9inC2 UdZddlmZddlZddlZddlZddlZddlZddlZddl Z ddl Z ddl Z ddl Z ddl Z ddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlZddlmZddlmZddlmZmZmZmZddlm Z m!Z!m"Z"m#Z#ddl$m%Z%dd lm&Z&ddl'Z'ddl(Z'dd l)m*Z*dd l+m,Z,dd l-m.Z.dd l/m0Z0ddl1m2Z2m3Z3m4Z4m5Z5ddl(m6Z6ddl7m8Z8m9Z9m:Z:m;Z;ddlm?Z?m@Z@ddlAmBZCddlDmEZEddlFmGZGddlHmIZImJZJmKZKddlLmMZMmNZNddlOmPZPmQZQmRZRmSZSmTZTddlUmVZVmWZWddlXmYZYmZZZm[Z[m\Z\m]Z]ddl^m_Z_m`Z`maZambZbddlcmdZdmeZemfZfmgZgmhZhmiZiddlRmjZjmkZkmlZlmmZmmnZnmoZompZpmqZqmrZrmsZsmtZtmuZumvZvmwZwddlxmyZydd lzm{Z{m|Z|m}Z}dd!l~mZdd"lmZmZdd#lmZmZmZdd$lmZdd%lmZdd&lmZmZmZmZmZdd'lTmZdd(lmZmZmZmZdd)lmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZmZdd*lmZd+ed,< ddlZn #e$rdZYnwxYwejr.dd-lmZdd.lmZdd/lmZdd0lmZdd1lmZdd2lmZmZmZdd3lmZe jeĦZe'jed4Ze'jed5ZejZe"d6Ze%d7ZGd8d9ZGd:d;ZeϦZeϦZdadzTracker.add..sDM4H4H4M4Mr)idrweakrefrefrappendadd)rrobjrs` @rrz Tracker.addstnn dm # #+j*M*M*M*M*MNNC I  S ! ! ! M  c " " " " " $ #ritemboolc.t||jvSr)rr)rrs r __contains__zTracker.__contains__s$xx4=((rcj|j|jdSr)rclearrrs rrz Tracker.clears.  rNrr)rrrr)rrrr)rrrrrrrrrrrrsd((((####))))rrzOptional[GlobalStateGuard]initial_global_statesrcstrglobalsdict[str, Any] co_fieldsOptional[dict[str, str]]rrcPt|||}t|j|Sr)original_forward_from_srcr=__code__)rrrresults rfx_forward_from_src_skip_resultrs* 'sGY ? ?F fo Mrcoderskipint list[str]ctjttjd|z}tj|}ttj fd||j |j tj|j dgztjdfdfd|Dd|j d |j d |j gz}|S) Nrc|dkS)Nfilenamer)fconvert_frame_interns rrz"log_dynamo_start..sa m';;r)linenamer dynamo_startc diS)Nstackr)rsrrz"log_dynamo_start.. s % rcZg|]'}|jk d|jd|jd|j(S)Line: , Name: , Filename: )rlinenor).0framers r z$log_dynamo_start.. sP  >1 1 1 POOuzOOu~OO 1 1 1rrrr)r intern_string__file__r&extractsummaryfrom_tracebacklist itertools takewhileco_firstlinenoco_name co_filenametorch_loggingtrace_structured)rr captured_tbframes_interned stack_stringsrrs @@rlog_dynamo_startrs3%3H==#+T:::BBDDK / <._fn s!##/11 H . . 0 0< <  ( * *< <  7 8 8< <  $)#B#D#D "'"L"N"N #QSSO==?? )"?,,L133K#l88::O!Nz&&(( @X2244@@%*Z%=%=%?%?N@@@@@@@@@@@@@@@$)H$G$G%% !"'!6!H 6UEH ! 3)++G#-//J  $ $(BBDD     $ $%H I I I & ( ( ( r4*6** x99;;q@@@gA@@  """**?;;;(>>?STTT2'? --- **?;;;' 444HAACC,03NNNH@@BBB!-6688AA 00@@@AAAAAAAAAAAAAAA33H&;;M%7||~~ffmmoofffu< < < < < < < < < < < < < < < < < < < < < < B x99;;q@@@gA@@  """**?;;;(>>?STTT2'? --- **?;;;' 444HAACC,03NNNH@@BBB!-6688AA 00@@@AAAAAAAAAAAAAAA33H&;;M%7||~~ffmmoofffu< < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < < sZY8CY!9E$  Y!$E( (Y!+E( ,CY!Q$EY! O = Y! O  Y!O A.Y!? Y8 Z$EY ; W'  Y 'W++Y .W+/A/Y Y!!Y% %Y8(Y% )Y8, Z8Y< <Z?Y< ZZZrrrrrr) functoolswraps_torchdynamo_orig_backend)rr#s` rpreserve_global_stater(sN_RDDDDDDL%'C! Jrrrrc"|jtvrdS|jjD]}||jvrv|j|}t |t r&|jds |turdStr'tj r|tust|rdSid fd |j D]}|rdStd|jj|jj|jjd S) z+Check if the frame has torch.* related bitsTztorch.robjectrrc "t}|vr|Sd|<ttjtjjfs4t tr,ttjjr d|<|Stj rHtrAt tj sttj r d|<|St ttfr&t!fdD|<|St t"rGt}t!fd|D|<|St t&t(t*tdt,fr d|<|St/rz:has_tensor_in_frame..has_tensor..s+">">Q::a==">">">">">">rc3.K|]}|VdSrrr-s rr0z:has_tensor_in_frame..has_tensor..s+"A"AQ::a=="A"A"A"A"A"ArN_fieldsc3JK|]}t|VdSr)getattr)rr.r/rs rr0z:has_tensor_in_frame..has_tensor..s5"T"T1::gc1oo#>#>"T"T"T"T"T"Tr)r isinstancerTensornnModulerotype issubclassr) trace_numpyrzndarraygenericrtupleanydictvaluesrrfloatrrnhasattrr2)robj_idrAr/rs` rr/z'has_tensor_in_frame..has_tensorsC X  F# #  cEL%(/: ; ;!  