//T/JK)EO$/S! - rbc .^^ ^ [T[RR5(deSSKn"U U4SjSUR 5m S[ S[[[ [44U 4SjjnS[[ S [S[ 4U 4S jjnS[ S[[[ [44U 4S jjnUm UnTRS -TR-nS nU 4Sjn[TT UUUUUS9 S[[ 4U 4Sjjn [TU 5 g)a Registers a module as a valid input type for :func:`torch.export.export`. Args: mod: the module instance serialized_type_name: The serialized name for the module. This is required if you want to serialize the pytree TreeSpec containing this module. Example:: import torch class Module(torch.nn.Module): def __init__(self): super().__init__() self.linear = torch.nn.Linear(3, 3) def forward(self, x): return self.linear(x) torch._export.utils.register_module_as_pytree_node(InputDataClass) class Mod(torch.nn.Module): def forward(self, x, m): return m(x) + x ep = torch.export.export(Mod(), (torch.randn(3), Module())) print(ep) rNc>>^\rSrSrUU4SjrSrU4SjrSrU=r$)=register_module_as_pytree_input_node..PrototypeModuleiWc>[TU]"U/UQ70UD6 [U[RR 5(de[ US5(aeTUlg)N _proto_cls)super__init__rZrVrrrnrF)rmrBrC __class__rs rKrHFregister_module_as_pytree_input_node..PrototypeModule.__init__XsO G Q 0 0 0a11 11t\22 22!DOrbc4URUR:H$rrF)rothers rK__eq__Dregister_module_as_pytree_input_node..PrototypeModule.__eq__^s??e&6&66 6rbc>T"U"55$rr)rr?PrototypeModules rK __deepcopy__Jregister_module_as_pytree_input_node..PrototypeModule.__deepcopy__as"46* *rbrM) __name__ __module__ __qualname____firstlineno__rHrOrS__static_attributes__ __classcell__)rJrRrs@rKrRrDWs "  7 + +rbrRrr/c>[UR55n[UR55n0UEUEn[UR 55[UR 55T"U5/4$r)rrrrrr)rrrr=rRs rKr@register_module_as_pytree_input_node..default_flatten_fndss 4 4 67S..01 >,> >N))+, $$& ' C /   rbrrc >^Uup#UbU"5c [S5eU"5nT ceT "U5upVS[RR4U4Sjjm[ S[ X555(aY[RR RRU[[ X 55SSS9 T"U5nSSS5 U$UnU$!,(df  W$=f)Nz!Module has been garbage collectedrAcP>[R"U5nURR5VVs0sH0up#[R"U5[R"U5_M2 snnUlUR5HupE[ XT"U55 M U$s snnfr)copy__dict__r8named_childrenr=)rAr_rrrEchild copy_modules rKrcWregister_module_as_pytree_input_node..default_unflatten_fn..copy_moduleys{))C.CCF<.default_unflatten_fn..s=&.default_unflatten_fnms! ;#%-BC Ce%%%!#  UXX__  =c&&<= = =))??T#j12T@"#&  C   s ; C C!c>T"U5untp#[X!5VVs/sHupE[U5U4PM snnU/UQ4$s snnfrr)rrrrBrrrs rKr Jregister_module_as_pytree_input_node..default_flatten_fn_with_keyssW)3C& &J/2:/IJ/ItqA"/IJ M M   Js>r1cTUtp[R"U/S/[U5-Q5$r)jsondumpsr)rrrs rKrAregister_module_as_pytree_input_node..to_dumpable_contexts*zz44D6CF?455rbc>[R"U5nT"[RR 55US'U$r)rsloadsrVrr)dumpablesrRs rKrCregister_module_as_pytree_input_node..from_dumpable_contexts/ JJx uxx01!rbr c>>T"U5up#X1R:XdeU$r)r)rrCflatsrrs rKdefault_flatten_fn_specEregister_module_as_pytree_input_node..default_flatten_fn_specs##C,,&&& rb) issubclassrVrrweakrefrnrrrrrrVrWrr) rrrrr rrrrr}rRrs ` @@rK$register_module_as_pytree_input_noder3s@ c588?? + ++ + + +'++ +  d3i.@(A Xc]W8 # %S 78J2K $J'L>>C/#2B2BB6  19/3d3i ! rbc0[U5 [U5 gr)rr)rs rK&deregister_module_as_pytree_input_nodersC #C(rbc[U[RR5(d SUS35e[U[RR5(d SUS35eURUlUR Ulg)Nr+z to be a nn.Module)rZrVrr _parametersr;)srcdsts rK _sync_staters     + 3%)*+      + 3%)*+ ooCO<[TU]5 [URU5 Xlgr)rGrHr__self__r)rmethodrJs rKrH_WrappedMethod.__init__s! FOOT* rb)r)rUrVrWrXrHrYrZ)rJs@rKrrs rbrcN[U5(d SUS35e[U5$)z Wrap a method as a module so that it can be exported. The wrapped module's forward points to the method, and the method's original module state is shared. r+z to be a method)rr)rs rK wrap_methodrs<   + 6(/*+  & !!rb)r/r)rhr)rlrrjr)Fr)NN)r/N)rr)rUr)rr_rrMrrsrr~rcollections.abcr contextlibrrrtypingrrr r r rV torch._guardsr torch._subclasses.fake_tensorr r#torch._subclasses.functional_tensorrtorch.fx._utilsrtorch.fx.passes.runtime_assertrr2r torch._opsr torch.exportrtorch.export.graph_signaturerrrrtorch.fx._pytreerrtorch.utils._pytreerrrrr r!r"r#r$r%r&r'r(r PARAMETERr5rdrrrr6rLrWr}rarirmrr~rrrboolrrrrrrrrrrrrrrr[rrr"r0r3r8rQrZr`rbrdrlrrprrrrrrr _subclassesr@rrrrrrr rr lru_cacherRrrr%rr'r1r8rArrrrrrrrbrKrs<   $%'@@ *D@?JH',AQQ$"  d t & OOW #(<%   %4 (  # .D  6'(<(<6'c3h6'r)!%c3h) )$) )8 $uxx33 $&! 38n  ://S/0 [ #u||# $[ [  [ } [  [ |)UXX]]+) )\)-,04 +/9==A$4 c4%4=)4 #3- 4 ""56 4$$9:44 4nE'Euxx}}EE    ((-- ehh  !  B(B B$B  ((--ell&R R ((--R R  ((--ell$ ehhmm_eEHHMM44G.HHI XX><UXX]]+<UXX]]@S< 33'UXX-A-A'd'TUXX%9%9D04     ehhmm 4tEHHMM*4s4 T%((--(T%((--=P(EHHMM(UXX]](t(7EHH$8$87uxx7SW7L L(588;O;O2PL^44GDUDU42E&'(<(<''T_$ _$2_$ _$CH~_$ _$DTTd*3 3 ""113< 9 9 9>d;~;$;.N.t.n W.4 QDs>7JD D**))"C$78F: ? ?    #uU\\588#5#556 67 *}d588??.C}}@)UXX__0E)$)  "$UXX__ "rb