shEN HSSKJrJrJrJr SSKrSSKJs Js J r SSK Js Js J s J r SSKJr SSKJr SSKJr \ R&\ R&\ R(\R(1rS\\\\4\\4S\S\S\\4S jrS \\\4S \\\4S\\\\\R4444S jrS+S \R8S\\\4S\SS4SjjrS+S \R8S\S\\\44Sjjr"SS\R85r"SS\5r "SS\5r!S\S\4Sjr"S\S\4Sjr#"SS\R85r$S\R8S\R8S\%\&S \SS4 S!jr'S"r(\ S#.S$\R8S%\R8S\%\&S\\\44S&jjr)S\R8S\R8S\\\\\R4444S'jr*\!S4S\R8S\R8SS4S(jjr+\!SS).S$\R8S%\R8S\\\\\R4444S*jjr,g),)AnyCallableOptionalUnionN)prepare)&get_default_compare_output_module_liststr_listkey_strpostfixreturncxURS5nUSU:XGaSRURS5SS5nUH[nSRURS5SS5nSRURS5SS5nXF:XaUs $XG:XdMYUs $ US:XaSRURS5SS5n[U5S:XagUH[nSRURS5SS5nSRURS5SS5nXF:XaUs $XG:XdMYUs $ gg)N.r_packed_params)splitjoinlen)r r r split_str match_strings2pattern1pattern2s n/Users/tiagomarins/Projetos/claudeai/copy_bank/venv/lib/python3.13/site-packages/torch/ao/ns/_numeric_suite.py _find_matchrs1  c"I}www}}S1!B78 Bwwrxx}Qr23Hwwrxx}Qr23H' '  & &777==#5a#; HH  !RS#%K :  !K (2(=K W %,:,?K [ )  1AB  !K (2(=K W %,:,?,BK [ )IIcN R=G # " 9M(MbME((9Sb>2K"-">"F "-">"F "05H5V#%  ,6,K  )"'446q9!.get_prefixs2v76C<7rstatsN)named_children isinstanceLoggerr=_get_logger_dict_helper)r5r6r7r;namechild module_prefixs rrArA|sz8))+  eV $ $8= K 6*W4 5 , ))+ 5; 6*T1 MB,rc`[RRS5 0n[XU5 U$)aTraverse the modules and save all logger stats into target dict. This is mainly used for quantization accuracy debug. Type of loggers supported: ShadowLogger: used to log the outputs of the quantized module and its matching float shadow module, OutputLogger: used to log the outputs of the modules Args: mod: module we want to save all logger stats prefix: prefix for the current module Return: target_dict: the dictionary used to save all logger stats z/quantization_api._numeric_suite.get_logger_dict)r)r*r+rA)r5r7r6s rget_logger_dictrFs,  HH  !RS#%KCf5 rc2^\rSrSrSrU4SjrSrSrU=r$)r@zBase class for stats loggingcZ>[TU]5 0Ul[RUlgN)super__init__r=r)quint8dtypeself __class__s rrLLogger.__init__s"  \\ rcg) Nr:rPxs rforwardLogger.forwardsr)rNr= __name__ __module__ __qualname____firstlineno____doc__rLrW__static_attributes__ __classcell__rQs@rr@r@s'"  rr@c2^\rSrSrSrU4SjrSrSrU=r$) ShadowLoggerzVClass used in Shadow module to record the outputs of the original and shadow modules. c^>[TU]5 /URS'/URS'g)Nr"r#rKrLr=rOs rrLShadowLogger.__init__s*   7"$ ;rc[U5S:aUSn[U5S:aUSnURSRUR55 URSRUR55 g)rTr(rr#r"N)rr=appenddetach)rPrVys rrWShadowLogger.forwardsf q6A:!A q6A:!A ;&&qxxz2 7""188:.rr:rYras@rrcrcs% / /rrcc2^\rSrSrSrU4SjrSrSrU=r$) OutputLoggerz+Class used to log the outputs of the modulec@>[TU]5 /URS'g)N tensor_valrfrOs rrLOutputLogger.__init__s #% < rcBURSRU5 U$)rTrq)r=rirUs rrWOutputLogger.forwards < ''*rr:rYras@rrnrns6&rrntcl[U5[LaUVs/sHn[U5PM sn$U$s snfrJ)typetuple_convert_tuple_to_listrurVs rryrys059!W5Eq 1q! "1 %q 1L1L 1s1c[U5[LaUVs/sHn[U5PM sn$UR(aUR 5$U$s snfrJ)rwlist_dequantize_tensor_list is_quantized dequantizerzs rr}r}sT 7d?.//Q  #Q/ >>\\^  /sAc^\rSrSrSrU4SjrS\R4SjrS\RS\RS\R4Sjr S\RS\ S\R4S jr S\RS\RS\R4S jr S\RS\ S\R4S jr SS\\RS \S\R4S jjrS\RS\RS\R4SjrSrU=r$)ShadowaShadow module attaches the float module to its matching quantized module as the shadow. Then it uses Logger module to process the outputs of both modules. Args: q_module: module quantized from float_module that we want to shadow float_module: float module used to shadow q_module logger_cls: type of logger used to process the outputs of q_module and float_module. ShadowLogger or custom loggers can be used. c>[TU]5 XlX l[R "5UlU"5UlgrJ)rKrL orig_module shadow_modulennq DeQuantizedequantlogger)rPq_module float_module logger_clsrQs rrLShadow.__init__s1 #)~~'  l rr c[U5nUR"U6n[U5nUR"U6nUR X55 U$rT)ryrr}rr)rPrVxloutputxl_float shadow_outputs rrWShadow.forwardsI $A &!!2&*2.**H5  F* rrVrkcURRX5nUR5nUR5nURRX5nUR X45 U$r)raddrrrrPrVrkrrs rr Shadow.addW !!%%a+ LLN LLN**..q4  F* rcURRX5nUR5nURRX5nUR X45 U$r)r add_scalarrrrrs rrShadow.add_scalarL !!,,Q2 LLN**55a;  F* rcURRX5nUR5nUR5nURRX5nUR X45 U$r)rmulrrrrs rr Shadow.mul*rrcURRX5nUR5nURRX5nUR X45 U$r)r mul_scalarrrrrs rrShadow.mul_scalar6rrdimcURRX5nUVs/sHoDR5PM nnURRX5nUR X55 U$s snfr)rcatrrr)rPrVrrrkrs rr Shadow.catAs^ !!%%a-%& 'Q\\^Q '**..q6  F*  (sA)cURRX5nUR5nUR5nURRX5nUR X45 U$r)radd_relurrrrs rrShadow.add_reluLsW !!**10 LLN LLN**33A9  F* r)rrrr)r)rZr[r\r]r^rLr)TensorrWrr"rrrr|intrrr_r`ras@rrrs # U\\  U\\ ell u||  ELL U u||  U\\ ell u||  ELL U u||  T%,,' c %,,  %,, 5<< ELL  rrrrmodule_swap_listrc[RRS5 0nUR5H upVXdU'M 0nUR5H\upVXT;aM XEn[ U5U;a [ XX#5 [ U5U;dM<[ Xh5(aMN[XhU5Xu'M^ UR5HupXRU 'M g)amPrepare the model by attaching the float module to its matching quantized module as the shadow if the float module type is in module_swap_list. Example usage:: prepare_model_with_stubs(float_model, q_model, module_swap_list, Logger) q_model(data) ob_dict = get_logger_dict(q_model) Args: float_module: float module used to generate the q_module q_module: module quantized from float_module module_swap_list: list of float module types to attach the shadow logger_cls: type of logger to be used in shadow module to process the outputs of quantized module and its float shadow module z8quantization_api._numeric_suite.prepare_model_with_stubsN) r)r*r+r>rwprepare_model_with_stubs_is_identical_module_typeritems_modules) rrrrfloat_module_childrenrBr5reassign float_modr.values rrrYs, HH  B!002 &)d#3H,,.   , )/  ?"2 2 $Y5E R  ?. .7P 8 8 $CJ?HN/ nn& !&#'rcUR5Vs/sHn[U5PM nnUR5Vs/sHn[U5PM nnX4:H$s snfs snfrJ)modulesrw)mod1mod2r5mod1_module_typesmod2_module_typess rrrsQ.2lln=nscn=.2lln=nscn=  11>=s AA)r float_modelq_modelc|[RRS5 [XX#5 U"U6 [ U5nU$)aRCompare quantized module in a model with its floating point counterpart, feeding both of them the same input. Return a dict with key corresponding to module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the output tensors of quantized and its matching float shadow module. This dict can be used to compare and compute the module level quantization error. This function first call prepare_model_with_stubs() to swap the quantized module that we want to compare with the Shadow module, which takes quantized module, corresponding float module and logger as input, and creates a forward path inside to make the float module to shadow quantized module sharing the same input. The logger can be customizable, default logger is ShadowLogger and it will save the outputs of the quantized module and float module that can be used to compute the module level quantization error. Example usage:: module_swap_list = [torchvision.models.quantization.resnet.QuantizableBasicBlock] ob_dict = compare_model_stub(float_model,qmodel,module_swap_list, data) for key in ob_dict: print(key, compute_error(ob_dict[key]['float'], ob_dict[key]['quantized'].dequantize())) Args: float_model: float model used to generate the q_model q_model: model quantized from float_model module_swap_list: list of float module types at which shadow modules will be attached. data: input data used to run the prepared q_model logger_cls: type of logger to be used in shadow module to process the outputs of quantized module and its float shadow module z2quantization_api._numeric_suite.compare_model_stub)r)r*r+rrF)rrrrdataob_dicts rcompare_model_stubrs9L HH  !UV[3CP TNg&G Nrc[RRS5 [U5n[U5n0nUHOn[ X5S5S:XaM[ [ USS9US5nUcM30XE'X&SXES'X5SXES'MQ U$) aFind the matching activation between float and quantized modules. Args: float_module: float module used to generate the q_module q_module: module quantized from float_module Return: act_dict: dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the matching float and quantized activations z8quantization_api._numeric_suite.get_matching_activationsrqrT)reverser=r"r#)r)r*r+rFrrsorted)rrrract_dictr.r/s rget_matching_activationsrs HH  B!.J$X.N "H ~"<0 1Q 6 z4 @#wO  HM%/%:<%HHM' ")7)<\)JHM+ & Orc[RRS5 Uc [5n[RR R USS9nX@l[USU0S9 XAl[USU[0S9 g)aiPrepare the model by attaching the logger to both float module and quantized module if they are in the allow_list. Args: float_module: float module used to generate the q_module q_module: module quantized from float_module logger_cls: type of logger to be attached to float_module and q_module allow_list: list of module types to attach logger z5quantization_api._numeric_suite.prepare_model_outputsN) activationr!T)inplace allow_listprepare_custom_config_dict)rrobserver_non_leaf_module_listr) r)r*r+rao quantizationQConfigqconfigr*NON_LEAF_MODULE_TO_ADD_OBSERVER_ALLOW_LIST)rrrr qconfig_debugs rprepare_model_outputsrs HH  ?;= HH))11ZPT1UM( dzVX% &P#% r)rrc[RRS5 Uc [5n[ XX#5 U"U6 U"U6 [ X5nU$)aCompare output activations between float and quantized models at corresponding locations for the same input. Return a dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the activations of quantized model and float model at matching locations. This dict can be used to compare and compute the propagation quantization error. Example usage:: act_compare_dict = compare_model_outputs(float_model, qmodel, data) for key in act_compare_dict: print( key, compute_error( act_compare_dict[key]['float'], act_compare_dict[key]['quantized'].dequantize() ) ) Args: float_model: float model used to generate the q_model q_model: model quantized from float_model data: input data used to run the prepared float_model and q_model logger_cls: type of logger to be attached to float_module and q_module allow_list: list of module types to attach logger Return: act_compare_dict: dict with key corresponding to quantized module names and each entry being a dictionary with two keys 'float' and 'quantized', containing the matching float and quantized activations z5quantization_api._numeric_suite.compare_model_outputs)r)r*r+rrr)rrrrract_compare_dicts rcompare_model_outputsrsTL HH  ?;= + G TN/ E r)r)-typingrrrrr)torch.ao.nn.quantizedrnnr#rtorch.ao.nn.quantized.dynamicdynamicnnqdtorch.nntorch.ao.quantizationr+torch.ao.quantization.quantization_mappingsrLinearLSTMrdictstrr|rrr4ModulerArFr@rcrnryr}rsetrwrrrrrrr:rrrs11 ##,,)  KKJJIIGG .*DcNDI-. c] BBS#XB04S#XB #tC%& &'BPC Cc3hC C C4Cc4i. RYY $/6/. 6  McMcMssbRYYbJ0'))0'ii0'$i0' 0'  0'f2 ** YY*$i*  #t)_ *Z))ii #tC%& &'F !))!ii!  !P // YY/  #tC%& &' /r