7h86SSKJr SSKrSSKJr SSKJrJrJrJ r J r SSK r SSK J r Jr SSKJrJr SSKJr \ (aSSKJr \ R,R/\S 5r\ R,R/\S 5r\\\ \\ R:4r\\\/\4r\R@"S S 9"S S55r!\R@"S S 9"SS55r"\R@"S S 9"SS55r#S+Sjr$S,Sjr%S-Sjr&S.Sjr'S/Sjr(S0Sjr)S1Sjr*S2Sjr+S3Sjr,S3Sjr-S4Sjr.\R@"SS55r/S5S jr0S6S!jr1"S"S#\5r2S7S$jr3S8S%jr4\R@"S S 9"S&S'55r5\R@"S S 9"S(S)55r6S9S*jr7g):) annotationsN)Enum)AnyCallableOptional TYPE_CHECKINGUnion)countersget_metrics_context)GraphPartitionMap InputType) OrderedSet)Sequence perf_hintscudagraph_static_inputsT)frozenc$\rSrSr%SrS\S'Srg) FunctionIDz9Unique counter of a function wrapped in cudagraphify_implintidN__name__ __module__ __qualname____firstlineno____doc____annotations____static_attributes__rY/var/www/fran/franai/venv/lib/python3.13/site-packages/torch/_inductor/cudagraph_utils.pyrrs ? Gr!rcB\rSrSr%SrS\S'S\S'S\S'S\S 'S rg ) PlaceholderInfo#z A serializable version of torch.fx.Node that contains information pertinent to placeholder stack traces. We use these in logging and error messages related to cudagraphs, and will cache these results. strname Optional[str] stack_tracelist[PlaceholderInfo]usersmutating_use_stack_tracerNrrr!r"r$r$#s  I  ++r!r$cV\rSrSr%SrS\S'S\S'S\S'S \S 'S \S 'S\S 'Srg)WrappedFunction2z Represents a function that you want to record for CUDA graph replay, with a little more metadata so we can identify if we have an applicable CUDA graph in our CUDA graph tree for it. zCallable[..., Any]model Sequence[int]static_input_idxsrrztuple[torch.Tensor, ...] constantsSequence[PlaceholderInfo] placeholdersmutated_input_idxsrNrrr!r"r.r.2s, $$N''++%%r!r.c[UR5S:Xa8[[UR55RR SS5$URHhnUR [RRRR:XdMAURR SS5=n(dMfUs $ g)Nr)) lenr+nextitermetagettargettorchopsatencopy_default)placeholder_nodeuser)s r"&get_mutating_use_stack_trace_from_noderFBs  ! !"a'D)//0166::=$OO%% ::--55 5!hhll=$??{?""& r!cUR$N)r,)placeholder_infos r"get_mutating_use_stack_tracerJQs  4 44r!cURnURRSS5n/nSnURS:Xa0URVs/sHn[ U5PM nn[ U5n[XX45$s snf)Nr) placeholder)r'r<r=opr+to_placeholder_inforFr$)rDr'r)r+r,is r"rNrNUs  D"''++M4@K E#m+1A1G1GH1GA$Q'1GH#I $   4e NN Is A9czURVs/sHoRS:XdM[U5PM! sn$s snf)NrL)nodesrMrN)graphnodes r"get_placeholder_inforTds9.3kk .9dWW =U!D!k  s88c SU3$)Nzskipping cudagraphs due to r)reasons r"format_default_skip_messagerWjs ( 11r!cSnUHnXn[U5=n(dM O [S[U5S35nU(aUSU3$U$)Nzmutated inputs (z instances). Found from : )rJrWr9)r5mutation_indicesr)idxrLmsgs r"get_mutation_stack_tracer^nsi"$K"' 6{C C; C  & 3/01= C( 66 Jr!c[RRRR(a?UR Vs/sH'nX0R ;aMU"X5(aM%UPM) nnO UR n[RSUR 5 [RSU5 U(a[URU5$S$s snf)Nz'check mutation static input indices: %sz#check mutation mutation indices: %s) r? _inductorconfigtritoncudagraph_treesr6r2static_inputs_logdebugr^r5)funcinputsis_cuda_graph_recorded_tensorr\r[s r"check_for_mutationris  $$44..+ .--- 1= . +  22143I3IACST  !!2!24DE!+ sC C*CcxURH*nURRSS5=n(dM(Us $ g)Nr))r+r<r=)rSrEr)s r"_get_use_stack_tracerks5zz((,,}d; ;; ;  r!cUR[R"S5S5 [RRR (a&UR[R"S5S5 UR [R"S55=n(a=SURS3n[U5=n(a[USU35$[U5$[U5S:Xa1[[UR555RS:XagSUR55n[S S RU535$) Nr<cpuz cpu device ()rZr8cudac38# UHn[U5v M g7frH)repr).0keys r" :check_multiple_devices_or_any_cpu_nodes..sA&@sc&@szmultiple devices: z, )popr?devicer`ragraph_partitionr=r'rkrWr9r:r;keystypejoin)device_node_mappingcpu_noder]r) keys_reprs r"'check_multiple_devices_or_any_cpu_nodesrsELL0$7 -- U 3T:&**5<<+>??x?X]]O1-.x8 8; 8.#6H /VW W*3//  A% )..01 2 7 76 AA&9&>&>&@AI &);DIIi?K $$44 !23 +=X+=CKAWC+=X+,1 +BHH5 ' GG>*a/ 't8[8Ys  B2B2cUR(a UR$URH"nUR(dMURs $ g)zE Gets the first non-empty stack trace of a placeholder or its users. N)r)r+)rLusers r"get_placeholder_stack_tracersF&&&!!    ## #" r!c.\rSrSrSrSrSrSrS SjrSr g) CheckInvariantStatusr8cURS:XagURS:XagURS:XagURSUR3$)NCudagraphManagedIdxMismatchz-cudagraph managed tensor data pointer changedStaticInputIdxMismatchz!static input data pointer changed&ExpectedDeadIndicesBeforeGraphMismatchz+expected dead indices before graph are livez: )r'r)rs r"__str__CheckInvariantStatus.__str__ sK 995 5B YY2 26 YYB B@ii[4::,/ /r!rNrr&) rrrrSUCCESSrrrrr rr!r"rrs$G#$./*0r!rc [U5[U5:Xa[U5[U5:XdS5eUVs/sHoQUPM nnUVs/sHoRUPM nnUS3n[[Xg55Hyunup[U [R 5(deX5n U R 5U :wdMCX n USU RSU SU R 5S[U 5S3 nM{ U$s snfs snf)zq Logs the mismatch between input data pointers and recorded data pointers. This checks only idxs in target_idxs. zClength mismatch between inputs, recorded_data_ptr, and placeholdersz. z input name: z. data pointer changed from z to z. input stack trace:  ) r9 enumerateziprr?Tensordata_ptrr'r) r5rgrecorded_data_ptr target_idxsmismatchrO t_tensors t_data_ptrs error_msgtensorrindexrLs r"log_data_ptr_mismatchrs v;#/0 0S[C DU5UM U%00KqKI01<=AQ'K=*C I!*3y+F!G F&%,,//// ??  (&-K+\+*:*:);<--5Jd6??;L:MN&&A+&N%OrS  "H 1=s C8C=cN^[UR55S-mSU4Sjjn[RRR R (aMT[RRR R :a[RU"55 gg)Nr8c>STS3$)NzCUDAGraph supports dynamic shapes by recording a new graph for each distinct input size. Recording too many CUDAGraphs may lead to extra overhead. We have observed a0 distinct sizes. Please consider the following options for better performance: a) padding inputs to a few fixed number of shapes; or b) set torch._inductor.config.triton.cudagraph_skip_dynamic_graphs=True. Set torch._inductor.config.triton.cudagraph_dynamic_shape_warn_limit=None to silence this warning.r)num_cudagraphssr"warn_msg4maybe_warning_due_to_dynamic_shape..warn_msg<s 00>/?@' ' r!TFr) r9ryr?r`rarb"cudagraph_dynamic_shape_warn_limitrr)fn_cache new_int_keyrrs @r""maybe_warning_due_to_dynamic_shaper6st)A-N   %%HH  // ' ' J J K hj) r!c8\rSrSr%SrS\S'S\S'S\S'S rg ) CudagraphCachedInfoiSz Info needed to realign inputs r4r5list[Optional[str]] stack_tracesz list[str]cudagraph_fail_reasonsrNrrr!r"rrSs,+%%%%r!rcL\rSrSr%SrS\S'S\S'S\S'S\S 'S \S 'S rg )CudagraphMetadatai^z& Metadata for recording a CUDA graph. r4r5OrderedSet[int]r2r6rrzdict[str, torch.Tensor]r3rNrrr!r"rr^s',+&&''%%&&r!rcv/n[5n[5n[UR5HupVXaR;aUR U5 XaR ;aUR U5 UbUR UnO[SURSU3S/SS9nURU5 M /nURH7n U b URURU 5 M&URS5 M9 URV s0sHoURU _M n n [UUUUU 5$s sn f)z Convert the cudagraph metadata at the graph level to the graph partition level, given the graph partition info (i.e., mapping from partition input/output index to graph input/output index). N partition_ _placeholder_)r'r)r+r,)rrinput_index_mappingr2addr6r5r$rappendoutput_index_mappingrconstant_namesr3r) partition_mapmetadatapartition_placeholderspartition_static_input_idxspartition_mutated_input_idxspartition_input_idxgraph_input_idxrLpartition_stack_tracesgraph_output_idxr'partition_constantss r" get_partition_cudagraph_metadatarks` 3=<4>L 09))1, 88 8 ' + +,? @ 99 9 ( , ,-@ A  &"//@K*!-"2"2!3=AT@UV )- K %%k2'1* )>>  ' " ) )(*?*?@P*Q R " ) )$ / ?4A3O3O3O4h  &&3O #$   s D6)rD torch.fx.Noderr()rIr$rr()rDrrr$)rRztorch.fx.Graphrr*)rVr&rr&)r5r4r[r1rr&)rfr.rglist[InputType]rhzCallable[[torch.Tensor], bool]rr()rSrrr()r|z!dict[torch.device, torch.fx.Node]rr()r]r&rr) rztorch.fx.GraphModulerzOrderedSet[str]r6rr2r1rr()rLr$rr() r5r4rgrrzSequence[Optional[int]]rr1rrrr&)rz)dict[tuple[int, ...], Callable[..., Any]]rrrbool)rr rrrr)8 __future__r dataclassesenumrtypingrrrrr r?torch._dynamo.utilsr r torch._inductor.utilsr r torch.utils._ordered_setrcollections.abcr_logginggetArtifactLoggerrrrdlistrr OutputType ModelType dataclassrr$r.rFrJrNrTrWr^rirkrrrrrrrrrrrrrr!r"rs"@@ =>/(00<H NN44 ' (5ell!234 5 d9o& 2 3  d#  $  d# , ,$ , d# & &$ & #  5 O 2+?L&  $B >T:TT6H:HH J    99#9(9% 9  90 0400+ /  #   >7 : d#&&$& d# ' '$ '3$333r!