3  ! ",S%(/"B"B!  $HV F# #    RZ(( -7sBJ,G,G  $HV F# # C$ ' ' "">">">">#">">">>>HV F# # C   #**,,''F""A"A"A"A&"A"A"AAAHV F# # C#sE4::t< = = $HV F# # 3   GC$;$; ""T"T"T"T"T "T"T"TTTHV F# # 5rz0skipping because no torch.* %s %s %sF)rr*rr)f_coder9co_names f_globalsr5r r startswithrrzr)r;r^f_localsrAlogdebugrrr)rrrvaluer/rs @@rhas_tensor_in_framerMksP |333t<(   eo % %/'*C#z**  ''11 58E\\tt f( cRii8C==itt "H(((((((V&&(( :e   44 II     #  5rFe ExceptionOptional[DynamoFrameType]exportrcd}t|dr,t||}t||j||_t ||dS)N exec_recordrQ)rCrjrxrSrecord_filenamer?)rNrrrQrUs rexception_handlerrVs] Oq-  ,.q$77_am<<<+&))))))rz/typing.Counter[Union[int, FrameStateSizeEntry]]FRAME_COMPILE_COUNTERfuncc<tjrt|S|Sr)r)cprofilecprofile_wrapper)rXs rmaybe_cprofiler\s &%%% KrcHtjdfd }|S) Nrrrrrrc  tj}|s Jdtd jdt |ddd}t j} tj}|j g|Ri|}tj|z }n4#t$r't dd} |i|}YnwxYwt d j||tj|} ||n+#t"$rt d |YnwxYwt d ||d } t'jd d ddddt |gt&j5} t'jdddt | g| jt d| dddn #1swxYwYn#t0$rt d|tjjd|tjjdYnwxYwt=t |x r#t>j !d fd|S)NzTrace id is Nonez/tmp/r/z.profilezfailed to enable cProfilerz7### Cprofile for %s trace id [%s] took %.3f seconds ###zCannot write to %szRaw profile at %sz.svg gprof2dotz-fpstatsz"--node-label=total-time-percentagez!--node-label=self-time-percentagez--node-label=total-time)stdoutdotz-Tsvgz-o)stdinz Generated SVG from profile at %szxFailed to generate SVG from profile -- dumping stats instead.Try installing gprof2dot and dot for a better visualizationlinkcddS)Ncprofile_manifold_url)rurlr) manifold_linksrrz;cprofile_wrapper..profile_wrapper..s!8OOr)"rcurrent_trace_idrrrreplacecProfileProfiletimeruncall ValueErrorrJ exceptionwarningraStats dump_statsOSError with_suffix subprocessPopenPIPE check_callrbFileNotFoundError sort_statsSortKeyTIME print_stats CUMULATIVErrrr) rrtrace_id profile_pathprofstart_tsretvalprofile_latencypssvg_pathgprof2dot_processrjrXs @rprofile_wrapperz)cprofile_wrapper..profile_wrappers!244++++++ MDM M MCMM$9$9#s$C$C M M M  !! +y{{H "T\$888888F"ikkH4OO + + + MM5 6 6 6OT4*6**FFF + E M      \$   > OOL ) ) ) ) > > > MM. = = = = = > '666++F33 E!87- %%"    J#%GT3x==9+2 >III! J J J J J J J J J J J J J J J"! E E E KKN    MM&.- . . : :2 > > > MM&.3 4 4 @ @ D D D D D  E@      =  N + +OOOO    s\2:B--.CCD..%EE 4H>AH HHHHHBJ32J3r$)r%r&)rXrs` rr[r[sE_TAAAAAAF rceZdZUdZded<dS)ConvertFrameBoxNzOptional[bool]error_on_graph_break)rrrr__annotations__rrrrrs$+///////rr frame_state*dict[str, Union[int, FrameStateSizeEntry]]rcd|vrt|d<tdz a|d}t|tsJt|}t|xxdz cc<d}t jx}r|j}t|||S)N_idr()compiled_autograd_idframe_idframe_compile_id) FRAME_COUNTERr5rrWrcurrent_compile_idrr)rrrrpriors rget_compile_idr"s K* E 5!H h $ $$$$,X6(###q(###1333u:$9 1)   rcFeZdZ d d!dZed"dZddd#dZdS)$ConvertFrameAssertTFN compiler_fnr} one_graphrrQexport_constraintsOptional[typing.Never]packageOptional[CompilePackage]rrct||_||_||_||_||_t |_dSr)rtr' _one_graph_export_export_constraints_packager_box)rrrrQrrs rrzConvertFrameAssert.__init__:sF &''')4&# #5  #%% r*Callable[[CompilerFn], ConvertFrameAssert]cfdS)NcFt|jjjSr)convert_frame_assertrrrbackendrs rrz8ConvertFrameAssert._clone_with_backend..Ms%3  O L  $   rrrs`r_clone_with_backendz&ConvertFrameAssert._clone_with_backendKs     rrrrr cache_entryOptional[CacheEntry]hooksrQrrrrr_ct|j}t||}t||t vrt Stj dr6tj d|j krt S|j dkr(|j drt S|j dkrt S|j dkrT|j tjt jjrt S|j dkr|j dkrt S|j dkr-|j dkr"t'|jst St+|r t-d d d d gt.j t3|st St5j}|r%d|jj vr|j}|rd|jj v|r#|jt:jurt dSt?a tC|} | j"} tGdd|j | tI| |j |j%|j&|j'd|j tPs8|j d|j d|j%} tRj*+| tYt[| 5t]|j|j/|j0|j|j1|j2|j3|j4|j5|||||| |dz|j6|j7} dddn #1swxYwYtpj9r'|j6 ddl:m;} | <|j6| S)NTORCHDYNAMO_DEBUG_FUNCTIONz )ztransformers/file_utils.pyztransformers/utils/generic.pyzdiffusers/utils/outputs.py __setattr__rzzzAttempt to trace generatorz.run_tracer/s ! M 5 5 7 7 7677  9N9N9P9P                                 .     ! ! # # #   *  . M        --// 0v}.///   M , , . . . FOOO M , , . . . FO se2BB A+ B+A/ /B2A/ 3B6 BBB B B D'A4DD''"E T)errorr)torch.fx.experimental.validatorrrrestartrZr[rR output_graphoutput_instructionsupdaterr3r0r/r.rO_torch_dynamo_tracer_output)rrrrrrrrrrrrQrrrrrr  tracer_outputrrNrr rs ` @@@r trace_framers(WVVVVVVV!##L " '!+#F(!!!!!!!!!,  *622 +!!!)))) 4 QQQF/000(666(66601A,1O1OPP QQQ (:6(N(N(N%  sBC,, D6D  DcPeZdZUdZded<ded<ded< dddZ dddZdS) DynamoOutputao Represents the core data returned from a single dynamo run, including: - Guards, wrapped inside tracer_output.output_graph.guards - Generated bytecode - Other information needed for compilation. This data structure should capture all the "interesting" information dynamo produces on the frontend side before it enters user backend. rRrrrzOptional[float]last_attempt_start_timeNFrrOptional[Hooks]saverrr strict_errorrrNc r|jj}|Jt||||r|jnd|r|jnd||SN) save_guardsr)rrrN guard_fail_fnguard_filter_fn)rrrrrrrs r build_guardszDynamoOutput.build_guardsesa)6 '''#   #( 2E  d%* 4E ! !%    rargdefsOptional[tuple[Any, ...]]GraphCaptureOutputc |jj}|Jtt||j|j|j|j|j|j |j |jj ||jj Sr) rrr#rSdump_guards_stateimport_sourcesrexport_metadatatracked_fakes_id_to_source traced_coderrrG)rr!rs rgraph_capture_outputz!DynamoOutput.graph_capture_outputys)6 '''! ..00+&,7     '  $ M   &    (   rNFNF rrrrrrrrrrrrNr)r!r"rr#)rrr__doc__rr r*rrrrrVs&%%%,,,, "&,0"      *48       rrcFeZdZUdZded<ded<ded<ded <d ed <d S) BackendInputa  Represents core data structure that dynamo will pass to a backend, including: - Graph module - Example inputs - The FakeTensorMode used for compiling graph. This data structure should capture all the information dynamo produces on for the user backend. r backend_idtorch.fx.GraphModulerr example_inputsz,torch._subclasses.fake_tensor.FakeTensorMode fake_moder|tensor_to_contextN)rrrr-rrrrr/r/sWOOO&&&&;;;;******rr/)frozenceZdZUded<ded<ded<ded<ded <eje Zd ed <d dddZddZ d S)GraphRuntimeEnvrrdict[str, str]r&r used_globalsr"rr!)default_factoryset[str] external_refsN extra_globalsr0r compiled_fnCallable[..., Any]r>Optional[dict[str, Any]]rcd|jD}i||j|pi||i}||t j|j||j|jS)Nc>i|]\}}|tj|Sr) importlib import_module)ralias module_names r z4GraphRuntimeEnv.forward_callable..s9   "{ 9*;77   rrr!) r&itemsr9_check_external_refstypesrrrr!)rr0r?r>r&rGs rforward_callablez GraphRuntimeEnv.forward_callables  &*&9&?&?&A&A      "      !!),,,! M LL     rrGrczg}|jD]}||vr|||rtd|ddS)Nz&Missing required external references: zM. Please load AOT compiled function with `f_globals=`)r<r RuntimeError)rrG missing_refsrs rrKz$GraphRuntimeEnv._check_external_refssw % ) )C)####C(((  ^^^^   r)r0rr?r@r>rArr@)rGrrr) rrrr dataclassesfieldrr<rMrKrrrr7r7s""""    &&&&&&&&/k/DDDMDDDD37       8      rr7ceZdZUdZded<ded<ded<ded <d ed <d ed <d ed< d!d"dZd#dZed$d ZdS)%r#z) Minimal version of DynamoOutput rSrr8r&zlist[CodeType]r)rrr"rr!rrGNFrrrrrrrrrrrNc \t||j||r|jnd|r|jnd||Sr)rNrrr)rrrrrrs rr zGraphCaptureOutput.build_guardssM$    #( 2E  d%* 4E ! !