3j lSrSSKrSSKrSSKrSSKrSSKrSSKrSSKJrJ r SSK J r SSK J r SSKJr SSKJrJrJrJr SSKrSSKJr SSKrSSKrSSKJrJr SS KJr SS KJ r SS K!J"r" SS K#J$r$ SS K%J&r&J'r'J(r(J)r)J*r* SSK+J,r, SSK-J.r.J/r/J0r0 SSK1J2r2 SSK3J4r4J5r5J6r6J7r7J8r8J9r9J:r: SSK;JJ?r?J@r@ SSKAJBrB SSKCJDrDJErEJFrFJGrGJHrHJIrIJJrJJKrKJLrLJMrMJNrN SSKOJPrPJQrQJRrR SSKSJTrT SSKUJVrVJWrW SSKXJYrY SSKZJ[r[ SSK\r]\(a SSK_J`r` SSKaJbrb SSKcJdrdJere SS KfJgrg SS!KhJiri \R"\k5rl\R\R\R\R\R\R\R\RS".ru\R\R\R\RS#.rv0\uE\vErw\xR\uR55r{\xR\vR55r|S$\}S%\~4S&jr\GRGRGR\GRGR-r"S'S(\R5r"S)S*\R5r"S+S,\5r"S-S.\5r"S/S0\5r"S1S2\[5r"S3S4\R5r"S5S6\R5rg!\^a Sr]GNf=f)7a This module contains variable tracker classes for handling tensors and tensor-related operations in Dynamo. The main class is TensorVariable which represents torch.Tensor inputs and intermediate values in the FX graph. It handles tensor operations, method calls, and maintains metadata about tensor properties like dtype, device, etc. Other key classes include: - SymNodeVariable: Represents symbolic scalars (int/float/bool) used for size computation and unspecialized values - NumpyNdarrayVariable: Handles numpy array interop through torch._numpy - UnspecializedPythonVariable: Represents unspecialized Python numeric values as 1-element tensors - TensorSubclassVariable: Handles tensor subclasses with __torch_function__ overrides - UntypedStorageVariable: Represents tensor storage objects - DataPtrVariable: Handles tensor data pointer operations These classes work together to track tensor operations and properties during Dynamo's tracing process. N)IterableSequence) nullcontext)chain)NoneType)AnyNoReturnOptional TYPE_CHECKING) sym_floatsym_int)compiled_autograd)is_opaque_reference_type) OpaqueBase) is_sparse_any) guard_scalarGuardOnDataDependentSymNodehas_free_symbols is_symbolicSymTypes)is_traceable_wrapper_subclass)configgraph_break_hints variables) trace_wrapped)ObservedAttributeErrorraise_observed_exceptionTorchRuntimeError unimplemented$UnknownPropertiesDuringBackwardTrace UserError UserErrorType)_ApplyBackwardHookcall_hook_from_backward_state) GuardBuilder install_guard) AttrSource) fqnget_custom_getattrget_fake_valueget_real_value guard_if_dynobject_has_getattributeproductproxy_args_kwargsraise_args_mismatchset_example_valuetensortype_to_dtype)AttributeMutationNewValueMutationNewVariableTrackerConstantVariable)ListIteratorVariable SizeVariable)TorchScriptObjectVariable)UserDefinedClassVariable) PyCodegen) OutputGraph)InstructionTranslatorInstructionTranslatorBase)UserFunctionVariableTensorWithTFOverrideVariable)>=z<===!=isis not)rIrJrGrHvaluereturnc`[[U5=(a [RRR U5(+=(a_ [ US5=(aL [UR[R5=(a! [URURS55$)N__self__) boolcallabletorch_dynamoutilsr.hasattr isinstancerNTensorgetattr__name__rKs X/home/wildlama/miniconda3/lib/python3.13/site-packages/torch/_dynamo/variables/tensor.pyis_bound_tensor_methodr[sy  : ##;;EB B : E: & : u~~u|| 4 : ENNENND 9  c# ^\rSrSrSrSSSSSSS S S S S SSSS1\R krS\R4Sjr SSSSS.S\RRS\RS\RS\RS\S \S \S \S\S\S\S\\S4S-S \\S4S-S \S-S\S-S\SS4"U4SjjjrSSSS\S-SS4SjjrS\S-4SjrSSS \S-S!\SS4S"jrS\4S#jrS\RR4S$jrS\4S%jrS\4S&jrSSS\4S'jr\S(\RS\\\44S)j5r SSS*\S\4S+jr!SSS\4S,jr"SSS\S-4S-jr#SSS\S-4S.jr$SSS\S-4S/jr%SSS\&S-4S0jr'SSS\4S1jr(SSS\&S-4S2jr)SSS\&S-4S3jr*SSS\&S-4S4jr+SSS\&S-4S5jr,SSS\-4S6jr.SSS\S-4S7jr/SSS\4S8jr0SSS\&S-4S9jr1SSS\4S:jr2SSS*\S\&4S;jr3SSS*\S\4S<jr4SSS\4S=jr5SSS\4S>jr6SSSS?\7\S-S\8\4S@jjr9SSSASBSC\SD\7\SE\\\4S\4 SFjr:S\4SGjr;\S\8\4SJjr?SSSK\S\4SLjr@SSS*\SM\7\SSNS\4 SOjrASSSM\S\S\S-4SPjrBSSSM\S\S\S-4SQjrCSS*\SR\S-S\S-4SSjjrDSSS\S-4STjrE\ErFSSS\S-4SUjrG\GrHSSS\&S-4SVjrISSS\&S-4SWjrJSSS\&S-4SXjrKSSSSY\S-S\&S-4SZjjrLSSSS\S-S\\S\S\S-4 S]jjrMSSS^\SS_4S`jrNSSS\S-4SajrOSSS\4SbjrPS[Sc.SSSd\\-SSe4SfjjrQSSS\4SgjrRSSh\S\Si\S\8\S-4SjjjrTSSSSk\S-Sl\S-Sm\S-Sn\S-S\S-4 SojjrUSSSM\S\SSp4SqjrVSSSr\S\4SsjrWSSSM\S\SS4StjrXSSS\4SujrYSSS\4SvjrZSSS\4Swjr[SSS\4Sxjr\SSS\4Syjr]SSSM\S\S\4Szjr^\\_RSS{j55raSSS\4S|jrbSSS\4S}jrcSSS\d4S~jreSS.SSS\S\S(\S-S\S-4 SjjrfSSSK\S(\S\4SjrgSSSM\S\S\-4SjrhSSSM\S\S\-4SjriSSSM\S\S\-4SjrjSSSM\S\S\-4SjrkSSSM\S\SS4SjrlSS.SSS\S\S-S\S-4SjjrmSS.SSSSSSS(\S-S\S-4 SjjrnSSS\S\4SjroSSSM\S\S\4SjrpSSSM\S\S\4SjrqSSS*\S\S\4SjrrSSSS \\-S\4SjjrsSSjrtS\-4SjruSSSM\S\S\S-4SjrvSSSSSM\S\S\S-4 SjrwSSjrxS*\SS4SjryS\4SjrzS\4Sjr{S\|S\4Sjr}Sr~U=r$)TensorVariablez=A torch.Tensor input or an intermediate value in the FX graphproxydtypedevicelayoutndimsizestride requires_grad is_quantized is_contiguous is_nested is_sparse class_typespecialized_value _is_name_setrLcj[URRURR5$)z Get the actual value represented by this variable if computation is run using the user-provided inputs. NOTE: this runs actual tensor computation and may be slow and memory-intensive. )r,r`nodetracerselfs rZr,TensorVariable.get_real_values#djjootzz/@/@AAr\N)_sizerfrirn has_grad_fnru.kwargsc &>[TU]"S0UD6 XlX lX0lX@lXPlXlXlX`l Xl Xl Xpl Xl XlXlUc#URR R"S:HnXlg)N placeholder)super__init__r`rarbrcrdrurfrgrhrirjrkrlrvrpoprn)rsr`rarbrcrdrgrjrhrkrlrvrurfrirnrw __class__s rZr|TensorVariable.