[Thv SrSSKrSSKrSSKrSSKrSSKrSSKrSSKrSSKrSSK r SSK r SSK r SSK r SSK r SSKrSSKJr SSKJr SSKJrJrJrJr SSKrSSKJr SSKrSSKJr SSKJr SSKJ r SS K!J"r" SS K#J$r$J%r% S S K&J'r' S S KJ(r(J)r) \RT"\+5r,\"S5r-\"S5r.\.R_5r0\0(aSSK1r2/r3Sr4\0(ah/SQr3\2RjRlRnRq5RsSS5r:SRw\3Vs/sH nSUS3PM sn5r4/SQr<"SS5r=Sr>Sr?"SS5r@\R"S5S5rBSS .S!jrCSS .S"jrDS#rES$rF"S%S&\G5rHS'rISES(jrJSFSSS).S*jjrKS+rLS,rMS-rNSFSSS).S.jjrOS/\S0S1S2S3S04S4jrPS5\-S3\\\-/\-44S6jrQ\Q"\R5rS\Q"\R"S755rU\Q"S5rV\Q"S5rW\Q"S5rX"S8S95rY"S:S;5rZ"S<S=5r[SGS>\\\\/\\\4S?\]S@\\^\]\_4SA\\_S3\^\]\44 SBjjr`SC\]S3\\-/\-44SDjrags snf)Ha Debug utilities for TorchDynamo compilation and execution. This module provides various debugging tools and utilities for TorchDynamo, including: - Minification support for reducing test cases while preserving bugs - Input/output handling via InputReader and InputWriter for reproducible testing - Accuracy checking between original and compiled models - Neural network module string conversion via NNModuleToString - Profiling tools and system information collection - Buck build system integration for Meta-internal testing Key classes: - InputReader/InputWriter: Handle serialization of model inputs/outputs - NNModuleToString: Converts nn.Modules to string representations - BuckTargetWriter: Manages Buck build system integration N)Counter) import_module)AnyCallableOptionalTypeVar)Tensor) rand_strided)is_float_dtype)StorageWeakRef)ContentStoreReaderContentStoreWriter)config) clone_inputs get_debug_dirTztorch._inductor.config)z1//caffe2/torch/fb/sparsenn:sparsenn_operators_gpuz-//caffe2/torch/fb/sparsenn:sparsenn_operatorsz///deeplearning/fbgemm/fbgemm_gpu:sparse_ops_cpuz+//deeplearning/fbgemm/fbgemm_gpu:sparse_opszfbcode:// ztorch.ops.load_library("z"))buck2runz@mode/dev-nosanc*\rSrSrSrSrSSjrSrg)BuckTargetWriterQc,[RR[RRU55uUlUlUR R SS5UlURR SS5SUR3UlURURRS5SUlURSSUlURnX"RS5SSSnSUS UR3Ul g) Nz.pyr/.zfbcode.zfbcode/r:) ospathsplitabspathsubdirpy_filereplacetargetfind cmd_line_path)selffilenametmps Q/var/www/auris/envauris/lib/python3.13/site-packages/torch/_dynamo/debug_utils.py__init__BuckTargetWriter.__init__Rs$&GGMM"''//(2K$L! T\ll**5"5 {{**345Qt{{mD IIdiinnY79: IIabM kk((9%'(,!#a }5c SR[Vs/sH nSUS3PM sn5n[R"SURSUR S[ SUSURS 3 5$s snf) Nrz "z",za load("@fbcode_macros//build_defs:python_binary.bzl", "python_binary") python_binary( name="z", srcs = ["z"], compile = False, deps = [ "//caffe2:torch", "//caffe2:libtorch", "//caffe2/functorch:functorch", "//triton:triton", "z", ], cpp_deps = [ z ], main_module = "z", par_style = "xar", ) )join extra_depstextwrapdedentr(r& cur_targetr")r+xextra_cpp_depss r.buildBuckTargetWriter.build`sz#Jz!is"$5z#JK  ;;-ll^  II;#   $KsA3cd[RRURS5n[ US5nUR UR 