Ei 2%SSKrSSKrSSKrSSKrSSKJr SSKJrJr SSK J r SSK J r J r JrJr SSKrSSKJr SSKJr SSKJr SS KJr S S KJr S S KJrJr S S KJr S SK J!r!J"r" S SK#J$r$J%r%J&r&J'r'J(r(J)r) /SQr*S r+Sr,Sr-Sr.Sr/Sr0Sr1Sr2Sr3\RhRk\2\.5r6\RhRk\1S5r7\ S\8S'"SS5r9\"SS9"SS55r:\"SS9"S S!55r;\"SS9"S"S#55r<\"SS9"S$S%\=55r>\"SS9\ "S&S'555r?\"SS9"S(S)\55r@\"SS9S2S*\RRS+\\ S,\\CS-\DS.\E\C\ 44 S/jj5rF"S0S15rGg)3N) defaultdict)IterableSequence) dataclass)AnyLiteral NamedTupleOptional)trace_structured) compatibility)map_arg)get_size_of_node) FxGraphDrawer)get_node_targetOperatorSupportBase) ShapeProp)!move_non_tensor_nodes_on_boundary split_by_tags)CALLABLE_NODE_OPSFxNetAccFusionsFinderis_node_output_tensorNodeListNodeSetTensors)FxNetAccNodesFinderFxNetSplitterInternalErrorSubgraph SplitResultgenerate_inputs_for_submodules NodeEventNodeEventTrackerF_fx_net_trackerz _nodes.txtz_all.txt FX_NET_ACC_SPLITTER_TRACKER_MODE%FX_NET_ACC_SPLITTER_TRACKER_DUMP_PATH)FX_NET_ACC_SPLITTER_TRACKER_TRACKED_NODES0)r'123 TRACKER_MODEc4\rSrSr\\\SS4S\S\4Sjjr Sr g) _SplitterSettingBaseLFmax_acc_splitsrcX[R"5nURSSS[SS9 URSSS[SS9 URS S SS S S 9 URSSSS SS 9 URSSSS SS9 UR 5upxUR (a UR OUUlUR (a UR OUUlUR(a UROUUlX@lUR(aURUl gUUl g)Nz--min-acc-module-sizez--min_acc_module_sizeFz.Minimum size limit of an accelerator subgraph.)requiredtypehelpz--max-acc-splitsz--max_acc_splitsz,Enforce a maximum number of split subgraphs.z --skip-fusionz --skip_fusion store_truezIf true then no fusion groups. Fusion group is used to enforce no non-tensor data flow between submodules. If we don't have this constrain, setting this to false is recommended as it can reduce overhead.)defaultactionr4z--allow-non-tensorz--allow_non_tensoraFor some backends non-tensor data flow between cpu and them are not allowed. Therefore, if a node supported by accelerator but it has non-tensor inputs or outputs to a cpu node we would want to consider it as a cpu node during splitting. However, for some backends we might not care about non-tensor data flow and we can set this option to true to disable the functionality that prevent non-tensor data flow.z#--move-non-tensor-nodes-on-boundaryz#--move_non_tensor_nodes_on_boundarya=AOTI does not support non-tensor nodes on acc->acc, acc->gpu and gpu->acc boundary. For non-tensor nodes on acc->acc boundary and acc->gpu, we move the nodes from upstream to downstream. For non-tensor nodes on gpu->acc boundary, it is handled by the pre-split process. (by method reduce_acc_nodes_non_tensor_input). )r2r7r4) argparseArgumentParser add_argumentintparse_known_argsmin_acc_module_size skip_fusionallow_non_tensorr0r) selfr=r>r?r0rparserargs_unknowns Z/home/mande/repo/quber/.venv/lib/python3.13/site-packages/torch/fx/passes/splitter_base.py__init___SplitterSettingBase.