__init__s( "6"       *(*""$&  ::??-->L".r\txr@ target_clscSSKJnJn Uc [U5nURR R RS5nU"X!XT"U55nUR5Hupx[XU5 M g)Nr4)get_specialized_propsinfer_subclass_type example_value) builderrrtyper`rpmetagetitemssetattr) rsrrrrrspecialized_propskvs rZsynchronize_attributes%TensorVariable.synchronize_attributessn H  dJ ,,00A 1 M+>}+M &++-DA DQ .r\cURRRRS5nUb UR$S$)zFGet the current version of self's fake tensor, or None if unavailable.rN)r`rprr_version)rs self_fakes rZ_get_fake_version TensorVariable._get_fake_versions6JJOO((,,_= %.%:y!!DDr\version_beforehas_tensor_argcUR5nUb>Ub:XB:a4U(aURU5 URRU5 gggg)zs Sync attributes if self was mutated by an inplace operation. See Note [Inplace ops and VariableTracker metadata] N)rroutput&check_input_mutation_on_current_stream)rsrrr version_afters rZ_sync_if_inplace_mutation(TensorVariable._sync_if_inplace_mutationsV..0  &).++B/ II < >%%b--/Q*>??r\rKc &^TRTRTR[TR5TR TR TRTR[T5S. nTRSLUS'[T5(a5[T5(d%[STR555US'U$[T5(d[STR555US'[TR!55US'["R$R&R)T5(aSUS'U$[U4S j["R*R,55US'U$![a SUS'GN f=f) N) rarbrcrdrgrjrhrkrlrvFc3\# UH"n[U5(a [U5OUv M$ g7frrint.0ss rZ ,TensorVariable.specialize..5s$#9EA+a..Aa/*,ruc3\# UH"n[U5(a [U5OUv M$ g7frrrs rZrr?s*#&A&a..Aa/%rrfric3T># UHnTRUS9(dMUv M g7f) memory_formatN)ri)rxrKs rZrrKs+/@***;A@s( ()rarbrcrrdrgrjrhrkrgrad_fn ExceptionrrtuplererfrQ_C _functorchis_batchedtensor _prims_common_memory_formats)rKpropss` rZ specializeTensorVariable.specialize sp[[llll O"00!..u+ !  )#(==##E'N6 1"%((## #E'N $ELLN3E(Oxx""33E::*.o& *//"00@@/*o&  C )$)E-  )s1E==FFnamec URRRSnUR(Gd][ U5(GaLUR 5upE[ UR5U5n[ X25nX$;a4[U[R[45(deSSK J n U"XUS9$[[U55(aT[R R"R%UR&R(U5n [*R,"Xi5$[U[R R"R.5(a[*R,"Xg5$[1U5(d[2R4"X5$UR(aURR75(d[8eUR&R:UR&R<S.n [?URR@U 5n U c[8e[EU 5(a[8e[GU 5(a[8e[ X5n [OURU5n[QW 5(aSS K)J*n U"XU[U 5S 9$[WURRY[ZR\"[^R`US 955 [2R4"XU5$![Ba n [8U eSn A ff=f![Ha- [K[HUS[U 5RLSUS3/S9 Nf=f) Nrr4 wrap_fx_proxy)rr`rLG'z' object has no attribute 'argsGetAttrVariable)sourcepy_typeattr)1r`rprrr__tensor_flatten__rWrrUrQrVrrrrr_libraryfake_class_registrymaybe_to_fake_objr fake_moder<rFakeScriptObjectrPr7rsubguards_allowedNotImplementedError local_scope global_scopeevalrrr.r*AttributeErrorrrXr(r[miscrr' make_guard functoolspartialr&HASATTR)rsrrfake_valattrs_ctxr`rrfake_script_objscope_input_associated_real_valueexc real_value attr_sourcers rZdynamic_getattrTensorVariable.dynamic_getattrRs::??''8 {{{>"'.."D"D"V"VII''#177OO22CC177MMm,,&,,R??  = = ? ?% %ii++"))2H2HI / ,0 0@0@%+H ( ( /% % "#? @ @% % : ; ;% %  !=DJ!d3  "* - - -";Z8H   KK " "9#4#4\5I5IPT#U V $$R[AAM /%3 . /  $9:CCDD_`d_eefg  s*7 K1 L 1 L;LL 4MMcURb [R"XR5$URUS/05$)Ndim)rdr7rrrsrs rZmethod_attr_ndimTensorVariable.method_attr_ndims8 99 "((YY7 7##Br26 6r\c^URb [R"XR5$gr)rar7rrs rZmethod_attr_dtype TensorVariable.method_attr_dtypes$ :: !"((ZZ8 8r\c^URb [R"XR5$gr)rbr7rrs rZmethod_attr_device!TensorVariable.method_attr_device$ ;; ""(([[9 9r\c^URb [R"XR5$gr)rcr7rrs rZmethod_attr_layout!TensorVariable.method_attr_layoutr r\cxURb-[R"XRRS:H5$g)Ncuda)rbr7rrrs rZmethod_attr_is_cuda"TensorVariable.method_attr_is_cudas1 ;; ""(([[-=-=-GH Hr\cUR5(a;URVs/sHn[R"X5PM nn[ U5$UR US/05$s snfNre) valid_sizerer7rr;r)rsrrsizess rZmethod_attr_shape TensorVariable.method_attr_shapesc ??  6:ii,6?%%b,i , & &##BB7 7 ,s A%c^URb [R"XR5$gr)rgr7rrs rZmethod_attr_requires_grad(TensorVariable.method_attr_requires_grads*    )"((-?-?@ @r\c^URb [R"XR5$gr)rhr7rrs rZmethod_attr_is_quantized'TensorVariable.method_attr_is_quantizeds*    ("((->->? ?r\c^URb [R"XR5$gr)rkr7rrs rZmethod_attr_is_sparse$TensorVariable.method_attr_is_sparse& >> %"((^^< <r\c^URb [R"XR5$gr)rjr7rrs rZmethod_attr_is_nested$TensorVariable.method_attr_is_nestedr$r\c$[SSUS3S/S9 g)Nz'Tensor.retain_grad() with AOTDispatcher var_getattr z retain_gradz8`Tensor.retain_grad()` does not work with AOTDispatcher.gb_typecontext explanationhintsr rs rZmethod_attr_retain_grad&TensorVariable.method_attr_retain_grads="4& 5R  r\cURRRUS5(a&URRRUS5$g)Ngrad)r side_effectshas_pending_mutation_of_attr load_attrrs rZmethod_attr_gradTensorVariable.method_attr_grads? 