55 SSS5 [UR/-nU(a%[RSSRU55 U$!,(df  NP=f)NTARGETSwzFFound an example that reproduces the error. Run this cmd to repro - %s ) r!r"r3r%openwriter:BUCK_CMD_PREFIXr*logwarning)r+ print_msg target_filefd cmd_splits r.rABuckTargetWriter.writezsggll4;; : +s #r HHTZZ\ "$$t'9'9&::  KKX# $ #s B!! B/)r*r"r&r%r(N)T)__name__ __module__ __qualname____firstlineno__r/r:rA__static_attributes__r1r.rrQs 6 4 r1rc([RR[5S5nUc-[R "5S[ R"53n[RRU5(d[R"USS9 U$)Nminifierz /minifier_T)exist_ok) r!r"r3rtempfile gettempdirgetpassgetuserexistsmakedirs)r"s r. minifier_dirrYse 77<<  4D |%%'( 7??3D2EF 77>>$   D4( Kr1c\rSrSr\R R \R R\R R\R R\R R\R R\R R\R R\R R\R R\R R \R R"\R R$\R R&\R R(\R R*\R R,\R R.\R R0\R R2\R R4/r\S5r\S5rSrg)NNModuleToStringc[5nUR5H5up#[U5[R;dM$UR U5 M7 [ U5S:a[RSU5 g)Nrz-We have not tested reprs of some modules - %sT) setnamed_childrentyper\ safe_reprsaddlenrCrD)gm cant_convert_modules r.can_convert_to_string&NNModuleToString.can_convert_to_strings`u **,IAF|#3#>#>>  (- | q KKG Vr1c$SSKJn Sn[R"S5nUR 5HYupEUR 5n[ UR5S5nUbUR(aUS3nX2S-SUSUS 3- nM[ URR5HupU cM U R5[::a(SS K Jn U R[:de[!U 5n Oh["R$"U 5(a'S ['U R(5S U R*S 3n O&S['U R(5S U R*S 3n U R(aU S3n UUS-SUSU S3- nM UR,R5HWupU cM SnU R(aSnS['U R(5S U R*US3n X2S-SU SU S 3- nMY X1"UR.S5S 3- nU$)Nr) _addindent z from torch.nn import * class Repro(torch.nn.Module): def __init__(self) -> None: super().__init__() z.cuda()zself.z = r) PRINT_OPTSz torch.randn(z, dtype=)ztorch.randint(1, size=zself.register_buffer('z', z) rz, device="cuda"ztorch.nn.Parameter(torch.randn(z))rZ)torch.nn.modules.modulerlr5r6r`__repr__next parametersis_cuda_buffersitemsnumelMAX_CONSTANT_NUMEL_INLINEtorch._tensor_strro thresholdreprtorchis_floating_pointlistshapedtype _parameterscode)rerltab model_str module_namerh module_str example_param buffer_namebufferro tensor_str param_nameparam maybe_devices r.convertNNModuleToString.converts$6OO  $#4#6 K"OO-.J!!2!2!4d;M(]-B-B *|73 !G9E+c*RH HI$7$&;;#4#4#6 K~||~!::8!++/HHHH!&\ ((00+D,>+?x ~UVW -T&,,-?,@VWX~~ *|73 7)1+c*SQ I#$7*"$!5!5!7 J}L}}0 :4 ;L:MXV[VaVaUbcobpprsJ !G9E*S BG GI"8  277A./r22 r1rON) rJrKrLrMr}nnLinearConv1dConv2dConv3d BatchNorm1d BatchNorm2d BatchNorm3d LayerNormDropoutSoftmaxReLUGELUIdentity MaxPool2d EmbeddingTanhConvTranspose1dGLULSTMFlattenAdaptiveAvgPool2drb staticmethodrirrNrOr1r.r\r\s0                            ""+J0  ==r1r\cJ[RR5(dgSn[R"SS/5nUR 5R S5nSRUVs/sHo3S;dM SUS 3PM sn5nXS3- n[S [[RR5555nUS - nUR5HupgUSUS US 3- nM US- nU$s snf![[R4a US - nNf=f)Nz:# torch.cuda.is_available()==False, no GPU info collected z# CUDA Info: nvccz --versionrr)r# z z# nvcc not found c3`# UH$n[RRU5v M& g7fN)r}cudaget_device_name).0is r. ,_cuda_system_info_comment..s&/O! ""1%%/Os,.z# GPU