CCjSrSSKrSSKrSSKrSSKrSSKrSSKJr SSKJ r J r J r J r J r SSKrSSKrSSKJr SSKJr SSKJrJrJrJrJr SSKJr SS KJrJr SS KJ r SS K!J"r" SS K#J$r$ SS K%J&r& SSK'J(r(J)r) SSK*J+r+ SSK,J-r-J.r. SSK/J0r0 SSK1J2r2J3r3J4r4J5r5 SSK6J7r7 SSK8J9r9 SSK:J;r; SSK SSK?J@r@ \ (a SSKAJBrB SSK%JCrC \R"\E5rF\RR\ES5rI"SS\R5rK"SS\ 5rLS\ S \ \"\0-4S!jrMS\ S \KS-4S"jrNS#S$S%\ S&\O\P\ 4S \L4S'jrQS%\ S \L4S(jrRS#S$S%\ S&\O\P\ 4S \L4S)jrSS*\ S \ 4S+jrTS,\LS \U\;4S-jrV\"S.S/55rW\"S0S155rX\"S2S355rY\"S4S555rZ\W\X-\Y-\Z-r[S6\ S \RS-4S7jr]S#S$S8\@\5S \^4S9jr_S]S#S$S:\ S,\LS;Sjjr`S#S$S,\LS8\@\5S \3S-4S?jraS@\U\5S-SA\U\5S-S \O\5\544SBjrbS#S$SCSDS,\LSE\4S \^4 SFjrcS#S$S6\ S \^4SGjrdS#S$S6\ S,\LS \4S-4SHjreS^S#S$S6\ S,\LSI\PSJ\RRrSK\7S-SL\g\ SM4S:\ SN\ SO\hSCSPSQ\hS-S S4SRjjriS#S$SS\ S%STS&\O\P\ 4S \h4 SUjrjS#S$S6\ SQ\hS \4S-4SVjrkS#S$S,\LSW\4S \ 4SXjrlS#S$SL\g\ SM4SY\U\g\K\ 4S \U\[4SZjrm"S[S\\&5rng)_z This module contains the InvokeSubgraphHigherOrderVariable class and its supporting helpers for subgraph reuse (auto-cache) in Dynamo's invoke_subgraph higher-order operator. N) dataclass)Anycast NamedTuple TYPE_CHECKING TypeGuard)graph_break_hints) unimplemented)extract_tensor_metadataGUARD_VALUE_DISPATCHGuardCheckSpec SKIP_GUARDUnsupportedGuardCheckSpec)SyntheticLocalSource) _make_inlinedunpack_iterable)VariableTracker)ConstantVariable)UserFunctionVariable)WrapHigherOrderVariable) ListVariable TupleVariable)UnspecializedNNModuleVariable)SymNodeVariableTensorVariable)UserDefinedObjectVariable)GuardInvokeSubgraphReuseConditionInvokeSubgraphReuseEntrySource)NestedCompileRegionOptions) GraphModule)Proxy)_pytree) OrderedSet)InstructionTranslatorBase)SubgraphTracingInfohierarchical_compilec$\rSrSrSrSrSrSrSrg)InputTagtensorsymnodeconstantobjectN) __name__ __module__ __qualname____firstlineno__TENSORSYMNODECONSTANTOBJECT__static_attributes__r0ڵC:\Users\johnn\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.13_qbz5n2kfra8p0\LocalCache\local-packages\Python313\site-packages\torch/_dynamo/variables/invoke_subgraph.pyr*r*s FGH Fr:r*c|\rSrSr%\\\\4\S'\\ S-\S'Sr \ \S'Sr \ RS-\S'Srg) InputFingerprintflat_vtsN arg_sourcesF has_unknowntreespecr0)r1r2r3r4listtupler*r__annotations__r rAboolrBpytreeTreeSpecr9r0r:r;r=r=sC5?2344ftm$$K'+Hfoo$+r:r=vtreturnc[U[5(ag[U[5=(a$ [UR[R 5$)zWhether a leaf VT is compared by value, via ``vt.value``. Enum members are UserDefinedObjectVariable (not ConstantVariable) in Dynamo, but they are immutable singletons, so value comparison is as sound as it is for ConstantVariable. T) isinstancerrvalueenumEnumrIs r;is_constant_likerQs7"&'' b3 4 XBHHdii9XXr:c<[U[5(a[R$[U[5(a[R $[ U5(a[R$[U[5(aURb[R$g)z5Return the tag for a leaf VT, or None if unsupported.N) rLrr*r5rr6rQr7rsourcer8rPs r; classify_vtrTso"n%% B ( ( "     B1 2 2ryy7L  r:txr& fn_args_vtkwargscU(d-SnUHn[U5bMSn O U(a [U5$[XU5$)aBuild an InputFingerprint by flattening (args, kwargs) via pytree. Flattens the argument structure into leaf VTs, classifying each leaf as tensor/symnode/constant/module. Also