99 ! ! > >tV L L99))33D&A Ar\c[R"[RRR 5R X/05$r)rTorchInGraphFunctionVariablerQr _autograd_get_data_attr call_functionrs rZmethod_attr_dataTensorVariable.method_attr_datas655 HH   - - -FB ' (r\cUR(a[SSUS3S/S9 g[RR S5$)NzTensor with grad_fn()r)z grad_fnz@Dynamo does not support tracing tensors with a grad_fn directly.r*)rvr rr9rrs rZmethod_attr_grad_fn"TensorVariable.method_attr_grad_fnsC    /&tfH5^  --44T: :r\c\SSKJn [R"U5R X/05$)Nr)_tensor_version)tensor_version_oprDrr:r=)rsrrDs rZmethod_attr__version#TensorVariable.method_attr__version s*755oFTT   r\c SSKJn U[;a[R"S5$[ R "U[5RX[ R "X5/05n[XC5(+nUR(aE[URR[R "["R$US955 [ R "X5$![[4a SnNf=f)Nr4rTFr)rall_tensor_attrsr9rr7rrWr=rUrrrr'rrrr&r)rsrrrvarret_vals rZcall_obj_hasattrTensorVariable.call_obj_hasattrs & # ##**40 0 !''G4BB?00:;RC %S::G ;;  &&%%l&:&:F  $$R11 67 G sAC%%C:9C:c^^^TRT5(aSTTR5;a[SSTST3STS3STS3/S9 O#TTR5;a[ S TS 35eTS :Xa%[ R "TTR55$[TS T3S5nUbU"T5OSnUbTR(a}TRR5(a^TS ;aUR5(dC[TR[R55 [!TRT5Ul TRb[#[$R&R(T5(a[[$R&R(T5n[#US5(a[#XUR+5S5(at[$R,R.[XUR+5S5R0;a2[2R4R7[!TRT5SS9$Uc TS:waS[ S-4UUU4SjjnU"5nUcTR9TT5nUc[:eU$)NzStrict mode banned opr) zGetattr invocation 'z"' in strict mode is not supported.zRemove `zj` from the list of banned ops by setting `torch._dynamo.config._autograd_backward_strict_mode_banned_ops`.r*zUnknown property z] during speculating backward, dynamo will insert contiguous call ahead and speculate it againr~ method_attr_)r3rg overloadsrz9Getting an inplace view on a graph input is not supported)rmsgr3rLc*>SSKJn SSKJn [R TS5nUcg[ U5[RLagURTR5T5nTRbU"TU[TRT5S9$U"TUS9$)Nr4rr)rr`rrr`) rrrrrJrrtypesGetSetDescriptorTypecreate_getattr_proxyrrr()rr static_attrr`rrsrs rZtry_generic_attr_handling=TensorVariable.var_getattr..try_generic_attr_handlingls21.224> & $E,F,FF'< >dfn =/E = =2/0F >))"d3F >% % r\c  UR(d[SSU3S/S9 UR(deURRURRS.nSn[ URR U5nUc[S SU3S /S9 [URR[R55 [U5n[R"X5$![an[SSU3S/US9 SnANSnAff=f) Nz$Unsupported call_id() without sourcezcall_id z6call_id() not supported for sourceless TensorVariable.r*rz#Error getting associated real valuezJDynamo encountered an error while trying to get the associated real value.r+r,r-r.from_excz'call_id() without associated real valuez>Dynamo could not find an associated real value for the tensor.)rr rrrrrrr'rr&ID_MATCHidr7r)rsrrrrid_values rZcall_idTensorVariable.call_ids{{ >"4&)T  {{{ii++"))2H2HI'+$ +/ 0@0@%+H ( ( / A"4&)\   dkk,,\-B-BCD23$$R22)  ="4&)1   s$ C)) D 3DD c URS:$Nr)rdrs rZhas_unpack_var_sequence&TensorVariable.has_unpack_var_sequencesyy1}r\idxesc XSSKJn SSKJn UR 5(a[ UR 5nO@URUS/05n[U[5(de[ UR5nUS:wdS5eUR 5(aUR SnOURUS[R"US5/05n[U[5(dUR5(de[U[5(aURUR 5nOUR#5nUc [%U5nO&[ U5U:XdSUS[ U5S 35e[X5(aeUV s/sHn U"[&XR)5U S 9PM! n n U V s/sH+n UR*"XUR,UR.5PM- sn $UV s/sH$n U"[1U5XR)5U S 9PM& sn $s sn fs sn fs sn f) Nr4wrap_fx_proxy_clsrCrerzCan't unpack scalar tensors.zCan't unpack a tensor of z rows into a tuple of z elements.rrr`)rr{torch_functionrDrlenrerrUr;rr7rrr_ evaluate_exprras_python_constantranger^rfrom_tensor_varrlrr) rsrrxr{rDsize_lensize_varlength dyn_lengthi base_varsrs rZunpack_var_sequence"TensorVariable.unpack_var_sequences /@ ??  499~H''FB;Hh 55 558>>*H1}<<<} ??  YYq\F))F_222q9:BJ :770022 3*o66#11"))<#668 =&MEu:' +F83I#e*U_` ' d 9 9  A"-"MMOA.s#DtMMOOts)1rrrr}rrr\r]r rJrrlistrr(rrWrQrVrUr=rr9rRrInternalTorchDynamoError)torch.distributed._functional_collectivesrrr create_proxyrr SUPPORTABLErrealize TypeErrorrr`rprrrlrTrXr0rrany)rsrrrrwrrrrrYis_base_tensor_methodrrrhandler_methodrrrealized_kwargsrher check_typer`rs rZrTensorVariable.call_method.sf @X   r " "t/K/K/M'M B&tfAdV1TF!F8D657   > ! =(a0QW  '**46 +4 7 (E4&4:2E,F O O%{{* :dkk;#GN-33B d8ST*eTFT$Z$78&   8  474L M M--44U; ; 6>vmm''@@H N .  &&#[==?2Db  9  +!$wz&!<>=&22  Z Ff$4 G!$wz&!<>5&22   $TWTF+;Dlln"Mnda1aiik>n"M'ETE_E!M ,,00A #0#?N8N>8N>> O0O++O03 PPc.UR"S/UQ70UD6$r_method_size_stridersrrrws rZ method_sizeTensorVariable.method_sizes''@@@@r\c.UR"S/UQ70UD6$)Nrfrrs rZ method_strideTensorVariable.method_strides''B4B6BBr\rc[U5nS[[S[S[4SjnUS:XaUO[R nUS:wa [ X5nO*US:Xa"UR5(a URnOSnUb#UcU"U5$[R "XR5$URRRRS5=nb}Uc:[ Xa5"5n[U5(dU"[SU555$g[ Xa5"U5n[U5(d[R "[U55$g)NroptionsrLc p[UVs/sHn[R"U40UD6PM sn40UD6$s snfr)r;r9r)rrys rZmake_const_size_variableDTensorVariable._method_size_stride..make_const_size_variables=@AB1!((6g6BFM Bs!3rerc38# UHn[U5v M g7frrrrs rZr5TensorVariable._method_size_stride..s,DVSVVV)r-rrr;r9rrWrrer`rprrrrr)rsrrr RetVariablerfakefake_rs rZr"TensorVariable._method_size_strides:3   # ,  )- $#A V^ 1 1 AA ={"1~%'..qv66JJOO((,,_= =D J{ ,.'//'u,DV,D'DEE0!,S1'//+223v;??r\cfUR5(a*[R"U[UR55$UR R RRS5=nb@UR5n[U5(d [R"U[U55$gr) rr7rr/rer`rprrnumelrr)rsrrrs rZ method_numelTensorVariable.method_numels ??  "((WTYY-?@ @JJOO((,,_= =D JZZ\F#F++&,,RV==r\c^URb [R"XR5$gr)rdr7rrs rZ method_dimTensorVariable.method_dims$ 99 "((YY7 7r\crURb*[R"XRR5$gr)rar7ris_floating_pointrs rZmethod_is_floating_point'TensorVariable.method_is_floating_points, :: !"