Hardware Info: z : )r}r is_available subprocess check_outputdecoder#r3FileNotFoundErrorCalledProcessErrorrrange device_countrw)rcuda_version_outcuda_version_linesscomment gpu_namesnamecounts r._cuda_system_info_commentrs- :: " " $ $L!I*%22FK3HI-446<^[U4SjT55=(a TT;$)Nc3,># UH oT;v M g7frrO)rstringkeys r.r;generate_env_vars_string..filter..s:zVS=zs)any)r allow_list skip_lists`r.filter(generate_env_vars_string..filters:z::Ss)?SSr1z os.environ['z'] = ''rz import os z )r!environrwr3)rrrvalue config_lines config_stringrrs @@r.generate_env_vars_stringrs<:JUIT ****,,JC #; +,se6%*, IIl+M  s A9 A9c hSSKnSSKnU(agURRRR 5nSUR RR 5SURRR 5SURRR 5SUS3 $)Nrz*# config omitted due to stable_output=Truez~import torch._dynamo.config import torch._inductor.config import torch._functorch.config import torch.fx.experimental._config r) torch._functorch.configtorch._inductor.configfx experimental_configcodegen_config_dynamor _inductor _functorch)rr}experimental_configs r.generate_config_stringr,s"!;((//77FFH $$&'(&&()*'')*+  r1cR[RR[5S5$)Nzminifier_launcher.py)r!r"r3rYrOr1r.get_minifier_repro_pathr@s 77<< (> ??r1cb[5n[RSU5 [(a[ U5R 5 [ US5nUR U5 SSS5 g!,(df  g=f![a&n[RS5 [S5UeSnAff=f)NzWriting minified repro to: %sr>rz(Could not write to {minified_repro_path}) rrCrDuse_buckrrAr@OSError exceptionNotImplementedError)contentsminified_repro_pathrGes r.helper_for_dump_minifyrDs13KK02EFx,-335U %s +r HHX , + + U b!"LMSTTUs6 A>A-$A>- A;7A>;A>> B.!B))B.c\rSrSrSrg) AccuracyErroriSrON)rJrKrLrMrNrOr1r.rrSsr1rc[U5n[[U55HEn[X[R 5(dM&XR XR5 MG U$)z This clone inputs is different from utils clone_input. In case of minifier, all the tensors are leaf tensors while creating a new graph. So, we set the requires_grad field w/o checking the leafness of the tensor. )rrrd isinstancer}r requires_grad_ requires_grad)example_inputs cloned_inputsidxs r.clone_inputs_retaining_gradnessrWsX !0MS() m(%,, 7 7   - -n.A.O.O P* r1cVSSKJnJnJn [R "U5nU(d [ U5n[US5(aURS5 [US5(aU"U5OU"U6nU(aU$U"U5(aU"U5nUR5 U"XSU5$)z Runs a forward and possibly backward iteration for a given mod and args. When disable_clone is True, we will use args as-is without cloning. This is higher fidelity but we may destroy the args in the process. r)collect_resultsreduce_to_scalar_lossrequires_bwd_pass zero_gradT _boxed_callN) testingrrrcopydeepcopyrhasattrrbackward) reargsonly_fwd disable_clonerrroutlosss r.run_fwd_maybe_bwdr dsSR r B .t4r; Tb-00"T(b$iC $S)  2D$ //r1 require_fp64 ignore_non_fpc SSKJn [XU5nSn[R(a8[ [ R"U5[U55up[XU5n[XU5n U"UU U[RSUS9n U $![a* U(a [S5e[RS5 N[f=f![a [RS5 gf=f) aI Check two models have same accuracy. require_fp64: if True, raise an error if we unable to calculate the fp64 reference ignore_non_fp: if