+ j\I Rt^RIHt^RIt^RIt^RIt^RIt^RIt^RIt^RI H t H t ^RI H t ^RIHtHtHt^RIt^RIt^RIHt^RIHt^RIHt^R IHt]!R 4t]'d ^R IHt^R I H!t!]PDPFt#]$]#,t%]$]#,t&]'](](])]'](](3,]'](](3,]PT](](](](3 ,t+]PX!]-4t.]P^Pa]-R 4t1RRlt2RRlt3!RR]44t5!RR]Pl4t7RRlt8RRlt9RRlt:RRlt;!RR4t<!R R!4t=!R"R#4t>R$R%lt?R&R'lt@R#)(a This module provides functionality for tracking and managing regions in computational graphs. It supports graph optimization by identifying and grouping similar regions based on their structure and behavior. The module implements algorithms for: 1. Tracking nodes and their relationships in the computational graph 2. Identifying identical or similar regions across the graph 3. Managing graph regions for optimization purposes 4. Supporting deduplication and other graph transformation passes The core functionality revolves around the GraphRegionTracker class which maintains mappings between nodes and their duplicates, enabling efficient graph analysis and optimization operations. ) annotationsN) defaultdictdeque)fields)Any TYPE_CHECKINGTypeVar) FakeTensor) OrderedSet) tree_flatten)_get_flat_args_uniqueT)Callable)InstructionTranslatorBasegraph_region_expansionc V^8dQhRRRR/#)msgstrreturnNone)formats"c/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_dynamo/graph_region_tracker.py __annotate__rDs**3*$*c6\P!V.VO5!R#N)graph_expansion_logdebug)rargss&*r debug_logr!Dsc)D)rc V^8dQhRRRR/#)rxz torch.Tensorrz(tuple[Callable[[T], T], tuple[Any, ...]]r)rs"rrrHs     -  rc^RIHpHp.pV!V4p\V4F'pVP \ WEP 44K) V\V43#))_ident%extract_tensor_metadata_for_cache_key)torch._inductor.codecacher&r'rappendgetattrnametuple)r#r&r'outmetadatafields& r&_extract_tensor_metadata_for_node_hashr0HsLX C4Q7H! 78ZZ01" E#J rc]tRt^UtRtR#)NodeHashExceptionrN)__name__ __module__ __qualname____firstlineno____static_attributes__rrrr2r2Usrr2c:a]tRt^YtRV3RlltRRltRtV;t#) InputPicklercV^8dQhRR/#rrrr)rs"rrInputPickler.__annotate__Zs$rc  <a^RIHo\P!4pWn\ SV`V4\PP4VnVPP\\\PV3Rl\PV3Rl\P V3Rl/4RVnR#)r%)r&c <S\V433#rrr#r&s&r'InputPickler.__init__..ds#a&(;rc <S\V433#rr?r@s&rrArBes&3q6))* +# * + LL  a LL ! !! $s#*A##B4A;;BB8B;)rGrKrR)r3r4r5r6rIrbr7 __classcell__)rUs@rr9r9Ys" % %rr9c V^8dQhRRRR/#)rargrrr)rs"rrrysssrc\V\4'dVPPR4#\V\P \ 34'dV#R#) example_valueN) isinstanceNodemetagetrNTensorint)rfs&r _extract_argsroys@#txx||O,, C%,,, - - rc V^8dQhRRRR/#)rnoderjrz.tuple[tuple[str, ...], tuple[Any | None, ...]]r)rs"rrrs HH H3Hrc\VP4wr\VPP 4\ P !^4R7p\\VPP444p\V4wrRW,pT\;QJd.RV4FNK 53#!RV443#)r%keyc38"TFp\V4xK R#5ir)ro).0rfs& r "_normalize_args..sFXc}S11Xs) r r sortedkwargsitemsoperator itemgetterr,keys)rq flat_args_ sorted_kwargs sorted_keys flat_kwargsall_argss& r_normalize_argsrs  *LI4;;,,.H4G4G4JKMt{{//123K!-0NK&H FXF GGFXFF GGrc$V^8dQhRRRRRR/#)r index_to_rankdict[int, int]regionszlist[list[Any]]rrr)rs"rrrs$ 8 8! 