%    rr7c2ddlm}i}|jjjD])}||}|||jvr|j|||<*||j}t|j|j ||j |j |S)Nr)get_global_source_name)rr&r9rr!r<) torch._dynamo.sourcerVrr'graph_input_idx_to_local_sourcerArG_get_external_refsrr7r&rr!)rrVr9source global_namer<s rget_runtime_envz"GraphCaptureOutput.get_runtime_envs??????   . N U U W W H H 0088K"dn,,,0N;,G [)// >> ].%LL'     rr;cddl}t}|j|D][}|jdkr"|jr||j/|jdkr!|jr||j\|S)Nr LOAD_GLOBAL LOAD_NAME)disrget_instructionsopnameargvalr)rr`r< instructions rrYz%GraphCaptureOutput._get_external_refs s "%%% 03/99 : :K!]22%:!%%k&8999#{22%:!%%k&8999rr+r,)rr7)rrrr;) rrrr-rr r\ staticmethodrYrrrr#r#s$###""""&&&&&&&& "&,0"      $    2\rr#c8eZdZUdZded<ded<dddddZdS) CaptureOutputa CaptureOutput should represent all the information produced from torch compiler for a single graph capture. This intends to be consumed by various compiler frontends so that we can share as much compiler internals as possible and avoid great divergence between different stacks. This data structure should eventually contain all the information compiler produces as more refactors happens to converge different compiler frontends. r#r*zOptional[BackendInput] backend_inputN)r?r>r?Optional[Callable[..., Any]]r>rArr@c|j}|jJ|jj}|p |jj}||||S)Nr=)r*r\rhr0rrM)rr?r> runtime_envr0s rrMzCaptureOutput.forward_callable0sf /??AA !---'2 !DT%7%D ++  =,   r)r?rir>rArr@)rrrr-rrMrrrrgrg s`-,,,)))) 5926         rrgmodr %tuple[FunctionType, Optional[object]]cddl}t|tjjr|j}t |dr|j}t|j dkrt|j dkrttjj j j dkrttjj j jdkrt|jdkrt|jdkrzttjj j jdkrNttjj j jdkr"|tjjjkr|j}n.t|tjjr|j}n|j}t |dr|j|jfS|j|r|dfSt3d|)z Utility function to get the function to trace, and optionally a bound self object, from a callable (nn.Module, function, or method). rN__self__zUnsupported model code type )inspectr5rr7r8forwardrC__func__len_forward_pre_hooks_forward_hooksmodulesmodule_global_forward_pre_hooks_global_forward_hooks_backward_pre_hooks_backward_hooks_global_backward_pre_hooks_global_backward_hooksr GraphModule _call_implrro isfunctionrO)rlrpresolved_forwards r get_traced_fnr@s NNN#ux''; #Z 0 0 9/8  & ' '1 , ,C&''1,,EH$+EFF!KKEH$+ABBaGGC+,,11C'((A--EH$+FGG1LLEH$+BCCqHH EHO$;;; +CC UX1 2 2 .CC,CsJA|S\))  C ADy?#??@@@rinspect.Signaturec.tj|dS)NF)follow_wrapped)rp signature)rs r_get_signaturerms  R 6 6 66rrtuple[Any, ...]rrA FrameInfocddl}t|\}}||f|z}|i}t|}|j|i|}||j}|jpd} |jj} | s| rOt| t| ksJ| dt| | Dt|j|j ||j|jpd|jS)zL Create a frame to trace, given a model, args, and optional kwargs. rNrc$i|] \}}||jSr) cell_contents)rrcells rrHz_get_frame..s! O O O*$T4% O O OrrI)rrrbindapply_defaults arguments __closure__r co_freevarsrsrzipr __globals____dict__ __defaults__) rlrrrrself_optrbound_argumentsrIrfreevarss r _get_framerqs%OOO %%LB{T! ~r""I$ind5f55O""$$$(Hn"G{&H 7 7||s8}},,,, O OHg8N8N O O O      $"    r constraints _is_export_deprecated_do_not_userOptional[list[Constraint]]rct|||}ttti5t |||cdddS#1swxYwYdS)a This API captures a full graph for a model, given example inputs to trace with. Specifically, it takes a callable (nn.Module, method, or function), args, and optional kwargs, and returns Dynamo-captured graph along with other important compile-time information. This serves as the common graph-capture mechanism for different torch compiler AOT frontends (e.g. AOT precompile, export). Note that this API doesn't apply context managers like metrics context, and the expectation is that the caller will apply them depending on the use case. The CaptureOutput is separated into two parts: 1. Frontend specific information, which includes: - guards - generated bytecode - other information tracked by OutputGraphCommon. 