"EigSSKrSSKrSSKrSSKrSSKrSSKJr SSKJ r J r J r J r SSK r SSKJr SSKJr SSKJrJrJrJrJrJr SSKJr SSKJr \S \4r\S \4r \RB"\"5r#"S S \RH5r%\%RLS \%RNS\%RPS\%RRS\%RTS\%RVS\%RXSSS0r-\."\-5VVs0sHupX_M snnr/"SS\RH5r0\0Vs0sHoURb_M snr2\Rf"SSSS9"SS 55r4\Rf"SS55r5\Rf"SSSS9"SS \455r6S\S\\\6S-\6S-44S jr7S\S\\64S!jr8S\S\\\6S-\644S"jr9S#\S-S\\S$44S%jr:"S&S'5r;"S(S)5r<"S*S+5r=\Rf"5"S,S-55r>"S.S/5r?"S0S15r@\Rf"S2S355rA\Rf"S4S555rB"S6S75rC"S8S95rDgs snnfs snf):N)Iterator)AnycastLiteralOptional)FunctionSchema)_ProfilerResult) _EventType_ExtraFields_Allocation_ExtraFields_TorchOp_ProfilerEvent_TensorMetadata RecordScope) _element_size)_utilsKey TensorKeyc\rSrSr\R "5r\R "5r\R "5r\R "5r \R "5r \R "5r \R "5r Sr g)CategoryN)__name__ __module__ __qualname____firstlineno__enumautoINPUT TEMPORARY ACTIVATIONGRADIENTAUTOGRAD_DETAIL PARAMETEROPTIMIZER_STATE__static_attributes__r\/home/mande/repo/quber/.venv/lib/python3.13/site-packages/torch/profiler/_memory_profiler.pyrrsO IIKE IJyy{HiikO IiikOr&r darkgreen goldenrodblack mediumpurplered mediumblue royalbluegreyc\rSrSr\R "5r\R "5r\R "5r\R "5r Sr g)Action7rN) rrrrrr PREEXISTINGCREATEINCREMENT_VERSIONDESTROYr%rr&r'r1r17s/))+K YY[F iikGr&r1TF)eq unsafe_hashfrozenc4\rSrSr%\R \S'Srg)rAdevicerN)rrrrtorchr<__annotations__r%rr&r'rrAs LLr&c\\rSrSr%Sr\\S'\\S'S\4SjrS\ S\ 4Sjr S\4S jr S r g ) _StorageFzBundle storage pointer and id. All profiling logic should use `allocation_id`, however it is useful to print storage pointers for debugging and unit tests sometimes look up values using the storage data pointer of a live Tensor.ptr allocation_idreturncN[UR5SSURS3$)Nz>18 ())hexrBrCselfs r'__repr___Storage.__repr__Qs'dhh-$Bt'9'9&:!<strrKobjectboolrRrVr%rr&r'r@r@FsB? H=#=YFYtY(#(r&r@c \rSrSr%Sr\\S'\\S'S\4Sjr SSS\ 4Sjr \ S \S -S \S -S \S -S \ RS\S4 Sj5r\S\S\S4Sj5r\S\S -S\S4Sj5r\S\\\\\44Sj5rSrg )r[agHashable identifier for a storage which has been assigned an ID. A detailed description of Tensor IDs and why they are needed is given in `torch/csrc/profiler/collection.h` when `TensorID` is declared. To summarize, multiple Storage buffers can map to the same logical Tensor. This dataclass is used to refer to a concrete in-memory StorageImpl of a Tensor. idstoragerDcjSURS[UR5SSURS3$)Nzid=z: z<24rFrG)r_reprr`r<rIs r'rKTensorKey.__repr__is1TWWIRT\\ 237r$++aHHr&rMc4URUR:$rO) _as_sortablerQs r'__lt__TensorKey.__lt__ls  5#5#555r& tensor_idN storage_ptrrCr<c@UbUbUb[X0[X55$grO)rr@)rhrirCr<s r'_makeTensorKey._makeos+  !')V0TU Ur&allocczURURURURUR5$rO)rkr_rBrCr<)clsrms r'from_allocationTensorKey.from_allocation~s)yy599e.A.A5<<PPr&tcUb<URURURURUR5$grO)rkr_storage_data_ptrrCr<)rorrs r' from_tensorTensorKey.from_tensors1 =99QTT1#5#5qQ Qr&cURURRURRURR 4$rO)r_r`rCr<typeindexrIs r'reTensorKey._as_sortables3ww 22DKK4D4DdkkFWFWWWr&r)rrrrrXrYr>r@rZrKr\rf staticmethodr=r<rrk classmethodr rprrupropertytuplerer%rr&r'rr[s G I#I6K6D6 : 4Z Tz   +   Q$;Q@UQQOd2x 7L XeCc3$67XXr&noderDc## URnURS[R:XGaURSR[ R :XaURS:XaU(aUSRS[R:XaUSRS;aUSRSR(ai[USRSRS[5(a7S[RUSRSRS54v gURS[R:XaURSnURbURb [!S5eURbLURR"H2up4n[RU5[RU54v M4 URbMURR"H2upEn[RU5[RU54v M4 ggg7f)Nrztorch::autograd::AccumulateGrad)z aten::detachz aten::add_'module and optimizer cannot both be set)childrentypedr TorchOpscoperBACKWARD_FUNCTIONnameinputsrPrrruPyCallmodule optimizerAssertionError parameters)rr typed_fields_pp_grads r'!