+ jMAa0t$^RIHt^RIt^RIt^RIt^RIt^RIt^RIH t ^RI H t ^RI H t Ht^RIHtHt^RIt^RIt^RIHt^RIHt^RIHuHt^RIHt^RIHt^R IH t ^R I!H"t"H#t#H$t$^R I%H&t&^R I'H(t(H)t)H*t*]'d^R I+H,t,H-t-H.t.^RI/H0t0^RI1H2t2]!R4t3]!R4t4]Pj!]64t7RRlt8] RRl4t9]$RRl4t:R4RRllt;RRlt<]$RRl4t=!RR]P|4t?]$R R!l4t@]$R"R#l4tAR$R%ltB]PPtD]DP]DP]DP]DP]DP]DP]DP]DP]DP]DP]DP]DP]DP]DPP]DPP]DP]DP]DP]DP]DP]DP]DP0t\]!]\4t\]$R&R'l4t]R(R)lt^R*R+lt_^s`R,]aR-&R.R/ltbR0R1ltcR5R2R3lltdR#)6) annotationsN)contextmanager)partial)Any TYPE_CHECKING) ParamSpecTypeVar)SymInt)get_decompositions) bind_symbols) aot_function aot_modulemake_boxed_compiler)strip_overloads)default_partition draw_graph#min_cut_rematerialization_partition)Callable GeneratorSequence)Node) IntLikeType_P_Rc V^8dQhRRRR/#)fx_gfx.GraphModulereturn)formats"[/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torch/_functorch/compilers.py __annotate__r#/s>cVPPR\PPP R7F,p\PPP VnK. VP4V#) call_functionoptarget) graph find_nodestorchopsaten_to_copytor) recompile)rnodes& r" _canonicalizer3/s[ %% 599>>#:#:&iinn''  NN Kr$cV^8dQhRR/#)rrzGenerator[None, None, None]r )r!s"r"r#r#9s??:?r$c#"\PPR4pRx\PPV4R# \PPT4i;i5i)FN)r,_C_jit_set_autocast_mode)old_jit_autocast_flags r"_disable_jit_autocastr98sL"HH;;EB?  ''(=>''(=>s A,A!A,!A))A,c$V^8dQhRRRRRR/#)rrrinpsz Sequence[Any]rtorch.jit.ScriptModuler )r!s"r"r#r#Ds"0 0 ^0 =0 =S0 r$c^RIHpV!V4p\4;_uu_4\V4VPP R\ PPPR7Fwp\VP4^8XgK\VP4^8XgK;RVP9gKN\ PPPVnKy VPPF^p/pVPP!4F5wrV\#V\ P$4'd VP&pWdV&K7 WCn K` VPP)4VP+4\ P,P/V4p\ P0P3VP4\ P,P5VP744p\ P,P9V4p\:;QJdRV4F 'gK RM RM !RV44'gV!V!RRR4V# +'giX#;i) z Compiles the :attr:`fx_g` with Torchscript compiler. .. warning:: This API is experimental and likely to change. Args: fx_g(fx.GraphModule): The input Fx graph module to be compiled. Returns: Torch scripted model. )_unwrap_lazy_graph_moduler&r'dtypec3t"TF.p\PPPV4xK0 R#5iN)r, _subclasses fake_tensoris_fake_tensor).0ts& r" ts_compile..rs*QDq5$$00??BBDs68TFN)torch.fx._lazy_graph_moduler>r9rr*r+r,r-r.r/lenargskwargsr0r)nodesitems isinstancedevicetypelintr1jitscriptr6_jit_pass_remove_mutationfreezeevaloptimize_for_inferenceany)rr;r>r2 new_kwargskvfs&& r" ts_compiler^CsF $T *D  JJ))uyy~~'>'>* D499~"s4;;'71'<DKKAW#iinn//  JJ$$DJ ))+a..A !1 ,%K %   II  T " **1773 II  QVVX & II , ,Q /sQDQsssQDQQQ tH; !< H= ! < Hs1A%I' I''I':E-I'(I'<I'I'' I8 c ,V^8dQhRRRRRRRRR R/#) rrr_rnamestr clear_metaboolrr )r!s"r"r#r#ws4  (+9=r$cJ\VP4\WVR7V#))rc)printcoder)rr`rarcs&&&&r"_draw_graph_compilerhws $))tj1 Kr$c V^8dQhRRRR/#)rrarbrz5Callable[[fx.GraphModule, list[Any]], fx.GraphModule]r )r!s"r"r#r#s HH H:Hr$c6\\\VR74#)ra)rrrhrks&r"draw_graph_compilerls w':F GGr$c$V^8dQhRRRRRR/#rrrr`rrr )r!s"r"r#r#s!  n   r$cV#)z Returns the :attr:`fx_g` Fx graph module as it is. This is a no-op compiler and can be used to check accuracy. .. warning:: This API is experimental and likely to change. r rr`s&&r"noprqs  Kr$cLa]tRt^tRRRR/RV3RllltRV3RlltR tV;t#) DebugInterpreter initial_envNenable_io_processingTc(V^8dQhRRRRRRRR/#)rrKrrtzdict[Node, Any] | Nonerurdrr )r!s"r"r#DebugInterpreter.