2. Backend specific information (indexed by unique backend id) such as: - fx graph - example inputs rN)rrrr_fullgraph_capture_frame)rlrrrrrs rfullgraph_capturers: sD& ) )E r(:(:;; < <  ' #-M                     sAA!AcLeZdZUded<ded<ded<ded<ded<d ed <d S) rrrrrrrrrr"r!N)rrrrrrrrrsZ&&&&&&rc ddlmdd fd } t|j|j|j|j|j|||d t }nz#t$rm}t|tj r|}|j (|j d|j }|j (|d|j d}~wwxYwt||jS) Nr)TracingContextgmr1r2list[torch.Tensor]rc}|j}|j}|Jt|jdt sJt |jd|||||S)Nr0)rr3r4r5metarr/)rr2rr3r4rrhs rfullgraph_compilerz4_fullgraph_capture_frame..fullgraph_compilersy),,..#- +=$$$"',/55555$ GL !2~yBS   rT)rrQrrrestart_reasons)rr1r2rrr1) torch._guardsr compile_framerrrrrrrLr?r)verbose __cause__with_tracebackrgr*r!) rrrr dynamo_outputrNcur_exnrrhs @@rrrs? -,,,,,,0M       6% J M L N M*3*EE      6 6 6A >  !"+   , ,T 2 2 2'G+t$$!+5 6 **5=99  s>A C A(CC rr;c td   f d } d}tjD] }|tj_ t d |d 5t| \}}|Jt||| cdddcS#1swxYwYt#tj $r}t|tj stjtd t#t$|jt)|dd}|r|||jpt1t3|t5j}|dkrt7dddgYd}~rd}~wtj$r}t|tj stjt)|dd}|r|td|jjj d}~wwxYwdS)a A helper function taking a frame and backend, then return the generated bytecode and guards as a common data structure. This is a shared interface for multiple compiler frontends (e.g. torch.compile, torch.export) that needs to capture a graph out of python code. rrrrrrRc tj}t ||| }|J|S)Nr)r overrides _get_current_function_mode_stackr)rrrrrrrrrrQrrrrrrrs r transformz compile_frame..transformsu O < < > > $           1#/!   &(((rNcompile_attempt_Tlog_pt2_compile_event)rrrzRestarting analysis due to %srdz!Excessive RestartAnalysis() callsrzDynamo attempted to trace the same frame 100+ times. Giving up on compiling as the compile time tradeoff is likely not worth the performance gain.rz*Skipping frame %s %s %s %s)rrrrrrR)!r]rcountrrattemptrhr4rr+RestartAnalysisr5rrrrJrrpr' __traceback__r4_cleanup_output_graphrrestart_reasonrr9rorKrrKrrr)rrrrrrrrrQrrrrrrrrrrNfailed_tracer_outputrs``````` ````` @rrrs0%&&O8#?$$33'.$0 :::RVWWW  *?i*P*P'-$000#"/%,C                   "   a!CDD '$&&& HH/11?CC    $+1.KT#R#R # =$::<<<    0 @CQLL A A A&*ikk #}}?!2 }   a!CDD '$&&&#*1.KT#R#R # =$::<<< II  #      I33sJB73)B+ B7+B/ /B72B/ 3B77IC)F55IBI  I)rrrrrQrrrr5rrOptional[ConvertFrameBox]r_cj :;<=ddlm}m}dd5dddn #1swxYwYd},|*r|*j7},|,ri|,_?|,j@x}-ri|-_Ad5d6l3mB}.t|.}/|%|,t|,jE}0t|,jFjE}1|,G}2t|,jHjI}3|,J}4t|,jK}5d7|,jLD}6d8|,jMD}7tbj2j3j4nQd}0d}1d}2d}3i}4d}5t%i}6t%i}7t%=t_j0|#z d9z | jNd?| jOd@|0dA|1dB|2dC|3dD|5dE|$dF|%dG|'dH|(dI|6dJ|7=<|)dutjPdt<|"t|4|&dK }8||8|r|*r|*jRn t|_RcdddcdddcdddcdddcdddcdddcdddS#t$r7}9t|9jV}$t|9}%tjXg}&t|9|| LtbjZ[dMdNdOPtj]|9\}'}(t?|9dQd}*t|9tttttt|t|tttf rtt|9jVdRt|9j|9jkdd}9~9wwxYw#tbj2j j;rWt#d1d235t:<d4t{j>d5dddn #1swxYwYd},|*r|*j7},|,ri|,_?|,j@x}-ri|-_Ad5d6l3mB}.t|.}/|%|,t|,jE}0t|,jFjE}1|,G}2t|,jHjI}3|,J}4t|,jK}5d7|,jLD}6d8|,jMD}7tbj2j3j4nQd}0d}1d}2d}3i}4d}5t%i}6t%i}7t%=t_j0|#z d9z | jNd?| jOd@|0dA|1dB|2dC|3dD|5dE|$dF|%dG|'dH|(dI|6dJ|7=<|)dutjPdt<|"t|4|&dK }8||8|r|*r|*jRn t|_RwxYw#1swxYwYdddn #1swxYwYdddn #1swxYwYdddn #1swxYwYdddn #1swxYwYdddn #1swxYwYddddS#1swxYwYdS)VNr)BisectValidationExceptionValidationExceptiong compile_inner) phase_namerrrrrrQr7tuple[ConvertFrameReturn, Optional[DynamoTracerOutput]]c tj5}|tjjtjttj |tj |||cdddS#1swxYwYtdfSr)rrrr_dynamocallback_handlerinstall_callbacksrDYNAMOrrrrfrecordr_)rrrr_compile_inners rrz_compile..compile_inners ! # # :u    .