+ j^RIHt^RIt^RIt^RIt^RIt^RIt^RIH t H t H t H t ^RI t ^RIHtHtHtHtHtHt^RIHt^RIHt] 'd^RIHt^RIHt^R IHt]R ] 3,t!]R ] 3,t"]PF!]$4t%!R R ]PL4t']'PPR]'PRR]'PTR]'PVR]'PXR]'PZR]'P\RRR/t/]0!]/4UUu/uFwrWbK uppt1!RR]PL4t2]2Uu/uFqVPfbK upt4]Pj!RRRR7!RR 44t6]Pj!RR44t7]Pj!RRRR7!RR ]644t8RR lt9R!R"lt:R#R$lt;R%R&lt<!R'R(4t=!R)R*4t>!R+R,4t?]Pj!4!R-R.44t@!R/R04tA!R1R24tB]Pj!R3R444tC]Pj!R5R644tD!R7R84tE!R9R:4tFR#uuppiuupi);) annotationsN)AnycastLiteral TYPE_CHECKING) _EventType_ExtraFields_Allocation_ExtraFields_TorchOp_ProfilerEvent_TensorMetadata RecordScope) _element_size)_utils)Iterator)FunctionSchema)_ProfilerResultKey TensorKeyc]tRt^%t]P !4t]P !4t]P !4t]P !4t ]P !4t ]P !4t ]P !4t Rt R#)CategoryN)__name__ __module__ __qualname____firstlineno__enumautoINPUT TEMPORARY ACTIVATIONGRADIENTAUTOGRAD_DETAIL PARAMETEROPTIMIZER_STATE__static_attributes__r`/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/profiler/_memory_profiler.pyrr%sO IIKE IJyy{HiikO IiikOr&r darkgreen goldenrodblack mediumpurplered mediumblue royalbluegreyc]tRt^=t]P !4t]P !4t]P !4t]P !4t Rt R#)ActionrN) rrrrrr PREEXISTINGCREATEINCREMENT_VERSIONDESTROYr%rr&r'r1r1=s/))+K YY[F iikGr&r1TF)eq unsafe_hashfrozenc"]tRt^Gt$R]R&RtR#)r torch.devicedevicerN)rrrr__annotations__r%rr&r'rrGs r&cT]tRt^Lt$RtR]R&R]R&RRltRRltR R ltR t R #) _StoragezBundle 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.intptr allocation_idcV^8dQhRR/#returnstrr)formats"r' __annotate___Storage.__annotate__Ws==#=r&c N\VP4R RVP R2#)z>18 ())hexr@rAselfs&r'__repr___Storage.__repr__Ws'dhh-$Bt'9'9&:!<rArOrSs&&r'__eq___Storage.__eq__Zs(%*XXt/A/AUEXEX/XXr&cV^8dQhRR/#rDrEr?r)rGs"r'rHrI]s((#(r&c ,\VP4#rW)hashrArNs&r'__hash___Storage.__hash__]sD&&''r&rN) rrrr__doc__r<rPrZr`r%rr&r'r>r>Ls)? H=Y((r&r>c]tRt^at$RtR]R&R]R&RRltRR lt]R R l4t ] R R l4t ] RRl4t ] RRl4tRtR#)ragHashable 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. r?idr>storagecV^8dQhRR/#rCr)rGs"r'rHTensorKey.__annotate__osII#Ir&c jRVP R\VP4R RVP R2#)zid=z: z<24rKrL)rdreprrer;rNs&r'rPTensorKey.__repr__os1TWWIRT\\ 237r$++aHHr&c V^8dQhRRRR/#)rDrSrrErUr)rGs"r'rHrgrs66I6$6r&c 4VPVP8#rW) _as_sortablerYs&&r'__lt__TensorKey.__lt__rs  5#5#555r&c ,V^8dQhRRRRRRRRRR/#) rD tensor_id int | None storage_ptrrAr;r:rETensorKey | Noner)rGs"r'rHrgvs:    "    r&c HVeVeVe\W0\W44#R#rW)rr>)rqrsrAr;s&&&&r'_makeTensorKey._makeus+  !')V0TU Ur&c V^8dQhRRRR/#)rDallocr rErtr)rGs"r'rHrgs QQ$;Q@PQr&c zVPVPVPVPVP4#rW)rvrdr@rAr;)clsrys&&r'from_allocationTensorKey.from_allocations)yy599e.A.A5<<PPr&c V^8dQhRRRR/#)rDt_TensorMetadata | NonerErtr)rGs"r'rHrgs27Gr&c Ve=VPVPVPVPVP4#R#rW)rvrdstorage_data_ptrrAr;)r{rs&&r' from_tensorTensorKey.from_tensors1 =99QTT1#5#5qQ Qr&cV^8dQhRR/#)rDrEztuple[int, int, str, int]r)rGs"r'rHrgsXX7Xr&c VPVPPVPPVPP 3#rW)rdrerAr;typeindexrNs&r'rmTensorKey._as_sortables3ww 22DKK4D4DdkkFWFWWWr&rN)rrrrrbr<rPrn staticmethodrv classmethodr|rpropertyrmr%rr&r'rrass G I6  QQ XXr&c V^8dQhRRRR/#)rDnoder rEz3Iterator[tuple[TensorKey | None, TensorKey | None]]r)rGs"r'rHrHs (N(N (N8(Nr&c#"VPpVP^,\P8XEd=VP^,P\ P 8XEd VPR8XdV'dV^,P^,\P8XdV^,PR9dV^,P^,P'd\V^,P^,P^,\4'dER\PV^,P^,P^,43xR#VP^,\P8XdVP^,pVPeVPe \!R4hVPeMVPP"F2wr4p\PV4\PV43xK4 VPeOVPP"F2wrEp\PV4\PV43xK4 R#R#R#5i)rztorch::autograd::AccumulateGradN'module and optimizer cannot both be set)z aten::detachz aten::add_)childrentypedrTorchOpscoper BACKWARD_FUNCTIONnameinputsrXr rrPyCallmodule optimizerAssertionError parameters)rr typed_fields_pp_grads& r'!