_extract_parameters_and_gradientsrs}}H 1 +++ JJqM  ;#@#@ @ II: :  QK  a J$6$6 6 QK   > > QK  a ' ' x{((+2215 G GI))(1+*;*;A*>*E*Ea*HIII A*++ +zz!}    *|/E/E/Q !JK K    * , 3 3 > > f++A. 0E0Ef0MMM!?  ! ! - , 6 6 A A 1++A. 0E0Ef0MMM!B . ,sHc#F# [U5HupUcM Uv M g7frOr)rr_p_grads r'extract_parametersrs 7=  =G>s! !c#H# [U5HupUcM X4v M g7frOr)rrrs r'extract_gradientsrs%7t<   )O=s" "event.c/nU(a^URS[R:Xa(URURSR5 UR nU(aM^[ U5$Nrr)rr rappendrparentr~)rscopess r' get_scopesrsV F  ;;q>Z// / MM%++a... /  % =r&c\rSrSrSr\S\S\\S-S44Sj5r \S\S\\ S44Sj5r \S\4S j5r \ S \S\\ S4S-4S j5rS rg) SchemaMatcheraLookup operator schema based on profiled name. When profiling we record the operator's name but not the schema. However some analysis requires that information. Fortunately we can look up registered schema from the recorded name. We do not, however, record the overload and so we must compare the profiled arguments with all overloads to determine viable matches. Note: Once https://github.com/pytorch/pytorch/issues/78871 is completed this code will be obsolete. rrrDN.c TSnURU5HhnU=(d URVs/sHnSPM snn[UR5H&upVX%==[URSS5-ss'M( Mj [ U=(d SUR 55$s snf)aDetermine which inputs may have mutated based on function schema. Note that we don't need to resolve down to a single schema to perform this analysis. An input is mutable if it is mutable in any overload. In practice, however, it is overwhelmingly common to match a single overload. If we cannot find any valid schema then we must be conservative and assume all inputs are mutable. NFis_writec3&# UHnSv M g7frOr).0rs r' 3SchemaMatcher.inputs_are_mutable..s 8x!xs) match_schemas arguments enumerategetattr alias_infor~r)rorrmutableschemariargs r'inputs_are_mutable SchemaMatcher.inputs_are_mutables&*''*FB1A1A!B1AA%1A!BG#F$4$45 gcnnj%HH 6+ W8 8qxx 899 "Cs B%c^^^[SUR55mS[4UU4Sjjm[U4SjTRUR5=(d S55$)Nc3# UHln[U[5(a[RU5O<[U[5(a&UVs/sHn[RU5PM snOUv Mn gs snf7frO)rPrrrulistrrjs r'r.SchemaMatcher.match_schemas..sr  !_--  ! !! $ !T""566Aq)''*A6  7sA A; A6*A;rDc >[UR5[T5:H=(a( [U4Sj[TURSS955$)Nc3^># UH"upTRXR5v M$ g7frO) _types_matchrx)robserved schema_argros r'r?SchemaMatcher.match_schemas..matches..s2C-(H  ??;;-s*-Tstrict)lenrallzip)rro signatures r'matches,SchemaMatcher.match_schemas..matchessMv''(C N:sC,/v//-C@ r&c3F># UHnT"U5(dMUv M g7frOr)rsrs r'rrsO @1GAJQQ @s! !r)r~rr\lookup_schemasr)rorrrrs` @@r'rSchemaMatcher.match_schemass[  XX    t  O 2 2166 : @b @OOOr&c [U[RR5(a,UR 5nUSL=(d UR X5$[U[RR 5(agUR[RRR55(a)[U[5=(a [SU55$[RR[4[RR[S54[RR ["4[RR$[&4[RR([*4[RR,[.4[RR0["[&[*[.444nUH"upE[X$5(dM[X5s $ USL$)NTc3B# UHn[U[5v M g7frO)rPrrrs r'r-SchemaMatcher._types_match..s62:Q 1i(((s)rPr=_C OptionalTypegetElementTyperAnyType isSubtypeOfListType ofTensorsrr TensorTyperNoneTyperxBoolTyper\IntTyperY FloatTypefloat ComplexTypecomplex NumberType)ror schema_typetype_mapjit_typepy_typess r'rSchemaMatcher._types_match so k588#8#8 9 9%446Kt#Ns'7'7'N N k588#3#3 4 4  " "588#4#4#>#>#@ A Ah-#62:63  XX ) , XX  T + XX   % XX  s # XX   ' XX ! !7 + XX 4eW"= >E #+ H+00!(55#+4r&rcSU;ag[[RRU55$![a gf=f)Nz::)r~r=r_jit_get_schemas_for_operator RuntimeError)rs r'rSchemaMatcher.lookup_schemas+s@ 4??EF F  s1'1 >>r)rrrrrXr|r r~r\rrrrr{rZrr%rr&r'rrs :#7:E$+sBR3+>%?