h'9 $SSKrSSKrSSKJr SSKJr SSKJrJr SSK r SSK J r SSK J r SSKJrJr SSKJr SS KJr S S KJr S rS rSr\R6"\5r\"SS9"SS55rS\ SS4SjrS\S\\4Sjr S\!S\!4Sjr"S\!S\!S\#4Sjr$S\\ RJ\ RJ/\ RJ4S\!S\!S\!4Sjr&"SS\ RNRP5r)S \S\S!\S\4S"jr*S \S\4S#jr+\"SS9"S$S%55r,\"SS9"S&S'55r-S(\!S\.4S)jr/S \S\0\\1\\.\!\2\!444S*jr3S+\0\\1\\.\!\2\ RJ44S,\0\\1\\.\!\2\ RJ44S\0\\-44S-jr4g).N)Sequence) dataclass)CallableOptional) compute_sqnr)ExportedProgram) GraphModuleNode) NodeSource) functional)bfs_trace_with_node_processnumeric_debug_handlecustom from_nodeT)frozenc.\rSrSr%Sr\\S'\\S'Srg)NodeSourceDebugInfoz Contains node source information for locating the node in the original graph. This replaces the numeric debug handle approach with direct node source info. namegraph_idN) __name__ __module__ __qualname____firstlineno____doc__str__annotations__int__static_attributes__re/home/james-whalen/.local/lib/python3.13/site-packages/torchao/quantization/pt2e/_numeric_debugger.pyrrs IMr"repreturnc0^[U[5(d[S[[535eSmS[R R SS4U4SjjnS[R R SS4U4Sjjn[X5 TS- m[X5 g) a: Attach numeric_debug_handle_id for all nodes in the graph module of the given ExportedProgram, like conv2d, squeeze, conv1d, etc, except for placeholder. Notice that nodes like getattr are out of scope since they are not in the graph. The graph nodes of input exported program are modified inplace. Here's an example of using debug handle quantize flow:: ep = torch.export.export(eager_model, example_inputs) generate_numeric_debug_handle(ep) m = ep.module() quantizer = XNNPACKQuantizer() m = prepare_pt2e(m, quantizer) m = convert_pt2e(m) z'Expected ep to be ExportedProgram, got rnoder%Nc>[TURR[05R[S55mg)Nr)maxmetaget CUSTOM_KEYNUMERIC_DEBUG_HANDLE_KEYr' unique_ids r# _find_max_id3generate_numeric_debug_handle.._find_max_idGs0 tyy}}Z4889QSTU r"c>[UR;a0UR['[UR[;a TUR[['TS- mgg)Nr )r,r*r-r.s r#_assign_debug_handle;generate_numeric_debug_handle.._assign_debug_handleMsQ TYY &$&DIIj ! #499Z+@ @>GDIIj !": ; NI Ar"r ) isinstancer ValueErrortypetorchfxr r)r$r0r3r/s @r#generate_numeric_debug_handler:,s( b/ * *5d?6K5L M  I 588== T 588==T 1 NI 9r"r'cS[SS4SjnS[S[4SjnURS:XdURS:XagU"U5(dgU"U5n[URUR S 9$) a Extract node source debug info from a node, or return None if the node does not need to be traced. Returns NodeSourceDebugInfo containing the name and graph_id from the node's greatest ancestor node source, or None if the node is not in the original graph. r'r%r cURR[5Sn[UR5S:a*URSn[UR5S:aM*U$)Nr)r*r+ FROM_NODE_KEYlenr)r' node_sources r#"_get_greatest_ancestor_node_sourceK_extract_node_source_debug_info.._get_greatest_ancestor_node_sourcelsYiimmM226 +''(1,%//3K+''(1,r"c[UR;dUR[cgUR[SRS:Xagg)NFr=z!ExportedProgram.module().unlift()T)r>r* pass_name)r's r#_is_node_in_original_graphC_extract_node_source_debug_info.._is_node_in_original_graphtsJ  )TYY}-E-M IIm $R ( 2 22 3r" placeholderoutputN)rr)r booloprrr)r'rArEgreatest_ancestor_node_sources r#_extract_node_source_debug_inforLbs|,$  ww-477h#6 %d + +$Ft$L!  * / /.77 r"xcSn[U[R5(aUR5nU$[U[[ 45(a-[ U5"UVs/sHn[U5PM sn5nU$[U[5(a0UR5VVs0sHup2U[U5_M nnnU$UnU$s snfs snnfN) r5r8Tensordetachlisttupler7_detachdictitems)rMdetachedeks r#rTrTsH!U\\""88: O Ae} % %721GAJ23 O At  ./ggi8idaAwqzMi8 O O 38s C$C ycx[U[R5(a8[U[R5(aURUR:H$[U[[ 45(a6[U[[ 45(a[ S[X555$[U[5(aF[U[5(a1SnUH'nU=(a X1;=(a [XX5nM) U$[RSX5 [U5[U5:H=(a X:H$)Nc3<# UHup[X5v M g7frO)_tensor_shape_equals).0e1e2s r# '_tensor_shape_equals..sHiFB'//isTz4Comparing non Tensors: %s and %s, they must be equal) r5r8rPshaperRrSallziprUr]logdebugr7)rMrZ all_equalrYs r#r]r]s!U\\""z!U\\'B'Bww!''!! Ae} % %*Qu *F*FHc!iHHH At  At!4!4 A!SafS2FqtQT2RI H!OAw$q'!,af,r"lossc [U[R5(ab[U[R5(aCU"UR[R5UR[R55$[U[ [ 45(aT[U[ [ 45(a9[U5"[X5VVs/sHup4[XU5PM snn5$[U[5(aG[U[5(a2UR5VVs0sHupVU[XX%5_M snn$gs snnfs snnf)a The returned loss will have the same structure as `x` and `y`, e.g. if both are Tensor, we'll return a Tensor if both are list, we'll return a list of Tensors if both are dict, we'll return a dict with the same key, and value being the loss between the two Tensors N) r5r8rPtofloat32rRrSr7re_loss_fnrUrV)rirMrZr_r`rYrXs r#rmrms!U\\""z!U\\'B'BADD'emm)<== Ae} % %*Qu *F*FAwSYGY622.YGHH At  At!4!478wwyAytq8DQT**yAA HAs E .Ec x^\rSrSrSrSrSS\S\\S\\ SS4U4S jjjr S \ S\ 4S jr S\4S jr S r U=r$) OutputLoggerz Base class for capturing output values for nodes in a GraphModule, it only captures Tensor output currently, but we can extend it to work for other types of inputs later if needed TN debug_info node_namenn_module_stackr%cT>[TU]5 X lX0lXl/UlgrO)super__init__rrrsrqstats)selfrqrrrs __class__s r#rvOutputLogger.__init__s' ".$#% r"rMcNURR[U55 U$rO)rwappendrT)rxrMs r#forwardOutputLogger.forwards '!*%r"c<SURSURS3$)Nz debug_info=z , node_name=zF, nn_module_stack={self.nn_module_stack}, num_stats={len(self.stats)}))rqrrrxs r#__extra_repr__OutputLogger.__extra_repr__s+$//*,t~~6FGS S r")rqrsrrrw)NN)rrrrr _is_impurerrrobjectrvr}rr! __classcell__)rys@r#rorost J $(,0 &' &C= &"&) &  & &F   r"romodelrqc SSKJn URRU5 U"URS35nU"U5n[ UU[ X!RURRS555 URRXQ405nSSS5 [URR55nUHnUWLaM URX5 M W$!,(df  NT=f)zFor a given node, adds an OutputLogger that observes the output of that node, and all its users use the OutputLogger output instead. The OutputLogger will contain the debug_info which can be used to compare graphs after transformsr)get_new_attr_name_with_prefix_loggerrsN)torchao.quantization.pt2e.utilsrgraphinserting_afterrsetattrror*r+ call_modulerRuserskeysreplace_input_with) rr'rqrget_new_attr_name logger_name logger_node orig_users user_nodes r#_insert_loggerrsN  $ $T *9TYYKw:OP'.    YY >O0P Q kk--k7BG  +djjoo'(J  # $$T7 ! + *s A5C%% C3c[R"U5nURRHn[ U5=ncM[ XU5 M! UR 5 U$)zAdd output loggers to unlifted node Args: model (GraphModule): original model Returns: a model with output loggers for all unlifted nodes )copydeepcopyrnodesrLr recompile)rnrqs r#"prepare_for_propagation_comparisonrsQ MM% E [[  9!< %u||U\\&BELL&PQ> >  # r"rcX\rSrSr%\\S'\\S'\\S'\\S'\\S'\\\S'Sr g ) NodeAccuracySummaryi=rqactual_node_nameactual_module_stack ref_node_nameref_module_stackresultsrN) rrrrrrrrrr!rr"r#rr=s,## 2 33r"r module_stackc[U[5(d [U5$[UR 55n[ U5S:aUSSn[U5$[U5$)zdSimplifies the stack from ("mod", "mod.foo", "mod.foo.0", "mod.foo.0.linear") to "mod.foo.0.linear" rr=)r5rUrrRvaluesr?)rmodule_values_list owning_modules r#_module_stack_to_strrGsc lD ) )<  l1134 "*2.q1 =!!<  r"c0nUR5Hfup#[U[5(dM[UR5S:dM7UR UR UR4XR'Mh U$)alFor a given model, extract the tensors stats and related information for each debug info. The reason we have a list of object, instead of Tensor is because the output of node may not be a Tensor, it could be (nested) list, tuple or dict as well. Returns: A dict is keyed by the NodeSourceDebugInfo and the values are a list of object recorded in loggers r)named_childrenr5ror?rwrrrsrq)rhandles_modules r#extract_results_from_loggersrUsmVXG))+  fl + +FLL0AA0E  && *G%% &, Nr" ref_resultsactual_resultsc 0nUR5HunupEnX1;a[RSU5 M&Xupxn [X5V V s/sHup[ XS9PM n n n [UU=(d S[U5U=(d S[U5U S9X#'M U$s sn n f![ an [ SUSUSU35U eSn A ff=f) aGiven two dict mapping from `NodeSourceDebugInfo` to list of tensors return a map from `NodeSourceDebugInfo` to `NodeAccuracySummary` that contains comparison information like SQNR, MSE etc. Args: ref_results (Dict[NodeSourceDebugInfo, Tuple[str, object, List[torch.Tensor]]]): reference results for each debug info actual_results (Dict[NodeSourceDebugInfo, Tuple[str, object, List[torch.Tensor]]]): actual results for each debug info Returns: Dict[NodeSourceDebugInfo, NodeAccuracySummary] zQCannot compare for debug info %s because it wasn't found in the transformed model)rrzFor debug_info=z from ref node z and actual node N)rqrrrrr) rVrfrgrer Exceptionr6rr)rr comparisonsrqref_name ref_stack ref_stats actual_name actual_stack actual_statsabrrXs r#compare_resultsrns&K8C8I8I8K4 4X)  + IIc  2@2L/ <  88DA-A=8 #6!(.B 4\ B".b1)< # )9L: ) !*_XJFWXcWde  s*B'B!$B'!B'' C 1CC )5rloggingcollections.abcr dataclassesrtypingrrr8torch.ao.ns.fx.utilsr torch.exportrtorch.fxr r torch.fx.tracebackr torch.nnr r graph_utilsrr-r,r> getLoggerrrfrr:rLrrTrIr]rPrmnnModulerorrrrrrrUrSrRrrrr"r#rsN $!% -(&)$41    ! $   3:o3:$3:l.$.8CINT & 588?? @ "0C >kk$ $&&&R $444 !v !# !  uXc]FDL%HI IJ21U8C=&$u||:L#LMM1U8C=&$u||:L#LMM 1 2 231r"