_extract_parameters_and_gradientsrs}}H 1 +++ JJqM  ;#@#@ @ II: :  QK  a J$6$6 6 QK   > > QK  a ' ' ' x{((+2215 G GI))(1+*;*;A*>*E*Ea*HIII A*++ +zz!}    *|/E/E/Q !JK K    * , 3 3 > > f++A. 0E0Ef0MMM!?  ! ! - , 6 6 A A 1++A. 0E0Ef0MMM!B . ,sA=J A.J /FJ c V^8dQhRRRR/#)rDrr rEzIterator[TensorKey]r)rGs"r'rHrHs^0Cr&c#J"\V4FwrVfK VxK R#5irWr)rr_p_grads& r'extract_parametersrs 7=  =G>s# #c V^8dQhRRRR/#)rDrr rEz,Iterator[tuple[TensorKey | None, TensorKey]]r)rGs"r'rHrHs 1r&c#L"\V4FwrVfK W3xK R#5irWr)rrrs& r'extract_gradientsrs%7t<   )O=s$ $c V^8dQhRRRR/#)rDeventz_ProfilerEvent | NonerEztuple[RecordScope, ...]r)rGs"r'rHrHs+0Gr&c.pV'daVP^,\P8Xd-VPVP^,P4VP pKh\ V4#r)rrrappendrparenttuple)rscopess& r' get_scopesrsM F  ;;q>Z// / MM%++a... /  =r&cr]tRt^tRt]RRl4t]RRl4t]RRl4t] RR l4t R t R #) 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. c V^8dQhRRRR/#)rDrr rEztuple[bool | None, ...]r)rGs"r'rHSchemaMatcher.__annotate__s::#7:3SchemaMatcher.inputs_are_mutable..s 8x!xs) match_schemas arguments enumerategetattr alias_inforr)r{rmutableschemariargs&& r'inputs_are_mutable SchemaMatcher.inputs_are_mutables&*''*FBB1A1A!B1AA%1A!BG#F$4$45 gcnnj%HH 6+ W88 8qxx 899 "Cs B.c V^8dQhRRRR/#)rDrr rEztuple[FunctionSchema, ...]r)rGs"r'rHrs PP2P7QPr&c aaa\;QJd.RVP4FNK 5M!RVP44oRVV3Rllo\;QJd:.V3RlSPVP4;'gR4FNK 5#!V3RlSPVP4;'gR44#)c3"TFmp\V\4'd\PV4M<\V\4'd&VUu.uFp\PV4NK upMTxKo R#uupi5irW)rXr rrlistrrjs& r'r.SchemaMatcher.match_schemas..sr  !_--  ! !! $ !T""566Aq)''*A6  7sA A< A7*A<cV^8dQhRR/#rDrErUr)rGs"r'rH1SchemaMatcher.match_schemas..__annotate__s  t r&c "<\VP4\S48H;'dg\;QJd5V3Rl\SVPRR74F 'dK R# R#!V3Rl\SVPRR744#)c3^<"TF"wrSPWP4xK$ R#5irW) _types_matchr)robserved schema_argr{s& r'r?SchemaMatcher.match_schemas..matches..s2C-(H  ??;;-s*-TstrictF)lenrallzip)rr{ signatures&r'matches,SchemaMatcher.match_schemas..matchessv''(C N:ssC,/v//-Css s sC,/v//-C@ r&c3H<"TFpS!V4'gKVxK R#5irWr)rsrs& r'rr sO @1GAJQQ @s" "r)rrlookup_schemasr)r{rrrsf&@@r'rSchemaMatcher.match_schemassE  XX  EE  XX     uO 2 2166 : @ @b @OuOuO 2 2166 : @ @b @OOOr&cV^8dQhRR/#rr)rGs"r'rHrs  D r&c  T\V\PP4'd.VP 4pVRJ;'gVP W4#\V\PP 4'dR#VP\PPP44'dM\V\4;'d5\;QJdRV4F 'dK R# R#!RV44#\PP\3\PP\R43\PP \"3\PP$\&3\PP(\*3\PP,\.3\PP0\"\&\*\.333pVF#wrE\W$4'gK\W4u# VRJ#)NTc3B"TFp\V\4xK R#5irW)rXrrrs& r'r-SchemaMatcher._types_match..s62:Q 1i(((sF)rXtorch_C OptionalTypegetElementTyperAnyType isSubtypeOfListType ofTensorsrr TensorTyperNoneTyperBoolTyperUIntTyper? FloatTypefloat ComplexTypecomplex NumberType)r{r schema_typetype_mapjit_typepy_typess&&& r'rSchemaMatcher._types_matchs k588#8#8 9 9%446Kt#NNs'7'7'N N k588#3#3 4 4  " "588#4#4#>#>#@ A Ah-##62:6## # #62:63  XX ) , XX  T + XX   % XX  s # XX   ' XX ! !7 + XX 4eW"= >E #+ H+00!(55#+4r&c V^8dQhRRRR/#)rDrrFrEz!tuple[FunctionSchema, ...] | Noner)rGs"r'rHr2sS%Fr&c RV9dR#\\PPV44# \dR#i;i)z::N)rrr_jit_get_schemas_for_operator RuntimeError)rs&r'rSchemaMatcher.lookup_schemas1s@ 4??EF F  s3'3 AArN) rrrrrbrrrrrrr%rr&r'rrsZ ::$PP2  Br&rcD]tRtRtRRltRRlt]RRl4tRtR #) OpTreeiGc V^8dQhRRRR/#rDresultrrENoner)rGs"r'rHOpTree.