8,; 8  8rcaV^,p\\\V44V3RlR7pVFpVUu.uF qTV,NK upVR&K R#uupi)r%c<SV,#rr)irs&rrA'_sort_with_ref_region..s -PQBRrrs:NNNN)ryrangelen)rr ref_regionsorted_indicesregionrsf& r_sort_with_ref_regionrsM JE#j/28RSN(671AYY7q 7sAcV^8dQhRR/#)rrGlobalStateKeyr)rs"rrrs  n rc \P!4\P!4\P!4\PP 4\PP 4\P!4\P!4\PP4\P!4\PPP43 #r) rNis_grad_enabledis_inference_mode_enabledget_num_threads_C2_get_cublas_allow_fp16_reduced_precision_reduction2_get_cublas_allow_bf16_reduced_precision_reductionget_default_dtype$are_deterministic_algorithms_enabled_get_cublas_allow_tf32-is_deterministic_algorithms_warn_only_enabled _autograd_saved_tensors_hooks_is_enabledrrrget_global_state_keyrs  '')  CCE CCE ! 224 '') ;;= ::<  rct]tRt^tRRlt]RRl4tRRltRRltR R lt R R lt R Rlt Rt R#)BackwardBfsArgIterc V^8dQhRRRR/#)roriginrjrrr)rs"rrBackwardBfsArgIter.__annotate__s22t22rc 0Wn\4VnR#r)_curr_queue)rSrs&&rrIBackwardBfsArgIter.__init__s!' */' rc V^8dQhRRRR/#)rrrjrrr)rs"rrrst 2rc \V4pVPV4VP4'g \R4hV#)z/expected origin node to be popped from iterator)r add_childrennextAssertionError)rits& rcreateBackwardBfsArgIter.creates6  ' wwyy !RS S rcV^8dQhRR/#rrz Node | Noner)rs"rrrskrc VPpVP'g RVnV#VPP4VnV#r)rrpopleft)rSrets& rrBackwardBfsArgIter.nexts=ii{{{DI  ++-DI rcV^8dQhRR/#rr)rs"rrrskrc VP#r)rrSs&rpeekBackwardBfsArgIter.peeks yyrc V^8dQhRRRR/#rrqrjrrr)rs"rrrs"""$"rc \V/4pVF,p\V\4'gKVPV4K. R#r)r rirj_append)rSrqrrfs&& rrBackwardBfsArgIter.add_childrens1)$3 C#t$$ S!rc V^8dQhRRRR/#)rrfrjrrr)rs"rrrs$$4$D$rc hVPf WnR#VPPV4R#r)rrr))rSrfs&&rrBackwardBfsArgIter._appends# 99 I KK  s #rcV^8dQhRR/#rrrr)rs"rrrsKKKrc <RVP RVP R2#)zBackwardBfsArgIter(cur=z, queue=)rrrs&r__str__BackwardBfsArgIter.__str__s( 8DKK=JJrrN) r3r4r5r6rI staticmethodrrrrrrr7rrrrrs;2" $ KKrrcz]tRt^tRtRRltRRltRRltRR ltR R lt R R lt RRlt RRlt Rt R#)GraphRegionTrackera GraphRegionTracker tracks each node added to the output graph and generates a key based on the source location, instruction pointer, input shapes, and global state at the time the node is inserted into the graph. Nodes with the same key are grouped together in a list of identical nodes (the value of node_to_duplicates). hash_to_duplicates: Dict[str, IdenticalNodes] - A dictionary mapping the key to a list of identical nodes node_to_duplicates: Dict[Node, IdenticalNodes] - A dictionary mapping a node to the list of identical nodes it belongs to input_pickler: InputPickler - An instance of InputPickler used to generate a node hash cV^8dQhRR/#r;r)rs"rrGraphRegionTracker.__annotate__s,,$,rc h\\4Vn/Vn/Vn\ 4VnR#r)rlisthash_to_duplicatesnode_to_duplicatesnode_to_mutated_arg_positionsr9 input_picklerrs&rrIGraphRegionTracker.__init__s)=H=N>@JL*)^rc ,V^8dQhRRRRRRRRR R/#) rfilenamerlinenorninstruction_pointerz int | Nonerqrjrr)rs"rrrs4 : : :%( :?I :QU :  :rc ^RIHp\4VVV\V43pV!VPP V44#)r%) sha256_hash)r(rrrrrb)rSrrrrqrrts&&&&& r _hash_nodeGraphRegionTracker._hash_nodesE : ! "    D !  4--33C899rc$V^8dQhRRRRRR/#)rn0rjn1rboolr)rs"rrrs!   $ 4 rc WP9;'dHW P9;'d2VPV,VPV,J;'dWJ#r)r)rSrrs&&&r _is_identical GraphRegionTracker._is_identicalsb )) )  ---  ''+t/F/Fr/JJ    rc$V^8dQhRRRRRR/#)rtxrrqrjrrr)rs"rrrs'KK6KdKtKrc bW P9doVPVPVPPVP VP V4,pVPV4W0PV&R#R# \d"p\PRY$4Rp?R#Rp?ii;i)z The main entry point for tracking a node. This function will hash the node argument and group nodes with the same hash together. It updates the hash_to_duplicates and node_to_duplicates dictionaries to track the new node. z(Unable to hash node %s with exception %sN) rrrf_code co_filenamerrr)r2logr)rSrrq duplicatesras&&& r track_nodeGraphRegionTracker.track_nodes K333!44OO --ryy":P:PRV !!$'0:''-4! K II@$ J J KsA@ >P@ rc H\VP4UUu/uFwr#W2bK ppp.p\V4pVPP 4Fp\ V4^8gK.p\ Pp VF/pW49gK \WV,4p VPV.4K1 \ V4^8gKnVPW34K VPRR7VU U u.uFwrV NK p p p \4p V FVp\VV VVP4V^,p\V4UUu/uFwppWV,bK ppp\VV4KX V Uu.uFp\ V^,4^8gKVNK up#uuppiuup p iuuppiuupi)a This function is responsible for extracting the largest regions of identical nodes from the given graph. **Note**: This function assumes the nodes that have been tracked with track_node are in the provided graph argument. The algorithm proceeds as follows: The nodes tracked via track_node above are organized into region groups. The initial region groups look like this: [[IdenticalNode1], [IdenticalNode2], [IdenticalNode3]] and each sublist is called a region. For each region group (starting at the topologically latest region group), the inner regions are gradually expanded one node at time from the flattened args and kwargs of the node in each region provided that for all regions in the group, the nodes being added are also identical (ie have the same key computed by track_node). This is checked by verifying that the two nodes have the same identical node list in node_to_duplicates. cV^,)#)r)rgs&rrA:GraphRegionTracker.get_identical_regions..\s RUFrrs)rnodes _populate_recursive_ancestor_maprvaluesrmathinfminr)sortsetfully_expand_region_grouprr)rSrrrqtopological_rankingregion_groups_with_ranknode_to_recursive_ancestorsgroup region_groupmin_rankr r region_groups seen_nodesrindexnrs&& rget_identical_regions(GraphRegionTracker.get_identical_regions5s7@ 6LM6L71tw6LM"$&Fu&M#,,335E5zA~! 88!D2#&xT1J#K$++TF3 " |$q(+22L3KL6 $$):$;)@A)@)@ A !$ )L %+""  &aJ>G >S>S(%1-->S  "- >*".; -:\c,q/>RUV>VLL]  cN6B"  s F , F:F&FFcV^8dQhRR/#rr)rs"rrrwsAAArc <RVP RVP R2#)z&GraphRegionTracker(hash_to_duplicates=z, node_to_duplicates=r)rrrs&rrGraphRegionTracker.__str__ws.78O8O7PPefjf}f}e~~A Ar)rrrrN)r3r4r5r6__doc__rIrrrrrrrr7rrrrrs;, : K(M,M@ DAArrcJ]tRtRtRtRRltRRltRRltR R ltR t R #) RegionWrapperi{zPHolds state for regions e.g. ancestors and new candidate nodes for considerationc$V^8dQhRRRRRR/#)rrRegionrdict[Node, set[Node]]rrr)rs"rrRegionWrapper.__annotate__~s$   ;P   rc \V4^8wd \R4hV^,pW n\P V4Vn\ V.4Vn\W#,4Vn Wn R#)r &all regions should start with one nodeN) rrrrriterr nodes_uniquer ancestorsr)rSrrrqs&&& rrIRegionWrapper.__init__~sa v;!  !IJ Jay+F(&--d3 &v.8>? rcV^8dQhRR/#rr)rs"rrr*s   rc 6VPP4#r)r-rrs&rnext_candidateRegionWrapper.next_candidatesyy~~rc V^8dQhRRRR/#)rrqrjrrr)rs"rrr*src VPUu.uFq"VP9gKVNK ppVFpW P9gKR# R#uupi)TF)usersr.r/)rSrquserexternal_userss&& rwill_inclusion_create_cycle)RegionWrapper.will_inclusion_create_cyclesI+/::W:4TEVEV9V$$:W"D~~%# Xs A A c V^8dQhRRRR/#rr)rs"rrr*sFFFFrc VPPV4VPPV4VPP V4VP PVPV,4R#r) r.rrr)r-rr/rMr)rSrqs&&rrRegionWrapper.addsW d# 4  t$ d>>tDEr)r/r-rr.rN) r3r4r5r6r$rIr3r:rr7rrrr&r&{s Z  FFrr&c ,V^8dQhRRRRRRRRR R /#) rrz list[Region]rz set[Node]rr)is_identical_fnzCallable[[Node, Node], bool]rrr)rs"rrrsHPDPD PDPD"7PD2 PD  PDrc \R4\RV4\;QJdRV4F 'dK RM RM !RV44'g \R4hVUu.uFp\WB4NK pp\\ ,!4pV^,P 4pVfR#V'EdV^,PV4'*pVP4VPV4VR,Fp V P 4p \R4\R W4V 'd V'gRpK>\R W94\R V!W44TW9;'daW9;'dUV PR 8g;'d>V PR 8g;'d'V!W4;'dV PV 4'*,pVPV 4\RV 24\R4K V'd\V4\V48wd \R4h\WV4FkwrV PV 4\RV 4\RV P\V PP!444VPV 4Km V^,P 4pEKVFpVP#4K \RV4\R4R#uupi)z2--------------------------------------------------zexpanding new region group: %sc3>"TFp\V4^8HxK R#5i)r N)r)rvrs& rrw,fully_expand_region_group..s6gFs6{agsFTr,N:r NNz--------------------z'considering candidate: %s, cur_node: %sz*candidate in previously claimed nodes?: %szis_identical: %s placeholderget_attrzadd_to_all_regions: z7Number of nodes to add must equal the number of regionszadding %s's childrenz%s %szend expand new region group: %s)r!allrr&r rjr3r:clearroprzipr rrzr{reverse) rrrr@rregion_wrappers nodes_to_add current_nodeadd_to_all_regionsregion_wrapper candidaterqs &&&& rrrs BC .8 36g63336g6 6 6EFFIPIPv f:d#%L"1%446L ,!0!3!O!O "   &-b11N&557I , - 99 $6%*" U  (/)*R S +NN1NNLLM1NNLLJ.NN$I< NN 'BB9MM      Y ' ,-?,@A B , -92< ?#s<'88$M),O(J$""4(0$7'499d4;;3D3D3F.GHt$ )K 'q)88: /9 BCIs!K+c V^8dQhRRRR/#)rrrrr)r)rs"rrrs ' 'N '?T 'rc0/pVPFp\4W&K VPFdp\V/4pVFOp\V\4'gKW,P W,4W,P V4KQ Kf V#r)r rr rirjrMr)rrrqrrfs& rr r s9; ,/E#) (r2C#t$$+188/4,155c:  '&r)Ar$ __future__rrJrEloggingrr|pickle collectionsrr dataclassesrtypingrrrtorch._loggingrNtorch.fxtorch._subclasses.fake_tensorr torch.utils._ordered_setr torch.utils._pytreer graph_utilsr r collections.abcrsymbolic_convertrfxrjrr(IdenticalNodesr,rrndtyper getLoggerr3r_logginggetArtifactLoggerrr!r0 Exceptionr2Picklerr9rorrrrrr&rr rrrrisX #   *..4/,. CL(; xx}} dd $* $* KK   !nn66 & *    %6>>%@H 8 ,'K'KT_A_ADFFBPDf 'r