__init__Msg((* # #A     ?     #    V    1 1>   002''  $ $$  6:5E5E!1!1;%)%:%:D ! !@P $255  2 2 .3 .)r?r0r=rr>N) __name__ __module__ __qualname____firstlineno__DEFAULT_MIN_ACC_MODULE_SIZEDEFAULT_SKIP_FUSIONDEFAULT_ALLOW_NON_TENSORr;boolrE__static_attributes__rGrDr-r-Ls58'1 27 G  G ,0 G G rGr-)is_backward_compatiblec\rSrSrSrS S\R RS\S\ \R R4Sjjr Sr S r g) r!z An event in graph split that happened on a node. source: Subject of the event desc: readable description dep: Optional dependency, usually the node that caused the event. Nsourcedescdepc(XlX lX0lgN)rUrVrW)r@rUrVrWs rDrENodeEvent.__init__s  rGcURRSURSUR(aURR3$S3$)N:  #)rUnamerVrWr@s rDto_strNodeEvent.to_strsE ++""#2dii[488$((--1UVVQT1UVVrG)rWrVrUrY) rHrIrJrK__doc__torchfxNodestrr rErarPrQrGrDr!r!sFPThhmm+.5=ehhmm5LWrGr!c\rSrSrSrSrSS\RRS\ S\ \RR4Sjjr SS jr S r SS jrS rS rg)r"z3 Tracks node events during the splitter execution. cNXlX l/Ul0Ul[UlgrY) tracker_mode dump_prefixevents node_eventsprintwriter)r@rkrls rDrENodeEventTracker.__init__s$(&  rGNnoderVrWc4[XU5nURRU5 URUR;a/URUR'URURR[ UR5S- 5 g)z! Add a new event to the tracker. rN)r!rmappendr_rnlen)r@rrrVrWevents rDaddNodeEventTracker.addss$c* 5! 99D,, ,*,D  TYY ' #**3t{{+;a+?@rGc@U(d URnURRU/5HknURUnU"X6R 5-5 U(dM3UR cMBUR UR RSSU-US9 Mm g)z Print a node and its events. @param recursive: if True, print nodes that caused the events on this current node. @param tab: Indentation for dependencies. @param writer: function to write to file. If None, use print. NTz|  recursivetabrp)rprngetrmrarW print_noder_)r@ node_namer{r|rpidxrvs rDr~NodeEventTracker.print_nodes[[F##'' 26CKK$E 3' (yUYY2IINNds 6  7rGc0nURHZn/X'URRU/5H3nURUnXRUR 55 M5 M\ U$)z! Create dict dump on all events. )rnr}rmrtra)r@retr_rrvs rDto_dictNodeEventTracker.to_dictsh%'$$DCI''++D"5 C(   06%  rGcU(d URnURH!nU"SUS35 URUSSUS9 M# g)zZ Print all nodes in a list. @param writer: function to write to file. If None, use print. zNode: :Fz rzN)rprnr~)r@rpr_s rD print_allNodeEventTracker.print_allsE [[F$$D VD6# $ OODEtFO K%rGc\^^[SSU4SjS9 SmTRS:a=[TR[-S5nTR T"U55 SSS5 UU4S jnTRS :XdTRS :XaTRS :Xa3[ RR[S 5RS 5O?TRR5VVs/sHup4[U5S:dMUPM snnnU"U5 gg!,(df  N=fs snnf)zm Function to be invoked at the end of the finder execution to printout tracked events specified by the mode. artifactc SSS.$)N!fx_net_acc_splitter_finder_eventsjson)r_encodingrQrQrGrD'NodeEventTracker.dump..s ;"!rGcL>[R"TR55$rY)rdumpsrr`srDrrstzz$,,.9rG) metadata_fn payload_fnc^U4SjnU$)Nc,>TRUS-5$)N )write)xfs rDfn2NodeEventTracker.dump..writeln..fnswwq4x((rGrQ)rrs` rDwriteln&NodeEventTracker.dump..writelns )IrGrwNc >[TR[-S5nUH3nT"SUS35 TRUSST"U5S9 T"SUS35 M5 SSS5 g!,(df  g=f)Nrz===== Tracking node z =====Tz|-rzz===== End of tracking node )openrl NODES_SUFFIXr~)nodesrrr@rs rDdump_selected_nodes2NodeEventTracker.dump..dump_selected_nodes std&&5s;q!&I29+VDEOO!TtGAJ$9)FKL "'<;;s :A"" A0,)r rkrrl ALL_SUFFIXrosenvironr}-ENV_FX_NET_ACC_SPLITTER_TRACKER_TRACKED_NODESsplitrnitemsru)r@rrr_rmrrs` @rDdumpNodeEventTracker.dumps  :       !d&&3S9Qwqz*: M    !T%6%6!%; $$) LbQWW .2-=-=-C-C-E-E\TVWXD-E   &&<:9(sD*D(D( D%)rlrmrnrkrprY)FrN)rHrIrJrKrcrErdrerfrgr rwr~rrrrPrQrGrDr"r"sTA ASAx 7NA"  L/'rGr"c~\rSrSrSrS\R RS\S\ 4Sjr S\ 4Sjr S r S rS \4S jrS rg)ri"a Finds a set of nodes that can be supported on ACC, excluding nodes that have non-tensor input/output to cpu nodes to prevent non-tensor data flow between backends and cpu. I.e. if we have a chain: ACC_NODE_1 -> ACC_NODE_2 -> ACC_NODE_3 -> CPU_NODE_1 where every ACC node produces non-tensor output, then they all should be treated as CPU nodes. This behavior can be turned off by passing allow_non_tensor=True. moduleoperator_supportr?cXlX lX0l[5Ul[ [ [5[5Ul grY) rrr?set acc_nodesr"r;r+ DUMP_PREFIXtracker)r@rrr?s rDrEFxNetAccNodesFinder.__init__1s1  0 0"%%'L(9;G rG cpu_worklistc~U(aURS5nURHnX0R;dMURRU5 URR USU5 [ U5(aM^URR US5 URU5 M U(aMgg)z Transitively excludes nodes from ACC supported set. For every node in the worklist: - removes its downstream ACC nodes from ACC supported set, - if any downstream ACC node produces non-tensor output, then it gets added into the worklist. rzacc_del|user_of_new_cpu_nodeznew_cpu_node|non_tensor_outputN)popusersrremoverrwrrt)r@rrrusers rD(reduce_acc_nodes_non_tensor_input_helpers##A&D >>)NN))$/LL$$T+I4P066 ((/OP$++D1 #lrGcD/nURRRHinUR[;aMX R ;aM*[ U5(aM<URRUS5 URU5 Mk URU5 g)zl Excludes nodes from ACC supported set that have direct upstream CPU nodes that produce non-tensor outputs. z&new_cpu_node|callable_non_tensor_inputN) rgraphroprrrrrwrtr)r@non_tensor_cpu_nodesrrs rD!reduce_acc_nodes_non_tensor_input5FxNetAccNodesFinder.reduce_acc_nodes_non_tensor_inputQs *,KK%%++Dww//~~%$T** LL  T#K L ' ' -, 556JKrGcv/nURHhn[U5(aMURHCnX0R;dMURU5 URR USU5 Mf Mj U(dgUHnURR U5 M URU5 M)zf Excludes nodes from ACC supported set that produce non-tensor outputs and have downstream CPU nodes. z'acc_del|non_tensor_output_with_cpu_userN)rrrrtrrwrr)r@ new_cpu_nodesacc_noder new_cpu_nodes rD"reduce_acc_nodes_non_tensor_output6FxNetAccNodesFinder.reduce_acc_nodes_non_tensor_outputds &(M NN(22$NND>>1%,,X6 (($&OQU ++! - %%l3!.  9 9- H)rGreturnc[URR55n[5UlURR R HnUR[;aURRUS5 M5URRX5(dURRUS5 MsURRUS5 URRU5 M UR(d UR5 UR5 URR!5 UR$)Nzinit_cpu|not_callablezinit_cpu|operator_supportz(init_acc|callable_and_operator_supported)dictr named_modulesrrrrrrrrwris_node_supportedr?rrr)r@ submodulesns rD__call__FxNetAccNodesFinder.__call__s$++3356 ""((Att,,   $;<((:::II   $?@ LL  Q J K NN  q !)$$  2 2 4  3 3 5 ~~rG)rr?rrrN)rHrIrJrKrcrdre GraphModulerrOrErrrrrrrPrQrGrDrr"sZ  H$$ H. H H2X2&L&I6'rGrc\rSrSrSrg)rirQN)rHrIrJrKrPrQrGrDrrsrGrc>\rSrSr%\\S'\\S'Sr\\ \S'Sr g)riis_accrNdevice_ordinalrQ) rHrIrJrKrO__annotations__rrr r;rPrQrGrDrrs L O$(NHSM(rGrcj\rSrSr%Sr\R R\S'\ \ \ 4\S'\ \S'Sr g)ria& Stores the results of the splitter. Attributes: split_module: root module after splitting. submodule_inputs: a dict that maps submodule name to its inputs. non_acc_submodule_prefix: the prefix for non acc submodules. For acc submodule the prefix is always "_run_on_acc_". split_modulesubmodule_inputsnon_acc_submodule_prefixrQN) rHrIrJrKrcrdrerrrrgrrPrQrGrDrrs-((&&&38n$!!rGrmodelinputstarget_submodulesdeepcopyrc ^^ ^ ^ /m 0m UR5VVs0sHupEXT_M snnm UU U 4SjnUR5HWupEXB;dM [U[RR5(aM7T R UR U55 MY U 4Sjn[R"5 U"U6 SSS5 U"5 T $s snnf!,(df  N=f![anU"5 UeSnAff=f)aZ Generate inputs for targeting submdoules in the given model. Note that if two submodules refer to the same obj, this function doesn't work. Args: model: root model. inputs: inputs to the root model. target_submodules: submodules that we want to generate inputs for. Returns: A dict that maps from submodule name to its inputs. cP>T(a[R"U5OUTTU'grY)copyr)r module_inputsrresultssubmodule_to_namess rD pre_forward3generate_inputs_for_submodules..pre_forwards#,4DMM- (- "6*+rGc8>THnUR5 M grY)r)hhandless rDclean_up_handles8generate_inputs_for_submodules..clean_up_handlessA HHJrGN) r isinstancerdjit ScriptModulertregister_forward_pre_hookno_grad Exception) rrrrr_modrrerrrs ` @@@rDr r s(GG5:5H5H5JK5J #)5JK ((*   $c599#9#9::s<<[IJ+  ]]_ 6N  N1L"_ s;C(C+=CC+ C($C+(C++ D5 C>>Dcp\rSrSrSrSrS+S\RRS\ \ S\ S\ S \ S \S \\4S jjrS \\ \ 44SjrS \\RR(\44SjrSrS\RRS\S \R2R44SjrS\RRS\S \ 4SjrS\RRS\4SjrS,S\4SjjrS,S\4Sjjr S-S\\!S \\RR(\44Sjjr"S\\RR(\44Sjr#S\ S \4Sjr$S\ 4S jr%S \&\\44S!jr'S \(\)4S"jr*S#\(\)S \(\)4S$jr+S#\(\)4S%jr,S,S&\S \RR4S'jjr-S \RR4S(jr.S \/4S)jr0S*r1g). _SplitterBaseia Splits a GraphModule into sub-GraphModules for execution on CPU or the accelerator. Output is a GraphModule with supported and unsupported operators grouped into as few sub-GraphModules as possible. Assumes that only "call_module", "call_function" and "call_method" from FX IR can potentially be executed on the accelerator. Given the following graph: ==> b ==> // \ a d \ // ==> c ==> class SimpleModule(torch.nn.Module): def forward(self, a): b = torch.sin(a) c = torch.cos(a) d = b + c return d and providing "operator_support" that indicates that 'b' and 'c' can be executed on the accelerator, we will get the following split result: main: def forward(self, a): run_on_acc_0_0 = self._run_on_acc_0_0(a) getitem = run_on_acc_0_0[0] getitem_1 = run_on_acc_0_0[1] run_on_cpu_1_1 = self._run_on_cpu_1_1(getitem, getitem_1) return run_on_cpu_1_1 _run_on_acc_0_0: def forward(self, a): sin_1 = torch.sin(a) cos_1 = torch.cos(a) return (sin_1, cos_1) _run_on_cpu_1_1: def forward(self, sin_1, cos_1): add_1 = sin_1 + cos_1 return add_1 ldNr sample_inputrsettingsnon_acc_submodule_name return_tuple nodes_finderc[U[RR5(d[ S[ U535eXl[UR 5R"U6 X@l X0l X l Uc5[UR URURR5nU"5UlURR(a0UlO[#XR5"5UlUR%5UlUR)5 XPl0UlX`l/Ulg)a Preprocesses graph before splitting: - finds nodes supported by ACC, - finds fusion groups for ACC nodes having non-tensor IO, - builds a graph of direct dependencies, - builds a map of fused nodes to their fusions. As a result we get self.acc_nodes, self.deps and self.fusions. zExpected GraphModule, got N)rrdrerAssertionErrorr3rr propagaterrrrr?rr>fusionsr find_depsdepsupdate_deps_for_fusionsr_node_submodule_map _return_tupletags)r@rrrrrrrs rDrE_SplitterBase.__init__s$&%(("6"677 #=d6l^!LM M $++((,7 0(  . T22DMM4R4RL& == $ $DL0HJDLNN$  $$&&<#35 )! rGrcUR$)zReturns a map from node name to submodule name, e.g. node: main_module_impl_impl_over_arch_unary_multiple_embedding _pooling_embedding_pooling_sparse_entity_equivalence_key _proxy_embedding_bag maps to submodule name of: _run_on_acc_1 )r r`s rDget_node_submodule_map$_SplitterBase.get_node_submodule_mapEs'''rGc[[5nURRRHQnUR [ ;aMURH(nUR S:wdMXRU5 M* MS U$)z Builds a graph of node dependencies. Leaf nodes don't have any dependencies and the "output" node doesn't have nodes depending on it. Resulting graph has only direct dependencies, i.e. there are no transitive dependencies. output) rrrrrrrrrw)r@rrrrs rDr_SplitterBase.find_depsNsg.9-=KK%%++Dww// 77h&JNN4(# , rGc&URHnURUnUHinURURURUU- 5 URH(nXB;dM URUR U5 M* Mk M g)z Updates graph of dependencies so that: - nodes from the same fusion depend on the same set of outer nodes, - outer nodes depending on a fusion depend on all nodes in that fusion. N)rrupdaterrw)r@rrfusionfused_neighborrs rDr %_SplitterBase.update_deps_for_fusions`sz LLD\\$'F"( $&&tyy'@6'IJ*00D) $++D11#)!rGrrcU$)z Lower the model to a backend. rQr@rrs rD_lower_model_to_backend%_SplitterBase._lower_model_to_backendss  rGcg)z When an error occurs during lowering or running the lowered mod, we use this function to find culprits in the `mod` that causes the error. zMUnable to find a culprit because _find_culprit() function is not implemented.rQrs rD _find_culprit_SplitterBase._find_culprit|s _rGsupported_nodesc^^SSSS.m"UU4SjS[5nU"USSS 9nUR5nURS 5 g) N AliceBlue chartreuse1crimson)r6 supported unsupportedc.>^\rSrSrUUU4SjrSrU=r$)E_SplitterBase._draw_graph_based_on_node_support..CustomDraweric>[TU]U5nUT;a TSUS'U$UR[;a TSUS'U$TSUS'U$)Nr% fillcolorr&r6)super_get_node_stylerr)r@rrtemplate __class__ color_mapr s rDr,U_SplitterBase._draw_graph_based_on_node_support..CustomDrawer._get_node_stylesk 7248?*,5k,BH[)   WW 11,5m,DH[) -6i,@H[)rGrQ)rHrIrJrKr,rP __classcell__)r.r/r s@rD CustomDrawerr(s  rGr2 node_supportTignore_getattrznode_support.dot)rget_main_dot_graph write_raw)r@rr r2drawer dot_graphr/s ` @rD!_draw_graph_based_on_node_support/_SplitterBase._draw_graph_based_on_node_supportsS#&$ = c>$G--/ ./rG dump_graphc ^[URR55n/n[[5n[[5nSmURR R GHCnUR[;aM[X&5nURVs/sH6n[U[RR5(aT"U5OSPM8 n n[U 5[!S[#[%U 555[U 55- n ['U SU 5n ['U4SjUR(R+555n UR,R/X&5(a(UR1U5 XGR3X45 GM/XWR3X45 GMF U(aUR5URU5 Sn UR+5H&upUHupXSU S[U 5S3- n M M( U S- n UR+5H&upUHupXSU S[U 5S3- n M M( [7U 5 U $s snf) NcRURRS5n[USS5$)N tensor_metadtype)metar}getattr)argr?s rD get_dtype5_SplitterBase.node_support_preview..get_dtypes#((,,}5K;6 6rGc34# UHupUcM Uv M g7frYrQ).0ir@s rD 5_SplitterBase.node_support_preview..s$CA$Cs  c3># UH<up[U[RR5(dM0UT"U54v M> g7frY)rrdrerf)rGkrCrDs rDrIrJs7'1FAc588==1$IcN#1s /AAz$ Supported node types in the model: z: (z, z) z& Unsupported node types in the model: )rrrrrrrrrrrBrrdrerfrunext enumeratereversedtuplekwargsrrrrtrwr:ro)r@r<rr supported_node_typesunsupported_node_typesrrtargetrC arg_dtypes last_indexarg_dtypes_tuplekwarg_dtypes_tuplereportstdtypesrDs @rDnode_support_preview"_SplitterBase.node_support_previewsI$++3356 $&*3/!,S!1 7KK%%++Dww//$Z6F  99$C#-S%((--"@"@ #dJ$  Z4$-hz.B$C J ,J %Z %<= !&'"kk//1'"  $$66zHH&&t,$,002B1WX&.22%:E,L   2 24;; P:-335IA8>4 S$4#5R=O8P7QQTUU9?6 ==/557IA8>4 S$4#5R=O8P7QQTUU9?8 g_s=Ic^^^SmUR5n[UVs/sHo3R(dMUPM sn5n[U5U- nTS[U5S3- mTSUSUS3- mURU5n[UVs/sHo3R(dMUPM sn5n[U5U- nTS[U5S3- mTSUSUS3- m[ U5HKupgTUR(aSUS 3OUR US 3- mT[UR 5S 3- mMM URU5 URS S 9nUR5 U(aH[US S S9n U R5n U R5HupU RU S35 M URn SnURR GHnUR S:XdMSUR";dM(TSUR"S3- m[%XR"5mSnU"UTUR&5n[)T5R*"U6 SnSmTS- mTRR H]nUR S:Xa6[-U5(dTSUR.S3- mOU[1TU5S- nUR S:XdM[UnM_ TS- mS[2R4R64UUU4Sjjn[9WR:U5 UR[=UT5- nTSUSTS 3- mTS!US"3- mUU :aUn UR"nUR?TU5nU"U6 TS$- mGM TS&U S 3- mTS'US(3- m[ET5 T$s snfs snf![@a TS#- mTURCTU5- mGMf=f![@a TS%- mTURCTU5- mGM f=f))Nrz+Before removing small acc subgraphs, total z subgraphs are created:r] acc subgraphs and z cpu subgraphs. z*After removing small acc subgraphs, total _run_on_acc_r\z node(s) T) remove_tagpreviewr4z.dot call_moduleaccz Processing acc submodule rcd^SmU4SjnURU5nU"U6 UR5 T$)Nc >UmgrYrQ)r@r sub_inputss rD get_inputsJ_SplitterBase.split_preview..get_submod_inputs..get_inputss%+ rG)rr)main_modsubmodexample_inputsrhhandlergs @rDget_submod_inputs6_SplitterBase.split_preview..get_submod_inputs s6!%J,$==jIFn-MMO%%rGrzChecking inputs... placeholderzInput = is not a tensor, this might cause problems during lowering! rzChecking outputs... rrcr>[U5(dTSURS3- mgT[TU5S- mg)NzOutput rqr)rr_r)rrrYrktotal_output_bytess rD get_bytes._SplitterBase.split_preview..get_bytes)s@166WTYYK7u#vv*.>vt.LQ.OO*rGzTotal input size in bytes is z , total output size in bytes is rzF theoretical max qps (bounds by PCIe bandwidth) for this submodule is z. z#Run into an error during lowering! zLowering and running succeed! z$Run into an error during inference! zB Theoretical max qps (bounds by PCIe bandwidth) for this model is z bottleneck is submodule .)#put_nodes_into_subgraphsrurremove_small_acc_subgraphsrNrrtagrevalrget_all_dot_graphsrr7PCIe_BWrrrTrBrrrrr_rrdrerfr rBmaxr RuntimeErrorrro)r@r< subgraphsgacc_subgraphs_numcpu_subgraphs_numrHsubgraph split_modr8 dot_graphsr_r9max_qpsbottleneck_modulerrrn submod_inputstotal_input_bytesr output_nodertqpslowered_submodrYrkrss @@@rD split_preview_SplitterBase.split_previewss113 I BIqI BC N->>@Y@PPghhQ())<=N I BIqI BC N->>?I?OOfggQ())<=Nc$((..*=b*A&BV&KL$$T* # , rGrcp[5nURR5Hup4X2;aM [5nUHnURX5 M UR U5 UHEnXQU'UR HnXt;dM XRU5 M UR U5 MG M grY)rrrrdifference_updateall_input_nodesrw)r@rprocessed_noderrrnew_deprrCs rDupdate_reverse_deps_for_fusions-_SplitterBase.update_reverse_deps_for_fusionsjs LL..0LD%eGtw'  % %f -!Q,,C( ((0-""1%1rGrycP[5nURRRHwnUR[ ;dMUR U:XdM+URH<nUR[ ;dMUR U:wdM+URU5 M> My U$)z Finds parent nodes of the `tag` subgraph. Traverse the inputs of nodes in the subgraph, if input doesn't belong to the subgraph and is not a placeholder, we consider it as the parent node of the subgraph. ) rrrrrrryrrw)r@ry parent_nodesrrrCs rDfind_parent_nodes_of_subgraph+_SplitterBase.find_parent_nodes_of_subgraphsyu KK%%++Dww++C//Cvv!22sww#~$((-0, rGc tUR[URSSS9S5S9nURU5 UR U5n[ 5nU(aSnUH nX&U::dM X`R ;dMUn O UcgXlURU5 URU5 XPR;a.URUHnXt;dM URU5 M URH1nUR[;dMX;dM URU5 M3 U(aMgg)zN Extend the acc subgraph with `tag` going the reversed topological direction. rr)maxsplitr/)rN)rr;rsplitrrrrryrrwrrrr) r@ryrr visited_nodesrrr fusion_noderCs rDextend_acc_subgraph!_SplitterBase.extend_acc_subgraphs %%SC!1LR1P-Q%R ,,T299#> !$ D"7m+^^0CD" |H    %   d #||##'<<#5K"7$((5$6 ++66..33K $$S),1lrGc[5n[5nURRRHnURS:XaK[ UR 5S:Xa2X0R;aURU5 OURU5 URS;aMpURH6nX@R;aURU5 M%URU5 M8 M X4$)z; Finds nodes that consume module inputs or get_attr nodes. call_functionr>get_attrrp) rrrrrrurrrwr)r@starter_cpu_nodesstarter_acc_nodesrrrs rD starter_nodes_SplitterBase.starter_nodess&)U%(UKK%%++Dww/)c$2F2F.G1.L>>)%))$/%))$/ww99 >>)%))$/%))$/ #,"!33rGc^^ TR5up[5m [U4SjU55(+n/n/nU(dU(GaaU(aUOUn[UU 4SjU5S5nUc5U(d [ S5eUR [ X4S95 U(+n/nMiURU5 T RU5 UR U5 UTR;aSUTR;a"URTRUT - 5 O!URTRUT - 5 URHMnUR[;aMUTR;aURU5 M<URU5 MO U(aGMWU(aGMaU(aUR [ X4S95 U(d [ S5eU$)Nc3Z># UH n[TRU5S:Hv M" g7f)rN)rur)rGrr@s rDrI9_SplitterBase.put_nodes_into_subgraphs..s%$WEVS1%6!%;EVs(+c3R># UHnTRUT::dMUv M g7frY)r)rGrr@rs rDrIrs"KMqTYYq\]-JMs' 'zSubgraph can't be empty)rrzCouldn't create subgraphs)rranyrMrrtrrrwrrrrrr) r@current_cpu_nodescurrent_acc_nodes acc_subgraphcurrent_subgraph_nodesr current_nodesrrrrs ` @rDrw&_SplitterBase.put_nodes_into_subgraphss/3/A/A/C,!$ "%$WEV$W!WW +-%' #41=-CTMKMKD |-45NOO  LO$0/ )+&   &   d # " ) )$ /t||#4>>)%,,T\\$-?--OP%,,T\\$-?--OP 77"334>>)%))$/%))$/#A #4#4V "    K ,-HI IrGrcv/nUGH/nUR(a[UR5URR:aUR U5 MU[ S[UR5SURR35 U(a*USRRUR5 MSUlUR U5 MU(a?USR(d+USRRUR5 GMUR U5 GM2 U$)zl This pass finds ACC submodules with less than specified size and merges them with adjacent CPU submodules. zBEliminating acc subgraph because it's smaller than the threshold: z < r/F)rrurrr=rtroextend)r@rrrs rDrx(_SplitterBase.remove_small_acc_subgraphss "$!Hx~~&$--*K*KKMM(+\x~~./s4==3T3T2UWr ((//?*/ h/&*"3"32J$$++HNN;MM(+%"& rGc/UlUHnUR(aS[UR53O"UR[UR53nURR U5 UR HAn[ US5(a[SUS35eX4lX0RUR'MC M g)Nr`ryzNode z was already tagged) r rrurrtrhasattrrryr r_)r@rrryrrs rDry_SplitterBase.tag:s !H??s499~./334S^4DE  II  S ! 4''4uTFBU5VWW69((3 '"rGrac[URURURS9nU(a<URRR Hn[ US5(dMU?M U$)N)rry)rrr r rrrry)r@rarrrs rDr_SplitterBase.splitJsZ$ KK1C1C   ))//4''0rGcxUR5nURR(a [U5 URU5n[ UVs/sHo"R (dMUPM sn5n[ U5U- n[ SUSUS35 URU5 UR5$s snf)NzGot r_z non-acc subgraphs) rwrrrxrurroryr)r@rsacc_subgraphs_countnon_acc_subgraphs_counts rDr_SplitterBase.__call__Ts113 == : : -i 833I> !i"Di881i"DE"%i.3F"F &'':;R:SSe f  zz| #Es B7)B7cPU"5n/nUR5Hup4URU5 M URRS:a.[ U5URR:a [ S5e[ XRU5n[XUR5$)NrziCannot fulfill max_acc_splits limit. This may cause split fragmentation and result in performance issues.) named_childrenrtrr0ru ValueErrorr rrr)r@rsubmodule_namesr__modrs rDgenerate_split_results$_SplitterBase.generate_split_resultsasv &557JD  " "4 (8 MM ( (1 ,O$t}}'C'CC0  : ++_ <4;V;VWWrG) r r rrrrrrrrr ) _run_on_cpu_FNFrY)2rHrIrJrKrcr|rdrerrrrr-rgrOr rrErrrfrrr rnnModulerrrr:r\rr;rrrrrPrlistrrwrxryrrrrrPrQrGrDrrsF(VG'5"6:."$$."sm.". ." ' ." !$ ."."23."h(S#X(4 w 67$ 2&88''18 _!5!5_w_3_088''0:B04AtAFmmh'+sm ehhmmW$ %,&D9O4P&6"(*s(*\4uWg%5642@$x.@DDNtH~6:T(^: 1E1E %((.. X XrGrr)Hr8rrr collectionsrcollections.abcrr dataclassesrtypingrrr r rdtorch._loggingr torch.fx._compatibilityr torch.fx.noder "torch.fx.passes.graph_manipulationr graph_drawerrrrr shape_propr split_utilsrr tools_commonrrrrrr__all__rLrMrNTRACKER_DUMP_PATHrr$ENV_FX_NET_ACC_SPLITTER_TRACKER_MODE)ENV_FX_NET_ACC_SPLITTER_TRACKER_DUMP_PATHrrr}rr+rr-r!r"rrrrrrrrgrOrr rrQrGrDrs #.!55 +1!?'B!I   &  'I$,S)/.jjnn-/@ -/JJNN(#- g() H H Ve,WW-W.e,o'o'-o'de,nn-nbe,  - e, )) -) e, "* "- " e,  - 88??- SM- }- -  #s(^ ---`P XP XrG