records the TreeSpec so that cache lookups can verify structural equivalence. Fast path: when kwargs is empty and all args are already leaf VTs (tensor/symnode/constant/module), skip the pytree flatten entirely. TF)rTbuild_fingerprint_fastbuild_fingerprint_with_pytree)rUrVrWall_leafrIs r;build_input_fingerprintr\sE B2&  )*5 5 ( @@r:c/n/nUHan[U5nUc"[S[U5RS35eUR XC45 UR [ USS55 Mc [ X5$)zCBuild fingerprint for the common case of flat leaf args, no kwargs.Nzclassify_vt returned None for z in fast pathrS)rTAssertionErrortyper1appendgetattrr=)rVr?r@rItags r;rYrYs~79H')K"o ; 0b1B1B0C=Q   " 72x67 H 22r:c^^^^^ ^ SSKJn /m/mSm S[SS4UUU 4SjjmS[S[R4UU U U4S jjm UR T[ U5U45nT "U5n[TTT U5$) aBuild fingerprint via pytree flatten for nested/kwargs cases. Recurses over the pytree structure natively (untraced), inlining/tracing only each container node's own ``flatten_fn`` rather than the full recursive tree_flatten dispatch around it. This is safe because node-type classification only depends on ``type()``, never on tensor values. Note: skipping the traced registry dispatch also means we no longer install guards on the pytree registry itself, for the plain-function flatten_fn case (the non-FunctionType fallback below still traces the full tree_flatten and so still installs them). That's fine either way: this fingerprint is built fresh from the live registry on every reuse-lookup (register_pytree_node isn't traceable, so the registry can't change mid-trace), and reuse is separately gated on treespec/tag equality. So a registry change between compiles can only make a cached subgraph ineligible for reuse (has_unknown / treespec mismatch), never silently wrong. rSourcelessBuilderFrIrJNc>[U5nUcSmgTRX45 TR[USS55 g)NTrS)rTr`ra)rIrbr@r?rAs r;add_leaf/build_fingerprint_with_pytree..add_leaf)s;"o ;K OOSI &   wr8T: ;r:node_vtc*>UR5n[R"U5(a[ R nU[R;aT "U5 [R"5$[RURn[U[R5(dW[T[T[R5"U55up4[TU5H nT "U5 M UR5$[T[TU5"U55upgUR5n[TU5V s/sH n T "U 5PM n n [R "XU 5$![a Sm [R"5s$f=fs sn f)NT) python_typeNotImplementedErrorrG treespec_leafis_namedtuple_class collections namedtupleSUPPORTED_NODES flatten_fnrLtypes FunctionTyperr tree_flattenas_python_constantrH)ri node_typerr leaves_vt treespec_vtleaf_vt children_vt context_vtcontextchild child_specsrgflattenrArUs r;r.build_fingerprint_with_pytree..flatten3s\ *++-I  % %i 0 0#..I F22 2 W '') )++I6AA *e&8&899&5M"f&9&9:7C& "I+2y9!:113 3"1  b*-g6#  //13B2{3ST3S%wu~3S Ty;???# *K'') ) *<UsE)>F)!F  F )torch._dynamo.variables.builderrerrGrHcreaterCr=) rUrVrWre container_vtrBrgr@r?rrAs ` @@@@@r;rZrZ s.B79H')KK<_<<<#@#@V__#@#@J%++Bj1A60JKL|$H Hk; IIr:sym_numc.URR$)a2Key for matching a symbolic input against a cached one. Compares the symbolic expression rather than the SymInt object. Each tensor holds its own SymInt objects, so two arguments carrying the same symbol (e.g. the atom count threaded through successive layers) are distinct objects. Equal expressions mean the same value, which is what reuse needs; distinct symbols still