((ZZ-I-IJ Jr\c2[R(a,[SSS/[RQ[R QS9 UR RRRS5=nb$[R"XR55$g)Nz0Encountered tensor.is_inference() during tracingrIz&tensor.is_inference() is not supportedr*r) r"fake_tensor_disable_inference_moder r FUNDAMENTALINFERENCE_MODEr`rprrr7r is_inference)rsrrs rZmethod_is_inference"TensorVariable.method_is_inference"s  4 4 JD&22&55  JJOO((,,_= =D J"((->->-@A Ar\crURb*[R"XRR5$gr)rar7r is_complexrs rZmethod_is_complex TensorVariable.method_is_complex3s* :: !"((ZZ-B-BC Cr\rcTUbUR5O[RnURb#[R "XUR;5$UR RRRS5=nb#[R "XRUS95$g)Nrr) rrQcontiguous_formatrir7rr`rprr)rsrrmemory_format_constrs rZmethod_is_contiguous#TensorVariable.method_is_contiguous8s (  , , .((     )"((DDVDV-VW Wjjoo**..??d L"((&&5H&I r\F non_blockingc ^UcTRb[TR[R5(a[ U4Sj[ R "555nTRRS:Xa$[R"USUR35$[R"USTRRSUR35$Ub[[UR555S:XatUR5n[R"U[U55nSSK Jn U(aSU0UEnU"UURR "S S /[#TU/U5Q765$g) Nc3P># UHupTRU;dMUv M g7fr)ra)rrrrss rZr-TensorVariable.method_type..Us#9daTZZ1_9s& &cpuztorch..ztorch.tensortyper4rrrr)rarUrbrQnextr3rrr7rrXr)rrrrrr0) rsrrarrw tensortype tensor_typetensor_type_constrs ` rZ method_typeTensorVariable.method_typeIsZ M &4;; 551779J{{5(&,,R6*:M:M9N1OPP&,,&!1!1 2!J4G4G3HI  D113459KK  224K / 5 5b#k:J K  .(,A&A  &&!'.?'@&I r\clsrDcb[U[5(awUR(afSSKJn UR 5nUR "XXBR5nURRR[5U[S9 U$[SUSUS3SS/[RQS 9 g) Nr4rCmutation_type_clszHArgument of `as_subclass` must be a non-dispatcher-style tensor subclassz .as_subclass(rCurrently not supportedz:Avoid this call or move it outside `torch.compile` regioner*)rUTensorSubclassVariablerr}rDrrrr4 _track_objobjectr5r rr)rsrr rDpy_clsrKs rZmethod_as_subclass!TensorVariable.method_as_subclassws c1 2 2szz D++-F.>>&**C II " " - -#1E . J^fM#a01L"..  r\c[UR[R5(aHURRS:waURROSn[ R "X5$g)Nr)rUrbrQrindexr7r)rsrrs rZmethod_get_device TensorVariable.method_get_devicesM dkk5<< 0 0)-)9)9U)BDKK%%E"((3 3r\cV[R"XRR5$r)r7rraitemsizers rZmethod_element_size"TensorVariable.method_element_sizes$$R)<)<==r\)forcerNumpyNdarrayVariablec,[R(d[SSUS3SS/S9 [(d[SSUS3SS /S9 UR[ R :wa[S URS 35eU(aWUR5(aBURUS /05nURRS SUR5405nO)URR"S S/[X/05Q76n[RX5$)Nz%Tensor.numpy() with trace_numpy=Falserz numpyzW`Tensor.numpy()` was called, but the `trace_numpy` configuration was manually disabled.zUSet `torch._dynamo.config.trace_numpy = True` to allow Dynamo to trace through NumPy.r*z&Tensor.numpy() without NumPy installedz_`Tensor.numpy()` was called, but the NumPy library is not available in the current environment.z5Ensure NumPy is installed in your Python environment.zcan't convert z4 layout tensor to numpy. Use Tensor.to_dense() firstdetachrrview_as)r trace_numpyr nprcrQstridedrrrrrrr0rr)rsrrtr`s rZ method_numpyTensorVariable.method_numpys!! ?&tfF375 r @&tfF3?L   ;;%-- '  -ab  U--//  Xr26AII**=%!**,RTUEII**y+S[S[RRS[4UU 4SjjnUR [R [R[R[R4;a[SSTS3SS /S 9 UR5S :XaU"X5$UR5S :Xa)[U5VVs/sHup4U"XAU5PM snn$[U5VVs/sHup5T"XQUS 9PM snn$s snnfs snnf)Nrr+rLc2>T"TUR55$r)r)rr+rrs rZwrap:TensorVariable.method_tolist..tolist..wraps$NN$r\z'Tensor.tolist() with non-integer tensorrz to_listzLDynamo currently does not support tracing `tolist()` on non-integer tensors.z[Ensure the input tensor to `tolist()` is an integer type (e.g., int8, int16, int32, int64).r*rr4)r+) rrQfxProxyr7raint8int16int32int64r r enumerate) r*r+r.rval sub_tensorrstolistrrs rZr9,TensorVariable.method_tolist..tolists   ?   ||    $ E*4&9!9B zz|q F..zz|q >G>OP>OFASA,/>OPP&/v%6%6MAzq\:%6 Qs D5Dr) rrrQrVr0r1rrrrprr7r)rsrr*outr9rs`` @@rZ method_tolistTensorVariable.method_tolists~*! 5<<! EHHNN! sTRUY! ! F%%**?;VT]]_-$$R--r\ vars_itererror_on_non_leafcSSKJn /n[5nUHn[U[5(dMUR (dM-UR (aU(a[SSU3SS/S9 MWMYU(a>UR(a[URU5(a[SS U3S S /S9 MURRnXu;dMURU5 URU5 M U$) a Collect unique leaf tensors from vars_iter for backward. Only collects leaf tensors (no grad_fn). Non-leaf tensors are skipped (or error if error_on_non_leaf=True) because when auto-detecting inputs, we must not stop gradients at non-leafs - they are intermediates, and the real leaf tensors (parameters) are further up the autograd graph. Deduplicates by proxy.node. Returns list of unique leaf tensor variables. r)SyntheticLocalSourcezbackward() with non-leaf tensorz-backward(inputs=[...]) with non-leaf tensor: zBbackward(inputs=[...]) with non-leaf tensors is not yet supported.zIOnly pass leaf tensors (parameters, graph inputs) to backward(inputs=...)r*z'backward() with in-graph created tensorz5backward(inputs=[...]) with in-graph created tensor: z^backward(inputs=[...]) with tensors created inside the compiled function is not yet supported.zSOnly pass tensors that are inputs to the compiled function or captured from outside) rrAsetrUr^rgrvr r`rpaddappend)rsr>r?rArh seen_nodesrKrps rZ_collect_backward_inputs'TensorVariable._collect_backward_inputss 2), C#~..33D3D3D??(%$E&STWSX$Y(l k# )' j=Q&R&R! I"WX[W\ ]%Br 99>>D-"t, c*OR r\gradient retain_graph create_graphinputsc X[R(d[SSUSUSUSUSU3 SS/S9 UR(dUR(d [ S5eUS LnU(aq[ URRURRR55nURU5nU(d[R"S 5$OR[U[R 5(a UR"OU/n URU S S 9nU(d [S S SS/S9 S[$R&"X50n UbX:S'UbXJS'[$R&"X5n X /n UbU R)U5 [$R&"U[*R,R.5n U R1XU 5nSSKJn URUSS S9nUR7U5 [$R&"U[*R8R:R<R>5nUce[AU5HEunnURCUS[$R&"UU5/05nUR1UUU/05 MG UREU5 [R"S 5$)aD Trace tensor.backward() by rewriting as autograd.grad() + accumulate_grad. Implementation: 1. Collect leaf tensors to compute gradients for 2. Call autograd.grad(loss, inputs) to compute gradients 3. For each leaf tensor, call accumulate_grad to update .grad Non-leaf tensor handling: - Auto-detect (inputs=None): Non-leaf tensors are silently skipped. This matches eager where only leaves get .grad. - User-provided (inputs=[...]): Errors if any non-leaf tensor is found. While eager backward(inputs=[non_leaf]) works, Dynamo cannot trace it because the accumulate_grad polyfill accesses .grad, and Dynamo creates a generic GetAttrVariable for .grad on non-leaf tensors (instead of a TensorVariable), which cannot be used in tensor operations. TODO: Support non-leaf tensors by fixing .grad access on non-leaf in Dynamo. z"Unsupported Tensor.backward() callrz backward rPz]Dynamo currently does not support tracing `Tensor.backward()` when trace_autograd_ops is off.z)Set torch._dynamo.trace_autograd_ops=Truer*z8tensor does not require grad and does not have a grad_fnNT)r?zbackward() with empty inputsz8backward(inputs=[...]) resulted in no valid leaf tensorsz?backward(inputs=[...]) requires at least one valid leaf tensor.zOEnsure at least one tensor in inputs is a leaf (requires_grad=True, no grad_fn) allow_unusedrIrJr4)GradModeVariableF) initializedr)#rtrace_autograd_opsr rgrvrrrleaf_var_creation_orderinput_source_to_varvaluesrFr9rrUrBaseListVariablerr7rrDrQautogradr3r= ctx_managerrNenterrainductoraccumulate_grad_defaultr6rexit)rsrrHrIrJrK auto_detectall_vars input_vars provided_vars grad_kwargs inputs_var grad_argsautograd_grad_fn grads_varrN grad_mode_varaccumulate_grad_fnidx input_vargrad_is rZmethod_backwardTensorVariable.method_backward*s6(( <&tfJxj,qQ]P^^_`f_gh{BC  !!$*:*:#J n  11 --446H66x@J(..t44 fi&@&@AA X  667J:V ai &'<' >r\cHURR"S[R/[ XU/05Q76nUR 5n[ RRR5 UR(a?URR(a$URRR5O [5 [URUSS9 SSS5 SSS5 UR!XUR#55 [$R&(d[$R((a0URR*R-URS5 [.R0"S5$!,(df  N=f!,(df  N=f)Nr=F)allow_non_graph_faker)rrrsetitemr0rrQrRrS+_disable_saved_tensors_hooks_during_tracingr shape_envignore_fresh_unbacked_symbolsrr+rprrruse_graph_deduplicationtrack_nodes_for_deduplicationregion_trackeradd_node_mutationr9r)rsrrrKr`rs rZmethod___setitem__!TensorVariable.method___setitem__zs"  &&     512 6 //1 MM   K K M|| 6 6 LL " " @ @ B  5::r F   N &&r5??;LM  ) )V-Q-Q II $ $ 6 6uzz1 E&&t,,   N Ms%1AF F"F F F F!c .[SSUSUSU3S/S9 g)Nz!Unsupported Tensor.resize_() callrz resize_ rPz=Dynamo currently does not support tracing `Tensor.resize_()`.r*r/rs rZmethod_resize_TensorVariable.method_resize_s+ 7"4& $qAW  r\c .[SSUSUSU3S/S9 g)Nz$Unsupported Tensor.resize_as_() callrz resize_as_ rPz@Dynamo currently does not support tracing `Tensor.resize_as_()`.r*r/rs rZmethod_resize_as_ TensorVariable.method_resize_as_s+ :"4& TF!F8DZ  r\c .[SSUSUSU3S/S9 g)Nz(Unsupported Tensor.sparse_resize_() callrz sparse_resize_ rPzDDynamo currently does not support tracing `Tensor.sparse_resize_()`.r*r/rs rZmethod_sparse_resize_$TensorVariable.method_sparse_resize_s, >"4&(8axH^  r\c .[SSUSUSU3S/S9 g)Nz2Unsupported Tensor.sparse_resize_and_clear_() callrz sparse_resize_and_clear_ rPzNDynamo currently does not support tracing `Tensor.sparse_resize_and_clear_()`.r*r/rs rZmethod_sparse_resize_and_clear_.TensorVariable.method_sparse_resize_and_clear_s, H"4&(B4&&Rh  r\c l[U5S:a%[SSUSUSU3S/[RQS9 g)Nr4zUnsupported Tensor.set_() callrz set_ rPzhDynamo currently does not support tracing `Tensor.set_()` overloads that include more than one argument.r*)r~r rrrs rZ method_set_TensorVariable.method_set_sL t9q= 8&tfF4&&BA6)556  r\)alphaotherrcUby[U[5(ad[R(aOSSKJn [ R"UR5nURXX0/05nURUSU/05$gr) rUr^rrrrrr:rr=r)rsrrrrrrhs rZ method_add_TensorVariable.method_add_sl  5.1133 6<<^=O=OPG**2u/CRHF##B&2> >r\c>Ub[R(a[R"[R 5R XU/05nSSKJn [R"UR5nUR XX@/05nURUSU/05$gr) rrrr:rQdivr=rrrr) rsrrrrKrhrr fma_results rZmethod_addcdiv_TensorVariable.method_addcdiv_s  !D!D;;EIIFTTg&F 7<<^=O=OPG ..rE3H"MJ##B*rB Br\rc[R"[R5R XU/05n[R"[R 5R X/05nUR US/05$)Nr)rr:rQeqr=rr)rsrrrhs rZmethod___contains__"TensorVariable.method___contains__ sm77AOO s R 77 BPP " !!"fb"55r\c0UR"US/UQ70UD6$)N register_hook_method_register_hookrs rZmethod_register_hook#TensorVariable.method_register_hooks! ))"oOOOOr\c0UR"US/UQ70UD6$)N"register_post_accumulate_grad_hookrrs rZ)method_register_post_accumulate_grad_hook8TensorVariable.method_register_post_accumulate_grad_hook$s- )) 4 7; ?E  r\hookc D^^^^^UR(Gd1[R(aURR U5umnS[ R S[RSS4UU4SjjnSSKJ n UR5nSURRS'U"UURRS UXt4055$UR5Rn[URR!55n ["R$"[&5n U R)XU/05n U R5Rn /m[+5mS [ R,R.S [ R,R.SS4UUU4S jjmT"X5 Un THnU R1U5 Un M U HnUR3X5 M [5U [65(deU R5UlUR;U5 ["R<"["R>RA5S 9$["R<"["R>RA5S 9nURRBREXUT5 U$)Nr*bw_staterLc h>[UT5nU"[R"[[UTS95 g)N)rir hook_name)rWrrrr%)r*rrrrs rZ_register_hook_trampolineGTensorVariable._method_register_hook.._register_hook_trampoline<s8%,FD$9M!!)))<%-&/  r\r4rThas_backward_hookr=rpstop_atc>UT;dXLagURH6n[U[RR5(dM.T"X!5 M8 UR R 5H6n[U[RR5(dM.T"X15 M8 TRU5 TRU5 gr) rrUrQr0NoderwrSrCrD)rprrkwarg collect_deps nodes_to_movevisiteds rZr:TensorVariable._method_register_hook..collect_depss7?do99C!#uxx}}55$S2%"[[//1E!%77$U42 D!$$T*r\r)#rrcompiled_autograd_enabledradd_backward_state_hookrQrV BackwardStaterrrrprrruserskeysrAutogradFunctionVariabler$ call_applyrBr0rrDreplace_input_withrUr^r`rRemovableHandleVariablebaser6r4r)rsrrrbw_state_proxyrr self_proxy tensor_nodeusers_to_replaceapply_hook_varrhtensor_prime_node insert_pointrpuserhandle_variablerrrrs ` @@@@rZr$TensorVariable._method_register_hook.sD {{{!::,.II,M,Md,S) > !LL 4E4S4S   "3!]]_ <@ $$%89$II**'1#4 P--/..K#K$5$5$:$:$<= '??@RSN#..r$<DF!' 1 6 6  24M*-%G +588== +588== +T + + * 8'L%##D)# & )'' G)fn55 55*DJ  ' ' + 44'nn==? $;;#..99;  ,,T$Or\crUSLaUR5nUR5RnURSnURU:wGa9UR (a[ SSUS3S/S9 URR(a[ SSUS3S S S /S9 [RRR5n[RRRS5 URRS [RRRS 05 URR!SSUR5405 URRS [RRRU405 [RRRU5 UR#U5 X lU(aURR$R'U5 UR (dnURR(R+U5(aDURR(R-US[.R0R3S55 U$![RRRU5 f=f)NTrz(Unsupported Tensor.requires_grad_() callrz requires_grad_zaDynamo does not support changes to a Tensor's `requires_grad` through calling `requires_grad_()`.r*z.requires_grad_() intermediate leaked as outputzAn intermediate tensor with requires_grad_() called on it (or a tensor derived from it) is returned from the compiled region. Graph breaking here to preserve partial acceleration.z8Call .detach() before returning if you only need values.zConsume the gradient inside the compiled function (call backward() and use .grad), or move requires_grad_() outside torch.compile.r=Trrequires_grad_r3)rrrprrgrr speculation_loggraph_break_on_requires_grad_rQrr!get_inplace_requires_grad_allowed!set_inplace_requires_grad_allowedr create_noderrrQrDr4is_attribute_mutation store_attrrr9r)rsrrgrpr prev_states rZmethod_requires_grad_$TensorVariable.method_requires_grad_s5  $)<<>M}}## /2  & &- 7{{F*4&@!J !!??L*4&@!$ SJ 0,,NNPJ HH   A A$ G R %%#HH''II   &&!$]]_&   %%#HH''IIM ##EEjQ  ( ( 7!.  1188> {{ryy'='='S'S((II**55fi&@&@&G&G&M ##EEjQs 0B.J +J6cSSKJn URRSSUR 5405nU"X5 UR U5 U$)Nr4rrdetach_)rrrrrr)rsrrr`s rZmethod_detach_TensorVariable.method_detach_sK* &&   ]]_     b  ##B' r\cF[SSUS3SS/[RQS9 g)Nz'Unsupported Tensor.share_memory_() callrz share_memory_zTDynamo does not support Tensor.share_memory_() which modifies tensor storage for IPCz7Move share_memory_() call outside the compiled region. r*)r rrrrs rZmethod_share_memory_#TensorVariable.method_share_memory_s2="4&7nI"..  r\c[U5S:Xa[US[5(d&[U5S:a*[SU55(aUR USX#5$g)Nr4rc3# UH:nUR5=(a [UR5[5v M< g7fr)r_rUrr)ras rZr,TensorVariable.method_new..2s8A$$&R:a6J6J6Lc+RRsAA new_empty)r~rUr;allrrs rZ method_newTensorVariable.method_new(s^ INz$q'<@@ IN ##B TB Br\rsc[U5S:wagUSnURS5nUb9UR5(dgUR5[R :wag[ U5nURSS5 URSURUS55 URSURUS55 [R"[R5RUU/U5$)Nr4rrcrarb)r~rr_rrQr%dictpop setdefaultrjrr:r*r=)rsrrrwdata_argrc fwd_kwargss rZmethod_new_tensor TensorVariable.method_new_tensor:s t9>7H%  ,,..((*emm;&\ x&gt'7'7G'DEh(8(8X(FG55ellCQQ  J   r\c~[XR5RRSR 55$r)UntypedStorageVariablerrpruntyped_storagers rZmethod_untyped_storage%TensorVariable.method_untyped_storageSs4& --/&&++O<LLN  r\c~UR(d,URRRU5 SUlgr)rnr`rp_rename)rsrs rZ set_name_hintTensorVariable.set_name_hintZs,  JJOO # #D ) $D r\cTUR5RRSSL$r)rrprrrs rZis_python_hashable!TensorVariable.is_python_hashable`s&}}##((9EEr\cb[UR5RRS5$r)hashrrprrrs rZget_python_hashTensorVariable.get_python_hashfs$DMMO((--o>??r\c[U[5(dgUR5RRSnUR5RRSnX#L$)NFr)rUr7rrpr)rsrr*bs rZis_python_equalTensorVariable.is_python_equalisR%11 MMO % %o 6 NN  ! ! & & 7v r\)rnrurlrbrarvrirjrhrkrcrdr`rgrfr)NF)F)NNNN)rLNr)rr@rLr^)rr@rLr9)rX __module__ __qualname____firstlineno____doc__r7_nonvar_fieldsrQrVr,r0r1rarbrcrrOrrrr|rrrstrrrrrr staticmethodr1rrrrr rr9rrrrr"r&r r0r7r>rArFrMrjrrrvrrrrrpropertyrer]r^rrrrrrmethod_nelementrmethod_ndimensionrrrrrrrrr'r<rrFrjrnrwr}rrrrrrrcacher|rrr:rrrrrrrrrrrrrrrr#r&r.r6r;r?rBrFrrJ__static_attributes__ __classcell__r~s@rZr^r^s G   ' '!N&B B.)-)-%)$(#&/xx~~&/{{ &/  &/  &/&/&/&/&/&/&/&/S#X%&/c3h$&&/ d{!&/"Tk#&/$%&/& '&/&/RFJ  )  7;d{    E3:E A #Ad A A  A*;C;%((..T4 @3 @ @/%,,/4S>//bXB)XB14XB XBt7#:77 $;RV@V %<SWAW %<SWAW ) D 8$;88) D ) D ) D ) D  *A h #:QU?U (#:(( ;) ; D  ; '> ? 2)2142 2>Y5YSY_Yv#31#3o#3J*AdJN9 )9 2:3-$2F9 o 9 v; #;,; ; ' ; c?23 ; ;&D&eCHoNcN T#Y !0 #!0!0  !0Fa #aa' a - a  aFA)A25AADA 4 A C)C25CADC 4 C ,0**!Dj* 4 *X 6 ?T;Q #O449O #) D ) D "$;@PSW@W TX):ID:P D (!" , #,Tz, ,  , 4  ,\ ) 0? ' 2$;RV@V >&=>/>OT(6)(65Dt5K(6 (6T(. 7(.O(.VOT;!/2;GK; o  %;@,0/3/3)- s- #s-"D(s-&, s- &, s-  $& s- 4 s-j% #%%" %  %- #- -  -& #"   , #   < #   8$!8$_$ #   2&#:&&( #((" (  (<__  &"!8"_"P3PP "9 >R !  #  Tz  t$- #-- -  -B  #    "      #    "      #    "      #    "     #"   6)-  # %  4  6)-  #""  %  4 & 6) 60? 6  6P #PP" P  P #  "    H)H14H:PU Un   h   #"  4  $  #  "  4   2 ) ! #$ FDF @@Vr\r^c ^\rSrSrSrSS1\R krS\4Sjr\ S%SS S\ S\ S-S \ SS 4 S jj5r S\ S\ S \ SS4U4Sjjr S\ 4SjrS\4SjrS\ 4SjrSS S\4SjrSS S\ S\4SjrS%S\SS\\-\-4SjjrSSS\S\\S \\\4S\4 SjrSSS\4SjrSSS\ S \ S\4SjrSSS\4SjrSSS\ S \ S\4SjrS\4S jr S\4S!jr!S"\"S\4S#jr#S$r$U=r%$)&riqz Represents a symbolic scalar, either int, float or bool. This is most commonly used to handle symbolic size computation, e.g., tensor.size(0), but it is also used to handle logic like float_tensor.item() or unspecialized float inputs. r`rrLc,[UR5$r)rrrrs rZrSymNodeVariable.debug_repr~sDLL!!r\NrrArr7c DUc[URU5nSURR;aURRSU:Xde[URU5 [ U[ R [[45(aB[ U[ R 5(a [U5OUn[R"X5$[X#40UD6nURRS:wa%URRRU5 U$)Nrry)r+rprr2rUsympyIntegerrrOr7rrr}rcurrent_tracerrecord_proxyable_vt)r rr`rrr;s rZrSymNodeVariable.creates ?$UZZ4G ejjoo -::???3w> >>%**g. g sD9 : :&0%--&H&Hc'lgG"((5 5e88 ::==M ) II $ $ 8 8 = r\rwc L>[TU]"S0UD6 XlX lSUlgNrz)r{r|r`r _tensor_var)rsr`rrwr~s rZr|SymNodeVariable.__init__s% "6"  26r\c[UR[5(a URRR$[ UR5$r)rUrrrppytyperrrs rZrSymNodeVariable.python_types8 dllH - -<<$$++ + % %r\cgrrzrrs rZis_symnode_likeSymNodeVariable.is_symnode_likerr\cUR$rrrrs rZrSymNodeVariable.as_proxyrr\c[UR[R5(aU$[UR[R[R 45(de[ RXR5S:g5$ru) rUrrQrSymIntSymFloatrrrrs rZrSymNodeVariable.bool_impls^ dllEMM 2 2K$,,u~~(FGGGG%%b--/Q*>??r\rac URcQ[R"U[R5R X/S[R"X505UlUR$)Nra)rer7rrQ scalar_tensorr=)rsrras rZ as_tensorSymNodeVariable.as_tensors[    #.44E'' mB/2G2G2R(ST  r\ output_graphr?c[UR5$![a^n[RR R R(ae[[RS[U53SS9eSnAff=f)Nz5Consider annotating your code using torch._check*(). constrain_as_size_example) case_name) rrrrQr0 experimental_configno_data_dependent_graph_breakr"r# ANTI_PATTERNrQ)rsrwrs rZrSymNodeVariable.evaluate_exprsi  - -* xx$$,,JJ**GAxP5   s A?AA::A?r@rrc rSSKJn U"UURR"SU/[ U/UQU5Q765$)Nr4rr)rrrrr0)rsrrrrwrs rZrSymNodeVariable.call_methodsF +  II " " #D=4=&9   r\c zSSKJn U"UURRS[UR 5405S9$Nr4rr=rU)rrrrr rrsrrs rZrSymNodeVariable.nb_int_impls> +))(("   r\c$URU5$rrrs rZrSymNodeVariable.method___int__s##r\c zSSKJn U"UURRS[UR 5405S9$r)rrrrr rrs rZrSymNodeVariable.nb_float_impls> +))(("   r\c$URU5$rrrs rZr SymNodeVariable.method___float__ s!!"%%r\cgrrzrrs rZrB"SymNodeVariable.is_python_hashable rr\c4[UR55$r)rErrrs rZrFSymNodeVariable.get_python_hash sD&&())r\rc[U[5(a!UR5UR5:H$[U[5=(a! UR5UR 5:H$r)rUrrr7r)rsrs rZrJSymNodeVariable.is_python_equal s] e_ - -%%'5+>+>+@@ @ uo . C""$(@(@(BB r\)rer`rr)&rXrLrMrNrOr7rPrQr classmethodrrr|rrrOrkrrr^rur rfloatrrr1rrrrrrBrFrrJrWrXrYs@rZrrqs   ' 'N "C" #  't     ,7c7C73747&T& # @ ' @  @ 7   7; $]3  e   #  '  S/)*    $ #   &$)$25$AD$ $  #   "&)&25&AD& & D**  V   r\rc ^\rSrSrSr\SSS\RRS\ SS4Sj5r SSS \ S\ 4S jr \S \ S \\ S \\ \ 4S\\\ \\ \ 444S j5rSSS \ S \\ S \\ \ 4S\ 4 U4SjjrS\4SjrSrU=r$)ri zq Represents a np.ndarray, but backed by torch Tensor via torch._numpy.ndarray. Use this for Tensor.numpy() call. rr@r`rrLc .SSKJn U"S[UUS.UD6$)Nr4rzr|rz)rr{r)rr`rr{s rZrNumpyNdarrayVariable.create% s, /  +    r\rcd^^^^ ^ SSKJm SSKJm SnTR 5R R Sn[R"U5nS[4UU UUU 4SjjnTS;aETRRS T TR 5T405n[RTU5nOTS ;a![R"T[UT55$TS ;aJ[![UT5=n5(d'[R"T[#S U555$U"5$TS :XaC[!UR$=n5(d [R"T['U55$U"5$TS;a[)SSTST3STS3/S9 OTS:Xa[)SSTST3STS3/S9 Uc[*eU$)Nr)numpy_attr_wrapperr4rrrLc n>T"TTRRSTTR5T4055$)Nr=)rrr)rrrsrrsrZinsert_into_graph;NumpyNdarrayVariable.var_getattr..insert_into_graph? s;  &&#%7$--/49PRT r\)Trealimagflatr=)rdr)shaperfc38# UHn[U5v M g7frrrs rZr3NumpyNdarrayVariable.var_getattr..a s6Iq!s1vvqrre)r flagsraz$Unsupported ndarray attribute accessr)rPz3Dynamo currently does not support tracing `ndarray.rr* __version__z&Unsupported ndarray.__version__ access)rSrrrrrprtnpndarrayr7rrrrrrWrrrerr r) rsrrrhrexample_ndarrayrr`rrrs ``` @@rZrj NumpyNdarrayVariable.var_getattr2 s /* ,,11/B ++m4 ?   0 0II**"$' E *00U;F) )"((W_d-KL L ( (#$)G$GAHH&,,R6Iq6I1IJJ$& & V^#)=)=$=A>>&,,RQ88$& & / / >&tfAdV4QRVQWWYZ   ] " @&tfAdV4QRVQWWYZ   >% % r\rrwcUS:Xa9SSS.nUR5VVs0sHupEURXD5U_M nnnX4$s snnf)Nclipminmax)a_mina_max)rr)rrrw kwargs_renamerrs rZ patch_argsNumpyNdarrayVariable.patch_argsy sO 6>&+eSSKJn SSKJn UR X#U5up4US:XaSSKJn SnSU;aUSnO[U5S:aUSnUSL=(a URS 5n [X5=(a UR[Ln U (dU (a%U"S S US US US U3S /[RQS9 US;a[T U]=XX45$US;aU"SS US US US U3SUS3/S9 UR R""SU"U5/[%U/['U5-U5Q76n [(R+X5$)Nrr/)numpy_method_wrapperastyper4)BuiltinVariablerarOzndarray.astype(object)rrPz`ndarray.astype('O')` or `ndarray.astype(object)` is not supported by torch.compile, as there is no equivalent to object type in torch.Tensor. This will be executed eagerly.r*)__len__rer9__iter__)tostringtobytes __delattr__zUnsupported ndarray method callz `ndarray.z&()` is not modelled in `torch._numpy`.r=)rr rSrrbuiltinrr~is_constant_matchrUrirrrr{rrrr0rrr) rsrrrrwr rr dtype_arg is_object_stris_object_typer`r~s rZr NumpyNdarrayVariable.call_method sy (0t6:  8  0I& "7O TQ G %T1Vi6Q6QRU6VM96Q9<<6;Q 4*4&$qaxH9;-99:  < <7&r> > 9 9 9&tfAdV1TF!F8D'v-ST    &&   & vT 2F ;  $**255r\c<[b[R$[$r)r$rrrrs rZr NumpyNdarrayVariable.python_type s >:: Or\rz)rXrLrMrNrOrRrQr0r1rrrQr7rjrr1rrrrrrWrXrYs@rZrr s   #  ,1HHNN  GJ      E5ESE_EN!/2<@oAU[TU]"U40UD6 X lX0lgr)r{r|rr)rsr`rrrwr~s rZr|$UnspecializedPythonVariable.__init__ s" )&)"&r\tensor_variablecH[S0[UR5DUUS.D6$)Nrrz)rr1__dict__)r rrrs rZfrom_tensor_variable0UnspecializedPythonVariable.from_tensor_variable s/+ ?++, #  r\)rrr)rXrLrMrNrOr^rPrQr0r1rrrOrr|rrrWrXrYs@rZrr s   & &N)- 'xx~~ '3;% '  '  '  ' ' !  '  3;%    '    r\rc^\rSrSrSrS1\R krS\RRS\ SS4U4Sjjr \ S \SS4S j5r S rU=r$) FakeItemVariablei zAn unspecialized python variable which prevents access to the underlying raw value. This is needed if item is called on a FakeTensor.rr`rwrLNc X>URSS5n[TU]"U40UD6 X0lg)NrF)r2r{r|r)rsr`rwrr~s rZr|FakeItemVariable.__init__ s+jj6  )&)&r\rc>[S0[UR5D6$rd)rr1r)r rs rZr%FakeItemVariable.from_tensor_variable s A$'?'?"@AAr\)r)rXrLrMrNrOr^rPrQr0r1rr|rrrWrXrYs@rZrr sl9   & &N 'ehhnn''' B,B BBr\rcL\rSrSrSSS\\S\\\4S\4SjrS\ 4Sjr S r g ) ri rr@rrwrLc ,SSKJn URRnU[R RLaSn[ U5S:XaUUSR5(a=[ U5S:Xa.USnUR"XURUR5nOd[SURSUSUS3SS /[RQS 9 O3[R"X5RX/[!U5-U5nUceURR"nU[R R"La'[R"X5RX/U5 UR$R&R)[+5U[,S 9 U$) Nr4rCrzCCalling subclass default constructor with more than tensor argumentrrrrzFAvoid this constructor call or move it outside `torch.compile` regioner*r )r}rDrK__new__rQrVr~rrrr rrr7rr=rr|rr4rrr5) rsrrrwrDnew_funcrKdata init_funcs rZr=$TensorSubclassVariable.call_function sm A::%% u||++ +C4yA~$q'"3"3"5"5#f+:JAw3BBdjj$++a#zzl&ixqI 92+66 "''5CCFT$Z'CJJ''  ELL11 1  ! !" 0 > >r5& Q )) Hc-A *  r\cUR$rrYrrs rZr)TensorSubclassVariable.as_python_constant* rr\rzN) rXrLrMrNrr7r1rQr=rrrWrzr\rZrr sI/ #/'/S/)* /  /bDr\rc ^\rSrSrS1\R krS\S\RS\ SS4U4Sjjr S\ 4Sjr S S S \ S \\S\\ \4S\4 U4S jjrSSjrSrU=r$)r9i. r from_tensorrwrLNc >>[TU]"S0UD6 XlX lgrd)r{r|rr)rsrrrwr~s rZr|UntypedStorageVariable.__init__4 s! "6"&*r\c"[R$r)rQUntypedStoragerrs rZr"UntypedStorageVariable.python_type? s###r\rr@rrc >US:XaU(dU(a&[UUS[U5S[U5S35 URR5n[ U5(d [ R "U[U55$SSKJ n SSK J n U"UURRS UURR54055$US :Xa[U5S:XaU(a[XS [U5S35 URRS [ R"R$R&URR5US R5405 U$[(TU]UXX45$) Nrez0 args and 0 kwargsz args and z kwargsr)untyped_storage_sizer4rr=resize_z0 kwargsr)r1r~rrerr7rrexternal_utilsrrrrrrrrQrarXresize_storage_bytes_r{r) rsrrrrwrhrrr~s rZr"UntypedStorageVariable.call_methodB sN 6>v#)4ykCK=@  '',,.F#F++&,,RV==A2$II**',))2246  9 Ta#BjS[M:QR II " " ""88!!**,d1g.>.>.@A   Kw"2T::r\clU"UR5 URS5 URS5 g)Nr:rr load_methodrrscodegens rZ reconstruct"UntypedStorageVariable.reconstructo s-  !-.Ar\)rrrr>rLN)rXrLrMrNr7rPr^rQrrr|rrrQrr1rrrWrXrYs@rZr9r9. s  ' 'N +# +++ + +  +$T$+; #+;+;?# +; S/)* +;  +;Zr\r9cP^\rSrSrS\S\SS4U4SjjrS\4SjrS Sjr S r U=r $) rliu rrwrLNc 2>[TU]"S0UD6 Xlgrd)r{r|r)rsrrwr~s rZr|DataPtrVariable.__init__v s "6"&r\c[$rrrrs rZrDataPtrVariable.python_type~ s r\clU"UR5 URS5 URS5 g)Ndata_ptrrrrs rZrDataPtrVariable.reconstruct s,  !J'Ar\)rr) rXrLrMrNr^rr|rrrrWrXrYs@rZrlrlu s:'#''  'Tr\rl)rOrloggingrrrrVcollections.abcrr contextlibr itertoolsrrtypingrr r r r^ torch._numpy_numpyrtorch.fxrQ torch.randomr r torch._dynamortorch._library.opaque_objectrtorch._opaque_basertorch._subclasses.meta_utilsr%torch.fx.experimental.symbolic_shapesrrrrrtorch.utils._python_dispatchrrIrrr_trace_wrapped_higher_order_oprrrrrr r!r"r#rr$r%guardsr&r'rr(rSr)r*r+r,r-r.r/r0r1r2r3r r5r6r7rr9listsr:r; script_objectr< user_definedr=numpyr$ModuleNotFoundErrortorch._dynamo.codegenr>torch._dynamo.output_graphr?torch._dynamo.symbolic_convertr@rA functionsrBr}rD getLoggerrXrgtltgelerneis_is_notsupported_tensor_comparison_opssupported_const_comparison_opssupported_comparison_opsr1fromkeysrS%supported_tensor_comparison_op_values$supported_const_comparison_op_valuesrrOr[r TensorBaserrVrJr^rrrrrr9rlrzr\rZr%sW" ."99 $+A)6G33:O0    JI&542 /6 0<!   ++ ++ ++ ++ ,,oo # ,,oo ++ ++ " %$)- #**,)%(,}}"))+($ &T88&&//%,,2G2GGX_Xv>k ok \[>[|# .# LB~B*353lD_DNoAQ Bs>J''J32J3