@@#*C0Q0S0S,T,T       = D F F G G G!>$ 599 : : : : : : : : : : : : : : : :    sBB55B9<B9-tuple[ConvertFrameReturn, DynamoTracerOutput]ctjx}}d#d }|d |j|j|j|d} t | !|#" }nV#t j$rD}|rtd|j Jt|j fcYd}~Sd}~wwxYwj Jd|j }|j } |j|j}|J|d|j|j|j|tt D]A\} } t%d| d5| ||} | | }dddn #1swxYwYB|t&|<t(|||z | jJ| j} d$d}|J||}||}d}|d|jd|dz }|d|jd|z }|jd||jd|ks J|d}|d|jdz }|d|jz }|j|jks J|d}|d|jdz }|d|jz }|j|jks J|| jr$| rt| fS| jJ| jt<j|<t%dd5| ||"du }dddn #1swxYwY"b|j!J""|j!|"#| j$j%"&| j'j(tSnd!}d"|z}tU||j+|}| s!|j,|,| jt[|| fS)%Nprefixrrrline_norrrrrc ttjr-tt |||||dSdSr) bytecode_log isEnabledForloggingDEBUGrKri)rrrrrs r log_bytecodez6_compile.._compile_inner..log_bytecodes[((77 ""#FD(GTJJ  rzORIGINAL BYTECODErz,No graph captured with export/fullgraph=Truez;compiler collective wasn't run before compilation completedzMODIFIED BYTECODEbytecode_hooks_Trcddl}|j|jzt|j|jzzt|j|jzzS)Nr)rp co_argcountco_kwonlyargcountrco_flags CO_VARARGSCO_VARKEYWORDS)rrps r count_argsz4_compile.._compile_inner..count_argssY NNN ()t}w'99::;t}w'==>>? rzarg mismatch: zold code object has args z, znew code object has args zfree var mismatch: zold code object has free var znew code object has free var zcell var mismatch: zold code object has cell var znew code object has cell var r )rrrUnknownzTorch-Compiled Region: ) rrrrrrrrrrrrrrrr).rorrrrr+rrJrKrr_ all_statesrrr enumeraterrArhrsrrr co_varnamesr co_cellvarsrQis_empty_graphrcleanupsreinstancer  guards_stateadd_guarded_codeadd_inlined_sourcerr)update_device_typecurrent_tracergraphrrP guard_managerguard_export_fnrb)$rrrr start_timerout_coderrNrrr hook_outputrrtotal_argcount_oldtotal_argcount_newmsgcheck_fncompile_id_strannotation_str guarded_coderrrrrrdynamo_time_before_restartrQrrrrrrs$ rrz _compile.._compile_inners04y{{:*       L          G)#5'"3MM} G G G J HIII0<<<%'')FF F F F F F F  G !(,=,H,T,T I-U,T,T!)%3  0 <&3&K ####  L         #?#9#9#;#;<< + +IC555TRRR + +"d422 **H + + + + + + + + + + + + + + + #' h"""%z_compile..s4!!!)5 tt$!!!rrecompile_user_contextsrc8d|jd|jd|jdS)N'z' (r))rrr)rs rformat_func_infoz"_compile..format_func_infos+U4<UUD,<UUt?RUUUUrztorch._dynamo hit config.%s (%s) function: %s last reason: %s To log all recompilation reasons, use TORCH_LOGS="recompiles". To diagnose recompilation issues, see %szK reached, because fail_on_recompile_limit_hit = True this is a HARD failurea1 reached with fullgraph=True. Excessive recompilations can degrade performance due to the compilation overhead of each recompilation. To monitor recompilations, enable TORCH_LOGS=recompiles. If recompilations are expected, consider increasing torch._dynamo.config.cache_size_limit to an appropriate value.z;pytorch/compiler:skip_code_recursive_on_recompile_limit_hitz reachedzDynamo cache limit exceededz Limit type: zLDynamo attempted to recompile the code object too many times, exceeding the zo cache size limit.Giving up on compiling as the compile time tradeoff is likely not worth the performance gain.rzBtorchdynamo start compiling %s %s:%s, stack (elided %s frames): %srrzsize mismatch at indexgc gc_time_us)rz run_gc_after_compile: running gcr() curr_framech|] }|j Srrrops rrz_compile..8s$X$X$XR_$X$X$Xrch|] }|j Srrrs rrz_compile..9s'((((*BO(((rgeA frame_keyrrrrr guard_countshape_env_guard_countgraph_op_countgraph_node_countgraph_input_count fail_type fail_reasonfail_user_frame_filenamefail_user_frame_linenonon_compliant_opscompliant_custom_ops) rdynamo_time_before_restart_shas_guarded_codespecialize_float is_forward%dynamo_compile_time_before_restart_us stack_tracegraph_node_shapesexception_stack_tracerTartifactc dddS)N dynamo_errorstringrencodingrrrrrz_compile..s* (%%rc(tjSr traceback format_excrrrrz_compile..s9#7#9#9r metadata_fn payload_fnrz: )rrrrrrQrr)rrrrrrQrr)rrrr)lr rrrr\rl code_contextr nullcontextrr)use_lazy_graph_modulerrrdr!guardrhrrrYr\r8rOr<inline_inbuilt_nn_modules update_outerrkr7rJrsr4rwfail_on_recompile_limit_hitrArrErKrKrrrjoinr&rformatrrotime_nsrrutilsReinplaceCountersrrVr has_outputsrUrTrun_gc_after_compilerrcollect local_scoperoot_txrIrrrsrr count_callsrnodesget_graph_sizes_structured placeholdersr!r"rrr%rvrrcrOr9rr3r4rVrrr+get_exc_messager5rLrIr@AssertionErrorrrrJrGrDrFrCrr)>rrrrrrrrQrrrrrrrrrrrrrmetrics_contextr8 compile_pgrreasonsrinbuilt_nn_reasonsinbuilt_nn_recompile_reasonr user_contexts_msgexceeded limit_typer r( start_time_nsrrr*rr rrrrr rrrrrrr)rr!r"metricsrNrrrrs> ````` `` ` `` ` @@@@rrrnsE.),"o>>>     ?> "\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>\>^\>|*++O&-&9T"""z?U?W?W v;<<ttz2233tt  d$   tt BCCIIKK tt tt  $-%H   tt tt%(EE')) ): % 0Z\\ J J   $ +/ J ' ' E =UL$=$=G?FU667ST: /4+/ ;E ;!HUL$=$=D""" /I4Fq4I ( ,7;.//7;3 $$$47Z     #F"G"G " N!!9P!!!     9;L M M M6z:NN* ,  V V V V KK;  ++  && #   1 -!nnn -!``` !M - -D-D-DEEE97:77!2%/!2!2!2 Q L     1H GG%-1t8<<<CCEE F F   4'tT22   #' %) 5926 04 -33555  ` *7-i*O*O 'L-     3%2%??\ 3!,,.. 3 ,D111#3(@DT(T(T0222f}#8 "!$NNN""HH?@@@JqMMM"""""""""""""""F 4&3 )%'"#^+6)&(FO ) ) ) ) ) )JI"v'9!&-00 (+F,<,C(D(D%!'!3!3!5!5#&v|'9#:#: $*$E$E$G$G!$'(;$<$<!$X$Xv?W$X$X$X!((.4.I((($ #599;;;;" (,%!%#' $&!$(!$'GG!'*2ww$"%%%.2lnn}.LPS-S*Y4<t/!$"5  jT  )**F {()>!.#$4$%6Y{+,D)*@ $%6!"'(<#$$30J$0$<$*$;"9B.:: +%():%;%;)>9GB  ( ( 1 1 1! %5M66244"6a ttttttttttttttttttttttttttttttttttttttttttttttttttV+ <+ <+ <Q,Ia&&K%.%9%;%;$< ! auV < < < < N + +:9 ,   @C?R:@@ < $&<$A'DdKKM%)",/'0-$ .  <"/Aww+77s1vv77 .11t!.#$4$%6Y{+,D)*@ $%6!"'(<#$$30J$0$<$*$;"9B.:: +%():%;%;)>9GB  ( ( 1 1 1! %5M66244"6a ttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttttts'l(l"k99k" <k j4 Jj'A!['j//Q* j*Q..j1Q.2Hj j4  k  k" ' k93 l? l( `#D2```(j/b4 jbjbHjjj!!j4 $j!%j4 ( k 4j88k ;j8<k ? k" kk" kk"  k9"k& &k9)k& *k9- l9k= =lk= l l(l l(l l((l,/l,c>eZdZ ddd Zedd Z dddZdS) ConvertFrameNrr}rrQrrrrcP||_t|d||_||_dS)NF)rr)r'r_inner_convert_hooks)rrrrs rrzConvertFrame.__init__{s7 *5&2 5'    r*Callable[[WrapBackendDebug], ConvertFrame]cfdS)Nc.t|jSr) convert_framer]rs rrz2ConvertFrame._clone_with_backend..s}  K  rrrs`rrz ConvertFrame._clone_with_backends     rrrrrrrrrrr_cbt|jtddxxdz cc< ||||||dz}tddxxdz cc<|S#t $r}t |tr|jjj du}|J|jjj rt |trt |t} |j} | rt tjrt!|drt!|drt#t%|j5|j} dt-j| } t1j| } d | d | t4jd d fd tdt<jdddn #1swxYwYt<j s| stC|dd}|j} tE|| ||}| rtF$|dntF%|dt |tLr7tOtQtRj*tRj*cYd}~St |tVr7tOtQtRj,tRj,cYd}~SYd}~nd}~wwxYwtOS)Nframestotalr(rokr real_stackrzGraph break: torch.compile cannot properly resume from this graph break, which results in a skip. torch.compile will skip tracing the frame zN and fall back to eager. The graph break occurred in the following user code: r+c dddS)Ndynamo_graph_break_reasonr.r/rrrrrz'ConvertFrame.