have distinct expressions. ``expr`` rather than ``_expr`` on purpose: it has the ShapeEnv's replacements applied, so a symbol that was specialized keys on the value it was specialized to. )nodeexpr)rs r; sym_num_keyr^s <<  r: fingerprintc0[5n/nURHyup4U[R[R4;dM+UR 5nUR U;dMMURUR 5 URU5 M{ U$)z8Collect deduplicated proxies from tensor/symnode leaves.) setr?r*r5r6as_proxyraddr`)rseen flat_proxiesrbrIproxys r;get_flat_proxiesrnst"uD "L'' 8??H$4$45 5KKMEzz%$##E* ( r:c$\rSrSr%Sr\\S'Srg) LiftedUserArgi{zVLifted arg that came from a user argument (intermediate activation or explicit input).indexr0N)r1r2r3r4__doc__intrEr9r0r:r;rr{s ` Jr:rc$\rSrSr%Sr\\S'Srg)LiftedCapturedSourceizRLifted arg that is a captured variable (e.g. a weight or parameter) with a Source.rSr0Nr1r2r3r4rrrEr9r0r:r;rrs \ Kr:rcR\rSrSr%Sr\\S'\\S4\S'\\S4S-\S'Srg) LiftedSyntheticObjectizCLifted arg that is a TorchScriptObject with a SyntheticLocalSource.ctor_fn. ctor_argsNctor_arg_sourcesr0) r1r2r3r4rrrErDr9r0r:r;rrs+M LS#XCHo,,r:rc$\rSrSr%Sr\\S'Srg)LiftedBoundSymboliaLifted arg that is a SymInt already bound as a graph input. SymInt graph inputs are created during tensor wrapping (not through VariableBuilder.wrap_symint), so they aren't registered in unspec_variable_map or variable_tracker_cache. Using LiftedCapturedSource for these would resolve the source to a concrete Python int via source.get_value() instead of reusing the existing symbolic proxy. rr0Nrr0r:r;rrs Ir:rfn_varc[U[5(aUR5R$[U[5(a*UR R RR$gN)rLr get_function__code__rrMforward__func__)rs r; get_fn_codersV&.//""$--- F9 : : LL ) ) 2 2  r:traced_sourcescURRRU-nU(a[R SU5 gg)aCheck if any source accessed by the subgraph has been mutated. SideEffects.mutated_sources records the exact AttrSource for every store_attr call. A simple set intersection with traced_sources tells us whether any source the subgraph read was later written to. z.subgraph_reuse: mutated sources detected -- %sTF)output side_effectsmutated_sourceshc_logdebug)rUroverlaps r;has_mutated_varsrs;ii$$44~EG <   r:body_r tracing_infor'has_reuse_hash_fncU(dnURbISSR[R"UR55-n[R SU5 gU(a[ X5(ag[U[5(aO[U[[45(aeURVs/sH/n[U[5(aM[U5RPM1 nnU(a[R SU5 gO*[R S[U5R5 gUR(a[R S5 ggs snf) u?Best-effort check for whether a traced subgraph result can be reused. It is possible that a subgraph is morally reusable but does not fall into the limited support that Dynamo has today. Current limitations: - The subgraph must not have side effects. - No sourceful variable accessed by the subgraph may have been mutated, because guards are snapshotted on source values at trace time — if the underlying object changed since then, the cached guards would silently evaluate against stale values. - Output must be a single tensor, or a tuple/list of plain tensors. - All flattened inputs must be one of: tensor, symnode, constant (including enum members), or a sourceful user-defined object (nn.Modules included) — for sourceless or other input types we rely on the treespec and tags for structural matching, so only types with well-defined comparison semantics are supported. When ``has_reuse_hash_fn`` is True, side-effect and mutation checks are skipped because the hash key replaces guards — there are no guards to go stale from mutations.  