__annotate__s2     ,  #   r$c `<\VP.VO5!Vn\SV`!VRVRV/#)rtru)r modulesymbol_mappingsuperrun)selfrtrurK __class__s&$$*r"r|DebugInterpreter.runsI + KK   w{  * AU  r$c V^8dQhRRRR/#)rnrrrr )r!s"r"r#rws55$535r$c D<aa a a a RV3Rllo RV 3Rllo RV 3Rllo RV V 3Rllp\SS`V4pRVP9d\P!VPR,4wrE\P!V4wrg\ V4\ V48wd#\ \ V4 R \ V4 24h\\\ V44WF4F6wo r\V \P4'gK(V!WV V3R l4K8 V#) c V^8dQhRRRR/#)rnirrintr )r!s"r"r#/DebugInterpreter.run_node..__annotate__s  [ S r$c<\V\4'gV#\P!VPP P SP44pVP'g\RV 24h\V4#)zexpected r to be a number, got ) rOr sympyexpandr2exprxreplacerz is_numberAssertionErrorr)rrr}s& r" subst_symint/DebugInterpreter.run_node..subst_symintsab&))  RWW\\2243F3FGHA;;;$'Fqc%JKKq6Mr$c V^8dQhRRRR/#)rnisztuple[IntLikeType, ...]rztuple[int, ...]r )r!s"r"r#rs 9 9$; 9 9r$ch<\;QJd.V3RlV4FNK 5#!V3RlV44#)c34<"TF pS!V4xK R#5irAr )rErrs& r"rGHDebugInterpreter.run_node..subst_symint_tuple..s8Cbb))Cs)tuple)rrs&r"subst_symint_tuple5DebugInterpreter.run_node..subst_symint_tuples%58C85 858C88 8r$c$V^8dQhRRRRRR/#)ra torch.Tensorbrrdr )r!s"r"r#rs!   , 4 r$c<S!VP44^8di\VP4FOpS!VPV44VPV48wgK0S!VP V44^8gKNR# R#)rFT)numelrangendimstridesize)rridxrs&& r"check_significant_strides.check_significant_stridessbAGGI&* =C$QXXc]3qxx}D(59$ ) r$c(V^8dQhRRRRRRRR/#)rnvrrvdesczCallable[[], str]rNoner )r!s"r"r#rs*  l  .checksD>>$'I$t*%VWWxx288#$xr"((4z%JKK!"''), 9$vhb 51CBGGI1N0OtTVT[T[T]S^_5R+DebugInterpreter.run_node..ss'$:M:M9N&Or$) r{run_nodemetapytree tree_flattenrJrziprrOr,Tensor)r}rrrn_vals_n_specr_vals_r_specrrrrrrr~sf& @@@@r"rDebugInterpreter.run_nodes   9 9     G Q  AFF?$11!&&-@OF$11!4OF6{c&k)$F }DV %FGG s6{!3VD 2!"ell33bOPEr$r)__name__ __module__ __qualname____firstlineno__r|r__static_attributes__ __classcell__)r~s@r"rsrss+  /3  &*   55r$rsc$V^8dQhRRRRRR/#)rrrr`rrzDCallable[[DebugInterpreter, Any, dict[Node, Any] | None, bool], Any]r )r!s"r"r#r#s$&& & &I&r$c,\V4P#)z Returns a (slow) interpreter over the FX graph module that also checks various debugging properties (e.g., that tracing strides matched real strides.) )rsr|rps&&r" debug_noprs D ! % %%r$c$V^8dQhRRRRRR/#)rrrr`rrr<r )r!s"r"r#r#s"  N s 7M r$c\V4\PPV4p\PP VP 44pV#rA)rr,rSrTrVrW)rr`r]s&& r"simple_ts_compilers=D A "A Hr$c V^8dQhRRRR/#)rr]zCallable[..., Any]rr )r!s"r"r#r#s..!.&8.r$c"\V\4#rA)r r)r]s&r"nnc_jitrs , --r$c$V^8dQhRRRRRR/#rnr )r!s"r"r#r#s!3>r$c0\VP4V#rA)rfrgrps&&r" print_compilers $)) Kr$c$V^8dQhRRRRRR/#)rfnzCallable[_P, _R] | nn.ModulerLrrr )r!s"r"r#r#s&$*$*$$*$*"$*r$c R\R\R\R\/pVPV4\ V\ P P4'd \V3/VB#\V3/VB#)a: Wrapper function over :func:`aot_function` and :func:`aot_module` to perform memory efficient fusion. It uses the :func:`min_cut_rematerialization_partition` partitioner to perform efficient recomputation. It uses NVFuser to compile the generated forward and backward graphs. .. warning:: This API is experimental and likely to change. Args: fn (Union[Callable, nn.Module]): A Python function or a ``nn.Module`` that takes one or more arguments. Must return one or more Tensors. **kwargs: Any other overrides you want to make to the settings Returns: Returns a ``Callable`` or ``nn.Module`` that retains the eager behavior of the original :attr:`fn`, but whose forward and backward graphs have gone through recomputation optimizations, and the graphs have been compiled with nvfuser. fw_compiler bw_compiler partition_fndecompositions) r^rdefault_decompositionsupdaterOr,nnModulerr )rrLconfigs&, r"memory_efficient_fusionrsd6 zz;0 F  MM&"ehhoo&&"'''B)&))r$c$V^8dQhRRRRRR/#)rrrr;zSequence[torch.Tensor]rr<r )r!s"r"r#r#=s$"" " 6""r$cVPR4\RVUu.uFq"PVP3NK up R24^RIHpV!4P 4!V!\W4#uupi)fooaQ ############################################################## # To minimize FX graph, copy and paste the below and run it # ############################################################## import torch import torch.fx as fx from functorch.compile import minifier, check_nvfuser_subprocess, check_nvfuser_correctness_subprocess inps = a? inps = [torch.ones(shape, dtype=dtype, device='cuda') for (shape, dtype) in inps] from foo import FxModule mod = FxModule().cuda() with torch.jit.fuser("fuser2"): # check_nvfuser_subprocess can be replaced with check_nvfuser_correctness_subprocess minifier(fx.symbolic_trace(mod), inps, check_nvfuser_subprocess) )FxModule) to_folderrfshaper?rrcudar^)rr;rrs&& r" debug_compiler=so NN5  &**T''177 T*+, * JOOt d !! +s A0 r graph_indexc V^8dQhRRRR/#)rinput_data_pathrbrlist[torch.Tensor]r )r!s"r"r#r#`s(:r$c t.p\VR4;_uu_4p\P!V4p.pVFp\V4^8XdTpV!\P!44pMVwrWrp V \ P \ P\ P\ P\ P \ P\ \09d\ P!^^WyV R7pM\ P!WyV R7pVPV4K RRR4V# +'giT#;i)zR Return a random input for the given inputs meta generated from _save_fx_default. rb)r?rPN)openpickleloadrJrandomr,rint32int64rduint8floatrandintrandappend) rinputsr] inputs_metarrQinput_r_strider?rPs & r" get_inputsr`s"$F ot $ $kk!n D4yA~fmmo.6:3WVIIKKKKJJIIKK  #]]1aFSF"ZZ6JF MM& !'  %. M/ % $. Ms DD&& D7 c 0V^8dQhRRRRRRRRRR R R /#) r current_namerb folder_namedump_example_inputrdgmztorch.fx.GraphModuleexample_inputsrrz nn.Moduler )r!s"r"r#r#sLfffff  f ' f  fr$c aaaa a ^RIHpRV 3Rllo RVVVV 3Rllo RV 3RllpRV 3R llpR V 3R llpV!VVVVV\R 7#) a The forward, backward, and joint computation graph will be stored in {folder_name}/{current_name}/{current_name}_forward_{graph_index}, {folder_name}/{current_name}/{current_name}_backward_{graph_index}, and {folder_name}/{current_name}/{current_name}_joint_{graph_index} respectively. The input shape of the graphs will be stored in the .input files. These files can be loaded with pickle, and is a list of format (type, shape, stride, dtype, device). In the case of type = int or float, it is just (type,). For joint graph input, it is a nested list [[],[]] where the two inner lists have the same format. If dump_example_input is True, example_inputs will be stored in .pt file. Since each function might produce multiple graphs, the graph_index is used to distinguish different graphs )aot_module_simplifiedc V^8dQhRRRR/#)rrKrrz list[Any]r )r!s"r"r#&_save_fx_default..__annotate__s  S Y r$c<.p\V4^8dL\V^,\4'd/VS!V^,4, pVS!V^,4, pV#VFp\V4\Jg\V4\ JdVP \V434KFVP \V4VPVP4VPVP34K V#)r) rJrOrrQrrrrrr?rP)rK input_metaargget_input_metas& r"r (_save_fx_default..get_input_metas t9q=ZQ77 .a1 1J .a1 1J CCyC49#5!!49,/!!#Y 3::<CJJO r$c(V^8dQhRRRRRRRR/#) r gm_to_saverrKr type_namerbrrr )r!s"r"r#rs,!!"!*-!:=! !r$cV<\VPP4^8Xd.\P\P RSV\ 4R#\P!V4pVPP\PPP44VP4S !V4p\P!S RS 2RR7VP!S RS RS RV R\ 2 4\#S RS RS RV R\ RS RV R\ R2R4;_uu_4p\$P&!WE4RRR4S'd:\P(!VS RS RS RV R\ RS RV R\ R 24R#R# +'giLS;i) rz!No nodes in graph {%s}_{%s}_{%s}.N/T)exist_okr`z.inputwbz.pt)rJr*rMlogloggingWARNINGrcopydeepcopy set_codegenr,fxCodeGenr1osmakedirsrrrdumpsave) rrKrrr r]rrrr s &&& r"graph_saver_helper,_save_fx_default..graph_saver_helpers z%% &! + GG3    ]]: & UXX^^3356 #D)  {m1\N3dC m1\N!L>9+Q{m T m1\N!L>9+Q{mSTUaTbbcdmcnnop{o|}C D    KK &   JJ-qa ~Qyk;-WXYeXffghqgrrsttAADE     s 6F F( c$V^8dQhRRRRRR/#rrrrrrr )r!s"r"r#rs$ ,> r$c<S!WR4V#)forwardr rrr!s&&r"graph_saver_forward-_save_fx_default..graph_saver_forwards 2y9 r$c$V^8dQhRRRRRR/#r$r )r!s"r"r#rs$ ,> r$c4<S!WR4\^, sV#)backward)rr's&&r"graph_saver_backward._save_fx_default..graph_saver_backwards 2z:q  r$c$V^8dQhRRRRRR/#)rrr joint_argsrrz%tuple[fx.GraphModule, fx.GraphModule]r )r!s"r"r#rs$11 1(:1 .1r$c,<S!WR4\W4#)joint)r)rr0r!s&&r"graph_saver_joint+_save_fx_default..graph_saver_joints 273 00r$)rrrr)functorch.compilerr) rrrrrrr(r-r3r r!s fff&& @@r"_save_fx_defaultr6s\,8  !!F 11 ! '(&-  r$c(V^8dQhRRRRRRRR/#)rrrbrrrdrz Callable[[bool, nn.Module], Any]r )r!s"r"r#r#s2TTT$'T=AT%Tr$c(^s\\WV4#)aK Dump the forward, backward, and joint computation graph. Example Usage: save_fx_func = graph_dumper_aot(current_name, folder_name, dump_example_input = False) optimize_ctx = torchdynamo.optimize( save_fx_func ) with torch.enable_grad(): with optimize_ctx: result = forward_and_backward_pass(model, example_inputs) )rrr6)rrrs&&&r"graph_dumper_aotr9sK #\@R SSr$)T)F)e__conditional_annotations__ __future__rrrrrr contextlibr functoolsrtypingrrtyping_extensionsrr rr,torch.fxrtorch.nnrtorch.utils._pytreeutils_pytreerr torch._decompr %torch.fx.experimental.symbolic_shapesr aot_autogradr rr compile_utilsr partitionersrrrcollections.abcrrr torch.fx.noder torch.typesrrr getLoggerrrr3r9r^rhrlrq Interpreterrsrrrr-r.detach gelu_backwardleaky_relu_backwardsigmoid_backwardthreshold_backwardhardtanh_backwardhardsigmoid_backwardhardswish_backward tanh_backward silu_backward elu_backwardcudnn_batch_normcudnn_batch_norm_backward masked_fillScalarrelu leaky_reluhardtanh hardswish hardsigmoid conj_physical is_same_sizerrrrr__annotations__rr6r9)r:s@r"rfsH""  %%0 $$,>GG*=="'t_ T] ! ??0 0 fH   Er~~EP&&  . yy~~KK""HHOOMMNN-4,,BC $*N"@ S>fTTTr$