$%Fr&rc\\rSrSrS\SS4SjrS\\4Sjr\ S\ \S44Sj5r S r g) OpTreeiAresultrDNc|UR5Ul[[UR 5SS95Ulg)NcUR$rO start_time_nsxs r'!OpTree.__init__..DsAOOr&key)experimental_event_tree _root_nodesr~sorteddfs _sorted_nodesrJrs r'__init__OpTree.__init__Bs.!99;"6$((*:S#TUr&c/h# [R"UR/UQ70UD6ShvN gN7frO)r traverse_dfsrrJargskwargss r'r OpTree.dfsFs)&&t'7'7I$I&IIIs (202.cUR$rO)rrIs r' sorted_nodesOpTree.sorted_nodesIs!!!r&)rr) rrrrr rrr rr}r~rr%rr&r'rrAsPVV4VJh~&>J"eNC$78""r&rcn\rSrSrS\SS4SjrS\S-SS4Sjr\S\ S\ \4S j5r S \ 4S jr S rg) SizeMapiNop_treerDNc0UlURGHqnURS[R:Xa8UR URS5HnUR U5 M M]URS[R:XdMURSnURbURb [S5eURbBURRH(upVnUR U5 UR U5 M* URcGMURRHDupgnUR U5 UR U5 UHupSUR U5 M MF GMt 0n URHnURS[R:XdM&URSn [RU 5n U (dMS[U R 5n U R#X5n X:wdMU SU 3n[$R'SU5 M URR)U 5 g)Nrrrz vs. z(Mismatch between allocation and free: %s)_valuesrrr r_flat_tensor_inputs_update_valuesrrrrr Allocationrrpabs alloc_size setdefaultlogwarningupdate)rJr rrrrrrrstate allocations alloc_fieldsrnew_size prior_sizedeltas r'rSizeMap.__init__Os-/ ((Dzz!} 2 2211$**Q-@A''*AA*"3"33#zz!}  ''3$..:()RSS&&2(4(;(;(F(F f++A.++F3)G ))5,8,B,B,M,M(5++A.++F3$)DA //2%*-N%)0-/ ((Dzz!} 5 55#zz!} // =3"<#:#:;H!,!7!7!FJ"-#-,eH: > $NPUV!)$ K(r&rrc [RU5nUbUbUR[R:Xa[ S[ UR=(d S/UR=(d S/SS955nU[UR5-nUS:a[SU35e[ URRUS5U5URU'gggg)Nc36# UHoSUS-v M g7frrNrrs r'r)SizeMap._update_values..s%W!qt %WsrTrrz$num_bytes must be non-negative, got )rrulayoutr=stridedmaxrsizesstridesrdtyperr get)rJrrrn num_bytess r'rSizeMap._update_valuess##A& ?q}U]]1J%(QC9IqcRV%WAM!''22I1}$'KI;%WXX #DLL$4$4S!$A!_--At$$ s>AA A Arc URU$rOr rJrs r' __getitem__SizeMap.__getitem__s||C  r&r/)rrrrrrrrr{r rr rr1r%rr&r'r r Nse.).)4.)` I$ 6 I4 I 4/9R!y!r&r cn\rSrSr%Sr\S-\S'Sr\S-\S'\ S\4Sj5r \ S\4Sj5r S r g) DataFlowEdgeiN input_versionFmutatedrDcURSL$rO)r5rIs r' is_allocationDataFlowEdge.is_allocations!!T))r&cURSL$rO)r6rIs r' is_deletionDataFlowEdge.is_deletions||t##r&r) rrrrr5rYr>r6r\r}r8r;r%rr&r'r4r4sR $M3:$ GTD[  *t**$T$$r&r4c\rSrSrS\SSSS4SjrS\\\44Sjr \ S\\\ \ \ 444S j5r\ S\\\ 44S j5r\ S\ \S 44S j5r\ S\ 4S j5rSrg) DataFlowNodeirgraph DataFlowGraphrDNc ,XlX lUR5UlURR 5HFup4UR (dMUR (aM+URRU5 MH URR 5VVs0sH"upVXVURRU54_M$ nnn[SUR555(d[SUSUR35egs snnf)Nc3.# UH upX:Hv M g7frOrrs r'r(DataFlowNode.__init__..s8&7da16&7szversion mismatch: z, ) _event_graph_determine_edges_edgesitemsr6r8bumpoutputslookuprvaluesr)rJrr?redgekvversionss r'rDataFlowNode.__init__s  595J5J5L **,IC|||D$6$6$6   %- ?Cll>P>P>RS>RdaA4;;--a011>RS8hoo&7888 #5hZr$++!OP P9Ts#)Dc[[R"UR/55n0nSU5Hn[ UR [ RU5SS9HupE[U[5(a@[RU5nURU[55RU5 MZ[U[5(dMqUHAn[RU5nURU[55RU5 MC M M [ R""[$5nUR'5HXupiUcM U(aUR(R+U5OSXlSU ;=(d [U 5S:Hn XUlMZ UHn U R0S[2R4:XdM&U R0SR6S:dME[R9U R0S5nXn UbU R.c[;SU35eSU lU(aUR(R+U5OSU lM UHon U R0S[2R4:XdM&U R0SR6S:dMESU[R9U R0S5lMq [=[?S UR'5555$) Nc3# UH6oRS[R:XdM%URSv M8 g7fr)rr rrs r'r0DataFlowNode._determine_edges..s-Sw!''!* @R@R2R:1771:ws $AATrrOrrzDouble delete: c36# UHupUcM X4v M g7frOrrrNrOs r'rrTsMmdaq6A6ms  ) r~rrrDrrrrrPrrrursetaddr collections defaultdictr4rHrErKr5r6rr rrrprdictr) rJsubtreemutable_by_keyr+op_inputrr op_input_iedges mutable_setr6rrMs r'rFDataFlowNode._determine_edgess\++T[[M:;DFSwSB%( =;;B?