True, do not compare outputs which are not floating point. This is mostly useful for the minifier (which wants to avoid quantizing floating point error into integer/boolean error) r)sameNzfCould not generate fp64 outputs, workaround with torch._dynamo.config.same_two_models_use_fp64 = FalsezCould not generate fp64 outputszWhile minifying the program in accuracy minification mode, ran into a runtime exception which is likely an unrelated issue. Skipping this graph.T)tol equal_nanr)utilsrr rsame_two_models_use_fp64 cast_to_fp64rrr Exception RuntimeErrorrCrDrrepro_tolerance) reopt_gmrr rrrreffp64_ref fp64_model fp64_examplesrespassings r.same_two_modelsr"s" B 9CH && ;(4 b!#B>#R) %J)HMH A    " "# G N7 ;"| KK9 :  ;   $ s#7B # C 1B>=B>C#"C#cURRGH@nURS:XaUR[R R RR:Xa}[UR5S:Xde[URS5(aEURS[R:wa$URS[R4Ul URS:XdMURRS5nUcM[U5(dGM[UR5n[RUS'X1l GMC URR!5 UR#5 U$)N call_functionrnrrr)graphnodesopr(r}opsprimsconvert_element_typedefaultrdrr float64kwargsgetdictlint recompile)modelnoder new_kwargss r.cast_dtype_args_to_fp64r5s !! GG & uyyCCKKKtyy>Q& &&diil++ !  0M!YYq\5==9 77o %KKOOG,E ^E%:%:!$++. &+mm 7#( " KK OO Lr1c^SSKJn URT5nT[R:Xa [ U5nU"U4SjU5nX4$)Nr)tree_mapc>[U[R5(a&UR5(aUR T5$U$r)rr}r r~to)r8rs r.cast_to..s< a & &1+>+>+@+@$$u+ r1)torch.utils._pytreer7r9r}r,r5)rr2inputsr7s` r.cast_tor>sI, HHUOE  (.    F =r1c6[[RX5$r)r>r}r,)r2r=s r.rrs 5==% 00r1c U"[R"U5[U55n[UUUUUUS9(+$![a [ R S5 gf=f)NrzWhile minifying the program in accuracy minification mode, ran into a runtime exception which is likely an unrelated issue. Skipping this graphF)rrrr"rrCr)rer compiler_fnr rr compiled_gms r.backend_accuracy_failsrCst! MM" >~N #    %'       # s8;AAstrideztorch._prims_common.StrideTyperztorch._prims_common.ShapeTypereturnc8UbU$[R"U5$r)rmake_contiguous_strides_for)rDrs r._stride_or_defaultrH s '6UU-N-Nu-UUr1dc^U4Sj$)Nc>UbU$T$rrO)r8rIs r.r:_mk_defaulter..s!-Q.Q.r1rO)rIs`r. _mk_defaulterrMs ..r1cpuc:\rSrSrS SjrSSS.SjrSrSrSrg) NopInputReaderi NcSUlg)Nrtotal)r+s r.r/NopInputReader.__init__!s  r1device dtype_hintc.U=RS- slg)NrrR)r+ storage_hashnbytesrVrWs r.storageNopInputReader.storage$s a r1cgrrOr+rr-s r.tensorNopInputReader.tensor' r1cgrrOr^s r.symintNopInputReader.symint*rar1rRrEN rJrKrLrMr/r[r_rcrNrOr1r.rPrP s7;t  r1rPcV\rSrSrS SS.SjjrSSS.SjrS SSSSS.SjjrS rS rg) InputReaderi0N)pbarcxUc[RS5 Ub [U5OSUl/UlX lg)Nz0no save_dir specified, will generate random data)rCrDr storerri)r+save_dirris r.r/InputReader.__init__1s7   KKJ K5=5I'1t   r1rUcURbURRS5 [U5n[U5nURbPUbMURR U5nX5R :wa [RSX5R 5 U$[RSU5 X$R-4n[SUS9n[XgXC5R5$![a NUf=f)Nrzdevice mismatch: %s != %sz1could not load %s, generating random data insteadr)riupdate_device_or_default_dtype_or_defaultrk read_storagerVrCrDritemsizerHr untyped_storage)r+rYrZrVrWr[rrDs r.r[InputReader.storage<s 99 II  Q #F+&z2 :: !l&> **11,?