__call__..s(C,411rc4dtjS)N r2)user_stack_tracesrrz'ConvertFrame.__call__..s2B/^/^iFZF\F\/^/^rr5T)exc_info stack_inforU)rl)frame_exec_strategy)-rrrErgr\rOr5rFrrrJrLgraph_break_logrrrrCrrrrfr?r3 format_listr+format_frame_inforrrrKr)rsuppress_errorsr4rBrJrrsrHr_rar`SKIPrERUN_ONLY)rrrrrrrrNr soft_failr user_stackuser_stack_formatted frame_inforU error_msgrks @rrzConvertFrame.__call__s   %%%7###q(###h (({E;TAX)F X t $ $ $ ) $ $ $Mb b b !788 #(=TI !(333"'<   !9:: "1k22I*>*@*@'))//BSBSTTUUFN  ( / /?? ?   !#'gg4Q:: $b3GGG    $,?F ( (+>F ( > > > > ()??????????????????s<AAA E A D:- E:EEE EctjdkrdStj|D]}|jdkr |jdzcSt d)N) rRESUMErz$RESUME instruction not found in code)r version_infor`rarboffsetrO)rinsts rfirst_real_inst_idxr/sc '!!q$T**$$ ;( " ";!# # # # # = > >>rceZdZddddZdS)ConvertFrameProtocolrrrrrrrrQrrrrrr_cdSrr)rrrrrrs rrzConvertFrameProtocol.__call__9s !SrNr)rrrrrrrrr8s7!!!!!!!!rrctSr)r$rrr&should_skip_due_to_torch_dispatch_moderDs : < < 5QQ ( .. (?"8KK &u|44?"0KK.5III?"wKK">K 9L(L,  &'' ' L $ 2 2u|7Ky7X7X L $ - -.G H H8Y $//&'' ' =  4 4 6 6/ I I0GIIJ !OOOOOO$0L)3)D+/+I+c%%%M#68Me 6JJ)4& -,{DJ #*  133  33{DJ !4F                                          sI5A3I55I9<I9K(!K7 K(K K(K K((K,/K,N)rrrrQrr)rrrrrrrr_)rrrrrrrrrrHsB OOOOOOrrrc"t||Sr)r)rrs rcatch_errors_wrapperrs h . ..rr)rrrrrrrrrz)rrrrrr)rrrr)rrrr)NF) rNrOrrrrPrQrrr)rXrrr)rrrrr) rr}rrrQrrrrrrr)rrrr)"rrrrrrrrrrrr}rrrrrr]rrrrrQrrrrrrrrrrrR)rlr rrm)rr rr)rlr rrrrArr) rlr rrrrArrrrrrg)rrrrrrrrg)rrrrrrrrrrrr}rrrr;rQrrrrrrrrrrr)NN)&rrrrrrrrrrrr}rrrQrrrrrQrrrr5rrPrrrrrrrrrrrr_)rr}rrQrrrrZ)rrrrr)rr)rrrrQrr)r- __future__r collectionsrrmrQr`r%rrDrprrrrarrxr threadingror3rLtypingrrpathlibrrrrr r r r r typing_extensionsrrrtorch._loggingtorch._C._dynamo.guardsrtorch._dynamo.callbackrtorch._dynamo.distributedrtorch._dynamo.symbolic_convertrrrrrrrtorch._utils_internalrrrrtorch.fx._lazy_graph_moduler%torch.fx.experimental.symbolic_shapesrrtorch.fx.graph_moduler r torch.monitorr!torch.nn.parallel.distributedr"torch.utils._python_dispatchr#r$r%torch.utils._tracebackr&r'rr)r*r+r,r-bytecode_analysisr.r/bytecode_transformationr0r1r2r3r4rr5r6r7r8 eval_framer9r:r;r<r=r>r?r@rArBrCrDrErFrGrHrIrJrKrLgraph_bytecode_inputsrMrrNrOrPrrQrrRrSpgorTrUrV replay_recordrWresume_executionrXsymbolic_convertrYrZr[r\r]r^r_r`rarbrBrcrdrerfrgrhrirjrkrlrmrnrorprqrrrsrtrurvrwrxvariables.torch_functionryrnumpyrzModuleNotFoundError TYPE_CHECKINGcollections.abcr{torch.utils.weakr|backends.registryr}rr~repro.after_dynamorrrrvariables.builderr getLoggerrrJrgetArtifactLoggerrroRLockrrrrrrrrr&rrr(#suppress_torch_distributed_warningsrMrVrCounterrWr\r[rrrrrtorch.utils.hooksrrrrrr/r7r#rgrrrrrrrrrZrarrProtocolrrrrrrrrs.#"""""      !!!!!!>>>>>>>>>>>>000000000000''''''!!!!!! 444444222222444444999999MMMMMMMMMMMM%%%%%% ?>>>>>QPPPPP&&&&&&AAAAAA MLLLLLLLEEEEEEEEEEEEEEGGGGGGGG  >===== ???????? +*****;;;;;;"!!!!!XXXXXXXXXXXX0JIIIII BBB  7((((((444444------''''''444444@@@@@@@@@@666666g!!~//*EE .228^LLy   WT]]Yt__        (gii wyy 377777*++MQ,+DQQQQh1KKK21Kb(, * * * * * K  EEEEP 0000000 0.ddddddddR15(,     $#####------,7;==888815HL48(,#OOOOOOd 5 5 5 5 5 5 5  5 p +++++++ +" $........b IIIIIII IX          >*A*A*A*AZ7777(,$$$$$T(,$ /3-2 $ $ $ $ $ $ N ''''''' '/3-2 555555D15HL48(,jjjjjjt(,HLI"(,48+IIIIIIXH$H$H$H$H$H$H$H$\)-=====????@???? ! ! ! ! !6? ! ! !====UUUUUUUUp//////s2G77HH