&!h88#//9C"--c359==gF $//&. '33J?&11#su=AA'J'/&T '' 5 . 4 4 6 CFI4;;+=+=c+Br (  ;.RE+4F'4Q%,c "!7AwwqzZ222qwwqz7L7Lq7P// ;z?t||';(?3%)@AA# @CT[[%7%7%<"AwwqzZ222qwwqz7L7Lq7PMQi// ;<J FMekkmMMNNr&c URR5VVs0sHHupUR(aMU[UR5[ [ UR54_MJ snn$s snnfrO)rGrHr8r\r6rrYr5rJrNrOs r'rDataFlowNode.inputss`  ))+  ,?? =AQYYc1??!;< <+    s A14A1cURR5VVs0sHXupUR(aUR(aUR(dM:XR cSOUR S-_MZ snn$s snnfr)rGrHr8r;r6r5res r'rJDataFlowNode.outputsse ))+ + !)) EAOO+q11D D+   s 9B"B.cV[SURR555$)Nc3v# UH/upUR(dMUR(dM+Uv M1 g7frO)r8r;rWs r'r-DataFlowNode.intermediates..s& -$!AQ]]AA-s 99 9)r~rGrHrIs r' intermediatesDataFlowNode.intermediatess) ++++-   r&c.URR$rO)rDrrIs r' start_timeDataFlowNode.start_times{{(((r&)rGrDrE)rrrrr rr\rr4rFr}r~r\rYrrJrlror%rr&r'r>r>s Qn Q_ Q Q2O$y,'>"?2Oh YdCi(889   in-   uY^4  )C))r&r>c\rSrSrS\SS4Sjr\S\\S44Sj5r SSjr \S\\ S44S j5r \ S\S\\ S44S j5rS \S\4S jrS \SS4S jrS \SS4SjrSrg)r@ir rDNcXlURU5Ul0UlURVs/sHn[ X 5PM snUlUR RSS9 UR5 gs snf)NcUR$rO)rors r'r(DataFlowGraph.__init__.. sALLr&r) _op_tree_extract_leaf_events _leaf_events_active_version leaf_eventsr> _flow_nodessortvalidate)rJr es r'rDataFlowGraph.__init__sm  55g><>;?;K;KL;KaL1;KL "89 MsA:.c,[UR5$rO)r~rzrIs r' flow_nodesDataFlowGraph.flow_nodessT%%&&r&c h[5nURHgn[URR55nX-nU(a2[ SUR R SURSU35eX-nMi 0nURHnURR5H0unupxURUS5n X:wdM![ SU SU35e URR5H0uphURXh5n X:a[ SUSU 35eXU'M2 M g)Nzduplicate outputs:  rz%version mismatch for input: expected z, got zversion regression: z < ) rXrrJrHrrDrrGrr') rJrJr node_outputs duplicatestensor_versionsrrversionexpected prior_versions r'r|DataFlowGraph.validates6.1eOODt||1134L /J$)$++*:*:);1T[[M:,W  #G$13OOD%)[[%6%6%8!\a*..sA6&(?zPWyY&9!% 2 2 4  / 3 3C A *(.wis=/J(/$ !5$r&cUR$rO)rwrIs r'ryDataFlowGraph.leaf_events1s   r&c^^/mS[S[4SjmS[4UU4SjjnURUS9HnM [[ TSS95$)a7Partially traverse the op tree and extract top level ops. Consider the following code: ``` with record_function("My annotation"): x.zero_() y.zero_() ``` The op tree (assuming no Autograd) will look like: TorchOp: "My annotation" TorchOp: zero_ TorchOp: fill_ TorchOp: zero_ TorchOp: fill_ The recursive structure of operator calls makes data flow unwieldy. In order to simplify analysis we would like to select the highest level ops to represent in the graph. In this case those are the `zero_` ops; the fact that `fill_` is called is an implementation detail. We also do not want to group everything under "My annotation" as this could create overly coarse bundles and lose critical semantics. To address this issue we walk over the graph and select the topmost torch ops ** which match at least one operator schema **. These form the leaves of the first pass through the op tree. (As well as any allocations or frees which do are not part of a kernel.) These events form the logical nodes in our data flow graph. r}rDcURS[R:H=(a\ URSR[R :H=(d+ [ [RURS55$r) rr rrrrr\rr)r}s r'leaf_op3DataFlowGraph._extract_leaf_events..leaf_opXs_771:!3!33   K$A$AAA 33AGGAJ?@ r&c>T"U5(dUR[R:XaTRU5 /$UR$rO)tagr rrr)r}ryrs r' children_fn7DataFlowGraph._extract_leaf_events..children_fn^s:qzzQUUj&;&;;""1% :: r&)rcUR$rOrrs r'r4DataFlowGraph._extract_leaf_events..hsqr&r)r r\rr~r)r rrryrs @@r'rv"DataFlowGraph._extract_leaf_events5s^B-/  ~ $   >  5A 6VK-FGHHr&rcbURRUS5nUc[SUS35eU$)Nrzversion for key  is Nonerxrr)rJrrs r'rKDataFlowGraph.lookupjs9&&11#q9 ? #3C5!AB Br&cURRUS5nUc[SUS35eUS-URU'g)Nzprior_version for key rr)rxr'r)rJrrs r'rIDataFlowGraph.bumppsJ,,00d;   #9#h!GH H$1A$5S!r&czURRUS5c[SUS35eSURU'g)Nrzcannot delete key z, already deletedrr0s r'deleteDataFlowGraph.deletevsA    * *3 2 : #5cU:K!LM M$(S!r&)rxrzrwrurDN)rrrrrrr}r~r>rr|r ryr{rvrrYrKrIrr%rr&r'r@r@s4'E,"34''/<!U>3#67!!2If2I~s7J1K2I2Ih) 6 6d6 ))))r&r@c\rSrSr%Sr\S-\S'\R"\ S9r \ \ \4\S'\R"\ S9r \ \ \4\S'\R"\S9r\\ \S'Srg) CategoryElementi|Nby_iddefault_factoryby_key by_version _by_id_keysetr)rrrrrrr> dataclassesfieldr\rrr TensorAndIDrXrr%rr&r'rr|sj!E8d?!(3(9(9$(OFDH$ %O.9.?.?PT.UJ[(*+U%0$5$5c$JM3y>Jr&rc\rSrSr%\R "SS9r\R\ \ 4\ S'S\ S\ SS4S jrS\ S\ SS4S jrS\ S \ S\ SS4S jrS\ S \ S\ SS4S jrS\S \ S\ S-4SjrSrg) CategoryDictic6[R"[5$rO)rZr[rrr&r'rCategoryDict.s 7 7 Hr&rr rcategoryrDNcX RURlURURRR U5 grO)r r_rrrYrJrrs r' set_by_idCategoryDict.set_by_ids6%- SVV" SVV**..s3r&cNX RURRU'grO)r r_rrs r' set_by_keyCategoryDict.set_by_keys+3 SVV##C(r&rcPX0RURRX4'grO)r r_rrJrrrs r'set_by_versionCategoryDict.set_by_versions:B SVV''7r&clURURRRX4U5 grO)r r_rrrs r'setdefault_by_version"CategoryDict.setdefault_by_versions) SVV''22C>8Lr&c0[U[5(a[U[5(dgURURnUR =(d@ UR RUS5=(d URRX4S5$rO) rPrrr r_rrr'r)rJrrelements r'r'CategoryDict.getsr c3   3 (B(B,,svv& MM <~~!!#t, <!!%%snd; r&r)rrrrrrr rZr[rYrr>rrrrrrrr'r%rr&r'rrs=H=N=NH>G[ $ $S/%9 :4Y4(4t44i4844C)CcCXCRVCMM'*M6>M M  s S X_ r&rc\rSrSrS\SS4Sjr\S\\\\ \ \4S44Sj5r S\ 4Sjr S\\\S-44S jrS\\4S jrSS jrSS jrSS jrSSjrSSjrSSjrSSjrSrg) MemoryProfileirrDNc[U5Ul[UR5Ul[ UR5Ul[ 5UlUR5 UR5 UR5 UR5 UR5 UR5 UR5 grO)rrur@_data_flow_graphr  _size_mapr _categories_set_gradients_and_temporaries#_set_parameters_using_python_tracer _set_inputs_set_parameters_using_data_flow_set_activations_set_optimizer_state_set_autograd_detailrs r'rMemoryProfile.__init__sv -dmm < /'> ++- 002  ,,.  !!# !!#r&.c ^/n0n0nTRR5GHcnURS[R:XdM'URSnUR nUS:nUR n[RU5n U bXX4'Mr[UR5n URU R4n U(aWX;a'URU[RU S4U45 MSX;'URU[RU S4U45 MURU[R U S4U*45 UR#U S5(aGM=URS[R$U S4U*45 GMf TR'5n [)[+U R-555n U V Vs/sH+upU S4U;dMUS:XdMS[R$X44PM- nn nTR.R0GH nUR2R55Hun nUR6(a+X*S4nURU[RU S445 OgUR8(aVUR:R nUR<nUc[?SU 35eURU[RX445 UR@(dMX*S4nURU[R XU 445 M GM URCU4SjU55 URESS 9 [GU5$s snn f) NrrTFrUzinput_version is None for key c3T># UHupup4XX44TRU4v M g7frO)r)rtimeactionrrrJs r'r)MemoryProfile.timeline..s1 06,nsC>4>>#+> ?06s%(c*USUSR4$r)valuers r'r(MemoryProfile.timeline..s1Q41"4r&r)$rurrr rrrrrprr<rBrr1r5r4r6popr3_category_snapshotr\rkeysrrrGrHr8r6rDr5rr;extendr{r~)rJoutputallocation_times live_unknownrrrr8rrtkeyrptr_and_devicesnapshot last_versionreventsrrMs` r'timelineMemoryProfile.timelines:<>@FH ]]&&(E{{1~!6!66${{1~ )44 *Q '' 00>#>?d%:;l112C&2&6&6 %CN$)9"MM!"F$<$!| 5R##c^ 4 ( 9 ))44D![[..0 T%%(t5AMM1fmmc1X">?\\ 11A"00G,/McU-STTMM1f&>&>"OP###(u6AMM1fnns.s'T@S91fq@Sc3(# UHoS4v M g7frNrrrs r'rrs&N;MCQx;Mc3(# UHoS4v M g7frrrs r'rrs&K?CQx?r)rXrrrrrHrlrJrr rLrrr')rJall_tensor_versionsrrrrs r'r MemoryProfile._category_snapshots03))44D  & &'T @Q@Q@S'T U  & &&N4;M;M&N N  & &t||'9'9'; <5 !!))002A  & &&K1??&K K3 !'': ; ; ND,,00> > ;   s>)D+cf^^[5m[T5nTRRHmn[ UU4SjUR R 555nU(dM:TRU5 TRSUR55 Mo [T5U:XaT$M)aExtract IDs of Tensors which depend or will depend on a gradient. Note that this weakened definition of "depends" requires us to loop over the data flow graph multiple times because it allows dependency information to flow backward through edges and removes the guarantee that nodes are topologically sorted. (Or indeed, even that a valid topological order exists.) Put another way, we have converted an acyclic data flow graph into a cyclic graph and we are attempting to partition cycles involving a gradient from the rest of the graph. c3># UHeunup#TRRX5[R[R4;dUR T;dMWUR v Mg g7frO)rr'rr!r#r_)rrrrdepends_on_gradientrJs r'rAMemoryProfile._any_version_depends_on_gradient..s_-@)\a''++C9 ))8+=+=>?vv!44 CFF-@s AA0A0c38# UHoRv M g7frO)r_rs r'rr(s.N#vvs) rXrrrr~rrHrrJ)rJ start_sizeridsrs` @r' _any_version_depends_on_gradient.MemoryProfile._any_version_depends_on_gradients),01J--88-1[[->->-@3'..s3'...N.NN9&&':5**-r&cvURR5HAn[U5H/up#URR U[ R 5 M1 MC URRH@nURH-nURRU[ R5 M/ MB g)z>Mark Tensors which are unambiguous and simple to reason about.N) rurrrrrr!rrrlrr)rJrrrrrs r'r,MemoryProfile._set_gradients_and_temporaries2s ]]&&(E.u5   **683D3DE6)))44D''  ++Ax/A/AB(5r&cURR5HDn[U5H2nUcMURR U[ R 5 M4 MF grO)rurrrrrr#)rJrrs r'r1MemoryProfile._set_parameters_using_python_tracerCsI]]&&(E'.=$$..q(2D2DE/)r&c^^TR5n[5m[TRR5HunUR R 5VVVs1sH unupEX54iM nnnnXbRR 5-n[UU4SjU55(dMpTU-mMw TR5nTRRHRn[R[UR5;dM,U[URR 55-nMT UHAup5URU;dMTRR!X5["R$5 MC gs snnnf)aMMark inputs based on which Tensors are updated using gradients. The process for differentiating between inputs and activations is more involved. Most Tensors in a training loop depend on at least one gradient: parameters depend on them through updates, and activations and optimizer state depend on them transitively through parameters. Critically, we do not need to know which Tensors are parameters to apply this method; we can simply walk the data flow graph to build the set of all values which depend on a gradient and then obtain the set of inputs from the conjugate set. There is, however, one hiccup. The first time we see a parameter is generally on the forward pass of the first step. We know from inspection of the data flow graph that v1 of that Tensor depends on a gradient (provided we profile an optimizer step), but not v0. To address this problem we weaken the definition of "depends on a gradient" to "any version of this Tensor depends on a gradient", which in turn strengthens the criteria for the input set enough to filter the activations in the forward pass of the first step.c3># UHKnTRR"U6[R[R4;=(d UT;v MM g7frO)rr'rr!r#)rrproduces_gradientrJs r'r,MemoryProfile._set_inputs..jsT!A  $$a(X->->@R@R,SS*))* sAAN)rrXreversedrrrrHrJanycopyrrrrDr_rrrr) rJrrrrrtensorsinput_candidatesrs ` @r'rMemoryProfile._set_inputsIs92#CCE/2eT22==>D?C{{?P?P?RS?R*;#|~?RGS ||))+ +G! "W,!?-113))44D,, 4;;0GG C (:(:(<$== 5-LCvv00  66sX^^T-!TsE9 c8^^ TR5n[5nUR5VVs1sHup4U[R:XdMUiM nnnTR R HnURR5VVV s1sH unupXy4iM n nnn [R[UR5;dM`[U4SjU 55(aM|[U4SjURR555(aMURU 5(dMXVRR5-nX*RU5-nM [5m [!TR R 5Hhn[UU 4SjURR555(dM8T R#SURR555 Mj UR%T 5 UVVs1sHupxUR&iM n nnU TR)5-n UHAupxUR&U ;dMTR*R-U[R.5 MC gs snnfs sn nnfs snnf)aNDeduce which Tensors are parameters. Consider the following code for the step of SGD with momentum (nesterov=False), where `d_p` is the gradient of `param` and `buf` is the momentum buffer. ``` buf.mul_(momentum).add_(d_p, alpha=1 - dampening) d_p = buf param.add_(d_p, alpha=-lr) ``` Both `param` and `buf` take a gradient and perform an in-place update. The python tracer will inspect calls to `nn.Module.forward` and `optim.Optimizer.step` to extract parameter and optimizer state respectively (including parameters), so this is generally a non-issue. However as a fallback we can also exploit several properties of parameters to distinguish them from other model state. First, they are directly used in the forward pass. (At this point we haven't established which parts of the graph correspond to the forward pass but we can deduce enough to suffice.) Some mutable state such as batch norm moving averages also contribute to the forward pass, but optimizer state does not. Second, a parameter is by definition used to compute at least one gradient and depends on at least one gradient. c3B># UHnTR"U6v M g7frOrrrrJs r'r@MemoryProfile._set_parameters_using_data_flow..sB6aD--q16c3B># UHnTR"U6v M g7frOrrs r'rrsP;OaD--q1;Orc3Z># UH nTR"U6=(d UT;v M" g7frOr)rrrJused_for_gradients r'rrs1-A!!1%?.?)??-s(+c32# UH unup#X4v M g7frOr)rrrrs r'rrs)++:K:K:MN:MsJQsl:MFN--Z 5LLB6BBBP4<<;M;M;OPPP*66v>>%););)==%$(9(9:O(PP$5/2eT22==>D++-"(()<@KK,>? ))44D:>++:K:K:MN:MsJQsl:MFNBHI&Q 0 0 4 4a 8& I",,) -DEE 11DKK9PP++-A$$::SASx?R?RS.5NIs 8E "EcURR5HnURS[R:XdM&URSR (dMFURSR R n[RRSU55HIup4[RU5nUcMURRU[R5 MK M g)Nrrc3,# UH u pUv M g7frOr)rrrs r'r5MemoryProfile._set_optimizer_state..s9j{q!Ujs)rurrr rrritchain from_iterablerrurrrr$)rJrrrrrrs r'r"MemoryProfile._set_optimizer_states]]&&(E{{1~!2!22u{{1~7O7O7O"[[^55@@ HH229j9DA$//2C((2238P8PQ )r&cS[R1nURRHn[R [ UR5;dM,URR5HYup4US:Xd$URRX4S- 5U;dM/URRX4[R5 M[ M gr) rr"rrrrrrDrJrHrr'r)rJpriorrrrs r'r"MemoryProfile._set_autograd_detailsx//0))44D,, 4;;0GG$(LL$6$6$8LC!|t'7'7';';C1'MQV'V((>>(*B*B%95r&)rrrurr)rrrrr rr}r~rYr1KeyAndIDrr\rr\rrrrXrrrrrrrrr%rr&r'rrs $ $4 $B%c68S&@ A3 FGBBHJtJ Dho)E$F "+#c("+HC"F 1UfHDTT& Rr&rcJ\rSrSrS SjrSrS SjrS SjrS S SjjrSr g) MemoryProfileTimelineiNcHURUlURUlg)aThe minimum representation of the memory profile timeline includes the memory timeline and categories. The timeline consists of [timestamp, action, (TensorKey, version), numbytes] elements, to denote any actions (pre-existing, create, destroy, or increment_version) that occurred to a specific Tensor for a chunk of memory. The categories help map each (TensorKey, version) pair into a category.N)rr categories)rJmemory_profiles r'rMemoryProfileTimeline.__init__s'// (44r&cb^^ [R"U5n/n/m SUU 4SjjnSnTRGHQupgupn URU:waMUS:wa[US- 5nUS:Xd Xe:aUS:aUn[ U5S:Xa<UR U5 T R S/[ V s/sHn SPM sn -5 O;XcS:wa3UR U5 T R T SR55 U[R[R4;a U"XU 5 MU[R:XaU"XU *5 U"XS-U 5 GM%U[R:Xa U"XU *5 GMF[SU35e UVs/sH ofS:aUOUPM nnUT 4$s sn fs snf)zConvert the memory timeline and categories into a memory plot consisting of timestamps and their respective sizes by category for a given device. Input: device Output: [timestamps, sizes by category] c>[U[5(aTRRX5OSn[US-nTSU==[ U5- ss'g)NrrU)rPrr0r'_CATEGORY_TO_INDEXrY)rrrrryrJr$s r'r8MemoryProfileTimeline._coalesce_timeline..updatesWc9--##C1  'x014E "Ie E * r&rUirrUnknown action: r)r=r<rrYrrr5rr1r3r4r5r6 ValueError) rJ device_strr<timesrt_minrrrrrnumbytesrr$s ` @r'_coalesce_timeline(MemoryProfileTimeline._coalesce_timelinesj)!# + +37== /A~xzzV#BwDM{qyQU5zQ Q aS/A#B/A!A/A#BBCBi Q U2Y^^-.&,,fmm<<sX.6333shY/saK26>>)shY/!#3F8!<==C4AF1661a%Q&6e|+$C(7s , F' F,cURU5up4SSKn[US5nURX4/U5 SSS5 g!,(df  g=f)zxSaves the memory timeline as [times, sizes by category] as a JSON formatted file to the given path for the given device.rNw)r=jsonopendump)rJpathr9r:r$rAfs r'export_memory_timeline,MemoryProfileTimeline.export_memory_timeline7s>..z:  $_ IIuna (__s A Ac ^[R"U5n/nU4SjnTRGHupgupn URU:waMU[R[R 4;a$UR U[UU U"X545 MdU[R:XaJUR U[UU *U"X545 UR U[UU U"XS-545 MU[R:Xa%UR U[UU *U"X545 M[SU35e SSK n [US5n U RXL5 SSS5 g!,(df  g=f)zSaves the memory timeline as raw memory event tuples in the form of (timestamp, action, numbytes, category) as a JSON formatted file to the given path for the given device.cz>[U[5(aTRRX5OSn[U$rO)rPrr0r'r5)rrrrJs r'get_category_indexLMemoryProfileTimeline.export_memory_timeline_raw..get_category_indexJs;c9--##C1  &h/ /r&rr7rNr@)r=r<rr1r3r4r_ACTION_TO_INDEXr5r6r8rArBrC) rJrDr9r< raw_eventsrJrrrrrr<rArEs ` r'export_memory_timeline_raw0MemoryProfileTimeline.export_memory_timeline_rawBsn j)68  048== /A~xzzV#&,,fmm<<!!(0 *38 6333!!(0! *38 !!(0 *3! < 6>>)!!(0! *38 !#3F8!<==_4Ab  $_ IIj $__s E"" E0c fSSKnURRS5nUc [S5 gSSKJn SSKJn SSKJ n SSK n URU5n U RU S5U RU S5p[U 5nX-n U RU SS9S - n[R "U5n[R"R%U5n[R"R'U5nU R)US S 9nUR+5n[,R/5H6unn[0UnUR3U S - USS2U4USS2US-4US S9 M8 UR5[,Vs/sHnUcSO UR6PM sn5 UR9S5 UR;S5 SR=U(aU/O/SUS - SSUS - SS3/-5nUR?U5 U"SSS9nURAUSS9 URCSS5 U"URE55RGS5nU(d [IS5eSUS3n[KUS SS!9nURMU5 SSS5 SSS5 gs snf!,(df  N=f!,(df  g=f)"zkExports the memory timeline as an HTML file which contains the memory timeline plot embedded as a PNG file.rN matplotlibzDexport_memory_timeline_html failed because matplotlib was not found.) b64encode)NamedTemporaryFiler)axisi@P)figsizedpig@@gffffff?)coloralphaUnknownz Time (ms)z Memory (GB)z zMax memory allocated: z.2fz GiB Max memory reserved: z GiBzw+bz.png)suffixpng)formatzutf-8z failed to encode image as base64z} GPU Memory Timeline HTML r@)encoding)'importlib.utilutil find_specprintbase64rRtempfilerSmatplotlib.pyplotpyplotnumpyr=arraymincumsumr=r<cudamax_memory_allocatedmax_memory_reservedfiguregca_CATEGORY_TO_COLORSrHr5 fill_betweenlegendr set_xlabel set_ylabeljoin set_titlesavefigseekreaddecoderrBwrite)rJrDr9rVtitle importlibmatplotlib_specrRrSpltnpmtr:r$r;stackedr<rlrmfigaxesrrXrtmpfileencodedhtmlrEs r'export_memory_timeline_html1MemoryProfileTimeline.export_memory_timeline_htmls #..22<@  " V  $/'  $ $Z 0xx1AuE  ))E)*W4j)$zz>>vF#jj<A+J")J;J" J J"" J0)r0rr)) N) rrrrrr=rFrNrr%rr&r'r.r.s3 5:x )D%N9=D DDr&r.)ErZrr itertoolsr%loggingcollections.abcrtypingrrrrr=torch._Crtorch._C._autogradr torch._C._profilerr r r r rr torch._utilsrtorch.profilerrr~rYr,r getLoggerrrEnumrr#r$rrr r!r"rprr5r1rrL dataclassrr@rrrrrrrr r4r>r@rrrr.)rcs00r'rs $// #.'!  K$%  !"tyy"   k NNG   | k& (11D'EF'Etqad'EFTYY)//1qwwJ/ $E$?@ (((( $D>.X.X?.Xb(N (N eI$i$&6678(NV^0C   eI$i/01nt+k36F0Good " "G!G!T  $ $ $])])@s)s)l KKK    >EEP ]]wG0s +I&I,