^^+KK ;V^^T GV...0#D6E:>NNPP%  sC%% C21C2)storage_offsetrris_leafc ,[X2S9n[U5n[U5n[U5n[ U5n[ R "/XQRUS9n [ R"5 U RXX#5 SSS5 U(dk[ R"5 U R[ RS9n SSS5 [ R"5 U RXX#5 SSS5 [ RRRU 5U:Xde[ R R#X5 UR$R'U 5 U $!,(df  N=f!,(df  N=f!,(df  N=f)Nro)rrVr) memory_format)rH_storage_offset_or_defaultrr_is_leaf_or_default_requires_grad_or_defaultr}r_rVno_gradset_ enable_gradclonepreserve_format _subclasses meta_utils safe_is_leaf_utilsset_tensor_metadatarappend) r+r[rrDrwrrrxmetadatats r.r_InputReader.tensorRs$F83NC!%(%g.1-@ LL eNN- ]]_ FF7E :""$GG%*?*?G@%w>!  ++88;wFFF ((5 _%$ s$+E#"E4F# E14 F Fc<URRU5 U$r)rr)r+vals r.rcInputReader.symintss  r1)rrirkrrfrOr1r.rhrh0s> d 7;tQ4  Br1rhcd\rSrSrSS.SjrSrSSS.S\4S jjrSS jrS r SS jr SS jr Sr g) InputWriteriF stable_hashc/Ul[R"5UlXlUb [ XS9OSUl0Ulg)Nr)_lines itertoolsrstorage_counterrlrrk seen_storages)r+rlrs r.r/InputWriter.__init__sC (0 # x A  r1crS/nURSUR55 URS5 U$)Nzdef load_args(reader):c3,# UH nSU3v M g7f)rmNrO)rls r.r$InputWriter.lines..s1[4s[szload_args._version = 0)extendrr)r+rs r.linesInputWriter.liness8 $  1T[[11 )*r1NrW device_hintrEc X[U5nURRU5nUbU$S[UR53nSn[ S5[ U5:waSU<3nSnUR nURS:XaUceUn[S5U:waSU<3nUR5n Sn URb5UR RS:waURRU5n URRUSU <SU <UUS35 XPRU'U$) Nbufrz , dtype_hint=metaz , device=z = reader.storage(, rp)r rr.rsrrrrVrarqrZrk write_storagerr) r+rurWrwsvmaybe_dtype_hintrrVrZrYs r.r[InputWriter.storages@ O ,    " "2 & =H$t++,- . T "&7 &C C!.zn=   '' ;;& * ** F d #v -&vj1L '') :: !o&<&<&A&AV&K::33ODL c#L#32fZ ~N^M__` a "#2r1c SSKJnJn URUR 5UR UR S9n/nU"U"[SURS9UR555(d1UR[[UR5555 [S5UR :waURSUR <35 U"[S5UR5:H5(d#URSUR5<35 [ R"R%U5nU(a&UR'SUR)555 [+S5UR,:waURSUR,<35 [ R.R0R3U5n[5S5U:waURS U<35 UR6RS S R9U[[UR55/UQ5-S U3-5 g) Nr)statically_known_truesym_eqrrozdtype=zstorage_offset=c36# UHupUSU<3v M g7f)=NrO)rkrs r.r%InputWriter.tensor..sI1H1#Qqe 1Hszrequires_grad=zis_leaf=zreader.tensor(r) # )%torch.fx.experimental.symbolic_shapesrrr[rurrVrHrrDrstrtuplerrr{rwr}rget_tensor_metadatarrwr}rrrrr|rr3) r+rrrrr[rtensor_metadatarxs r.r_InputWriter.tensorsW,,   AGG $ %d!'':AHHJ G   KKE!((*-. / T "agg - KK& , -$ &t ,0@0@0B B   KK/!*:*:*<)?@ A,,::1=  KKI1F1F1HI I $T *aoo = KK.(;< =##..;;A> t $ / KK(7+. /  ii#eAGGn"5==> ?dVn  r1c*URRSUS[U535 [U[[ 45(aURRS5 [ U5Hup4USUS3n[U[R5(aURXT5 M?[U[[R45(aURXT5 MwURXT5 M URRS5 gg)Nrz# was unsupported type for dumping: z"""[])rrrarrr enumerater}r r_intSymIntrc unsupported)r+rargraname_is r.rInputWriter.unsupporteds Rv%Hc TU cD%= ) ) KK  u %!