__annotate__HsVVV4Vr&c VP4Vn\\VP 4RR74VnR#)cVP#rW start_time_nsxs&r'!OpTree.__init__..JsAOOr&keyN)experimental_event_tree _root_nodesrsorteddfs _sorted_nodesrOrs&&r'__init__OpTree.__init__Hs.!99;"6$((*:S#TUr&cV^8dQhRR/#)rDrEzIterator[_ProfilerEvent]r)rGs"r'rHr LsJJ&>Jr&c/ j"\P!VP.VO5/VBRjxL R#L5irW)r traverse_dfsrrOargskwargss&*,r'r OpTree.dfsLs)&&t'7'7I$I&IIIs (313cV^8dQhRR/#rDrEtuple[_ProfilerEvent, ...]r)rGs"r'rHr Ps""8"r&c VP#rW)rrNs&r' sorted_nodesOpTree.sorted_nodesOs!!!r&)rrN) rrrrrrrr(r%rr&r'rrGs"VJ""r&rcP]tRtRtRRltRRlt]RRl4tRR ltR t R #) SizeMapiTc V^8dQhRRRR/#rDop_treerrEr r)rGs"r'rHSizeMap.__annotate__Us.).).)4.)r&c /VnVPEFpVP^,\P8Xd=VP VP^,4FpVP V4K KfVP^,\P8XgKVP^,pVPeVPe \R4hVPeCVPPF(wrVpVP V4VP V4K* VPfEK)VPPFDwrgpVP V4VP V4VFwrSVP V4K KF EK /p VPFpVP^,\P8XgK+VP^,p \PV 4p V 'gK]\V P 4p V P#W4p W8wgKV RV 2p\$P'RV4K VPP)V 4R#)rNrz vs. z(Mismatch between allocation and free: %s)_valuesr(rrr_flat_tensor_inputs_update_valuesrrrrr Allocationrr|abs alloc_size setdefaultlogwarningupdate)rOr.rrrrrrstate allocations alloc_fieldsrnew_size prior_sizedeltas&& r'rSizeMap.__init__Us-/ ((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&c V^8dQhRRRR/#)rDrrrEr r)rGs"r'rHr/s I I 6 I4 Ir&c  \PV4pVeVeVP\P8Xd\ R\ VP;'g^.VP;'g^.RR744pV\VP4,pV^8d\RV 24h\ VPPV^4V4VPV&R#R#R#R#)Nc3N"TFq^,V^,,xK R#5irNrrs& r'r)SizeMap._update_values..s %W!qt %Ws#%Trz$num_bytes must be non-negative, got )rrlayoutrstridedmaxrsizesstridesrdtyperr1get)rOrrn num_bytess&& r'r3SizeMap._update_valuess##A& ?q}U]]1J%(QC9I9IqcRV%WAM!''22I1}$'KI;%WXX #DLL$4$4S!$A!_--At$$ s?AA A AcV^8dQhRR/#)rDrrr)rGs"r'rHr/s!!y!r&c (VPV,#rWr1rOrs&&r' __getitem__SizeMap.__getitem__s||C  r&rWN) rrrrrr3rr2rYr%rr&r'r+r+Ts,.)` I!!r&r+c`]tRtRt$RtR]R&RtR]R&]RR l4t]R R l4t R t R#) DataFlowEdgeiNrr input_versionFz bool | NonemutatedcV^8dQhRR/#rr)rGs"r'rHDataFlowEdge.__annotate__s**t*r&c VPRJ#rW)r]rNs&r' is_allocationDataFlowEdge.is_allocations!!T))r&cV^8dQhRR/#rr)rGs"r'rHr`s$$T$r&c VPRJ#rW)r^rNs&r' is_deletionDataFlowEdge.is_deletions||t##r&r) rrrrr]r<r^rrbrfr%rr&r'r\r\s< $M:$ G[  **$$r&r\c]tRtRtRRltRRlt]RRl4t]RR l4t]R R l4t ]R R l4t Rt R#) DataFlowNodeic$V^8dQhRRRRRR/#)rDrr graph DataFlowGraphrEr r)rGs"r'rHDataFlowNode.__annotate__s& Q Qn Q] Qt Qr&c  WnW nVP4VnVPP 4FHwr4VP 'gKVP 'dK-VPPV4KJ VPP 4UUu/uF"wrVWVVPPV43bK$ ppp\;QJd*RVP44F 'dK RM RM!RVP444'g\RV RVP 24hR#uuppi)c3."TF wrW8HxK R#5irWrrs& r'r(DataFlowNode.__init__..s8&7da16&7sFTzversion mismatch: z, N) _event_graph_determine_edges_edgesitemsr^rbbumpoutputslookuprvaluesr)rOrrkredgekvversionss&&& r'rDataFlowNode.__init__s  595J5J5L **,IC|||D$6$6$6   %- ?Cll>P>P>RS>RdaA4;;--a011>RSs8hoo&78sss8hoo&7888 #5hZr$++!OP P9Ts%(EcV^8dQhRR/#)rDrEzdict[TensorKey, DataFlowEdge]r)rGs"r'rHrms2O2O"?2Or&c v\\P!VP.44p/pRV4Fp\ VP \ PV4RR7FwrE\V\4'dA\PV4pVPV\44PV4K[\V\4'gKsVFAp\PV4pVPV\44PV4KC K K \ P"!\$4pVP'4FcwriVfK V'dVP(P+V4MRW,nRV 9;'g\V 4R8Hp WV,nKe VFp V P0^,\2P48XgK+V P0^,P6^8gKO\P9V P0^,4pW,p VeV P.f\;RV 24hRV nV'dVP(P+V4MRV nK VFp V P0^,\2P48XgK+V P0^,P6^8gKORV\P9V P0^,4,nK \=\?RVP'4444#)c3"TF?qP^,\P8XgK*VP^,xKA R#5irE)rrrrs& r'r0DataFlowNode._determine_edges..s/Sw!''!* @R@R2R:1771::ws 'A A TrNzDouble delete: c38"TFwrVfK W3xK R#5irWrrr{r|s& r'rrsMmdaq6A6ms rW) rrrrqrrrrrXr rrr7setaddr collections defaultdictr\rurrrxr]r^rrr4r6r|rdictr) rOsubtreemutable_by_keyrRop_inputrr op_input_iedges mutable_setr^rrzs & r'rsDataFlowNode._determine_edgess_++T[[M:;DFSwSB%( =;;B?&!h88#//9C"--c359==gF $//&. '33J?&11#su=AA'J'/&T '' 5 . 4 4 6 CFI4;;+=+=c+Br (  ;.RRE+4F'4Q%,c "!7AwwqzZ222qwwqz7L7Lq7P// ;z?t||';(?3%)@AA# @CT[[%7%7%<"AwwqzZ222qwwqz7L7Lq7PMQi// ;<J FMekkmMMNNr&cV^8dQhRR/#)rDrEz!dict[TensorKey, tuple[bool, int]]r)rGs"r'rHrms  9 r&c  VPP4UUu/uFIwrVP'dKV\VP4\ \ VP43bKK upp#uuppirW)rtrurbrUr^rr?r]rOr{r|s& r'rDataFlowNode.inputss`  ))+  ,?? =AQYYc1??!;< <+    s A14A1cV^8dQhRR/#)rDrEzdict[TensorKey, int]r)rGs"r'rHrms  - r&c VPP4UUu/uF`wrVP'dVP'dVP'gK=YP f^MVP ^,bKb upp#uuppirW)rtrurbrfr^r]rs& r'rwDataFlowNode.outputssi ))+ + !)) EAOO+q11D D+   s*B B'BcV^8dQhRR/#)rDrEztuple[TensorKey, ...]r)rGs"r'rHrms  4 r&c \;QJd,.RVPP44FNK 5#!RVPP444#)c3z"TF1wrVP'gKVP'gK-VxK3 R#5irW)rbrfrs& r'r-DataFlowNode.intermediates..s& -$!AQ]]AA-s ;; ;)rrtrurNs&r' intermediatesDataFlowNode.intermediatessNu ++++- u u ++++-   r&cV^8dQhRR/#r]r)rGs"r'rHrms))C)r&c .VPP#rW)rqrrNs&r' start_timeDataFlowNode.start_times{{(((r&)rtrqrrN) rrrrrrsrrrwrrr%rr&r'riris_ Q2Oh      ))r&ric]tRtRtRRlt]RRl4tRRlt]RR l4t] R R l4t R R lt RRlt RRlt RtR#)rli c V^8dQhRRRR/#r-r)rGs"r'rHDataFlowGraph.__annotate__ s4r&c WnVPV4Vn/VnVPUu.uFp\ W 4NK upVnVP PRR7VP4R#uupi)cVP#rW)rrs&r'r(DataFlowGraph.__init__..sALLr&rN) _op_tree_extract_leaf_events _leaf_events_active_version leaf_eventsri _flow_nodessortvalidate)rOr.es&& r'rDataFlowGraph.__init__ sm  55g><>;?;K;KL;KaL1;KL "89 MsA=cV^8dQhRR/#)rDrEztuple[DataFlowNode, ...]r)rGs"r'rHrs''4'r&c ,\VP4#rW)rrrNs&r' flow_nodesDataFlowGraph.flow_nodessT%%&&r&cV^8dQhRR/#rDrEr r)rGs"r'rHrs//$/r&c  \4pVPFpp\VPP44pW,pV'd3\ RVP P RVP RV 24hW,pKr /pVPFpVPP4F1wpwrxVPV^4p W8wgK"\ RV RV 24h VPP4F1wrhVPWh4p W8d\ RV RV 24hWV&K3 K R#)zduplicate outputs:  z%version mismatch for input: expected z, got zversion regression: z < N) rrrwrurrqrrtrrM) rOrwr node_outputs duplicatestensor_versionsrrversionexpected prior_versions & r'rDataFlowGraph.validates6.1eOODt||1134L /J$)$++*:*:);1T[[M:,W  #G$13OOD%)[[%6%6%8!\a*..sA6&(?zPWyY&9!% 2 2 4  / 3 3C A *(.wis=/J(/$ !5$r&cV^8dQhRR/#r%r)rGs"r'rHr8s!!7!r&c VP#rW)rrNs&r'rDataFlowGraph.leaf_events7s   r&c V^8dQhRRRR/#)rDr.rrEr&r)rGs"r'rHr<s2I2If2I1K2Ir&c aa.oRRloRVV3RllpVPVR7FpK \\SRR74#)a4Partially 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 are not part of a kernel.) These events form the logical nodes in our data flow graph. c V^8dQhRRRR/#)rDrr rErUr)rGs"r'rH8DataFlowGraph._extract_leaf_events..__annotate__^s  ~ $ r&c$VP^,\P8H;'dfVP^,P\P 8H;'g0\ \PVP^,44#r) rrrrr rrUrr)rs&r'leaf_op3DataFlowGraph._extract_leaf_events..leaf_op^sj771:!3!33   K$A$AAAA 33AGGAJ?@ r&cV^8dQhRR/#)rDrr r)rGs"r'rHrds  > r&c<S!V4'g VP\P8XdSPV4.#VP#rW)tagrr4rr)rrrs&r' children_fn7DataFlowGraph._extract_leaf_events..children_fnds:qzzQUUj&;&;;""1% :: r&)rcVP#rWr rs&r'r4DataFlowGraph._extract_leaf_events..nsqr&r)rrr)r.rrrrs& @@r'r"DataFlowGraph._extract_leaf_events;sIB-/     5A 6VK-FGHHr&c V^8dQhRRRR/#)rDrrrEr?r)rGs"r'rHrps)r&c dVPPV^4pVf\RV R24hV#)rzversion for key  is Nonerr7r)rOrrs&& r'rxDataFlowGraph.lookupps9&&11#q9 ? #3C5!AB Br&c V^8dQhRRRR/#rDrrrEr r)rGs"r'rHrvs66 6d6r&c VPPVR4pVf\RV R24hV^,VPV&R#)Nzprior_version for key r)rrMr)rOrrs&& r'rvDataFlowGraph.bumpvsJ,,00d;   #9#h!GH H$1A$5S!r&c V^8dQhRRRR/#rr)rGs"r'rHr|s)))))r&c ~VPPV^4f\RV R24hRVPV&R#)rNzcannot delete key z, already deletedrrXs&&r'deleteDataFlowGraph.delete|sA    * *3 2 : #5cU:K!LM M$(S!r&)rrrrN)rrrrrrrrrrrrxrvrr%rr&r'rlrl s_''/<!!2I2Ih 6 ))r&rlc]tRtRt$RtR]R&]P!]R7t R]R&]P!]R7t R]R &]P!] R7t R ]R &R t R#) CategoryElementiNCategory | Noneby_iddefault_factoryzdict[TensorKey, Category]by_keyzdict[TensorAndID, Category] by_versionzset[TensorKey] _by_id_keysetr)rrrrrr< dataclassesfieldrrrrrr%rr&r'rrsP!E?!(3(9(9$(OF %O.9.?.?PT.UJ+U%0$5$5c$JM>Jr&rc]tRtRt$]P !RR7tR]R&RRltRR lt R R lt R R lt RRlt Rt R#) CategoryDictic 6\P!\4#rW)rrrrr&r'rCategoryDict.s 7 7 Hr&rz-collections.defaultdict[int, CategoryElement]r1c$V^8dQhRRRRRR/#rDrrcategoryrrEr r)rGs"r'rHCategoryDict.__annotate__s!44Y4(4t4r&c W PVP,nVPVP,PP V4R#rW)r1rdrrrrOrrs&&&r' set_by_idCategoryDict.set_by_ids6%- SVV" SVV**..s3r&c$V^8dQhRRRRRR/#rr)rGs"r'rHrs!44i4844r&c XW PVP,PV&R#rW)r1rdrrs&&&r' set_by_keyCategoryDict.set_by_keys+3 SVV##C(r&c(V^8dQhRRRRRRRR/# rDrrrr?rrrEr r)rGs"r'rHrs/CC)CcCXCRVCr&c ZW0PVP,PW3&R#rW)r1rdrrOrrrs&&&&r'set_by_versionCategoryDict.set_by_versions:B SVV''7r&c(V^8dQhRRRRRRRR/#rr)rGs"r'rHrs2MMM'*M6>M Mr&c vVPVP,PPW3V4R#rW)r1rdrr7rs&&&&r'setdefault_by_version"CategoryDict.setdefault_by_versions) SVV''22C>8Lr&c$V^8dQhRRRRRR/#)rDrrrr?rErr)rGs"r'rHrs!  s S _ r&c B\V\4'd\V\4'gR#VPVP,pVP ;'gBVP PVR4;'gVPPW3R4#rW) rXrrr1rdrrrMr)rOrrelements&&& r'rMCategoryDict.gets| c3   3 (B(B,,svv& MM < <~~!!#t, < <!!%%snd; r&rN)rrrrrrr1r<rrrrrMr%rr&r'rrs==H=N=NH>G :44CM   r&rc]tRtRtRRlt]RRl4tRRltRR ltR R lt R R lt RRlt RRlt RRlt RRltRRltRRltRtR#) MemoryProfileic V^8dQhRRRR/#rr)rGs"r'rHMemoryProfile.__annotate__s $ $ $4 $r&c \V4Vn\VP4Vn\ VP4Vn\ 4VnVP4VP4VP4VP4VP4VP4VP4R#rW)rrrl_data_flow_graphr+ _size_mapr _categories_set_gradients_and_temporaries#_set_parameters_using_python_tracer _set_inputs_set_parameters_using_data_flow_set_activations_set_optimizer_state_set_autograd_detailrs&&r'rMemoryProfile.__init__sv -dmm < /'> ++- 002  ,,.  !!# !!#r&cV^8dQhRR/#)rDrEz-tuple[tuple[int, Action, KeyAndID, int], ...]r)rGs"r'rHr sBBGBr&c  a.p/p/pSPP4EFqpVP^,\P8XgK,VP^,pVP pV^8pVP p\PV4p V eWW3&K|\VP4p VPV P3p V'dZW9d(VPV\PV ^3V34KRW;&VPV\PV ^3V34EK VPV\P V ^3V)34VP#V R4'dEKKVPR\P$V ^3V)34EKt SP'4p \)\+V P-444p V U Uu.uF-wrV R3V9gKV^8XgKR\P$W33NK/ pp pSP.P0EFpVP2P54Fwp pVP6'd0W*R3,pVPV\PV ^334MiVP8'dXVP:P pVP<pVf\?RV 24hVPV\PW334VP@'gKW*R3,pVPV\P WV ,334K EK VPCV3RlV44VPERR7\GV4#uupp i)rTFzinput_version is None for key c3\<"TF!wrwr4WW43SPV,3xK# R#5irW)r)rtimeactionrrrOs& r'r)MemoryProfile.timeline..s1 06,nsC>4>>#+> ?06s),c:V^,V^,P3#r)valuers&r'r(MemoryProfile.timeline..s1Q41"4r&rr)$rrrrr4r6rrr|rr;r@rr1r4r3r5popr2_category_snapshotrrkeysrrrtrurbr^rqr]rrfextendrr)rOoutputallocation_times live_unknownrr=r6rbrtkeyrptr_and_devicesnapshot last_versionreventsrrzsf 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("TFq^3xK R#5irErrrs& r'rr8 s&N;MCQx;Mc3("TFq^3xK R#5irErr;s& r'rr8 s&K?CQx?r<)rrrr:rrurrwrr1ryrrrM)rOall_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*cV^8dQhRR/#)rDrEzset[int]r)rGs"r'rHr s"+"+("+r&c aa\4o\S4pSPPFp\;QJd0.VV3RlVP P 44FNK 5M)!VV3RlVP P 444pV'gKqSPV4SPRVP44K \S4V8XgKS#)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<"TFgwpwr#SPPW4\P\P39gVP S9gKYVP xKi R#5irW)rrMrr!r#rd)rrrrdepends_on_gradientrOs& r'rAMemoryProfile._any_version_depends_on_gradient..#s_-@)\a''++C9 ))8+=+=>?vv!44 CFF-@s AA2A2c38"TFqPxK R#5irW)rdr;s& r'rrD.s.N#vvs) rrrrrrrur:rw)rO start_sizeridsrCsf @r' _any_version_depends_on_gradient.MemoryProfile._any_version_depends_on_gradients),01J--88e-1[[->->-@ee-1[[->->-@3'..s3'...N.NN9&&':5**r&cV^8dQhRR/#rr)rGs"r'rHr 8sCCCr&c xVPP4FAp\V4F/wr#VPP V\ P 4K1 KC VPPF@pVPF-pVPPV\ P4K/ KB R#)z>Mark Tensors which are unambiguous and simple to reason about.N) rrrrrrr!rrrrr)rOrrrrrs& r'r,MemoryProfile._set_gradients_and_temporaries8s ]]&&(E.u5   **683D3DE6)))44D''  ++Ax/A/AB(5r&cV^8dQhRR/#rr)rGs"r'rHr IsFFTFr&c VPP4FEp\V4F3pVfK VPP V\ P 4K5 KG R#rW)rrrrrrr#)rOrrs& r'r1MemoryProfile._set_parameters_using_python_tracerIsI]]&&(E'.=$$..q(2D2DE/)r&cV^8dQhRR/#rr)rGs"r'rHr Os1U1UT1Ur&c jaaSP4p\4o\SPP4FpVP P 4UUUu0uF wpwrEW53kK ppppWbPP 4,p\;QJd VV3RlV4F 'gK RM RM!VV3RlV44'gKSV,oK SP4pSPPFWp\P\VP49gK-V\VPP 44,pKY VFBwr5VPV9gKSPP!W5\"P$4KD R#uupppi)aMMark inputs based on which Tensors are updated using gradients. The process for differentiating between inputs and activations is more involved. Most Tensors in a training loop depend on at least one gradient: parameters depend on them through updates, and activations and optimizer state depend on them transitively through parameters. Critically, we do not need to know which Tensors are parameters to apply this method; we can simply walk the data flow graph to build the set of all values which depend on a gradient and then obtain the set of inputs from the conjugate set. There is, however, one hiccup. The first time we see a parameter is generally on the forward pass of the first step. We know from inspection of the data flow graph that v1 of that Tensor depends on a gradient (provided we profile an optimizer step), but not v0. To address this problem we weaken the definition of "depends on a gradient" to "any version of this Tensor depends on a gradient", which in turn strengthens the criteria for the input set enough to filter the activations in the forward pass of the first step.c3<"TFMpSPP!V!\P\P39;'gVS9xKO R#5irW)rrMrr!r#)rrproduces_gradientrOs& r'r,MemoryProfile._set_inputs..psY!A  $$a(X->->@R@R,SS**))* s AA ATFN)rHrreversedrrrrurwanycopyr rrrqrdrrrr) rOrCrrrrtensorsinput_candidatesrSs f @r'rMemoryProfile._set_inputsOsP2#CCE/2eT22==>D?C{{?P?P?RS?R*;#|~?RGS ||))+ +Gs!sss! "W,!?-113))44D,, 4;;0GG C (:(:(<$== 5-LCvv00  66sX^^T-!TsF. cV^8dQhRR/#rr)rGs"r'rHr sHDHDHDr&c aa SP4p\4pVP4UUu0uFwr4V\P8XgKVkK pppSP P EFgpVPP4UUU u0uF wpwrWy3kK p ppp \P\VP49gKa\;QJdV3RlV 4F 'gK RM RM!V3RlV 44'dK\;QJd7V3RlVPP44F 'gK RM, RM(!V3RlVPP444'dEKVPV 4'gEK0WVPP4,pW*PV4,pEKj \4o \!SP P 4Fp\;QJd8VV 3RlVPP44F 'gK RM- RM)!VV 3RlVPP444'gKwS P#RVPP444K VP%S 4VUUu0uFwrxVP&kK p ppV SP)4,p VFBwrxVP&V 9gKSP*P-V\P.4KD R#uuppiuup ppiuuppi)aNDeduce which Tensors are parameters. Consider the following code for the step of SGD with momentum (nesterov=False), where `d_p` is the gradient of `param` and `buf` is the momentum buffer. ``` buf.mul_(momentum).add_(d_p, alpha=1 - dampening) d_p = buf param.add_(d_p, alpha=-lr) ``` Both `param` and `buf` take a gradient and perform an in-place update. The python tracer will inspect calls to `nn.Module.forward` and `optim.Optimizer.step` to extract parameter and optimizer state respectively (including parameters), so this is generally a non-issue. However as a fallback we can also exploit several properties of parameters to distinguish them from other model state. First, they are directly used in the forward pass. (At this point we haven't established which parts of the graph correspond to the forward pass but we can deduce enough to suffice.) Some mutable state such as batch norm moving averages also contribute to the forward pass, but optimizer state does not. Second, a parameter is by definition used to compute at least one gradient and depends on at least one gradient. c3D<"TFpSP!V!xK R#5irWr3rrrOs& r'r@MemoryProfile._set_parameters_using_data_flow..sB6aD--q16 TFc3D<"TFpSP!V!xK R#5irWr^r_s& r'rr`sP;OaD--q1;Orac3^<"TF"pSP!V!;'gVS9xK$ R#5irWr^)rrrOused_for_gradients& r'rr`s4-A!!1%??.?)??-s--c32"TF wpwr#W3xK R#5irWr)rrrrs& r'rr`s)++:K:K:MN:MsJQsl:MFN--Z 5LLB6BB6BBBP4<<;M;M;OPP4<<;M;M;OPPP*66v>>%););)==%$(9(9:O(PP$5/2eT22==>Ds++-sss++-"(()<@KKL>M M cV^8dQhRR/#rr)rGs"r'rHr sTT$Tr&c \P\P0p\P\P0pVP P FpVPP4UUUu0uF wpwrVWF3kK ppppVUu0uFqPP!V!kK p pW,'gKnWV,, 'dK\P\VP49gKVPP4F0pVPP !.VO\PN5!K2 K R#uupppiuupi)z(Flood the graph to identify activations.N)rrr r#rrrrrurrMr rrrqrwr) rOrequired also_allowedrrrr!rrinput_categoriess & r'rMemoryProfile._set_activationssNNH$7$78 **H,>,>? ))44D:>++:K:K:MN:MsJQsl:MFNBHI&Q 0 0 4 4a 8& I",,) -DEE 11DKK9PP++-A$$::SASx?R?RS.5NIs 8E "E!cV^8dQhRR/#rr)rGs"r'rHr s R Rd Rr&c VPP4FpVP^,\P8XgK+VP^,P 'gKPVP^,P P p\PPRV44FJwr4\PV4pVfK VPPV\P4KL K R#)rc3,"TF wrVxK R#5irWr)rrr;s& r'r5MemoryProfile._set_optimizer_state..s9j{q!UjsN)rrrrrrritchain from_iterablerrrrrr$)rOrrrrrs& r'r"MemoryProfile._set_optimizer_states]]&&(E{{1~!2!22u{{1~7O7O7O"[[^55@@ HH229j9DA$//2C((2238P8PQ )r&cV^8dQhRR/#rr)rGs"r'rHr sdr&c R\P0pVPPFp\P \ VP49gK-VPP4F_wr4V^8Xg*VPPW4^, 4V9gK5VPPW4\P4Ka K R#rW) rr"rrr rrrqrwrurrMr)rOpriorrrrs& r'r"MemoryProfile._set_autograd_detailsx//0))44D,, 4;;0GG$(LL$6$6$8LC!|t'7'7';';C1'MQV'V((>>(*B*B%95r&)rrrrN)rrrrrrr/r3r$rHrrrrrrrr%rr&r'r r s` $BBHJ "+HC"F 1UfHDTT& Rr&r cP]tRtRtRRltRtRRltRRltR R R lltR t R #)MemoryProfileTimelineicV^8dQhRR/#rr)rGs"r'rH"MemoryProfileTimeline.__annotate__s 5 5$ 5r&c JVPVnVPVnR#)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)r/r categories)rOmemory_profiles&&r'rMemoryProfileTimeline.__init__s'// (44r&c aa \P!V4p.p.o RVV 3RllpRpSPEFowrgwrp VPV8wdKVR8wd\VR, 4pVR8XgWe8d V^8dTp\ V4^8Xd@VP V4S P ^.\ U u.uFp ^NK up ,4MDWcR,8wd8VP V4S P S R,P44V\P\P39d V!WV 4EKV\P8XdV!WV )4V!W^,V 4EKBV\P8XdV!WV )4EKd\RV 24h VUu.uF qf^8dTMTNK ppVS 3#uup iuupi)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] cV^8dQhRR/#rr)rGs"r'rH>MemoryProfileTimeline._coalesce_timeline..__annotate__ s + +4 +r&c<\V\4'dSPPW4MRp\V,^,pSR,V;;,\ V4, uu&R#)Nr)rXrrrM_CATEGORY_TO_INDEXr?)rrr@rrrOrJs&&& r'r:8MemoryProfileTimeline._coalesce_timeline..update sWc9--##C1  'x014E "Ie E * r&iUnknown action: r)rr;r/r?rrrrWr1r2r3r4r5 ValueError) rO device_strr;timesr:t_minrrrrnumbytesrrJs f& @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|+$C(7s 7 G -G cV^8dQhRR/#rr)rGs"r'rHr=s ) )$ )r&c VPV4wr4^RIp\VR4;_uu_4pVPW4.V4RRR4R# +'giR#;i)zxSaves the memory timeline as [times, sizes by category] as a JSON formatted file to the given path for the given device.Nw)rjsonopendump)rOpathrrrJrfs&&& r'export_memory_timeline,MemoryProfileTimeline.export_memory_timeline=sB..z:  $__ IIuna (___s A  A cV^8dQhRR/#rr)rGs"r'rHrHsD%D%dD%r&c  0a\P!V4p.pV3RlpSPEFwrgwrp VPV8wdKV\P\P 39d)VP V\V,V V!W434KjV\P8XdWVP V\V,V )V!W434VP V\V,V V!W^,434KV\P8Xd+VP V\V,V )V!W434EK\RV 24h ^RI p \VR4;_uu_4p V PWL4RRR4R# +'giR#;i)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.c<\V\4'dSPPW4MRp\V,#rW)rXrrrMr)rrrrOs&& r'get_category_indexLMemoryProfileTimeline.export_memory_timeline_raw..get_category_indexPs;c9--##C1  &h/ /r&rNr)rr;r/r1r2r3r_ACTION_TO_INDEXr4r5rrrr) rOrrr; raw_eventsrrrrrrrrs f&& r'export_memory_timeline_raw0MemoryProfileTimeline.export_memory_timeline_rawHsr j)68  048== /A~xzzV#&,,fmm<<!!(0 *38 6333!!(0! *38 !!(0 *3! < 6>>)!!(0! *38 !#3F8!<==_4Ab  $__ IIj $___s (F F NcV^8dQhRR/#rr)rGs"r'rHrsDD Dr&c  ^RIpVPPR4pVf\R4R#^RIHp^RIHp^RIH p ^RI p VPV4p V PV ^,4V PV ^,4r\V 4pW,p V PV ^R7R, p\P !V4p\P"P%V4p\P"P'V4pV P)V^PR7pVP+4p\,P/4FIwpp\0V,pVP3V R, VR V3,VR V^,3,VR R 7KK TP5\,Uu.uFpVfR M VP6NK up4VP9R 4VP;R4RP=V'dV.M.RVR, R RVR, R R2.,4pVP?V4V!RRR7;_uu_4pVPAVRR7VPC^^4V!VPE44PGR4pV'g \IR4hRV R2p\KVRRR7;_uu_4pVPMV4RRR4RRR4R#uupi +'giL;i +'giR#;i) zkExports the memory timeline as an HTML file which contains the memory timeline plot embedded as a PNG file.N matplotlibzDexport_memory_timeline_html failed because matplotlib was not found.) b64encode)NamedTemporaryFile)axis)figsizedpig@@:NNNgffffff?)coloralphaUnknownz Time (ms)z Memory (GB)z zMax memory allocated: z.2fz GiB Max memory reserved: z GiBzw+bz.png)suffixpng)rGzutf-8z failed to encode image as base64z} GPU Memory Timeline HTML r)encodingi@)'importlib.utilutil find_specprintbase64rtempfilermatplotlib.pyplotpyplotnumpyrarraymincumsumrr;cudamax_memory_allocatedmax_memory_reservedfiguregca_CATEGORY_TO_COLORSrur fill_betweenlegendr set_xlabel set_ylabeljoin set_titlesavefigseekreaddecoderrwrite)rOrrrtitle importlibmatplotlib_specrrpltnpmtrrJrstackedr;rrfigaxesrrrtmpfileencodedhtmlrs&&&&& r'export_memory_timeline_html1MemoryProfileTimeline.export_memory_timeline_htmls #..22<@  " V  $/'  $ $Z 0xx1AuE  ))E)*W4j)$zz>>vF#jj<rrrrrrrr+r\rirlrrr r)rcs00r'rs{" 44 '!('2  K$%  !"tyy"   k NNG   | k& (11D'EF'Etqad'EFTYY)//1qwwJ/ $E$?@ (((( $D>.X.X?.Xb(NV ood " "G!G!T  $ $ $])])@s)s)l KKK    >EEP ]]wG0s  I+I