z;subgraph_reuse: not eligible -- subgraph has side effects%sFzDsubgraph_reuse: not eligible -- output contains non-tensor types: %szJsubgraph_reuse: not eligible -- output type %s is not tensor or tuple/listz:subgraph_reuse: not eligible -- unsupported input VT typesT)side_effect_stackjoin traceback format_listrrrrLrrritemsr_r1rA) rUrrrrr stack_msgitem non_tensors r;is_reuse_eligiblers-8   ) ) 5rww%%l&D&DE I LLM  .rBB&.)) F]L9 : :  $dN3 DJ  $   LLV    X L ! !  H  1 s 5EEc SSKJn /nURGH)upVU[R:Xa[ U[ 5(d![S[U5R35eURRRRSS5nUc[RS5 gUR![R[#U545 MU[R$:Xam[ U[&5(d![S[U5R35eUR![R$[)UR*545 GMGU[R,:Xat[/U5(d![S[U5R35eUR![R,[UR05UR0445 GMU[R2:Xa$UR![R2S45 GM[S US [U5RS 35e [5U5nUR7S UR855 [55n UH7n U R7UR:R<R?U 55 M9 /n U Hn U R@n U RC5n [DR"U 5nU[FLaM@Ub[ U[H5(a[KS U SU RLS35eUR:ROU 5n[S[TU5nURWX5nU R!XUU 45 M [RS[YU 55 U"UU URZUS9$![Pa [KSU RLS U S35Sef=f)aBuild an InvokeSubgraphReuseCondition from a traced subgraph. A reuse condition is a mix of two kinds of checks: 1. **Input tag checks** (from flat_vts): For each flattened leaf VT, we record its tag (InputTag.TENSOR/SYMNODE/CONSTANT/OBJECT) and metadata (e.g. tensor shape/stride/dtype/device/requires_grad). At lookup time, the treespec ensures structural equivalence, and then we compare tags and metadata leaf-by-leaf. 2. **Guard checks** (from traced_sources): During the subgraph trace, every source accessed via VariableBuilder is recorded. We look up all guards installed on those sources (and on the arg_sources) to build the set of guards that must be re-evaluated on cache hit. This is more robust than guard diffing because it catches guards that were already installed before the subgraph trace began. Raise if any guard type is unsupported, as a feedback for compiler developers to support that guard type. r)r,expected TensorVariable for TENSOR tag, got example_valueNzKsubgraph_reuse: cannot build condition -- tensor input has no example_value.expected SymNodeVariable for SYMNODE tag, got 0expected constant-like VT for CONSTANT tag, got zUnexpected input tag 'z' for z/ -- is_reuse_eligible should have rejected thisc3.# UH ocMUv M g7frr0).0ss r; (build_reuse_condition..NsK"9Qqq"9s z(subgraph_reuse: unsupported guard type 'z ' on source ''z*subgraph_reuse: failed to resolve source 'z guardzNumber of guards %s) input_checksguardsrBr). torch._guardsrr?r*r5rLrr^r_r1rrmetagetrrr`r r6rrrr7rQrMr8rupdater@rrget_guards_for_sourceoriginating_sourcecreate_fn_namer rr RuntimeErrornameresolve_source_value Exceptionrr get_metadata_fnlenrB)rUrrrrrbrIexample all_sourcesall_relevant_guardsrS guard_tuplesguardtype_strhandlerrMexpecteds r;build_reuse_conditionr sc2;24L'' (// !b.11$B48CTCTBUVhhmm((,,_dCG a   2I'2R S T H$$ $b/22$DT"XEVEVDWX   !1!1;rzz3J K L H%% %#B''$FtBxGXGXFYZ   !2!2T"((^RXX4N O P HOO #   $ 7 8 (VDH4E4E3FG>? ?(Ln%KK+"9"9KK&)e""299#3#3#I#I&#QRHJL$))'')&**84 j  ?j2KLL:8*MRXR]R]Q^^_`  II226:E~w/**58Vh>?7%: LL&L(9: '!%%%  ! [RS[UR5[UR 55 g[ [URUR 55GHunuupVupxXW:wa[RSUUU5 gU[R:Xa[U[5(d![S[U5R35eURR R"R%SS5n U c[RSU5 g['U 5n X:wa[RS U5 gMU[R(:Xav[U[*5(d![S [U5R35e[-UR.5U:wa$[RS UUUR.5 gGMvU[R0:XdGM[3U5(d![S [U5R35e[5[6[[84U5up[UR:5U La-[RS UU [UR:55 gUR:U :wdGM5U[UR<5:aUR<UOSn U[UR<5:aUR<UOSnU bUbGM[RSUUUR:5 g [?UR<UR<5mS[@S[@4U4SjjmT(a [CU4SjURD55nO URDn[GX5(agT(dgURHRJRLURHRJRNS.n0n0nURPHunnnnURST5nUU:XaM!URUUUU5nURcUU5(aMN[RSURY5UURZUUUR\(a)SR_UR\Ra55OS5 g g![Vao [RSURY5UURZUR\(a)SR_UR\Ra55OS5  gf=f)ulCheck if a cached subgraph can be reused for the current call. Three-phase check: (1) Verify that intermediates (tensor metadata, symnode types, constant values) match the cached input_checks — these are lightweight structural comparisons that don't require source resolution. (2) Check for mutations on the remapped traced_sources — if any source the subgraph read has been mutated since the original trace, the cached guards would evaluate against stale values. (3) Build a source replacement mapping (old sources → new sources) and re-evaluate the snapshotted guards under the new sources. Nz1subgraph_reuse: reuse failed -- treespec mismatchFzMsubgraph_reuse: reuse failed -- input count mismatch: cached %d vs current %dzRsubgraph_reuse: reuse failed -- input %d tag mismatch: cached '%s' vs current '%s'rrzDsubgraph_reuse: reuse failed -- input %d tensor has no example_valuezAsubgraph_reuse: reuse failed -- input %d tensor metadata mismatchrzVsubgraph_reuse: reuse failed -- input %d symnode mismatch: cached '%s' vs current '%s'rz\subgraph_reuse: reuse failed -- input %d constant type mismatch: cached '%s' vs current '%s'zqsubgraph_reuse: reuse failed -- input %d constant mismatch with no source to replace: cached '%s' vs current '%s'rrJc&>TRX5$rrrsource_replacements r;replacement_fn#is_reusable..replacement_fn s!%%a++r:c3D># UHoRT5v M g7fr)clone)rrrs r;ris_reusable..sX?W!ggn55?Ws TGLzsubgraph_reuse: reuse failed -- cannot resolve source guard type: %s guard source: %s guard source name: %s user stack: %srz zsubgraph_reuse: reuse failed -- guard type: %s guard source: %s guard source name: %s expected: %s got: %s user stack: %s)2rBrrrrr? enumeraterr*r5rLrr^r_r1rrrrr r6rrrr7rQrrDrrMr@rr r%rrrroot_tx f_globalsf_localsrr get_valuerrr user_stackrformateval_fn)rUrrri cached_tag cached_valcur_tagcur_vtrcur_meta cached_type cached_value cached_srcnew_srcremappedresolve_globalsresolve_locals resolve_cacherSrrr new_sourcerMrrs @@r; is_reusablers&%+*>*>)BTBT*T ?  9 ! !"c+*>*>&?? [  && '  $$ % >*j8 lNN  98,, ,#F++$FtF|G\G\F]^)-U49-=z(J %KFLL!4 <& |||+ 3|7788!,,Q/3{6677 ++A. %LLQ" !c=f2  +"9"9,&,V,Xy?W?WXX++%%  YY   ( ( YY   ' ''O&(N')M,5,<,<(5\\.1     ((.-XE"uh// LL$$$&##((//12! "W-=Z I  LL$ $$&##((//12!  s>TA5VVcSSKJn URRRR [ RR5n[X25(dg[U5nUSL=(a XCR;$)z?Cheap check: does the cache have any entries for this function?rInvokeSubgraphCacheFN) rrrtracing_contexthop_dispatch_set_cache get_cachetorch_higher_order_opsinvoke_subgraphrLrsubgraph_reuse_cache)rUrrinvoke_subgraph_cachefn_codes r;has_reuse_entriesrSsk 2II55LLVV // + A A&!G $  X7.X.X#XXr:c0^^SSKJn TRRRR [ RR5n[XC5(dg[U5nUcgSSS[S[4UU4SjjnURXV5$)NrrcondrentryrJc >[TUTU5$r)r)r!r"rrUs r; evaluator#find_reuse_match..evaluatorvs2t[%88r:)rrrrrrrrrrLrrrFfind_reuse_entry)rUrrrrrr$s` ` r;find_reuse_matchr'cs 2II55LLVV // + A A&!G 9,95M9 99 ! 1 1' EEr: body_name body_gmodconfigp_args.rmax_reuse_entriesz#InvokeSubgraphReuseCondition | Nonehash_keyc 0^SSKJn U SLU SL:wd [S5eURRR R [RR5n [X5(dg[U5nUcg[XUR5n[U[5n/m[R "U4SjU5 [#T5nUVs/sHan[U[R$5(dM$UR&UR)5UR*UR,UR.4PMc nn[1UUUUUUUR2US9nU bU R5XUU 5 gU c [S5eU R7XUU 5 gs snf)aSave a traced subgraph into the reuse cache for future cache hits. Builds an InvokeSubgraphReuseEntry with the freevar mapping (how each lifted arg maps back to user inputs or captured variables), output metadata, and arg sources. On a future cache hit, stamp_out_subgraph uses this entry to emit a new invoke_subgraph call without re-tracing. Exactly one of ``condition`` or ``hash_key`` must be provided. ``condition`` stores the entry in the guard-based cache (linear scan); ``hash_key`` stores it in the hash-key cache (O(1) lookup). rrNz5Exactly one of condition or hash_key must be providedc~>UR5(d[U[5(aTRU5$S$r) is_tensorrLrr`)rIuser_output_vtss r;"save_reuse_entry..s7 <<>>ZO<<#))"- r:)r(r)r*subgraph_input_mappingsingle_tensor_outputoutput_metadatar@num_user_outputsz0hash_key must not be None when condition is None)rrr^rrrrrrrrLrbuild_subgraph_input_mappingr?rrvisitrTensorshapestridedtypedevice requires_gradrr@add_reuse_entryadd_reuse_entry_by_key)rUrrr(r)r*r+rrr,rr-rrrr4r5r7tr6r"r1s @r;save_reuse_entryrC~s22 $ H$4 5TUUII55LLVV // + A A&!G9 K((&fn= .0O    ?+A a & B!((*aggqxxA %51' ++) E-- '8    !ST T44 u&7 5s #F=AF reuse_hash_fnlist[VariableTracker]cSSKJn URRRR 5 [ X5"U0UD6nSSS5 [W[5(a[UR[5(d[SU35eUR$!Uan[SU35UeSnAff=f!,(df  Ny=f)uTrace the user's reuse_hash_fn to get a constant integer hash key. Guards installed during the hash function tracing are skipped — the hash key itself is the reuse condition, not the guards. r) UnsupportedzAreuse_hash_fn must be fully traceable without graph breaks. Got: Nz2reuse_hash_fn must return a constant integer, got ) torch._dynamo.excrGrrguards_contextskip_guard_installrrrLrrMr)rUrDrVrWrGresultes r;trace_reuse_hash_fnrMs.  " " 1 1 D D F "25zLVLF G f. / /z&,,PS7T7T@ I   << STUSVW   G Fs(B<BB9%B44B99B<< C cSSKJn URRRR [ RR5n[XC5(dg[U5nUcgURXR5$)Nrr) rrrrrrrrrrLrfind_reuse_entry_by_key)rUrr-rrrs r;rOrOsk 2II55LLVV // + A A&!G 8 8 KKr:cachedc^ SSKJn SSKJnJn [ U5nUR n[UR U5m /nURRRURRRS.n 0n 0n URGHXn [U [5(aURXlR 5 M8[U ["5(aSSKJn URR(R*U R,n[X5(a3U"5nXRR(R*U R,'URU5 M[U [.5(aU R0nU R2nU(aT (a~/n/n[5UU5HTunnUb)UR7U 4Sj5nUR9XU 5nUnURU5 URU5 MV [;U5n[;U5nURR=U R>UU5nURURA55 GM[U [B5(dGMU RDnT (aUR7U 4Sj5nUR9XU 5nU"UU5"U5nURURA55 GM[ URFc [IS 5eURF [;S URJ55nSSS5 U"XRL5nUURL/UQ7nU"U[NRPRR[;U50WURT5nURV(aMURXn[US[Z5(d$[IS []US5R^35eUS$URXnUR`nUS:a3U[cU5:a$S S K2J3n URiU[;USU55$U$!,(df  GN=f)a.Emit a new invoke_subgraph call by stamping out a cached subgraph. Sources in the cached entry are parameterized: they refer to the original call's sources and must be rewritten to the current call's sources via source replacement before we can look up or create the corresponding graph placeholders. r)VariableBuilder)add_call_function make_attrr) LazyProxyNc&>TRX5$rrrs r;r2$stamp_out_subgraph..=s:L:P:PQR:Vr:c&>TRX5$rrrs r;r2rWMs8J8N8Nq8Tr:z4tx.fake_mode must not be None for stamp_out_subgraphc 3Z# UH!upp4n[R"UUUUUS9v M# g7f))r=r>r?N)r empty_strided)rr;r<r=r>req_grads r;r%stamp_out_subgraph..Xs>  ;Q6uh   &  ;Qs)+zExpected tensor output but got rd)5rrR(torch._dynamo.variables.higher_order_opsrSrTrr@rrrrrr4rLrr`rrtorch._dynamo.output_graphrUcurrent_tracer bound_symbolsrrrrrrrrDsynthetic_graph_inputrrrrS fake_moder^r6r(rrrr*r5rrr_r1r7rbuilderrer)!rUrrPrRrSrTrrnew_lifted_argsrrrsubgraph_inputrUrrr new_ctor_argsnew_ctor_arg_sourcesvalarg_srcr rIrrMr body_noder+ flat_variablernrers! @r;stamp_out_subgraphrn s@U#K0L!--O1&2D2DoVOYY   ( ( YY   ' ''O&(N')M!77 nm 4 4  " "<0D0D#E F (9 : : <II,,::>;N;NOE%++NS ((66~7J7JK  " "5 ) (= > >&00I->> $6 " ')$$' 3C$DLC*")--0V"W%//+]#*!((-(//8%E"-0 #()=#> 00&& 3CB  " "2;;= 1 (< = ='..J!'--.TU ((-XE Z07B  " "2;;= 1Q8V ||STT   ;A:P:P    "../I)) v*F GQ"R "!r:c^\rSrSrSrSrSrSrSrSr SSS\ SS S \ \ \ 4S \ S \ S \ 4U4SjjrSSSS S \ \ \ 4S \ 4SjrSrU=r$)!InvokeSubgraphHigherOrderVariableiz&torch.ops.higher_order.invoke_subgraphFTrUr&fn_vtrVrErWr) attr_namerJc >[U[[45(d%[S[ U5S/[ R QS9 URRRR[RR5n[U[5(a5UR5RnUR5R n O[U[5(d![#S[%U5R 35eUR&R(R*RnUR&R(R n /n U(aUR-U5n Un [/U 5Hn XRR0;a[#SU S35eURR0U n UR2(d [#S5eSS KJn U"XXR25(dMU s $ [8TU]uXX4US 5n[<R?S U UU [AU 5S -5 U(aURCX5 U$) NzIEncountered non user function variable during invoke_subgraph HOP tracingz;invoke_subgraph does not support non user function variablegb_typer} explanationhintsz,expected UnspecializedNNModuleVariable, got z submodule 'z' not found in nn_modulesz0tx.fake_mode must be set for subgraph comparisonr)are_same_graph_modulessubgraphzZ%s: Installing subgraph with identifier '%s', bringing total count for '%s' function to %sr])"rLrrr strr SUPPORTABLErrrrrrrrrr1r^r_rMrrget_dynamo_installed_submodulesreversed nn_modulesrcr^rsuper install_subgraph_in_output_graphrrradd_dynamo_installed_submodule)selfrUr~rVrWr)rrrfn_namepreviously_installed_submodules current_modsubmodule_name previous_modrr( __class__s r;rBInvokeSubgraphHigherOrderVariable.install_subgraph_in_output_graphs%"?AU!VWW cE Y6)556   II % % < < F F''77   e1 2 2((*33G((*33Ge%BCC$B4;CWCWBXYkk))22;;Gkk))22G*,' %EEgN ,$K#++J"K!)