# 61#Qa..KK*C#677KK*$$V/' KK  u % *r1cJURRSU<SUS35 g)Nz reader.const(rz!, filtered out during compilation)rr)r+rs r.constInputWriter.consts' D85.O P r1c[U[R5(aURRnUR R SU<SU35 g)Nzreader.symint(r)rr}rr3hintrr)r+rrs r.rcInputWriter.symints? c5<< ( (((--C ^C7%v>?r1)rrlrrrkre) rJrKrLrMr/rrr[r_rrrcrNrOr1r.rrs905  6:tPS8 @&$ @r1rfuncrV sym_shapesdefault_sym_shapec ^^^^SSKJn UR5VVs0sHupVXe_M nnnSRUR 55n[ R "U5n SUS3n SUS3n Sn "S S 5n 0nT=(d 0mUU4S jmS [4UU4S jjnURnUR5HunnUS :XaM[R"U U5nU(a>UR5unn[URS55nUUnU"UU5UU'[R"U U5nU(dMT"URS55UU'M S[ R"U5R ;aqU "5nUUS'[R""X5HKnUR5unnnn[URS55nUUn[%UUU"UU55 MM U$s snnf)a Takes in a function which has been printed with print_readable() and constructs kwargs to run it. Handles Tensor inputs, Symints, and a graph module which might have tensor constants. Consider a function `forward` defined as follows: def forward(self, primals_1: "f32[1001, 6]", primals_2: "f32[s0]", primals_3: "Sym(s0)",): _tensor_constant0: "i64[4190]" = self._tensor_constant0 # Further implementation kwargs = aot_graph_input_parser(forward) forward(**kwargs) r) dtype_abbrs|z(_tensor_constant\d+): \"(z0)\[\s*(.*?)\s*\]\" = self\.(_tensor_constant\d+)(z)\[\s*(.*?)\s*\]z Sym\((s\d+)\)c\rSrSrSrSrg)/aot_graph_input_parser..TensorContaineriz#Container for tensors as attributesrON)rJrKrLrM__doc__rNrOr1r.TensorContainerrs-r1rc|>^[R"TT;=(d TSLU4Sj5 TRTT5$)Nc>TS3$)Nz; not in symbolic_shapes and default sym shape not passed inrO)rcsr.r:=aot_graph_input_parser..get_sym_int..'s vhYZr1)r}_checkr.)rcrrs`r. get_sym_int+aot_graph_input_parser..get_sym_int$s9 j A$5T$A Z ~~f&788r1rEc>/n/n[U5HjupEUR5nSU;a,T "U5nURU5 URU5 MGU(dMPUR[U55 Ml UR(a[ R O[ RnU"X!T S9nUH"n [ RRX5 M$ U$)Nr)rrV) rstriprrr~r}randnzerosr mark_dynamic) rrresolved_shape dynamic_dimsrdimr constructorr rIrVrs r. gen_tensor*aot_graph_input_parser..gen_tensor+s &FA))+Ccz$%%a(##A&3"))#c(3'&+%<% Iy)//#./Ei(E&ue4F5M /:6 5' A7F5M1 ""4(333#% "v[[!8AE16 .Iy)Q)//#./Ei(E Iy*UE*B C B MSCsG:r,c^^[R"5m[RR [RR T55mU4SjnUU4Sjn[ R"U5 U$)z Decorator to cProfile a given function and save the result to disk on process exit. Args: filename: filename to save profile to cJ>^[R"T5UU4Sj5nU$)Nc>TR5 T"U0UD6TR5 $!TR5 f=fr)enabledisable)rr-fnprofs r.wrapper3profile_to_file..decorator..wrapperhs1 KKM 4*6*  s+=) functoolswraps)rr rs` r. decorator"profile_to_file..decoratorgs%     r1c >TRT5 [RR[R "STSTS355 g)Nz! Wrote profile to z+, view with: snakeviz z ) dump_statssysstderrrAr5r6)r,rsr.save_it profile_to_file..save_itrsK !  OO""*,&Z(  r1)cProfileProfiler!r"r$ expanduseratexitregister)r,r rrs` @r.profile_to_filer]sR    Dwwrww11(;rrCrHrMfloat32rrrVrqr{r}r|rPrhrrrr/rrr)r8s0r.r1s$   #33 ##..;M.! CL 89  $ $ &    J $$..==?GG SWXJIITA!9!B?TUM644nccL T0/42-2(@ U I  0> 7 7t*$1  RV 5 6V +V& V/Q/8Xa[M1$45/"%--0"5<<#67*1-)%0#E*   EEhs@s@p+/'+ ` DL>4</ 0` `c3h(` } `  #s(^ `F#c#hsAv&6#eUs(I: