Lj@UddlmZddlZddlZddlZddlZddlZddlmZddl m Z ddl m Z m Z ddlmZmZddlZddlZddlmZddlmZddlmcmZddlmZddlmZdd lmZd d l m!Z!m"Z"m#Z#d d l$m%Z%d d l&m'Z'm(Z(m)Z)e rddl*m+Z+m,Z,m-Z-ddl.m/Z/ddl0m1Z1edZ2edZ3ejhe5Z6d&dZ7ed'dZ8e#d(dZ9 d) d*dZ: d+dZ;e#d,dZ<GddejzZ>e# d-dZ?e#d.dZ@d/dZAejjZCeCjeCjeCjeCjeCjeCjeCjeCjeCjeCjeCjeCjeCjeCjjeCjjeCjeCjeCjeCjeCjeCjeCjhZ[ee[Z[e#d,dZ\ d0dZ] d1d Z^da_d!e`d"<d2d#Za d3d$Zb d4 d5d%Zcy)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|jjdtjjj D]+}tjjj |_-|j|S)N call_functionoptarget) graph find_nodestorchopsaten_to_copytor recompile)fx_gnodes Y/var/www/html/ia-enhanced/venv/lib/python3.12/site-packages/torch/_functorch/compilers.py _canonicalizer,/s` %% 599>>#:#:&(iinn'' ( NN Kc#Ktjjd} dtjj|y#tjj|wxYww)NF)r#_C_jit_set_autocast_mode)old_jit_autocast_flags r+_disable_jit_autocastr28sM"HH;;EB?  ''(=>''(=>s A+A A+!A((A+cDt5t||jjdtj j jD]l}t|jdk(st|jdk(s5d|jvsDtj j j|_ n|jjD]X}i}|jjD]0\}}t|tj r |j"}|||<2||_ Z|jj%|j'tj(j+|}tj,j/|jtj(j1|j3}tj(j5|}t7d|Ds||ddd|S#1swYSxYw)a 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. rrr dtypec3dK|](}t|tjj*ywN) isinstancer# _subclasses FakeTensor).0ts r+ zts_compile..ns#M1:a!2!2!=!=>Ms.0N)r2rr!r"r#r$r%r&lenargskwargsr'r nodesitemsr7devicetypelintr(jitscriptr/_jit_pass_remove_mutationfreezeevaloptimize_for_inferenceany)r)inpsr* new_kwargskvfs r+ ts_compilerQCs  JJ))uyy~~'>'>*  0D499~"s4;;'71'<DKKAW#iinn//  0 JJ$$ %DJ ))+ "1a.A ! 1  "%DK  %   II  T " **1773 II  QVVX & II , ,Q /MMM tH;< H=< HsA&H2H HE1HHcLt|jt||||S)N) clear_meta)printcoder)r)_namerSs r+_draw_graph_compilerXss! $))tTj1 Kr-c6ttt|S)NrW)rrrXrZs r+draw_graph_compiler[{s w':F GGr-c|S)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)rVs r+nopr_s  Kr-cBeZdZddd dfdZdfd ZxZS)DebugInterpreterNT initial_envenable_io_processingcZt|jg||_t||||dS)Nrb)r modulesymbol_mappingsuperrun)selfrcrdr> __class__s r+rizDebugInterpreter.runs> + KK   w{ {AU  r-c  dfd d fd d fd d fd }t|}d|jvrtj|jd\}}tj|\}}t |t |k7r"t t |dt |ttt |||D]/\ }} t| tjs"||| fd1|S) Nct|ts|Stj|jj j j}|jstd|t|S)Nzexpected r to be a number, got ) r7r sympyexpandr*exprxreplacerg is_numberAssertionErrorint)nirrjs r+ subst_symintz/DebugInterpreter.run_node..subst_symints\b&)  RWW\\2243F3FGHA;;$'Fqc%JKKq6Mr-c,tfd|DS)Nc3.K|] }|ywr6r])r:rurws r+r<zHDebugInterpreter.run_node..subst_symint_tuple..s8bb)8s)tuple)nisrws r+subst_symint_tuplez5DebugInterpreter.run_node..subst_symint_tuples8C88 8r-c|jdkDr`t|jD]H}|j||j|k7s-|j |dkDsHyy)Nrr FT)numelrangendimstridesize)abidxrws r+check_significant_stridesz.check_significant_stridessgAGGI&* =%C$QXXc]3qxx}D(59$ % r-c dt|stdt||j|jk7r,t|d|jd|j|j |j k7rKt|d|j d|j d|j ||}|sKt|d|j d|j d|j y)Nz"expected desc to be callable, got z:  != z aka )callablersrCr4rr)nvrvdesc same_stridesrr|s r+checkz(DebugInterpreter.run_node..checksD>$'I$t*%VWWxx288#$xr"((4z%JKK!"''), 9$vhb 51CBGGI1N0OtTVT[T[T]S^_5Rz+DebugInterpreter.run_node..ss'$:M:M9N&Or-)rurreturnrt)r{ztuple[IntLikeType, ...]rztuple[int, ...])r torch.Tensorrrrbool)rrrrrzCallable[[], str]rNone) rhrun_nodemetapytree tree_flattenr=rsziprr7r#Tensor)rjnrrvn_vals_n_specr_vals_r_specrrrrrwr|rks` @@@@r+rzDebugInterpreter.run_nodes  9   G Q  AFF?$11!&&-@OFG$11!4OFG6{c&k)$F }DV %FGG s6{!3VVD Q 2r!"ell3b"OP Qr-)r>rrczdict[Node, Any] | Nonerdrrr)rrrr)__name__ __module__ __qualname__rir __classcell__)rks@r+raras@/3%)    ,  #   55r-rac,t|jS)z Returns a (slow) interpreter over the FX graph module that also checks various debugging properties (e.g., that tracing strides matched real strides.) )rarir^s r+ debug_noprs D ! % %%r-ct|tjj|}tjj |j }|Sr6)rr#rErFrHrI)r)rVrPs r+simple_ts_compilers=D A "A Hr-c"t|tSr6)rr)rPs r+nnc_jitrs , --r-c0t|j|Sr6)rTrUr^s r+ print_compiler s $)) Kr-c ttttd}|j|t |t j jr t|fi|St|fi|S)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) rQrdefault_decompositionsupdater7r#nnModulerr)fnr?configs r+memory_efficient_fusionrsW6"!;0 F  MM&"ehhoo&"'''B)&))r-c|jdtd|Dcgc]}|j|jfc}dddlm}|j |t||Scc}w)NfooaQ ############################################################## # 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) r)FxModule) to_folderrTshaper4rrcudarQ)r)rLrrs r+ debug_compiler9sq NN5  &**!''177 *+, *HJOOt dD !! +sA/ rt graph_indexc Jg}t|d5}tj|}g}|D]}t|dk(r|}|t j}n|\}}}} } | t j t jt jt jt j t jt thvrt jdd|| | }nt j|| | }|j| ddd|S#1swY|SxYw)zZ Return a random input for the given inputs meta generated from _save_fx_default. rbr r)r4rBN)openpickleloadr=randomr#rtint32int64ruint8floatrandintrandappend) input_data_pathinputsrP inputs_metarrCinput_r_strider4rBs r+ get_inputsr\s"$F ot $"kk!n  "D4yA~fmmo.6:3eWeVIIKKKKJJIIKK  #]]1aeFSF"ZZU6JF MM& !' "". M/". Ms C>DD"c ddlm}d fd d  fd d fd } d fd } d fd }||||||tS) aO 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 difference graphs r)aot_module_simplifiedcg}t|dkDr1t|dtr||dz }||dz }|S|D]}t|tust|t ur|j t|fA|j t||j|j|j|jf|S)Nrr ) r=r7rzrCrtrrrrr4rB)r> input_metaargget_input_metas r+rz(_save_fx_default..get_input_metas t9q=ZQ7 .a1 1J .a1 1J  CCyC49#5!!49,/!!#Y 3::<CJJO  r-ct|jjdk(r,tjtj d|t ytj|}|jjtjjj|j |}tjdd|j!ddd|dt t#ddd|dt dd|dt dd5}t%j&||dddr7tj(|ddd|dt dd|dt d yy#1swYCxYw) Nrz!No nodes in graph {%s}_{%s}_{%s}./T)exist_okrVz.inputwbz.pt)r=r!r@logloggingWARNINGrcopydeepcopy set_codegenr#fxCodeGenr(osmakedirsrrrdumpsave) gm_to_saver> type_namegmrrP current_namedump_example_input folder_namers r+graph_saver_helperz,_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 A &  '  JJ-qa ~Qyk;-WXYeXffghqgrrsttAADE    ' 's )FF c||d|S)Nforwardr]rexample_inputsrs r+graph_saver_forwardz-_save_fx_default..graph_saver_forwards 2~y9 r-c.||dtdz a|S)Nbackwardr )rrs r+graph_saver_backwardz._save_fx_default..graph_saver_backwards  2~z:q  r-c0||dt||S)Njoint)r)r joint_argsrs r+graph_saver_jointz+_save_fx_default..graph_saver_joints 2z73 Z00r-r)r>rrz list[Any])rfx.GraphModuler>rrstrrr)rrrlist[torch.Tensor]rr)rrrrrz%tuple[fx.GraphModule, fx.GraphModule])functorch.compilerr) rrrrrrrrrrrs ``` @@r+_save_fx_defaultr{s,8 !"!*-!:=! !!F ,>   ,> 1 1(:1 .1 ! '(&-  r-c*datt|||S)as 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) r)rrr)rrrs r+graph_dumper_aotrsK #\;@R SSr-)r)rrr)rzGenerator[None, None, None])r)rrLz Sequence[Any]rtorch.jit.ScriptModule)T) r)rrVrrWrrSrrr)rWrrz5Callable[[fx.GraphModule, list[Any]], fx.GraphModule])r)rrVrrr)r)rrVrrzDCallable[[DebugInterpreter, Any, dict[Node, Any] | None, bool], Any])r)rrVrrr)rPCallable[..., Any]rr)rCallable[_P, _R] | nn.Moduler?rrr)r)rrLzSequence[torch.Tensor]rr)rrrr) rrrrrrrztorch.fx.GraphModulerrrz nn.Module)F)rrrrrrrz Callable[[bool, nn.Module], Any])d __future__rrrrrr contextlibr functoolsrtypingrrtyping_extensionsrr rnr#torch.fxrtorch.nnrtorch.utils._pytreeutils_pytreerr torch._decompr %torch.fx.experimental.symbolic_shapesr aot_autogradrrr compile_utilsr partitionersrrrcollections.abcrrr torch.fx.noder torch.typesrrr getLoggerrrr,r2rQrXr[r_ Interpreterrarrrr$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__rrrr]r-r+r4s4"  %%0 $$,>GG*=="'t_ T]g! ??, , `AE  (+9=H H:H   Er~~EP& & &I&&  . yy~~KK""HHOOMMNN-4,,BC $*$$*$*"$*N" " 6""@ S>ffff  f ' f  fVEJTT$'T=AT%Tr-