'}h?UddlZddlZddlZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddl Z ddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlmZddl mZmZddlm Z ddlm!Z!ddlm"Z"m#Z#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+m,Z,m-Z-m.Z.m/Z/m0Z0m1Z1m2Z2ddl3m4Z4 ddl5Z6 ddl8Z9ddl:m;Z<dd l=m>Z>dd l8m?Z?d d l@mAZAe6re6e6je6je6jfZEe/ejd feGd<e6eZd?Zd@ZdAZdBZe j:dCdDZddEZdFZejZGdGdHZdIZej`eJae(eeGdK<dLefdMZdNedOdfdPZddQZdOe,efdRZejZGdSdTZGdUdVeZee_dWZddXdYZdZZd[e9jxfd\Zed]Zd^Zdd_Zd`ZdaZdbZe j:d dcZddZddeZdfZedgdhZeefeee]eede̐jeje9je9je9je9jh ZdiZdjZdkZdlZdmZdnZeeګjZe0e+e"eGdo<eeګjZe0e2e"eGdp<eejjZe0e2e"eGdq<eeeIZe0e*e"eGdr<ejZejZdsZdtZduZeZdvZdwZdxZdyZddzZd{Zd|ZdOe]fd}ZdOe]fd~ZdZddlmZdZdZdZdd&dd&dd&d&edjfdZdZej8dZeZejeZe'e"e,e"feGd<eZe,deGd<eZGddZe j:ddZdZddZdZ dZ ddZ ejZ dZedZdZdZdZde"fdZdZdejfdZde"fdZde"fdZGddej0ZdefdZde1e9jxe"fdedddOe/ee-effdZddZdZdZdZddgZ ddgZ!gdZ"e e!ze"zZ# ddZ$ ddZ%dZ&dZ'GddZ(dZ)GddZ*GddZ+dZ,dZ-dZ.dZ/d%e]dedOefdÄZ0ejZGdĄdūZ1de]dOe-e1fdDŽZ2dejdejfdOe]fdʄZ4d˄Z5d̄Z6d̈́Z7d΄Z8dddOefdфZ9d҄Z:dӄZ;ej8dԄZy#e7$rdZ6YwxYw#eN$rYOwxYw)N)contextmanager) lru_cachewraps)Path)MethodWrapperType)AnyCallablecastClassVarCounter DefaultDictDequeDictIteratorKeysViewListOptionalSetTupleTypeUnion ValuesView)RemovableHandle)detect_fake_mode LazyString)config.NP_SUPPORTED_MODULES) FakeTensoris_fakemaybe_get_fake_mode)fx)enable_python_dispatcher)log_compilation_event)LazyModuleMixin) tree_map_onlycountersoptimus_scuba_logzztabulate..}s$ ),DIIc#sm $ s)+)tabulate ImportErrorr3 itertoolschain)rowsr/r:s r8r:r:wsZ   w 77  yy 09 40P    s6AAc<tjr t|S|SN)rcprofilecprofile_wrapper)funcs r8maybe_cprofilerDs %% Kc.tfd}|S)Nc tt}t j|dz}t j }|j tj}|j g|i|}tj|z }|jtd jd|d|ddtj|}|j||jd} tj dd d d d d t#|gtj$} tj&dddt#| g| j(tdt#| |S#t*$rtd|j-tj.j0j3d|j-tj.j4j3dY|SwxYw)Nz.profilez### Cprofile for z iter z took z.3fz seconds ###z.svg gprof2dotz-fpstatsz"--node-label=total-time-percentagez!--node-label=self-time-percentagez--node-label=total-time)stdoutdotz-Tsvgz-o)stdinzGenerated SVG from profile at zxFailed to generate SVG from profile -- dumping stats instead.Try installing gprof2dot and dot for a better visualization)next timer_counterr__name__cProfileProfileenabletimeruncalldisableprintrIStats dump_stats with_suffix subprocessPopenr5PIPE check_callrJFileNotFoundError sort_statsSortKeyTIME print_stats CUMULATIVE) argskwargs profile_cnt profile_pathprofstart_tsretvalprofile_latencypssvg_pathgprof2dot_processrCs r8profile_wrapperz)cprofile_wrapper..profile_wrappers=) DMM{m8,DDE ! 99;d4T4V4))+0   f[MX[G\\h i \\$   %++F3 E * 0 087- %" !   ! !s8}5'..  23x=/B C ! E N  MM&..-- . : :2 > MM&..33 4 @ @ D  Es9A:E55BG=<G=)r)rCrps` r8rBrBs" 4[))V rEctdzayNr) curr_framerEr8increment_framerus aJrEcXtjtjdayNr)r,clearr+rsrtrEr8reset_frame_countrys""$JrEct|z ayr@)op_count)cnts r8increment_op_countr}s  OHrEc d}i}tjD]6}|jD]!\}}||z }||vr|||<||xx|z cc<#8d}|jD]\}}|d|dt|d}t |y)NgzTIMING: :)r,valuesitemsroundrW)total total_by_keytimingskeytimingoutvalues r8print_time_reportrs EL%,,.,"==? ,KC VOE,&$* S!S!V+!  ,, C"((*/ UQse1U5!_-./ #JrEc&fd}|r||S|S)NcVtjrStfd}|S)Ncj}|tvr gt|<tjj |d5t j }|i|}t j |z }dddt|j rMtt}|tvr it|<t|vr|t|<St|xx|z cc<S#1swYrxYw)Nz (dynamo_timed)) __qualname__r+torchprofilerrecord_functionrTappendr5rsr,) rerfrt0r time_spent frame_keyrC phase_names r8 time_wrapperz>dynamo_timed..dynamo_timed_inner..time_wrappers##C2202(-//3%0GH .YY[$)&)!YY[2-  . %S ) 0 0 < O $6646&y1%7 %BB@J&y1*=H'y1*=K=H . .s 4C&&C/)rrAr)rCrrs` r8dynamo_timed_innerz(dynamo_timed..dynamo_timed_inners- ??K t   &rErt)original_functionrrs ` r8 dynamo_timedrs4!"344 rEr5FcFdffd }|dk(r9tDcgc]}||t|df}}d}|t|dz }|S|d k(rKtjDcgc] }||d }}ttj }||fSy cc}wcc}w) a Get metrics about torchdynamo frontend/backend compilation times. Accumulates information from functions tagged with `@dynamo_timed`. repr='str' returns a printable string for user interaction, and 'csv' returns headers, rows which can be logged for output aggregate causes values from multiple compilations (e.g. split graphs) to be accumulated into one value. If false, expect more than one value per metric. c|Sr@rtxs r8zcompile_times..(srEc`r|t|Sdjt||S)Nr2)sumr3r4)ritem_fn aggregates r8fmt_fnzcompile_times..fmt_fn(s+ 3v;' 'yyWf-..rEr5c |dS)Nz.4frtrs r8rzcompile_times../s 3rE)rz!TorchDynamo compilation metrics: )Functionz Runtimes (s)r.csvc |dS)Nz.6frtrs r8rzcompile_times..7s 1S'rEN)r+r:rlistkeys) reprrrkr>rvrr/s ` r8 compile_timesrs +/  u}. /2sHH] 67rEc eZdZddZdZdZy)DuplicateWarningCheckerc2||_|jyr@)maxsizereset)selfrs r8__init__z DuplicateWarningChecker.__init__Rs  rEc6tj|_yr@) collections OrderedDictsetrs r8rzDuplicateWarningChecker.resetVs**,rEc^||jvr/|jj|dtjsyyd|j|<t |j|j kDr?|jj dt |j|j kDr?y)NT)lastF)r move_to_endrverboselenrpopitemrrs r8addzDuplicateWarningChecker.addYs $((? HH 4 0>> !DHHSMdhh-$,,.  e ,dhh-$,,.rEN))rP __module__rrrrrtrEr8rrQs- rErc"tjjdddk(}|rWtjj t jt jt jdtStjS)NTORCH_COMPILE_DEBUG01T)dynamoaotinductor output_code) osenvirongetr_loggingset_logsloggingDEBUGadd_file_handler contextlib ExitStack) compile_debugs r8setup_compile_debugrhshJJNN#8#>#EM == ]]   !!    !!rEc,tjyr@)graph_break_dup_warning_checkerrrtrEr8reset_graph_break_dup_checkerrws#))+rEctjjtd}tj|dt j tjj|dt jdjtj}|jfd|S)N torchdynamoTexist_okz debug.logz torch._dynamoc&jSr@ removeHandlerlog_file_handlerloggersr8rz"add_file_handler..sv334DErE) rpathr3 get_debug_dirmakedirsr FileHandler getLogger addHandlerrrcallback)log_path exitstackrrs @@r8rr{sww||MO];HKK4(**277<<++NO    /F &'$$&I EF rEcBtj}tjxt j tjt jjjD](j|jfd*|S|S)Nc&jSr@rrsr8rz setup_log_file..sv';';>( # KKKX  KK1D Ah% $  # $ $ $ G ;XFGs55BAB8B BBBBCCgcTd}|jD]}d|jvs|dz }|S)Nrcallr)nodesop)rcns r8 count_callsrs5 A WW QTT> FA HrEc|Sr@rtrs r8identityrs HrEcN t|y#t$rYyt$rYywxYw)NTF)hash TypeError ValueErrorrs r8hashablers. Q  s  $$$cyr@rtrerfs r8nothingr"srEc<eZdZdZdZdZd dZdZdZdZ d Z y) ExactWeakKeyDictionaryz\Similar to weakref.WeakKeyDictionary, but use `is`/`id` rather than `==` to compare equalityc@t|_t|_yr@)dictrrefsrs r8rzExactWeakKeyDictionary.__init__sf F rEc2|jt|Sr@)ridrs r8 __getitem__z"ExactWeakKeyDictionary.__getitem__s{{2c7##rENcL|jjt||Sr@)rrr))rrdefaults r8rzExactWeakKeyDictionary.gets{{r#w00rEc0t||jvSr@)r)rrs r8 __contains__z#ExactWeakKeyDictionary.__contains__s#w$++%%rEct|jvr'tj|fdj<|j<y)Nc&jSr@) _remove_id)refidxrs r8rz4ExactWeakKeyDictionary.__setitem__..s$//#:NrE)r)r'weakrefr2r)rrrr3s` @r8 __setitem__z"ExactWeakKeyDictionary.__setitem__s?g dii $[[.NODIIcN  CrEcr||jvr |j|=||jvr|j|=yyr@)rr')rr3s r8r1z!ExactWeakKeyDictionary._remove_ids6 $++  C $))  # rEcl|jj|jjyr@)r'rxrrs r8rxzExactWeakKeyDictionary.clears   rEr@) rPrr__doc__rr*rr.r5r1rxrtrEr8r$r$s(f$1&!  rEr$clt|tttfr t ||vSt ||uS)zisinstance() without subclasses) isinstancetuplerrr)obj allowed_typess r8istyper>s1-%s!34CyM)) 9 %%rEc^t|tjxs|tjuSr@)r:typing_FinalGenericrs r8 is_typingrDs# eV]] + Fu/FFrEc tsyt|tjtjtjtj tj tjtjtjfSNF) npr>int8int16int32int64uint8uint16uint32uint64rCs r8is_numpy_int_typerPsU   GG HH HH HH HH II II II   rEctsyt|tjtjtjfSrF)rGr>float16float32float64rCs r8is_numpy_float_typerU s2   JJ JJ JJ  rEct|xs{t|tjxrtt j |dxs>t|t jjt jjfS)N __wrapped__) is_functionr: functools_lru_cache_wrapperinspectgetattr_staticr_opsOpOverloadPacket OpOverloadrCs r8is_function_or_wrapperr`skE S eY99 : F ..umD E S eejj995::;P;PQ R rEct|tjtjtjtj t jjfSr@) r:types FunctionTypeBuiltinFunctionTypeMethodDescriptorTypeWrapperDescriptorTyperjitScriptFunctionrCs r8rXrX sH      % %  & &  ' ' II $ $    rEct|dSrw) unwrap_with_attr_name_if_wrapper)fns r8unwrap_if_wrapperrl-s +B / 22rEcvt|tjrtj|d}d}||fSt |r4tj|ddrtj|d|}d}||fSt |r4tj|ddrtj|d|}d}||fSd}||fS)NrW_torchdynamo_inlineF__script_if_tracing_wrapper __original_fn)r:rYrZr[r\rX)rk attr_names r8rjrj1s"i223  # #B 6!  y= RW33B8MuU  # #B(=r B)  y= RW33 )5 # #B <#  y= y=rEcDtsyt|tjSrF)rGr>ndarrayrCs r8is_numpy_ndarrayrtEs  % $$rEctjtjjgtj }|tj jfz}t||S)zCheck of obj is a tensor) rTensornn Parameterrtraceable_tensor_subclasses _subclassesr!r>)r< tensor_lists r8istensorr|LsW     + +K !2!2!=!= ??K #{ ##rEc"t|tSr@)r:r')mods r8is_lazy_modulerWs c? ++rErct|yr@)rW)res r8 print_oncer[s  4LrEc||fd}|jt|jdk(sJ|jdS)zNSome black magic to create a cell object that usually only exists in a closurecSr@rtrsr8rzmake_cell..fdsrErr) __closure__r)valrrs @r8 make_cellr`s> A == $Q]]);q)@@ @ == rEc 0 td|D}|jDcic]\}}||j}}}||fScc}}w#t$r=}ddlm}ddlm}|d||d|t|j|d}~wwxYw)Nc3<K|]}|jywr@)as_proxyr6args r8r9z$proxy_args_kwargs..ms:c3<<>:r unimplemented)typestrzcall_function args: r) r;rrNotImplementedErrorrrvariables.baserrr) rerf proxy_argsrr proxy_kwargserrs r8proxy_args_kwargsrks :T:: "2!GT&--/=R4S3T U   s(%AA A A B8BBcDeZdZUeed<eed<eed<eed<eed<eed<eeed<eeed<eeed <eeed <eeed <eed <eeed <eeed<eeed<eeed<eeed<eeed<eeed<eeed<eeed<eeed<y)CompilationMetricsrr co_filenamer cache_sizeaccumulated_cache_size guard_countshape_env_guard_countgraph_op_countgraph_node_countgraph_input_count start_timeentire_frame_compile_time_sbackend_compile_time_sinductor_compile_time_scode_gen_time_s fail_type fail_reasonfail_user_frame_filenamefail_user_frame_linenonon_compliant_opscompliant_custom_opsN) rPrrr5__annotations__intrfloatrrtrEr8rrysN LO##C=(SM!sm#}$!)%0$UO+%e_,e_$}#&sm+$SM)3xc("rEr@maxlen_compilation_metricscompilation_metricscftj|tjr t |yyr@)rrrlog_compilation_metricsr&)rs r8record_compilation_metricsrs( 34 %%12&rEnew_sizereturnctt|kDr'tjtt|kDr'tjt|}|ay)Nr)rrpopleftrdeque)r new_deques r8set_compilation_metrics_limitrsE " #h .$$& " #h .!!"6xHI$rEc,tjyr@)rrxrtrEr8clear_compilation_metricsrs rEc ttSr@)rrrtrEr8get_compilation_metricsrs $ %%rEcFeZdZUdZeeefed<eed<dZe dZ y) CleanupHookz,Remove a global variable when hook is calledscopenamecdtxjdzc_|j|j=yrr)CleanupManagercountrr)rres r8__call__zCleanupHook.__call__s#! JJtyy !rEcb||vsJtxjdz c_|||<t||Srr)rrr)rrrs r8createzCleanupHook.creates75   !d 5$''rEN) rPrrr8rr5rrr staticmethodrrtrEr8rrs16 S> I"((rErc4eZdZUdZeded<fdZxZS)rrinstancecZ|j|D] }| t| |yr@)rsuperr1)rr3hook __class__s r8r1zCleanupManager._remove_ids-KK$ D F  3rE)rPrrrr rr1 __classcell__)rs@r8rrs E'((  rErc|jj|j}|jr+|j|jj|_|S)z!Clone the tensor and its gradient)clonerequires_grad_ requires_gradis_leafgrad)rys r8 clone_tensorrsB    1AyyQVV' HrEdtypec Lt|r|Sfd}tj5|jjdk(r||cdddSt dt |j|jD}|jrtj|dzf|}n3tj|dzxs |j|j}|j|jz dz|jz}|j|j|j| |j!|j#|j$r|j'|j(|j$r(|j*t-|j*|_t1|dr|j2j5|_|cdddS#t.$r||cYcdddSwxYw#1swYyxYw) zcopy while preserving stridesc>tj|}|jr|j|j|jr(|j t |j |_t|dr|jj|_|S)Nr_dynamo_dynamic_indices) rrrrrr clone_inputhasattrrcopy)rrrs r8 torch_clonez clone_input..torch_clonesr KKN 99  Q__ - 99+ u5AF 1/ 0()(A(A(F(F(HA %rExlaNc32K|]\}}|dz |zyw)rNrt)r6shapestrides r8r9zclone_input..s" %2UFUQY&   )rdevicerr)r"rno_gradrrrzipsizer is_quantizedempty_quantizedemptyrdata_ptr element_size as_strided_copy_rrrrrr RuntimeErrorrrr)rrr needed_sizeresultcache_line_offsets ` r8rrsqz  88==E !q>  69!&&(AHHJ6O   >>**K",<+>BF[[b (8FZZ\FOO- - 3 ^^  1668QXXZ1BC " LL #yy%%aoo6yyQVV/)!&&> 1/ 0-.-F-F-K-K-MF *?0 "q> !90 "1s7!HC6H A:G:,H:H HHHH#ct|tur|t|}|jD]\\}}t|trt |||<%t|t jsJt|t|||<^|St|}tt|D]1}t||t js!t||||<3|Sr@) rr&rr:r; clone_inputsrrvrrranger)example_inputsresrris r8rrs Nt#>"))+ .JC%''.C!%6CU C6&u-C  .  ~ C 3s8_) c!fell + Q(CF) JrErc |jy#t$r7}ddlm}t j ddt |}|d||d}~wwxYw)Nr) SkipFramez\(.*z(torch.compile cannot be run in context: )rrrrresubr)rrrfunctorch_subclass_names r8skip_frame_if_in_functorch_moder!sX   "#%&&"d3i"@67N6O P   s A2AAc#Ktjj}tjjj }|5|5tj tjj}t|tjjr1tj tjj}dddddd dtjjj 5tjjtjjrtjjdddy#1swYxYw#1swYxYw#1swYyxYw#tjjj 5tjjtjjrtjjdddw#1swYwxYwxYwwr@) r_C_DisableFuncTorchutils_python_dispatch_disable_current_modesrrandom get_rng_statercuda is_available set_rng_state)disable_functorchdisable_current_modes rng_statecuda_rng_states r8preserve_rng_stater.s22!KK88OO  E"3"5EKK : : <= ' 2 :: " " $"[[)A)A)CDN EE 9 [[ ) ) @ @ B 9 LL & &y 1zz&&( ((8 9 9EEEE 9 9U[[ ) ) @ @ B 9 LL & &y 1zz&&( ((8 9 9 9sAIF B E?F $I-F#1(IAF6 I?F F  FIF I#)H> AH2) H>2H; 7H>>Ic t|tjjjtjj j tjjtjjfSr@) r:rrg_traceTopLevelTracedModule_scriptRecursiveScriptModulerh ScriptModule)model0s r8 is_jit_modelr@sW  II   1 1 II   3 3 II $ $ II " "   rEc6t|r|S tjj||S#t$r` tjj |cYS#t$r1|rt jdnt jdYYywxYwwxYw)Nz jit errorz0Both torch.jit.trace and torch.jit.script failed) rrrgtracer scriptrr error)modelrrs r8 torchscriptr#LsE  Nyyun55 N N99##E* * N k* LM   NNs,/ BAB6BBBBcX tj|S#ttf$rYywxYwr@)r[getfilerOSErrorr<s r8r%r%^s-s## w s ))c*tt|S)LTest if an object is a namedtuple or a torch.return_types.* quasi-namedtuple)is_namedtuple_clsrr's r8 is_namedtupler+es T#Y ''rEc t|trMt|dgxsdg}t|dd}|dk(xs'|dtuxrt|dxr t|dS y#t$rYywxYw) r) __bases__Nrtorch.return_typesr_make_fieldsF) issubclassr;getattrrr)clsbasesmodules r8r*r*js c5 !Cb1;dVES,5F11aE!Wgc7&;WY@W  "     sAA!! A-,A-c|turgdSt|tsJt|dr |jSt j Gdd}|jdk(sJ|t|t|j}dg|jz}t|D];}|ddk7s tt|||s#||t||j<=|S) zIGet the fields of a namedtuple or a torch.return_types.* quasi-namedtuple)startstopstepr0ceZdZUeed<y)!namedtuple_fields..MarkerindexNrPrrrrrtrEr8Markerr;s rEr>r.Nrr)slicer1r;rr0 dataclasses dataclassrr4rn_fieldsdirr:r2r<)r3r>r<fieldsrs r8namedtuple_fieldsrExs e|(( c5 !! !sI{{ >>1 11 1 c&% -. /C#'&3<<"7FC4 7c>jd);VD/3F73%++ ,4 MrEc Btj5tjtjj tj j r1tjtj j gtj|j|jD]3}j||jtj|f5 dddfd}|S#1swYxYw)Ncjtj5tjjtjj rtjjD]'\}}}|j |k7s|j|) dddy#1swYyxYwr@)rrr rrr_versionr)paramversionoriginal_valuerr saved_states r8restorez"checkpoint_params..restores ]]_ 0 LL & &y 1zz&&( ((82= 0.w>>W,KK/ 0  0 0 0sA5B) B))B2) rrrr r rrr<r= parametersbuffersrrH)gmrIrMrrrLs @@@r8checkpoint_paramsrQs LKK : : <= :: " " $"[[)A)A)CDN __R]]_bjjlC LE   u~~u{{57IJ K L L0 N#LLs C+DDcvtjjrtjj}nt}|t j tjdtj}t|D]}||}|tj}||z fS)Ni9) rrr synchronizer"gccollect manual_seedrT perf_counterr)r"rtimesrSrrrt1s r8timedrZs zz jj,,  MJJL d   B 5\'     B 27?rEc ltdtj||jdDS)Nc34K|]}|jywr@)is_cudar6rs r8r9z check_is_cuda..sWQqyyWsT)allr<r=rN)rPrs r8 check_is_cudar`s( W)//."--PTBU"VW WWrErc|dkDsJt|Dcgc]}d| }}t|dd|ddz}tddj|ddj|d}d|d |_|Scc}w) Nrrzlambda ,z: ()rot__helper)rreversedevalr3rP)rrvarsrotatedrks r8 rot_n_helperrks q5L5"1X &asG &D &tBCy49,-G 's388G+<*=Q? @B7#BK I 's A4ct|ttfrtt t |St |tjtfxst|tthzSr@) r>r; frozensetr_r4is_safe_constantr:enumEnumrcommon_constant_typesr?)rs r8rnrnsU a%#$3'+,, a$))T* + v '0rEclddlm}m}t||r|j |j S|S)Nr)ConstantVariableSymNodeVariable) variablesrsrtr:r evaluate_expr)rrsrts r8specialize_symnoderws0<#'&&s'8'8':;; JrEc`ddlm}t|}t||r|j S|S)Nrrs)rursrwr:as_python_constant)rrss r8 guard_if_dynr{s.+ S !C#'(%%'' JrEcjtdtj||jDS)Nc3<K|]}|jywr@)is_python_constantr^s r8r9z&check_constant_args..sV!q##%Vr)r_r<r=rr!s r8check_constant_argsrs$ VytV]]_/UV VVrEcddlm}ddlm}d}t j ||j D]$}t||r|dz }t|||fsy&|dkDS)Nrry)UnspecializedPythonVariablerF)variables.constantrsvariables.tensorrr<r=rr:)rerfrsr unspec_countrs r8check_unspec_python_argsrsf4=L __T6==? 3 a4 5 A LA ;=MNO   ! rEc|ddlmtfdtj||j DS)Nr)NumpyNdarrayVariablec36K|]}t|ywr@)r:)r6rrs r8r9z+check_numpy_ndarray_args.. s!  1*+s)rranyr<r=r)rerfrs @r8check_numpy_ndarray_argsrs16 v}}7 rE dict_keys dict_values odict_valuestuple_iteratorcnt|}tjjj ||Sr@) object_newrrwModuler)r3r<s r8 nn_module_newrs& S/C HHOOS! JrEcLtjtj|dSrr)rYreduceoperatormul)its r8productrs   HLL"a 00rEc>|j\}\}}|||zSr@) __reduce__)rr<rr<r7s r8tuple_iterator_getitemr#s%}}Avu uu} rEc$|j|Sr@) as_subclass)tr3s r8 to_subclassr+s == rEcZttjt|||dzSrr)rNr<isliceiter)drs r8dict_keys_getitemr/s$    a!QU3 44rEcj|jj}|j}|rdnd}|d|d|}|S)NLGz["z"].)rrPr)rlocalrrr local_names r8 enum_reprr3sA ?? # #D **CCcE7"TF#cU+J rEc|jjjjd}t |sddlm}|d|S)N example_valuerrz:Cannot check Tensor object identity without its fake value)rnodemetarr"rr)vt fake_tensorrs r8_get_fake_tensorr=s>++-$$))--o>K ; &RS rEc ddlm}m}m}m}j r%t fd|D}|j|Sd} |rt|r d} td} |D]} | r.t| |st| us |jdcS|tjj|| gi} | | } a|tjj|| | gi} | |jd} | S)Nr)BuiltinVariablersTensorVariableVariableTrackerc3K|]7}|jxr!|jjk(9ywr@)r~rz)r6rsearchs r8r9z iter_contains..OsJ   " F$$&&*C*C*EE F =AFT)rurrsrrr~rrr:rreq call_functionor_) rrtxcheck_tensor_identityrrsrr found_constmust_check_tensor_idfoundrchecks ` r8 iter_containsrFs  "    &&{33 FN!C#!&)'+E   !^,-a00+22488#HKK0>>rAv;PRSE}' 5CC  } ''. LrEctt|tjtjjt fS)z4Returns whether it indexes dictionaries using its id)r:rrvrwrr)rs r8 key_is_idros# a%,,9JK LLrEct|jDcgc]}t|r t|n|c}Scc}wr@)rrr))rrs r8 key_to_idrts+27**, ?QYq\BqEq ( ?? ?s5cddlm}t|ttfr\dj fd|D}t|trd|dSt|tsJt |dk(rd|dSd|d St|tjrt| jd d S||r |jSt|tr d }||S|S)Nr)is_builtin_callablercc38K|]}t|ywrN const_reprr6srs r8r9zconst_repr..|sDQj%88D[](z,)rdr'rcx|j}|j}|dk(r |jS|dz|jzS)Nbuiltins.)rrr)oklassr5s r8fullnamezconst_repr..fullnames?KKE%%F#)))C<%"4"44 4rE) trace_rulesrr:rr;r3rrorprreplacerPr)rrr elems_reprrs ` r8rrxs0!dE]#XXD!DD a zl!$ $a' ''1v{:,b)):,a(( Atyy !%(00b99 Q zz At  5{rEcHdjfd|D}d|zdzS)Nrcc38K|]}t|ywrrrs r8r9z!dict_keys_repr..sGq 1E22Grrr)r3) const_keysrkeys_strs ` r8dict_keys_reprrs&xxGJGGH >C rE __dict_key)UnsupportedFakeTensorExceptionc |S#t$r9}ddlm}d|jd}tj ||||d}~wwxYw)Nrrz Unsupported: z with fake tensor propagation.)rrrreasonrr)rkrrmsgs r8wrap_fake_exceptionrsM(t )(&ahhZ'EF CC a' (s A 4AA ctjjj|5t fdcdddS#1swYyxYw)Nc.tjSr@)rdeepcopyr'sr8rz)deepcopy_to_fake_tensor..s4==+=rE)rrzr FakeCopyModer)r< fake_modes` r8deepcopy_to_fake_tensorrs;    & & 3 3I >?"#=>???s AA c~tjtjtj||z S)z+ Calculate root mean squared error )rsqrtmeansquare)r2rs r8rmsers) ::ejjcCi!89 ::rEg-C6?Tc  ttttjj tj frtttfsJdtdtttk7r  dyttk(xr%t fdtDSttrttsJtjtjk(s;JdtjdtjtjD]+} t| | |  r" d | yy ttj t"frttj$j&rJttj$j&rJd fd fD\j(r.j(sJj+j+ttj sJdtdtrj,j,k7r dj,j,yj,tj.k(rery tj0j3tj4j3tj4} | s d| Srj7j3tj8j7j3tj8tj0d ry tjj:j=dd} | dkr@t>jAd| jCjEjG| dk\Ssj3j,tj0ry j,tjHk(rtKjG} tMjN| rt>jAdtKjG}j,tjPk(rdnd}jSdks;jTdk(r'jVdjVdcxk(rdk(sndk\rd}||| zd z zk}|s d!|| jY|Sry  d"yttZt\tdt.tj^fr ry k(} | s d#t| Stas tcrSr&tastcsjGttuxrk(} | s  d$| SterSttuxr;ttjftjf Stjhd%vrMttusJt f d&jjjDStmd'tjh)(z-Check correctness to see if ref and res matchNztype mismatch rzLength mismatchFc 3PK|]\}}}t|||   yw)) log_errorN)same) r6aibi fp64_reficos_similarity equal_nan exact_dtype ignore_non_fprrelax_numpy_equalitytols r8r9zsame..sI, "B  $#  , s#&zkeys mismatch z == rrrrrrrzAccuracy failed for key name %sTcdt|tjr|Stj|Sr@)r:rrvtensorrs r8 to_tensorzsame..to_tensors""1ell31 Ha HrEc3.K|] }|ywr@rt)r6rrs r8r9zsame..sMinMszdtype mismatch %s, %sr)atolrtolrz+Accuracy failed: uint8 tensor did not matchrgư>)dimepsgGz?zSimilarity score=%sz=Found nan in reference. Consider running in higher precision.g@g@irbrg{Gz?g$@z4RMSE (res-fp64): %.5f, (ref-fp64): %.5f and shape=%sz+Accuracy failed: allclose not within tol=%szAccuracy failed (%s): %s != %sz!Accuracy failed (numpy): %s != %s) MaskedLMOutputSeq2SeqLMOutput!CausalLMOutputWithCrossAttentionsLongformerMaskedLMOutput InstancesSquashedNormalBoxesNormal TanhTransformFooVariablec 3 K|]7}tt|t |t|   9yw)rN)rr2) r6rrrrfp64_refrrr2rrrs r8r9zsame..sX  S!S!#&-#'%9+#   rzunsupported type: )7r:rr;rrw ParameterListSizerrr_rr&rrsortedrrvrrzr! is_sparseto_denserboolallclosetorLflattenrS functionalcosine_similarityrrcpudetachitemrTrmathisnanbfloat16numelndimrrr5rrrPrUrt as_tensorrP__dict__r)r2rrrrrrrrrrrscore ref_error res_error multiplier passes_testrs`````````` @r8rrsD#eUXX%;%;UZZHI#e}-WS {!DQTI;/WW- s8s3x  ' (3x3s8# , , &)c8%<, )   C #t$$$388:# HHJ#   C C O,DSXXZ0A B C  # AFFQK#1' +)="/' ;Q?! " C%,,. /c5#4#4#?#?@@@c5#4#4#?#?@@@ INc88LMS( ==== =,,.C,,.C#u||,VtCyk4PS9+.VV, yyCII%1399ciiHyyEJJ& NNFFF-FFF-' KL ++-""5==1C++-""5==1C~~c3SsdKHH''99#st9TEt| 1599;3E3E3G3L3L3NOD= ffSYY'~~c3SsiP~~. 3/446 ::i(KKW!3/446 %(II$?SS NN$t+A #))B-399R=*MA*Mh "%J'J,BS4Z,OP "N!! #" CS I C#sDJellC D  3J 6S 3 L 3 #6s#;  c "&9#&>((*C #Y$s) # 5#* 93 D # S T#Y&  D OOC OOC )#!5' -  c    CyDI%%%  ||((*   /S 0B0B/CDEErEc~|jjdd}d|jd|d|jdS)N/rbrz' (rrd)rsplitrr)rshort_filenames r8format_func_infor8sD%%++C04N t||nC/q1D1D0EQ GGrEc#Ktj}tjt_tj}tjt_ d|t_|t_y#|t_|t_wxYwwr@)rcache_size_limitsysraccumulated_cache_size_limit)priorprior_acc_limits r8disable_cache_limitr?s_  # #E!kkF99O*-++F'> "'.=+#(.=+sAB A1B 1B  B guard_failuresz-torch._dynamo.output_graph.GraphCompileReasongraph_break_reasonscbeZdZdZdZdej jfdZdZ dZ dZ dZ y ) CompileProfilerzUtility for profiling how and what dynamo would compile. Can be used for * diagnosing recompilation issues * determining an appropriate compile cache limit * (TODO)confirming which functions got compiled/skipped c6d|_d|_t|_yrw) frame_countr{r?backend_ctx_ctorrs r8rzCompileProfiler.__init__s 3rErPc|xjdz c_|jjD]&}d|jvs|xjdz c_(|j S)Nrr)rEgraphrrr{forward)rrPrrs r8rzCompileProfiler.__call__sP AHHNN #D  "  #zzrEc|Sr@rtrs r8 __enter__zCompileProfiler.__enter__s rEcyr@rt)rtypr tracebacks r8__exit__zCompileProfiler.__exit__s rEcdtiS)Nr@)r@rs r8 get_metricszCompileProfiler.get_metricss  .11rEcJ |j}|dfdfd}Dcgc]}t||||gc} d} fd}tjd}||xsdz }|tjdz }||xsd z }|Scc}w) Nr@c t|Sr@r)rgfs r8num_recompilesz.CompileProfiler.report..num_recompilessr$x= rEcddj|Dcgc] }t|c}Scc}w)Nr0)r3r5)rrrUs r8recompile_reasonsz1CompileProfiler.report..recompile_reasonss(99bh7c!f78 87s-ctdtvr+td}t|Dcgc] }|||g c}ddgSycc}w)N graph_breakzGraph Break ReasonCountr.)r)r:) graph_breaksrs r8graph_break_reportz2CompileProfiler.report..graph_break_reportsL(' 6 9EF#c<,-F17;)Gs5ctrJtDcgc] }| c}}tgd}|tjd|dzSycc}w)N)r RecompileszRecompile Reasonsr.zC Set torch._dynamo.config.cache_size_limit to z/ to avoid being cache limited. )rmaxr:textwrapdedent)rmax_recompiles recomp_tablerUrV summarized_gfs r8recompilation_reportz4CompileProfiler.report..recompilation_reportsm2w!$r%JtnT&:%J!K'!K  $hooBBPAQR' %JsAa Torchdynamo Profiler Report =========================== Graph Breaks ------------ Graph breaks happen when torchdynamo encounters code it can't safely trace. If you want to find out why breaks are happening, check below for each break reason You may gain additional insight by passing `fullgraph=True` to torch.compile, to stop at the first break. zNo graph breaks detected.a+ Recompilation ------------- These subgraphs were recompiled more than once due to guard failures Guard failures indicate some condition assumed to be static by the tracer changed, making it unsafe to reuse the compiled program. zNo recompilation detected. )rQr8rarb) rmetricsrXrr]rfreportrUrVres @@@r8rhzCompileProfiler.reports""$ % & ! 9  d #^D%9;LT;R S       $&E*EE(//    &(J,JJ m s!B N) rPrrr8rrr$ GraphModulerrKrOrQrhrtrEr8rCrCs74 588// 2@rErCcdtjjjdzdztt j z}tj j||S)Nrun_z%Y_%m_%d_%H_%M_%S_%fz-pid_)datetimenowstrftimer5rgetpidrr3)root_dirdir_names r8_get_debug_dirrrsf       ! * *+A B C   biik     77<<( ++rEc8tj}t|Sr@)rdebug_dir_rootrr) debug_roots r8rr$s&&J * %%rEcd|jvr't|jdr|jdS|rddlm}|dyy)Nrrrz9`FakeTensor` example value was required but not available)rr"torch._dynamo.excr)rrequiredrs r8extract_fake_example_valuery)s?$))# /0J(Kyy)) 3QRrEc8t||jusJ|Sr@)r#r)rrs r8ensure_graph_faker{4s q !R\\ 11 1 HrEcdtjjffd }tjjj ||S)Nrc|jdk(rd|jvr t|S|jd}s&t|tj r t |S|S)Nrr)rrget_fake_valuer:rrvr{)rrallow_non_graph_fakers r8visitz)get_fake_values_from_nodes..visit:s\ 44? "aff'D"!R)=> >ff_%# 3 (E$S"- - rE)rr$Nodermap_arg)rrrrs` ` r8get_fake_values_from_nodesr9s4   88==  ..rEc $ddlm}ddlm}m}m}m}m}j} djvr'tjdrjdStjjf|\d| dk(r^tdkDrPtdt j"j$r)t'dj(ft+ddz| dk(rbj,j.j0t3rt5d rj7t'j( j(5t95t;fd } dddddd|s4tkt jltojptr } S#1swYIxYw#1swYMxYw#|$rt<$r(} | } | j> | j>} t| t j@jBjDr|d | jFd nt| t j@jBjHr|d | jFdnRt| t j@jBjJr| jF} d} t| t jLjNr[t jPjS| jTjV| jTjX}||\}}d|d|d|d} |d| jFd| dn}t| t jZj\j^j`r||jbd| dt| |r!||jb| jd| |te| jg| jhdd} ~ wwxYw)az Run the computation represented by `node` using fake tensors and return the result. allow_non_graph_fake: whether to allow the return result to be: 1. non-fake or 2. fake that is not created by this instance of Dynamo. If `True`, you must be prepared to deal with such return values, ideally by further wrapping them as this graph's fakes. r)ValueRangeErrorr)TorchRuntimeErrorr Unsupported UserError UserErrorTyperN call_method call_module_initialize_hookc6tjSr@)run_nodeoutput)rerfnnmodulerrsr8rz get_fake_value..ysD$IrEzdata dependent operator: zC; to enable, set torch._dynamo.config.capture_scalar_outputs = Truezdynamic shape operator: zM; to enable, set torch._dynamo.config.capture_dynamic_output_shape_ops = Truerz:It's possible that the support was implemented in module `z` and you may need to `import z`(z ), otherwise zunsupported operator: z (z~see https://docs.google.com/document/d/1GgvOe7C8_NVOMLOCwDaYV1mXXyHMXY7ExoewHqooxrs/edit#heading=h.64r4npvq0w0 for how to fix)aTried to use data-dependent value in the subsequent computation. This can happen when we encounter unbounded dynamic value that is unknown during tracing time. You will need to explicitly give hint to the compiler. Please take a look at constrain_as_value OR constrain_as_size APIs. constrain_as_size_example) case_name)r):torch.utils._sympy.value_rangesrrrrrrrrrr"rrerfrr:rrwrrrr;r nn_modulestargetrr_infer_parametersr%rr __cause__rzrDataDependentOutputExceptionrCDynamicOutputShapeExceptionUnsupportedOperatorExceptionr]r_r_dispatch_pystub_schemar overload_namer$ experimentalsymbolic_shapesGuardOnDataDependentSymNodeCONSTRAINT_VIOLATIONr5with_traceback __traceback__r(rvrYpartialr{)rrrrrrrrrrret_valrcauseimport_suggestion maybe_pystubr5ctxrrerfrs`` @@@r8r~r~Hs@ B$))# /0J(Kyy))- TYY $&:LD&H ]s4y1}DGUXX__1U'Q>@5ab?R ]99'' 4 ( #:L(M  & &x 6+8R\\B<R \\ 35 )IG  z  LL)++,="Ew  NC      5R  ;; "KKE  5$$00MM  +EJJ<8TT  5$$00LL  *5::,7^^  5$$00MM B " "ejj334$xx88JJOORZZ%=%=   +".KFC##)(*HQ5 /& ( B7H6IJ##   588((88TT 22BCHJ6   /M>>q JPQ QA'66qGTQk5RsI G G %G8G GG G  GGP'H#P  Pc$ttddS)Nr)r2 _current_nodertrEr8get_current_noders ='4 00rEc#vKt}|t_ d|t_y#|t_wxYwwr@)rrr)rolds r8set_current_noders.  CM" ! c s9) 9 69cjt5fd} dk(rjicdddSdk(r+tdjddicdddSdk(r|J|icdddSdk(r$|j jcdddSd k(r(d j vsJj d cdddS dddt#t tf$r}dd lm }||||d}~wt$r0}t||j|j|d}~wwxYw#1swYtxYw) a Runs a given node, with the given args and kwargs. Behavior is dictated by a node's op. run_node is useful for extracting real values out of nodes. See get_real_value for more info on common usage. Note: The tracer arg is only used for 'get_attr' ops Note: The nnmodule arg is only used for 'call_module' ops Nodes that are not call_function, call_method, call_module, or get_attr will raise an AssertionError. c Nddjddd t|zS)NzFailed running rz(*z, **z): )rr5)rrerfrrs r8make_error_messagez$run_node..make_error_messages4$RD$++bd6($ORUVWRXX XrErNrrrrget_attr placeholderrr)rrrr2 get_submodulerrrrrr rrrAssertionError) tracerrrerfrrrrrs ``` @r8rrs B $  Y _$"t{{D3F3}$4wtAw 4d12hI&I}$+++00z!++DKK8}$&$))333yy1#%:  $%CD > * 21 56A = 1!45DD  1:  sL EC:&C:6C:C:9#C::E D E*+EEEE0c,ddlm}j}||vr||S|j}tj j j|j|jffd\}}|dk(rMjj|j}t|stj|}n ||i|nd} t!||||}|||<|S#t"$r0} |t%| j'| j(dd} ~ wwxYw)z Run the actual computation represented by `node` and return the result. This will execute any dependent nodes in the graph as well. r)rct|Sr@)get_real_value)rrs r8rz get_real_value.. s.F+rErN)rrreal_value_cacherrr$rrrerf output_graphrrrrrrrr5rr) rrrcacherrerf nn_module real_valuers ` r8rrs '  # #E u}T{ B88==(( DKK +LD&  ]''224;;? i( i0I t &v & RfdD&)D  d   RA'66qGTQRsC D#+DDcXddlmfd}|jD]<\}}t|tj j s+Jd|d|||jD]<\}}t|tj j s+Jd|d||y)Nr)FakeTensorConfigcbjr"ddl}d|j|jSy)Nrz"FAKE TENSOR CREATION TRACEBACK: zNEnable TORCH_FAKE_TENSOR_DEBUG=1 to get creation stack traces on fake tensors.)debugrN format_list _debug_trace)rrNrs r8 stack_or_hintz7assert_no_fake_params_or_buffers..stack_or_hint$s/  ! ! 89N9Nq~~9^8_` `crEzUnexpected fake buffer rzUnexpected fake param )torch._subclasses.fake_tensorr named_buffersr:rrzr!named_parameters)rPrrbufferrIrs @r8 assert_no_fake_params_or_buffersr!s>d((*D f E%%00  D $TF!M&,A+B C D D**,B e 5$$//  B #D6=+?*@ A B BrEr<c8|jd|jS)z9 Returns the fully qualified name of the object. r)rrr's r8fqnr6s!nn Qs//0 11rEcRtjjjr|S|Sr@)r_dynamorassume_static_by_default)count1count2s r8ifdynstaticdefaultr=s }}44  rEr~c 8ttjtjj t t |jD]D}|jds|ddk7stj|jd|ddFy)z@ Ensure all the files in a given submodule are imported z.pyrrrN) rrlistdirrrr r5__file__endswith importlib import_modulerP)r~r s r8import_submodulerDsw2::bggood3 6M&NOPG   U # s(:  # #s||nAhsm_$E FGrErc ttjt|dtj ry y#t $rYywxYw)N__getattribute__TF)r:r[r\rrbrcAttributeErrorrCs r8object_has_getattributerMsS   " "4;0B C           s7< AAc tjt|d}|tj j jurd}|S#t$rd}Y7wxYw)N __getattr__)r[r\rrrrwrr)r getattr_fns r8get_custom_getattrrYsV++DKG UXX__000   sA AAceZdZdZdZdZy)TensorStaticReasonrr rN)rPrr PARAMETER NOT_TENSORNN_MODULE_PROPERTYrtrEr8rrdsIJrErrc|tjk(ry|tjk(ry|tjk(ryt d|)Nz>mark_dynamic on parameter, parameters are always static today.z2mark_dynamic on a non tensor, how did this happen?z4tensor is static because it is nn module associated.zIllegal reason )rrrrr)rs r8tensor_static_reason_to_messagerjsJ #---O #...C #666E ?6(3 44rEr is_tensor guard_sourceztorch._guards.GuardSourcec|jr"tjrdtjfSt |t jjur"tjrdtjfS|sdtjfSy)a@ Given a tensor, source, and is_tensor flag, determine if a shape should be static. Args: tensor - the real tensor to evaluate, parameters force a static shape. is_tensor - internal dynamo check, essentially "is_tensor": target_cls is TensorVariable, tensors not in a TensorVariable for whatever reason are forced static. Returns a tuple, where the first element is the bool of whether or not this tensor should have a static shape. The second element is a TensorStaticReason, useful for passing to tensor_static_reason_to_message if needed. T)FN) is_nn_moduler&force_nn_module_property_static_shapesrrrrrwrxforce_parameter_static_shapesrr)rrrs r8tensor_always_has_static_shapertsr   "v'T'T':::: F|uxx)))f.R.R'1111 '2222 rEc4fdtfdS)NcdSS)Nrrt)maybe_idrsr8 format_namez+lazy_format_graph_code..format_names  V1XJ' 'KrEctddjjjj dS)Nz===== z ===== F) print_output)_format_graph_coderI__code__rprint_readable)rrPsr8rz(lazy_format_graph_code..s?"[]O8 , JJ   + +   5  1 rEr)rrPrrs```@r8lazy_format_graph_coders   rEcd|d|d|dS)NzTRACED GRAPH rr0rt)rr  graph_strs r8rrs TF!H:Qyk <.inners  )AC ;+>+J+JIVV ~,Wb9:;    sBAC!B>=B>r)rrPrs`` r8lazy_format_graph_tabularrsW" e rEc f|d|d|d|dtj|jd S)Nrz line z r0)disBytecode)prefixrr line_nors r8format_bytecoder s;XQtfAhZvgYc#,,t:L:P:P:R9SSU VVrE_forward_pre_hooks_forward_hooks_backward_pre_hooks_backward_hooks)_state_dict_pre_hooks_state_dict_hooks_load_state_dict_pre_hooks_load_state_dict_post_hookscztjjjj} g}| xr| xr| }|s|r|j t |s|r|j t|r|j tg}|D],}t||g} | D]} | | } |j| .|Sr@) rrr eval_frame reset_codeextendforward_hook_namesbackward_hook_namesstate_dict_hook_namesr2r) r~check_forward_hookscheck_backward_hookscheck_state_dict_hooksrhook_dicts_to_checkcheck_all_hooks all_hookshook_dict_namehooks hook_namers r8nn_module_get_all_hooksr#s !!,,77J  '$ $ '& & o""#56""#67""#89I-#^R0 #I#D   T " ## rEc6t||||}t|S)zL Helper function to check if a module has any hooks attached to it. )rrr)r#r)r~rrrr!s r8nnmodule_has_hooksr%s& $ /15  E ;rEc2t|r|St|tjrt |j St|t jr|jdSt|ttfrt|d|DS|S)z0Convert tensor and tnp.ndarray to numpy.ndarray.T)forcec32K|]}t|ywr@)to_numpy_helperr6r<s r8r9z"to_numpy_helper..AC?3/Ar) r"r:tnprsr)rrrvnumpyr;rrrCs r8r)r)stu~ %%u||,, E5<< ({{{&& EE4= )tE{A5AAA rEctJt|tjrtj|St|t jr |j St|ttfrt|d|DS|S)zeConvert tnp.ndarray to tensor, leave other types intact. If a list/tuple, loop through it to convert.c32K|]}t|ywr@)numpy_to_tensorr*s r8r9z"numpy_to_tensor..r+r) rGr:rsrr-r,rr;rrrCs r8r0r0sk >>%$u%%%%|| EE4= )tE{A5AAA rEceZdZdZdZdZy)numpy_to_tensor_wrappercN||_d|jjz|_yNwrapped_rrP)rrs r8rz numpy_to_tensor_wrapper.__init__s"TVV__4 rEc6d|jjdS)Nz>r6rs r8__repr__z numpy_to_tensor_wrapper.__repr__s-dffoo->bAArEc<|j|i|}t|Sr@)rr0rrerfrs r8rz numpy_to_tensor_wrapper.__call__s"dffd%f%s##rEN)rPrrrr9rrtrEr8r2r2s5B$rEr2ct|tjrt||}t |St|t j r*ttj||}t |Syr@)r:r,rsr2r0rrv)r<rrs r8numpy_attr_wrapperr=!sZ#s{{#c4 s## C &ckk#&-s## 'rEc(eZdZdZdefdZdZdZy)numpy_method_wrapperzgConvert obj from torch.Tensor to tnp.ndarray and call method. Then convert result back to torch.Tensor.methodc:||_d|jz|_yr4)r@rP)rr@s r8rznumpy_method_wrapper.__init__-s "T[[0 rEc"d|jdS)Nz.Js0 KN 3 =CKK 3 F s;=)rr0r;s r8rznumpy_operator_wrapper.__call__Gs5z RV dggtns##rEN) rPrrr8r rrr9rrtrEr8rGrG=s%[18CH-1@$rErGcBt|ts|S|jrg}|jD]v}t|tj rI|j |jjj|jjf|j |xg}|jD]v}t|tj rI|j |jjj|jjf|j |xn |j}|j}t j|||j|j|j}|j!|S)N)rrr)r:r!_has_symbolic_sizes_stridesrrSymIntrr shape_env size_hintexprr empty_stridedrrrzero_)rrrrrs r8defakerTQs* a $ $$ A!U\\* AFF,,66qvv{{CD A    !A!U\\* aff..88EF a  ! vvx  ggxxoo  AGGI HrEcNddl}||jjjuSrw)torch.utils.checkpointr  checkpoint)r<rs r8is_utils_checkpointrXqs ! %++((33 33rEc ddlmcm}ddlm}|j }t jjjr |j}|j|fi|S)Nrr) TorchHigherOrderOperatorVariable) torch._higher_order_ops.wrap_higher_order_opswrapvariables.higher_order_opsrZtag_activation_checkpointr _functorchrfunctionalize_rng_opswrap_activation_checkpointmake)optionshigher_order_opsrZactivation_checkpoint_ops r8build_checkpoint_variablergxs_;;L 22 44#3#N#N 0 + 0 0    rEc`ddlm}|}|dk(r |S|dk(r|rddlm}|}|Sd}|S)Nr)is_dynamo_supportedr%rr) has_tritonF)rritorch.utils._tritonrj) device_typericompile_supportedrjs r8is_compile_supportedrnsO/+-e    #42&L " rEoffsetcb|jd}t|d|jddS)z Convert byte offset `offset` of `str` into character offset. Byte offset is used for 3.11+ instruction column data. Takes things like unicode characters into consideration. Unchanged from CPython implementation. zutf-8Nr)errors)encoderdecode)r5roas_utf8s r8 _fix_offsetrus4jj!G ww&&wy&A BBrEc6eZdZUeed<eed<eed<eed<y)_Anchorsleft_end_linenoleft_end_offsetright_start_linenoright_start_offsetNr=rtrEr8rwrwsrErwsegmentctjdk\sJddl} |jd|zdz}t |j dk7ry|jdfd}fd fd }fd }|j d}t||jr|j}t||jrtt|jjd z }|||jj } || \}} || x} j#s| d vr;| dvr ||| \}} n ||| \}} || x} j#r6| d vr;| dz} | t |kr!|| x} j#s | dvr| dz } t%|| || St||j&rtt|jjd z } || |jj } | | \} } | | dk7r|| | \} } | | dk7rtt|jd z }|||j } t%| | || St||j(rtt|j*jd z } || |j*j } | | \} } | | dk7r|| | \} } | | dk7rtt|jd z }|||j } t%| | || Sy#t$rYywxYw)a Given source code `segment` corresponding to a bytecode instruction, determine: - for binary ops, the location of the binary op - for indexing, the location of the brackets. `segment` is expected to be a valid Python expression ) rNz( z )rr0c"t||Sr@)ru)linenoroliness r8 normalizez-_extract_anchors_from_expr..normalizes5=&11rEc|tkr8|t|k\r'd}|dz }|tkr|t|k\r'|tkr|t|ksJ||fSrDrT)rcolrs r8next_valid_charz3_extract_anchors_from_expr..next_valid_charsys5z!cSv-?&?C aKFs5z!cSv-?&?E "sSv-?'???s{rEcp|dz }||\}}|tkr|t|ksJ||fSrrrTrrrrs r8 incrementz-_extract_anchors_from_expr..incrementsH q%fc2 E "sSv-?'???s{rEctd}|dz }||\}}|tkr|t|ksJ||fSrDrTrs r8nextlinez,_extract_anchors_from_expr..nextlinesM! %fc2 E "sSv-?'???s{rErz)\#z\#rr)r; version_infoastparse SyntaxErrorrbodyr6r:ExprrBinOpr rleft end_linenoend_col_offsetisspacerw SubscriptCallrC)r|rtreerrr statementrQ cur_linenocur_colch right_col left_linenoleft_col right_linenorrs @@r8_extract_anchors_from_exprrsQ   w && &yy501 499~ MM$ E2  ! I)SXX& dCII & c499#7#781",;*2:w*G'J*3J*H'J z*7332<<>", ! ICj 122$Z0;;DDFeOQ JYG G cmm , sDJJ$9$9:Q>K djj.G.GHH$3K$J !K $X.#5(1+x(H% X $X.#5 T__59L!,0C0CDIK<K K chh ' sDII$8$89A=K dii.F.FGH$3K$J !K $X.#5(1+x(H% X $X.#5 T__59L!,0C0CDIK<K K K sL77 MMrinstc |jJ|jjytj|j|jjj }|jj |S|jj|jj|St||jj}d}d}g}|jj |jjk(rDt||jj}|||}|jd|zd||z zzn||ddz}|jd|zdt||z zztj|j|jj j }t||jj}t|jjdz|jj D]}tj|j|j } || dzz }t| t| jz } |jd| zdt| | z zz||d|z }t|t|jz } |jd| zd|| z zzd} t|} | !|D cgc]} | j!dd}} n%|D cgc] } t#| } } | j$dk(r| xj&|z c_| j(dk(r| xj*|z c_tt|D]}tt| |D]t}|| j$kr|| j$k(r|| j&kr2|| j(k(r|| j*k\rQ|| j(kDra| ||dk(smd| ||<v| D cgc]} dj-| }} d}tt|D]X}|tj|j|jj|zj dzz }|||dzz }Z|S#t$rYwxYwcc} wcc} wcc} w) a Python 3.11+ only. Returns lines of source code (from code object `code`) corresponding to `inst`'s location data, and underlines relevant code to `inst`. Example: CALL on `g`: f(g( ^^ h(x))) ^^^^^ We need our own implementation since `format_frame_summary` in Python's `traceback` module doesn't handle multi-line expressions (and their anchor extraction code is not completely correct). Nrr~r0r^r) positionsr linecachegetlinerrstripr col_offsetrrurrrlstriprrrrrxryrzr{r3)rr first_line start_offset end_offsetr|markers last_linerline num_spacesanchorsmarkermutable_markersrrrs r8get_instruction_source_311r, s >> %% % ~~$""4#3#3T^^5J5JKRRTJ ~~  ( ~~  (DNN,I,I,Qz4>>+D+DELJGG ~~  DNN$9$99 T^^-J-JK \*5s\)C: 3L,MMN\]+d2s\)C3z?\3Q,RRS%%   dnn77 &( !DNN,I,IJ DNN11A5t~~7P7PQ NF$$T%5%5v>EEGD td{ "GTS%77J NN3+cSY5K.LL M  N 9[j))^c)*:*:*<&== sZ'#j1H*IIJ"&G ,W5 :AB6>>#s+BBHO+OVDL+O+O  " "a '  # #| 3 #  % % *  & &, 6 &CL) 7FS!89: 7G333W444w?V?V9Vg888w999G666"6*3/3636OF+C0 7 7"2AAv2776?AA F 3w< $   d..0E0E0I J Q Q S   '!*t## $ MW    C,P4Bs$ R%0R5R:R?% R21R2cRt|tjr t|ddSy)N_dynamo_static_input_type)r:rrvr2rs r8get_static_address_typer s#!U\\"q5t<< rEctjjjtjjtj tj j f}tjjjtjjtjtj jf}|g||vSr@) rr Generator get_statedefault_generatorr r set_stater)rgetterssetterss r8is_rng_state_getter_or_setterr s $$ ))    G $$ ))    G (g(( ((rEct|tjxrA|jdk(xr0|jj t jjuS)N__get__) r:rbrrP__self__ __objclass__rr _TensorBaserCs r8is_tensor_base_attr_getterr sK5%112 @ NNi ' @ NN ' '588+?+? ?rEct|dS)N__torch_function__)rrCs r8is_torch_function_objectr s 5. //rErz,torch._dynamo.variables.base.VariableTrackerc~ddlm}ddlm}t ||xs$t ||xrt |j dS)Nr)UserDefinedObjectVariable)TensorWithTFOverrideVariabler)torch._dynamo.variablesr&torch._dynamo.variables.torch_functionrr:rr)rrrs r8has_torch_functionr s<AS b6 7 201 4 BHH2 3rEcd}d}tjjjx}r)||jvr|j|}|j }|j |d||S)NF) static_shapessymbolic_contextsource)r_guardsTracingContexttry_gettensor_to_context tensor_source from_tensor)rrrrtracing_contexts r8to_fake_tensorr sw F--66>>@@@ 11 1.@@C %33F   1A& ! rEcd|D]}t||st||cSt|d|)zX Return the first available attribute or throw an exception if none is present. z% does not has any of the attributes: )rr2r)r<attrsattrs r8get_first_attrr sC& 3 3% %& C5 EeWM NNrEc#Kd}|r7tjjj|5}|dddydy#1swYyxYww)Nc:d}tj||ddS)Nctjjjddxxdz cc<tjj ||S)Ncompiled_autogradcompilesr)rrr r) _inductorcompile)gm_example_inputs_s r8inner_compilerzKmaybe_enable_compiled_autograd..compiler_fn..inner_compiler s? MM   ( ()< =j IQ N I??**3@ @rET)backend fullgraphdynamic)rr)rPrs r8 compiler_fnz3maybe_enable_compiled_autograd..compiler_fn s  A}}R4QUVVrE)rrrrS) should_enablerrs r8maybe_enable_compiled_autogradr sNW ]] , , 3 3K @ CI     s/AAAA  Ac@Gddt}t|S)Nc eZdZy)*invalid_removeable_handle..InvalidN)rPrrrtrEr8Invalidr s rEr)r&r)rs r8invalid_removeable_handler s $  79 %%rEc t|tjjs|St|tjj r|S|j j|j }|j|_|Sr@) r:rrwrr$rir__new__r.)r~proxys r8nn_module_proxyr sZ c588?? + #uxx++, MM ! !#-- 0E\\EN LrE)NN)r5Fr@)rN)F)r)T)FFF(?atexitrrrrQr@rlrrorYrTr[r<rrr(rrrIrr[r;ra threadingrTrbr@r4rrrpathlibrrrr r r r r rrrrrrrrrrr utils.hooksrr-rGModuleNotFoundErrortorch._loggingr torch._numpy_numpyr, torch._guardsrrrrfftlinalgr r ModuleTyperNP_TO_TNP_MODULEr;rr!r"r#r;rtorch._functorch.config%torch.fx.experimental.symbolic_shapesr$torch._dispatch.pythonr%torch._utils_internalr&torch.nn.modules.lazyr'torch.utils._pytreer( defaultdictr)r5r*troubleshooting_urlnnmodule_doc_urlnnmodule_doc_url_msgrrPrr+rr,rrOr:rDrBrsruryr{r}rrrregisterr FloatTensorrS DoubleTensorrTdouble HalfTensorrRhalfBFloat16Tensorr* ByteTensorrL CharTensorrH LongTensorrKlong IntTensorrJr ShortTensorrIshort BoolTensorrtensortype_to_dtyperrrrrrrrGraphrrrr"r$r>rDrPrUr`rXrlrjrtr|rrrrrAr!DEFAULT_COMPILATION_METRICS_LIMITrrrrrrrrrrrrrvrrrr#r%r+r*rErQrZr`rkcomplexbytesrEllipsisrCodeTyperr memory_formatlayoutrqrnrwr{rrrr&rrrrrrrr__length_hint__tuple_iterator_lenobjectrrrrrrN iter_nextrrrrrrrrrGLOBAL_KEY_PREFIXtorch._subclassesrrrrr!rr8r? orig_code_maprr@rA seen_code_maprCrrrryr{rr~rrrrrrrrrrrrrprrrrrrr rrrall_hook_namesr#r%r)r0r2r=r?rGrTrXrgrnrurwr InstructionrrrrrrrrrrrrtrEr8r4s:         %&#(* .)  FF II II > eE$4$4c$9:   FFCGG IIszz IIszz   %wVV ,;71-,C;+B+B;CVCV+W+c73<' (W$&4S>&TK.//UVg!46$sDK/0535Dd3:../4 !   -`   H@!H88   u{{3  5 u}}ejj1 5>>+ u{{n uzzm u{{EJJ/ OOekk599-  U[[1 uzzm (#:"; ",  ? G 288   B& G&  3(%$,T  ###2%'!3D+2C2C ,3e./ 34F3%C%D%! &&8!9& ((($ + )* !0f&  99" $( Q0*"X 2   J  NN LL KK  LL"W  "&dfkkm!4 4 4%)$&--/%: T*S/ ":&*+B;+B+B+D+K+K+M&N d:c?#N&*4=&9Xc]#9#33 ^^  1  5&RM @sB C !=(? ; iiVFrH   > >'( /Fk.E.Ed.K CcN+KNRVTIJS'( __FT,,&  /pf ! 1"".b!HB*2S2G%**G 3 c 5,>5 %,,# $. 4,- -. 2  =*W+,<=,.?@ $&99B0b5'b21b25b>=b>