=)==(%n%55NO "yy33NC ||(J*; *)##L&G< z9j   h    / 01 4   ! ! @ @ Tr:argscSSKJn SSKJn USnUSSnSnSn Sn [ US5(a8UR 5n [ U SS5n[ U SS5n [ U S S5n URR(+n U (a|U byU"S 5 [XXs5n SSS5 [XW 5nUbL[RSU UUR 5 [#XU5nU"S5 [%XU5sSSS5 $OU (a}['X5(amU"S5 [#XU5n[)UUU5nSSS5 Wb@[RSUUR 5 U"S5 [%UWU5sSSS5 $UR*c [-S5eU"S5 UR/XXsUR*5unnnnnnnnSSS5 [1W5S:a"[3SSUSU3S/[4R6QS9 [9U[:5(aUWR<S'WSW/USSQ7nU (a[#XU5nU b>WR>n[AUWUUUS S9(d [CS5e[EUUUUWUUUWU W S9 OCWR>n[AUWUUU5(a#[GUUU5nUb[EUUUUWUUUWU US9 U"U[HRJRL[OU5UWWWUS9$![a [RS S S 9 ef=f!,(df  GN=f!,(df  GN=f!,(df  GN=f!,(df  GN=f!,(df  GN=f)Nr) dynamo_timed)*_call_function_with_auto_output_flatteningr]r __marked_compile_region_config__+__marked_compile_region_max_reuse_entries__'__marked_compile_region_reuse_hash_fn__zYFailed to extract nested_compile_region() config from InvokeSubgraphHigherOrderVariable. T)exc_infoinvoke_subgraph_reuse_hash_fnz?subgraph_reuse: hash key %d hit for '%s', reusing subgraph '%s'invoke_subgraph_reuse_stamp_outinvoke_subgraph_reuse_lookupz9subgraph_reuse: cache hit for '%s', reusing subgraph '%s'z_HOP_NAME must not be Noneinvoke_subgraph_tracez"invoke_subgraph: kwargs unexpectedzargs: z , kwargs: z3kwargs should have been flattened into lifted args.rnested_region_config)rzreuse_hash_fn was provided but the subgraph is not eligible for reuse. Check the logs with TORCH_LOGS='+hierarchical_compile' for details.)r-)r)r*)(torch._dynamo.utilsrr^rhasattrrrarlogwarningrexportrMrOrrr(r\rnrr' _HOP_NAMEr^create_wrapped_noderr r DYNAMO_BUGrLr!rrrrrCrrrrrD)rrUrrWrrrrVr*r,rDfnreuser-rPrmatchr+p_kwargsrrr)r(body_graph_output_vtsrrrs r;_call_function0InvokeSubgraphHigherOrderVariable._call_function&s 5 a!"X  6> * * ((* %GN$+Eq%!!(A4! II$$$ ].=>.r*U?-RBF! U$$  6bfM !"CD-bvFED"(44<=5bfM (>  OOO ""CD-b+uEED >> ! !=> > 1 2((ZX %3 x=1  < j9Q&11   f8 9 95;INN1 2 1I  ABZ  1"&IK(!-!!6#&!I !,("'%%""")-&/ :   # # 3 3 &M    !  U  o!  ?>ED>=ED 3 2sG7K> L!- L3#M1 M('M)> L! L03 M M M&) M8r0)r1r2r3r4r_ALLOW_FALLBACK_TO_EAGERsupports_input_mutationsupports_aliasingallow_side_effectsfilter_aliased_intermediatesrdictrr"rrr9 __classcell__)rs@r;r}r}s8I$"$( O 'OO, O S/)* O  OO Obn 'n &n S/)* n  n n r:r})NF)rNN)orrorNloggingrrs dataclassesrtypingrrrrrrtorch._higher_order_ops torch._dynamor rHr torch._dynamo.guardsr r r rrtorch._dynamo.sourcerrrrtorch._dynamo.variables.baser torch._dynamo.variables.constantr!torch._dynamo.variables.functionsrr^rtorch._dynamo.variables.listsrr!torch._dynamo.variables.nn_modulertorch._dynamo.variables.tensorrr$torch._dynamo.variables.user_definedrrrrrr 'torch._higher_order_ops.invoke_subgraphr!torch.fx.graph_moduler"torch.fx.proxyr# torch.utilsr$rGtorch.utils._ordered_setr%torch._dynamo.symbolic_convertr&r' getLoggerr1r_logginggetArtifactLoggerrrOr*r=rQrTrrr\rYrZrrCrrrrrLiftedArgOriginCodeTyperrFrrrrrrr'fxrDrrCrMrOrnr8r}r0r:r;rs   !BB ++6>8=BLEKJJ O- )/HL !  ) )(4J Kptyy,z, Y  Y";;< YCHtO$A#AA cNA A63s3/?3"OJ#OJOJ cNOJ OJd    "2 tE{        -- -     ((+@@CTT  5#v& 215# E#E E"E( E v&- E  E EPm#m!mv&m"D( m` &4-( &4-(  &&. I#I-I"I+ I  IX Y# Y  Y  Y F#F F"F$ FJ7;[ #[ [ "[  [ xx## [ ' - [  #s(O[  [ [ [ 5[ Dj[  [ |#( o% &   :L#L LL$ L$u#u!u %u upC"#C" #s(OC"5?234C" / C"LK (?K r: