+ "ݳj[ R^RIt^RIt^RIt^RIHtHtHtHt^RIt^RI H t H t ^RI H t HtHtHt^RIHt^RIHtHt^RIHtHt^RIHtHt^R IHt^R IHtH t H!t!H"t"H#t#^R I$H%t%H&t&H't'^R I(H)t)H*t*^R I+H,t,^RI-H.t.^RI/H0t0^RI1H2t2H3t3H4t4H5t5H6t6^RI7H8t8^RI9H:t:H;t;Ht>^RI?H@t@HAtA^RIBHCtC^RIDHEtE^RIFHGtG^RIHHItI^RIJHKtK^RILHMtM.R=OtN]P]P]P]P]P]P]P]P]P]P. tY]P]PPPP]P]PPPP/taRRltbRRltcR>RR lltdR>R!R"llteR#R$ltfR%R&ltgR'R(lthR)R*ltiR+R,ltjR-R.ltkR/R0ltlR1R2ltmR3R4ltnR?R5R6lltoR@R7R8lltpRAR9R:lltqRBR;R<lltrR#)CN)AnyCallableOptionalUnion) BackendConfigget_native_backend_config)get_fused_module_classesget_pattern_to_dtype_configsget_qat_module_classes-get_root_module_to_quantized_reference_module)quantized_decomposed_lib)convert_eq_obsupdate_obs_for_equalization)ConvertCustomConfigPrepareCustomConfig)_is_observed_module_is_observed_standalone_module)lower_to_fbgemm))_compare_prepare_convert_qconfig_mappings_generate_node_name_to_qconfig&_is_qconfig_supported_by_dtype_configs_update_qconfig_for_fusion_update_qconfig_for_qat) _get_module graph_module_from_producer_nodesnode_arg_is_weight) QConfigAnyqconfig_equals)QConfigMapping)_remove_qconfig) DeQuantStub) _parent_nameget_qparam_dictis_per_channelto_underlying_dtypeweight_is_quantized) GraphModule)ArgumentGraphNode)_USER_PRESERVED_ATTRIBUTES_KEY) NodeSourceNodeSourceAction)type_before_parametrizations) FROM_NODE_KEY)collect_producer_nodes)_is_activation_post_process)create_getattr_from_value)_assert_and_get_unique_devicecV^8dQhR\\\PP3,R\ \ \3,RR/#)modelpreserved_attrsreturnN)rr'torchnnModuledictstrr)formats"b/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/quantization/pt2e/convert.py __annotate__r@gs@ ( ( ehhoo- . (#s(^ (  (c\P!V4VP\&VP\,P4Fwr#\ WV4K R#)zXStore preserved attributes to the model.meta so that it can be preserved during deepcopyN)copymetar+itemssetattr)r6r7 attr_nameattrs&& r?attach_preserved_attrs_to_modelrIgsH 26?1KEJJ-.!::&DEKKM $'NrAcPV^8dQhR\PPRR/#)r5r6r8N)r9r:r;)r>s"r?r@r@ss"  %((// d rAc\V\4'g3\R\\ V44,R,R,4hR#)z#input model must be a GraphModule, z Please make zsure to follow the tutorials.Nz,input model must be a GraphModule, Got type:) isinstancer' ValueErrorr=type)r6s&r?_check_is_graph_modulerOssK e[ ) ) $u+  .  .   *rAcfV^8dQhR\PPR\R\\ \P P3,R\\ \\ \3,3,R\\ \3,R\\P,RR/# r5r6nodemodulesnode_name_to_scopenode_name_to_qconfig model_devicer8N r9fxr'r*r<r=r:r;tuplerNrrdevice)r>s"r?r@r@~sP#P# 88  P# P##uxx& 'P#S%T "223 P# sJ/ P# 5<<( P# P#rAc  aVPpVfQh\VP\4'gQh\ WS4wrxW!P,p \ V R4'dV P W4R#\;QJdZV3Rl\VP4\VPP44,4F 'dK RMO RMK!V3Rl\VP4\VPP44,44p V 'g\V 4'gXVPV4;_uu_4VPVP^,4VPV4RRR4R#V P p Rp \ V R4'd V P"p Rp RpV \$9EdV 'EgR pRpV P'4wpp\)V P*4'd\-V P.4p\0P2P4P6P8p\0P2P4P:P8pV P<pV P>p\AV 4pR VR VR VR VRVRV/pM\0P2P4PBP8p\0P2P4PDP8p\GV4p\-V4pV P<pV P>p\AV 4pR VR VR VRVRV/pVPV4;_uu_4VP^,pV.pVPI4FiwppVR9dL\V\F\,34'g0\KVVWx,V,VV4pVPMV4KXVPMV4Kk VPOVV\QV4/4pV!VV4V.VR,,pVPSV\QV4V !VV44p VPV 4V!V V4VPV4RRR4R#V 'EdR p\0P2P4PBPTp\AV 4pV\0PV\0PX39gQR4hV P<pV P>p\[V R\0P\4p!\[V R\0P^!\0P`4Pb4p"R VRVRV"RV/p\dV!,p#VPV4;_uu_4VP^,pVP^,.p$VPI4Fwpp%V$PMV%4K VPOR V#\QV$4/4p&VPOR \fPhV&^3/4p'VPOR \fPhV&^3/4p(VR ,pVR,pVR,p R V'R V(R VRVRV /pRRR4VPV4;_uu_4VP^,pV.pVPI4F3wppVR9dTpVPMV4K"VPMV4K5 VPOVV\QV4/4pV!VV4V.VR,,p\0P2P4PDPTpVPSV\QV4V !VV44p VPV 4V!V V4VPV4RRR4R#V \0Pj8Xd\0P2P4PlPnp)VPV4;_uu_4VP^,pVPOR V)V\0Pj3/4p*VPOR V)V*\0PF3/4p+VPV+4VPV4RRR4R#R# +'giR#;i +'giR#;i +'giELS;i +'giR#;i +'giR#;i)aUReplace activation_post_process module call node with quantize and dequantize node working with decomposed Tensor Before: ... -> observer_0(x) -> ... After: ... -> torch.ops.quantized_decomposed.quantize_per_tensor(x, ...) -> torch.ops.quantized_decomposed.dequantize_per_tensor() -> ... or quantize_per_channel and dequantize_per_channel Nconvertc3<<"TFp\VS4xK R#5iN_has_none_qconfig.0nrUs& r? M_replace_observer_with_quantize_dequantize_node_decomposed..#:A !122:FT is_dynamicc/pRVP9d7VPR,PpV\P8wdRV/pV#)val out_dtype)rDdtyper9float32) dequantize_op input_nodedequantize_op_kwargs dq_out_dtypes&& r?add_dequantize_op_kwargs\_replace_observer_with_quantize_dequantize_node_decomposed..add_dequantize_op_kwargssE! JOO #%??5177Lu}},(3\'B$##rAcb\VR\P.4.VP\&R#).replace_observer_with_quantize_dequantize_nodeN)r,r-CREATErDr/)qdq_node original_nodes&&r?!add_quantize_dequantize_node_infoe_replace_observer_with_quantize_dequantize_node_decomposed..add_quantize_dequantize_node_infos. @!(() (  m$rA call_function_scale_ _zero_point__axis_ _quant_min_ _quant_max__dtype_:NNzVonly uint8 and int8 are supported in reference flow for dynamic quantization right nowqschemeeps_eps_r|r})8graphrLtargetr=_get_module_path_and_prefixhasattrr\alllistargsuserskeys_is_conversion_supportedinserting_beforereplace_all_uses_with erase_noderlrhSUPPORTED_QDTYPEScalculate_qparamsr$rintch_axisr9opsquantized_decomposedquantize_per_channeldefaultdequantize_per_channel quant_min quant_maxr%quantize_per_tensordequantize_per_tensorfloatrEr2append create_noderYr{tensoruint8int8getattrper_tensor_affinefinformr_QSCHEME_TO_CHOOSE_QPARAMS_OPoperatorgetitemfloat16convert_element_typeno_fuse),r6rRrSrTrUrVr module_pathprefixactivation_post_processskip_replacementrlrhrrry node_type quantize_opscale zero_pointrrnrrdtype_qparamsroquantize_op_inputskey value_or_node qparam_nodequantized_node dq_inputsdequantized_noderrchoose_qparams_opchoose_qparams_op_inputsvaluechoose_qparams_node scale_nodezero_point_nodedtype_convert_opconvert_fp16_nodeconvert_fp32_nodes,&&&&f& r?:_replace_observer_with_quantize_dequantize_node_decomposedr~s& KKE    dkk3 ' '' '5 "6K&kk2& 22''4sdii4 (9#::sssdii4 (9#::78OPP # #D ) )  & &tyy| 4   T "*  $ ) )EJ& 55,77 $  !!::$ *. 3EEGz 199 : :199:G))88MMUUK ..EEMM 099I/99I(/F5 'yy6 G ))88LLTTK!II::PPXXM%LEZJ/99I/99I(/F5 yy6 G # #D ) )1J", &-mmo"]55"=5#,??#<#,s2%$ #K'--k:'--m<1'64#..;.@(A2N .nd C((+=b+AAI$22i ( C    & &'7 8 -.> E   T "[* )\ $ ii44HHOO %U+%++uzz22  - 2,55 +55 19e>U>UV-uekk%--6P6T6TU 9 9 S v  :'B  # #D ) )1J(, ! ~ $%mmo U)//6.#("3"3!2E:R4SUW# **!1!14G3KRJ$//!1!14G3KRO .I .II&E:yy5 G)*: # #D ) )1J", &-mmo"]55#0K&--k:'--m<'6#..;.@(A2N .nd C((+=b+AAI"II::PPWWM$22i ( C    & &'7 8 -.> E   T "M* )N %--  9999NNVV  # #D ) )1J % 1 1!1J 3NPR! !& 1 1!14Eu{{3SUW!   & &'8 9   T "* ) }* ) T* ) )Z* ) ):* ) )T* ) )s@4b!1Db5 Cc  DcA>c1! b2 5 c c  c. 1 d cfV^8dQhR\PPR\R\\ \P P3,R\\ \\ \3,3,R\\ \3,R\\P,RR/#rQrW)r>s"r?r@r@sO#O# 88  O# O##uxx& 'O#S%T "223 O# sJ/ O# 5<<( O# O#rAc > aVfQh\VP\4'gQhVPp\ WS4wrxW!P,p \ ;QJdZV3Rl\ VP4\ VPP44,4F 'dK RMO RMK!V3Rl\ VP4\ VPP44,44p V 'g\V 4'gXVPV4;_uu_4VPVP^,4VPV4RRR4R#V Pp Rp \V R4'd V P p V \"P$\"P&\"P(\"P*\"P,39EdV 'Eg~Rp RpV P/4wpp\1V P24'd1\5V P64pRVRVR VR V /p\"P8pM.\;V4p\5V4pRVRVR V /p\"P<pVPV4;_uu_4VP^,pV.pVP?4FMwppVR9d0\AVVWx,V,VV4pVPCV4K<VPCV4KO VPEW\GV4/4pVPIR V3R 7pVPV4VPV4RRR4R#V 'dRp \"PJp\"PLPNPPR9pR V R V/pVPV4;_uu_4VP^,pV.pVP?4FwppVPCV4K VPEW\GV4/4pVPIR V3R 7pVPV4VPV4RRR4R#V \"PR8XdRp RpR V /pVPV4;_uu_4VP^,pV.pVP?4FwppVPCV4K VPEW\GV4/4pVPIR V3R 7pVPV4VPV4RRR4R#R# +'giR#;i +'giR#;i +'giR#;i +'giR#;i)zReplace activation_post_process module call node with quantize and dequantize node Before: ... -> observer_0(x) -> ... After: ... -> torch.quantize_per_tensor(x, ...) -> x.dequantize() -> ... Nc3<<"TFp\VS4xK R#5ir^r_ras& r?rdB_replace_observer_with_quantize_dequantize_node..rfrgFTrhr{r|r}r~r dequantize)r_reduce_range_ call_methodtor)fbgemmx86)*rLrr=rrrrrrrrrrrrlrrhr9quint8qint8qint32 float8_e5m2 float8_e4m3fnrr$rrrrrrrEr2rrrYrquantize_per_tensor_dynamicbackends quantizedenginer)r6rRrSrTrUrVrrrrrrlrhrrrrrrrorrrrrr reduce_rangers&&&&f& r?/_replace_observer_with_quantize_dequantize_noders    dkk3 ' '' ' KKE5 "6K&kk2sdii4 (9#::sssdii4 (9#::78OPP # #D ) )  & &tyy| 4   T "*  $ ) )EJ& 55,77         z$ *. 3EEGz 199 : :199:G5 '5 G  44K%LEZJ %YPUVG33K # #D ) )1J", &-mmo"]55#<#,s2%$ #K'--k:'--ms"r?r@r@.s$ < <  < <  s"r?r@r@:s"  ehhoo $ rAcVPpRp\VR4'd VPpV\9;'dV'*;'gT;'gV\P 8H#)Frh)rlrrhrr9r)rrlrhs& r?rr:sh # ) )EJ& 55,77  # # 8 8Z " "  " " EMM !rAc^V^8dQhR\R\\\3,R\/#)r5rRrUr8)r(r<r=rr)r>s"r?r@r@Hs-    *.sJ*?   rAc\V\4;'d,VPV9;'dWP,RJ#)zTCheck if a node has a qconfig of None, i.e. user requested to not quantize the node N)rLr*name)rRrUs&&r?r`r`HsC 4 4 4 II- - 4 4  +t 3rAc4V^8dQhR\R\RR/#)r5observedbackend_configr8N)r'r)r>s"r?r@r@Us"))K))SW)rAc VPPFipVPR8wdKVPFCpV'gK \ W#4'gK \ V4pVfK1\ W4pV!4KE Kk R#)zExtract the subgraph that produces the weight for dynamic quant or weight only quant node and run the subgraph to observe the weight. Note that the observers of dynamic quant or weight only quant ops are run during the convert step. r{N)rnodesoprrr0r)rrrRnode_argweight_observer_nodesweight_observer_modules&& r?_run_weight_observersrUso $$ 77o %  Hx.t>>(>x(H%(0)I*&'("%rAc@V^8dQhR\R\R\RR/#)r5argrRrr8N)rr*r))r>s"r?r@r@ks(  C t E d rAc\V\4'dIVPR8Xd8VPR8Xd'VP^,pVP W4R#\V\ \34'dVFp\WAV4K R#\V\4'd&VP4Fp\WAV4K R#\P!R\V4 24R#)zwIf the arg is a dequantize Node, or a list/tuple/dict of dequantize Node, we'll recursively remove the dequantize Node rrz6Unsupported node type in recursive remove dequantize: N)rLr*rrrreplace_input_withrrY"_maybe_recursive_remove_dequantizer<valueswarningswarnrN)rrRr quantize_node arg_elements&&& r?rrks#t=!8SZZ<=W  3 C$ ' 'K .{% H C  ::s"r?r@r@sT666S%T "2236sJ/6 38_ 6rAcLVP^,p\V\4'g QRV 24hVPV9dW#P,RJMRpV'd\ VP 4pV'd V^,MRpRpVFMpVP R8XgKVP\PPP8XgKKTpM V'dTpRp MTpRp V'd)VPV9dWP,wrW3#Rp W3#)aQGiven and observer node, get the `Scope` or the fully qualified name for the submodule containing the observed node, also return a prefix of "_input" when the observed node is an input of a F.linear op, and not the output of another quantized op. TODO: this logic is hacky, we should think about how to remove it or make it more general z.Expecting observed node to be a Node, but got Nr{_input) rrLr*rrrrrr9r: functionallinear) rrTrU observed_nodeis_input_observer_onlyrfirst_linear_use_or_first_use linear_nodercrr_s &&& r?rrs!MM!$M mT * * 8H *   !5 5 //0D8  X^^$49at% Att&188uxx7J7J7Q7Q+Q  ,7 ))6% & ) . .2D D+,N,NO    rAc4V^8dQhR\R\RR/#rr)r>s"r?r@r@s&DD$DuDDrAcVPV4;_uu_4VPRV34p\VP4FpW2JgK VP W4K RRR4R# +'giR#;i)z-Inserts dequantize node for `node` in `graph`rN)inserting_afterrr<rr)rRrdequantize_node user_nodes&& r?rrs]   t $ $++L4'Bdjj)I/,,TC* % $ $ $s.A, A,, A= cV^8dQhR\R\\\PP 3,R\ \PP ,/#)r5rRrSr8)r*r<r=r9r:r;r)r>s"r?r@r@sC    c588??23  ehhoo rAcVPP4FGpVPR8XgKV\VP4,p\ V4'gKEVu# R#)zP If the node is observed, return the observer instance. Otherwise, return None. call_moduleN)rrrr=rr1)rRrSmaybe_obs_node maybe_obss&& r?_maybe_get_observer_for_nodersR**//+    -N$9$9 :;I*955  , rAc V^8dQhR\R\\\PP 3,R\P PR\R\\,RR/#)r5rRrSr6 is_referencerr8N) r*r<r=r9r:r;rXr'rrr)r>s"r?r@r@se=<=< =< #uxx& '=< 88  =< =< ]+ =<  =Q&KK**3/  :>>& -- #$q('*a/K1KK/( kk2#-"#%T[[1K G $D ;C rAc(V^8dQhR\R\\\PP 3,R\ \,R\\\3,R\R\R\R\\P,R R / #) r5rRrSobserved_node_namesrUr is_decomposedrrVr8N) r*r<r=r9r:r;setrrrrrZ)r>s"r?r@r@sQ9Q9 Q9 #uxx& 'Q9SQ9sJ/ Q9 " Q9  Q9Q95<<(Q9 Q9rAcV\VP4,pVPp Rp \V4p \ W4'dFVP p VP 4p\VP4wr\W,W4VPV9pV e\W4'g V'gR#\V4pVP\V4.4p\V V4'gR#\V 4pV'gR#RpTp\ V\ P"P$P&P(4'd TpV^,pRV/p\ V\ P$P*4'drV P-4pV P-4pV!VP.4V!VP04\3V4p\3V4pVP5RVRV/4EMh\ V\ P$P6\ P$P834'dVP:Fp\=VV4'gKVP?R4'gK0\AVV4pV P-4p V PB\ PD8Xd V !V4\3V 4VV&K MV RJpV'd<V P-4p VeTpM \GV4pV'dV PIV4V'*pV'dV'd V'gV !VP,4VP5\3V 44\KV4pVP\MV4R4pVfQR\MV4 24hVPOVV4p VeV V^&R#\VP4wr\W,V V 4R#)aConvert a weighted module to reference quantized module in the model If the QConfig of a QAT module is not set, the module will still be converted to a float module. Args: - node: The call_module node of the observed standalone module - modules: named_module of original model - observed_node_names: names for the set of observed fx node, we can skip this conversion if the node is not observed Nr/ weight_ih weight_hhweightz3No reference quantized module class configured for )(r=rqconfigr rLweight_fake_quantto_floatr"rFrr`r getrNrr&r9rr: intrinsic _FusedModule RNNCellBaser4r2r3r#updateLSTMGRU_flat_weights_namesr startswithrrlrr3rr r. from_float)!rRrSr.rUrr/rrVoriginal_moduler5weight_post_processqat_module_classesr+r is_observedpattern_to_dtype_configs dtype_configsis_weight_quantized fused_module float_module wq_or_wq_dictweight_post_process_ihweight_post_process_hhweight_qparams_ihweight_qparams_hhwnr4is_ptqrZis_qat)root_module_to_quantized_reference_moduleref_qmodule_cls ref_qmodules!&&&&&&&& r?convert_weighted_modulerVsO(c$++./O)11G/?/66.??)224(5 $d<))22K  T 8 8 Z[g>h=ij &",,\=IK% Q(5 $dK8rAcbV^8dQhR\R\R\\\\ \ 3,R3,R\R\R\\\\ \ 3,R3,R\\\\ \ 3,R3,R \R \R \/ #) r5r6rconvert_custom_configNis_standalone_module_remove_qconfig_flagqconfig_mappingrr/keep_original_weightsr8 r'rrrr<r=rrr)r>s"r?r@r@sFF FF!!4d38nd!JKF F  F >4S>4?@ F-c3h=>FF FFrAc  Vf \4p\V\4'd4\P!R\ ^R7\P !V4p\V\4'd>\P!R\ ^R7V'd\P !V4MRp\P!V4pVe\V\4'gQh\V\4'd4\P!R\ ^R7\P !V4pVf \4p\V4'gQR4hVPR,p V Pp V Pp V P p V P"p \VP%RR 74pV'dV P&p\P!V4pV P('d \+WV4\-W4\/W4\1VVVP2WZ4pV P54FUwppVV9g QR V R 24hVV,fK%\7VVV,4'dK@QR V R V RVV, 24h Tp V P8e\;W4p\=WV4\?W4^pV P@pV PBp\EV4p\GVPI44p\KV4p\MV4p\OV4p\QVP2PR4EFLpVPTR8Xd-TpV^, pVV9d\WVVP24K?KAVPTR8Xd\YV4^8XdKdTpVPZ^,p \V \P\F34'd*VF!p!\]V V!,VVP24K# K\V \^\34'd#^V9d\]V VVP24KK\P!R\aV 4 24EK!VPTR8XgEK5\cVV4p"V"fQh\eV"4'd/V'd\gVVVV V V4EKs\iVVVV V V4EK\V"\j4'd\mVVP24EK\oV"4'd\qVWW4EK\sV"4\uV4PwV4PwV49gEK\sV"4V9d\sV"^,4V9dEK:\yVVV V VVVV4EKO VP2P{4\}WP24pV'g \W W4pV'd \V4VP4VPPRR4V#)a+ We will convert an observed model (a module with observer calls) to a reference quantized model, the rule is simple: 1. for each observer module call in the graph, we'll convert it to calls to quantize and dequantize functions based on the observer instance 2. for weighted operations like linear/conv, we need to convert them to reference quantized module, this requires us to know whether the dtype configured for the weight is supported in the backend, this is done in prepare step and the result is stored in observed_node_names, we can decide whether we need to swap the module based on this set Args: * `is_standalone_module`: when this flag is True, it means we are quantizing a submodule that is not inlined in parent module, and will be quantized separately as one unit. * `is_decomposed`: a boolean flag to indicate whether we want to use the quantize operator for decomposed quantized tensor (torch.ops.quantized_decomposed.quantize_per_tensor) or default/standalone quantized tensor (torch.quantize_per_tensor) Returns: a quantized standalone module, whether input/output is quantized is specified by prepare_custom_config, with input_quantized_idxs, output_quantized_idxs, please see docs for :func:`~torch.ao.quantization.prepare_fx` for details NPassing a convert_custom_config_dict to convert is deprecated and will not be supported in a future version. Please pass in a ConvertCustomConfig instead. stacklevelzPassing a QConfig dictionary to convert is deprecated and will not be supported in a future version. Please pass in a QConfigMapping instead.zPassing a backend_config_dict to prepare is deprecated and will not be supported in a future version. Please pass in a BackendConfig instead.z-incoming model must be produced by prepare_fxrF)remove_duplicatez Expected key z in convert node_name_to_qconfigz Expected k zD to have the same value in prepare and convert QConfigMappings, but z was updated to placeholderoutputz1Unsupported node type for output_quantized_idxs: r)CrrLr<rr FutureWarning from_dictrrCdeepcopyrrrrDrTprepare_custom_configr.rU named_modulesr[rRrrrrrrEr!equalization_node_name_to_qconfigrrrinput_quantized_indexesoutput_quantized_indexesr rYrr r r3rrrrr#rrr*rNrr1rrr!rrr,r.r0unionrVeliminate_dead_coder'rr delete_all_unused_submodulespop)#r6rrXrYrZr[rr/r\observed_graph_module_attrsrTrhr.rUrSprepare_qconfig_mapping modules_copyconvert_node_name_to_qconfigkvweight_eq_obs_dictplaceholder_node_seen_cntinput_quantized_idxsoutput_quantized_idxsrSroot_module_classesrDfused_module_classesrVrRcur_placeholder_node_idx return_noderdr(mods#&&&&&&&&& r?r\r\sL$ 3 5'..  Q   !4 = =>S T/4((  L   :IN $ $_ 5d mmO4O  "j.&Q&QQ Q.$''  K   '00@24 u % %V'VV %"'**-K"L#66 $99%@$S$S#885&&&>?G ' 7 7 }}W- & - - - #O D"5:1 # (F <o( $ )..0DAq44 s"BC 4,A.:%a)Ea)HII!!%#-.J1.M-NPI 1 <"DDP9Hu'9:%0!"&;&S&S'<'U'U 6nE. I N N PQ/?3NC07LU[[&&' 77m #'@ $ % * %'+?? (ekk: @ WW ()Q.KYYq\F&4-000C6s [%++1FT4L11--6v{EKKX. GV ~VWW %dG,C? "?*3// N*,$ D*,$ C--F%++044)', .c2c:M6N6T6T"7e()* 159MM4SVs"r?r@r@s----!!4d38nd!JK- -  - >4S>4?@ --c3h=>-- --rAc Vf \4p\V\4'd4\P!R\ ^R7\P !V4p\V4VPp V U u/uF"p \W 4'gKV \W 4bK$ p p \VVVVVVVVVR7 p \W4V #uup i)z_`is_standalone_module`: see docs in :func:`~torch.ao.quantization.prepare_standalone_module_fx`r_r`)rZr[rr/r\) rrLr<rrrerfrOpreserved_attributesrrr\rI) rrrXrYr r[rr/r\preserved_attr_namesrHr7rs &&&&&&&&& r? _convert_fxrs$ 3 5'..  Q   !4 = =>S T<(0EE)(D < & *gl))( ,'%#3 I$I? 's 3B> B>c &V^8dQhR\R\\\\\ 3,R3,R\\ \\\ 3,R3,R\\\\\ 3,R3,R\/#)r5rrXNr[rr8)r'rrr<r=rrr)r>s"r?r@r@sw111 !4d38nd!JK1>4S>4?@1-c3h=> 1  1rAc &\VRVRVVRR7#)aConvert a calibrated or trained model to a reference quantized model, with decomposed representation for quantized Tensor see https://github.com/pytorch/rfcs/blob/master/RFC-0019-Extending-PyTorch-Quantization-to-Custom-Backends.md for more details, reference quantized model is a standard representation of a quantized model provided by FX Graph Mode Quantization, it can be further lowered to run on the target hardware, like accelerators Note: this is not public API Args: * `graph_module` (GraphModule): A prepared and calibrated/trained model (GraphModule) * `convert_custom_config` (ConvertCustomConfig): custom configurations for convert function. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `_remove_qconfig` (bool): Option to remove the qconfig attributes in the model after convert. * `qconfig_mapping` (QConfigMapping): config for specifying how to convert a model for quantization. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. * `backend_config` (BackendConfig): A configuration for the backend which describes how operators should be quantized in the backend. See :func:`~torch.ao.quantization.quantize_fx.convert_fx` for more details. Return: A reference quantized model (GraphModule) with operators working with decomposed quantized Tensor Example:: # prepared_model: the model after prepare_fx/prepare_qat_fx and calibration/training # TODO: add backend_config after we split the backend_config for fbgemm and qnnpack # e.g. backend_config = get_default_backend_config("fbgemm") reference_quantized_model = _convert_to_reference_decomposed_fx(prepared_model) TF)rrXr r[rr/)r)rrXr[rs&&&&r?#_convert_to_reference_decomposed_fxrs'R 3'% rA)r\r,rVr^)FFN)FNFTNNFF)NFTNNFF)NNN)srCrrtypingrrrrr9$torch.ao.quantization.backend_configrr*torch.ao.quantization.backend_config.utilsr r r r $torch.ao.quantization.fx._decomposedr "torch.ao.quantization.fx._equalizerr&torch.ao.quantization.fx.custom_configrr%torch.ao.quantization.fx.graph_modulerr(torch.ao.quantization.fx.lower_to_fbgemmr.torch.ao.quantization.fx.qconfig_mapping_utilsrrrrrtorch.ao.quantization.fx.utilsrrrtorch.ao.quantization.qconfigrr%torch.ao.quantization.qconfig_mappingrtorch.ao.quantization.quantizer torch.ao.quantization.stubsr!torch.ao.quantization.utilsr"r#r$r%r&torch.fxr'torch.fx.graphr(r)r*torch.fx.graph_moduler+torch.fx.tracebackr,r-torch.nn.utils.parametrizer.torchao.quantization.pt2er/%torchao.quantization.pt2e.graph_utilsr0"torchao.quantization.pt2e.observerr1torchao.quantization.pt2e.utilsr2 torchao.utilsr3__all__rrrrruint16int16int32rrrrrrchoose_qparamsrper_tensor_symmetricchoose_qparams_symmetricrrIrOrrrrr`rrrrrr,rVr\rrrAr?rs 11 JE E@:3!00@;C3HJE7  LL KK LL KK JJ LL KK KK    UYY;;JJQQ  > > W W ^ ^! ( P#lO#t <  ), *6rD =<@Q9hFR-`11rA