iTSSKJr SSKJrJr SSKJr SSKJrJ r J r J r J r J r Jr SSKrSSKJr SSKJrJr SSKJr SS KJr SS KJrJrJr SS KJr SS KJ r SS K!J"r" SSK#J$r$ SSK%J&r&J'r'J(r(J)r)J*r*J+r+J,r,J-r-J.r.J/r/J0r0J1r1J2r2 \&r3/SQr4\"SS\55r5\5RlS4\&\"\/\Rn\RpS9S4\5RrS4\)\"\/\Rn\RtS9S4\5RvS4\'\"\/\Rn\RxS9S4\5RzS4\'\"\/\Rn\RxS9\"\.\Rn\RxS94\5R|S4\,\"\/5S4\5R~S4\+\"\/\R\RxS9S4\5RvS4\(\"\1\Rn\RxS9S4\5RzS4\(\"\1\Rn\RxS9\"\0\Rn\RxS94\5R|S4\-\"\15S40 rA\"SS55rB"SS\$5rCSrDSrEg)) dataclass)IntEnumunique)partial)CallableDictListOptionalSequenceSetTupleN)QnnPassManager)constraints_loaderget_backend_opinfo)"load_backend_rules_and_constraints)NormalizedConstraints)_soc_info_table QcomChipsetQnnExecuTorchBackendType) OpOverload) GraphModule)UniformQuantizationObserverBase) Quantizer) get_16a16w_qnn_ptq_configget_16a4w_qnn_ptq_configget_16a4w_qnn_qat_configget_16a8w_qnn_ptq_configget_16a8w_qnn_qat_configget_8a4w_qnn_ptq_configget_8a8w_qnn_ptq_configget_8a8w_qnn_qat_configget_ptq_per_block_quant_config get_ptq_per_channel_quant_configget_qat_per_block_quant_config get_qat_per_channel_quant_configQuantizationConfig) QnnQuantizer QuantDtyperrrrr!r"r rr#c0\rSrSrSrSrSrSrSrSr Sr S r g ) r)Fz' bits of activation and bits of weight rrN) __name__ __module__ __qualname____firstlineno____doc__ use_16a16w use_16a8w use_16a4wuse_16a4w_blockuse_8a8wuse_8a4w__static_attributes__r0o/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/executorch/backends/qualcomm/quantizer/quantizer.pyr)r)Fs'JIIOHHr=r)F) act_dtype weight_dtypeTc\rSrSr%\R r\\S'Sr\ \S'Sr \ \S'Sr \ \S'Sr \ \\S'Sr\ \S 'Sr\ \\S 'S rS rg) ModuleQConfig quant_dtypeFis_qatis_conv_per_channelis_linear_per_channelN act_observer act_symmetricepsc l URUR4[;a&[SURSURS35e[URUR4unnnUR(a&U"UR URUR S9OU"UR UR S9UlSn/Ul[U5HonURRUR(a'U"UR URUUR S9OU"UR XPR S95 Mq [RRRRS [RRR RS [RRR"RS [RRR$R&S [RRR(R&S [RRR*RS 0Ul0UlUR0(Ga-[RRRR[RRR R[RRR"R[RRR$R&[RRR(R&/nUR.R3UVs0sH#owUR,;dMXpR,U_M% sn5 UR4(ay[RRR*R/nUR.R3UVs0sH#owUR,;dMXpR,U_M% sn5 U(aq/Ul[U5HZnUR6RUR(aU"UR URUS 9OU"UR US 95 M\ ggs snfs snf) Nz the quant config, (quant_dtype: z , is_qat: z) is not support)rIrHrJ)rIrJr,)rIrHch_axisrJ)rIrLrJrr)rIrHrL)rIrL)rDrEQUANT_CONFIG_DICT RuntimeErrorrHrIrJ quant_configper_channel_quant_config_listrangeappendtorchopsatenconv1ddefaultconv2dconv3dconv_transpose2dinputconv_transpose3dlinear op_axis_dict use_per_channel_weight_quant_opsrFupdaterGper_block_quant_config_list) selfquant_config_funcper_channel_quant_config_funcper_block_quant_config_funcpotential_axisiconv_opsk linear_opss r> __post_init__ModuleQConfig.__post_init__s   dkk *2C C243C3C2DJt{{m[kl  t//= >   ) '  "00!..HH  #1C1CR -/*~&A  . . 5 5(( 2&*&8&8%)%6%6 ! HH 7&*&8&8! '$ IINN ! ! ) )1 IINN ! ! ) )1 IINN ! ! ) )1 IINN + + 1 11 IINN + + 1 11 IINN ! ! ) )1  13-  # # # %%-- %%-- %%-- //55 //55 H  1 1 8 82:U(Q4CTCT>T(%%a(((U   % %))..//778J  1 1 8 82<W*QTEVEV@V(%%a((*W  '/1D ,>*0077 ,, 4*.*<*<)-):):$% 9*.*<*rBrBs`(11K1FD %%"'4'>BL(:;BM4C%Sr=rBcV^\rSrSrSr\R \RS4S\S\S\ 4U4Sjjjr \ S5r S \ S \ 4S jrS \ S \ 4S jrS \ S S 4SjrS\ S S 4SjrS\ S S 4SjrS\R(R*S \\4SjrS\R(R*S \\\4SjrS\R(R*4SjrS\\S S 4SjrS\\ S S 4Sjr!S\\"S S 4Sjr#S \$\"4Sjr%S\&\ \'4S S 4Sjr(S%S\)S S 4S jjr*S!r+S"\\'\\,4S S 4S#jr-S$r.U=r/$)&r(i a QnnQuantizer is a quantization annotator designed for QNN backends. It utilizes the rules_map found in the respective {backend}_rules.py file, which is a dictionary that links OpOverload to both annotator and validation functions. This mapping guides how each node is annotated and validated for quantization. During validation, the backend_opinfo pybind library containing operation details from the QNN SDK is used to verify quantization constraints and maintain backend compatibility. This library is available with QNN SDK version 2.41 or later. If the library is unavailable, QnnQuantizer will not validate quantization constraints for operations. Args: backend: QnnQuantizer uses the backend_type to dynamically load the appropriate backend rules as needed. soc_model: QnnQuantizer checks each operation according to the soc_model. For example, LPBQ requires V69 or a newer version. strict: When enabled (default), the validation stage raises a ValueError if quantization constraints are not met. In this mode, all quantization constraints must be satisfied to fully delegate to the QNN Backend. When disabled, only warnings will be logged. Example usage: quantizer = QnnQuantizer( backend=QnnExecuTorchBackendType.kHtpBackend, soc_model=QcomChipset.SM8750 ) quantizer.set_default_quant_config( quant_dtype=QuantDtype.use_8a8w, is_qat=False, is_conv_per_channel=True, is_linear_per_channel=True, act_observer=MovingAverageMinMaxObserver, ) quantizer.set_block_size_map({"conv2d": (1, 128, 1, 1)}) quantizer.set_submodule_qconfig_list([ (get_submodule_type_predicate("Add"), ModuleQConfig(quant_dtype=QuantDtype.use_16a4w)) ]) quantizer.add_custom_quant_annotations(...) quantizer.add_discard_nodes([node.name to skip annotation]) quantizer.add_discard_ops([node.target to skip annotation]) Tbackend soc_modelstrictc>[TU]5 X0l[U5Ul[ UUl[UR5uUlUl [URR55Ul URR5Ul[URU5Ul[#5Ul/Ul0Ul/Ul[5UlSUlgN)super__init__rsstrrqrsoc_infor _rules_map_constraint_cachesetkeys supported_opscopy quant_opsrbackend_opinforBdefault_quant_configsubmodule_qconfig_listblock_size_mapcustom_quant_annotations discard_nodes_recipe)rbrqrrrs __class__s r>rwQnnQuantizer.__init__5s  7| ' 2 3U LL3 ///2$//2F2F2H.I*.*<*<*A*A*C1yI$1O!  #!<>%'*u r=cUR$ru)rrbs r>recipeQnnQuantizer.recipeTs ||r=modelreturncUR(a'URRXR5 U$URU5 UR U5 U$)a Annotates GraphModule during prepare_pt2e. If a recipe is provided, it will be used to annotate the model. Otherwise, fallback to the default annotation flow. Args: model (GraphModule): The FX GraphModule to annotate. Returns: GraphModule: The annotated model. )rannotaterz _annotate_annotate_custom_annotationrbrs r>rQnnQuantizer.annotateXsI << LL ! !% 9   NN5 !  , ,U 3 r=c4[5RU5$)z Applies QNN-specific transformation before annotation during prepare_pt2e. Args: model (GraphModule): The FX GraphModule to transform. Returns: GraphModule: The transformed model. )r!transform_for_annotation_pipeliners r>transform_for_annotation%QnnQuantizer.transform_for_annotationmsAA%HHr=NcURRHnURUR;aMUR U5nU(dM9UR UR RX#UR5nUR(dMU(aM[SURSUR 35e g)NzValidation failed for node z with target ) graphnodesnamer!_get_normalized_quant_constraintsrztarget validate_fnryrs ValueError)rbrnodenormalized_constraints_listvalids r>validateQnnQuantizer.validateysKK%%DyyD...*.*P*PQU*V '** 4@@t}};;;uu$5dii[ dkk][&r=gmcURRHanURUR;aMUR U5nU(dM9UR UR RX#5 Mc g)z Annotates the nodes of the provided GraphModule in-place based on user defined quant configs during prepare_pt2e. For each node in the graph, nodes without quant config or those explicitly listed in `self.discard_nodes` are not annotated. N)rrrr_get_quant_configrzr annotate_fn)rbrrrOs r>rQnnQuantizer._annotates[ HHNNDyyD...11$7L| ,88L #r=c:URH nU"U5 M grur)rbrannotation_funcs r>r(QnnQuantizer._annotate_custom_annotations#<rQnnQuantizer._get_quant_configs<[[ b#   ,,T2,,00; ;: ;))--dkk1=GF667'A :7)C}~ A77@F * M 88 8==bAGF889GC Cr.QnnQuantizer._get_normalized_quant_constraintss[[ b#  &*# .. ??..r2999""88:;+/*@*@*D*DV*L '*2.@''/+&&**6O**r=cfURHup#U"U5(dMUs $ UR$)a Retrieves the `ModuleQConfig` for a given node by matching the first applicable callable function in the `submodule_qconfig_list`. You can add submodule-specific quant config using the `set_submodule_qconfig_list` method. Args: node (torch.fx.Node): The node for which to retrieve the quant config. Returns: ModuleQConfig: The matched submodule config, or the default config if no match is found. )rr)rbrfuncqconfigs r>r#QnnQuantizer._get_submodule_qconfigs2"88MDDzz9(((r=rcXlg)z Add custom annotation functions to be applied during prepare_pt2e. Args: custom_quant_annotations (Sequence[Callable]): A sequence of functions that take a GraphModule and perform custom annotation. Nr)rbrs r>add_custom_quant_annotations)QnnQuantizer.add_custom_quant_annotationss )A%r=rc$[U5Ulg)z2 Specifies node IDs to exclude from quantization. N)r|r)rbrs r>add_discard_nodesQnnQuantizer.add_discard_nodess!Zr=rTcLUHnURRU5 M g)z5 Specifies OpOverloads to exclude from quantization. N)rremove)rbrTrs r>add_discard_opsQnnQuantizer.add_discard_opss B NN ! !" %r=cUR$)zj Returns the set of supported OpOverloads for quantization. Returns: Set[OpOverload]: Supported ops. )r~rs r>get_supported_opsQnnQuantizer.get_supported_opss!!!r=rcXlg)z Set the mapping from node names to block sizes for per-block quantization. Args: block_size_map (Dict[str, Tuple]): Mapping from node name to block size. N)r)rbrs r>set_block_size_mapQnnQuantizer.set_block_size_map s -r=rDc ,[UUUUUUUS9Ulg)a Set the default quant config for quantizer. Args: quant_dtype (QuantDtype): Specifies the quantized data type. By default, 8-bit activations and weights (8a8w) are used. is_qat (bool, optional): Enables Quantization-Aware Training (QAT) mode. Defaults to Post-Training Quantization (PTQ) mode. is_conv_per_channel (bool, optional): Enables per-channel quantization for convolution operations. is_linear_per_channel (bool, optional): Enables per-channel quantization for linear (fully connected) operations. act_observer (Optional[UniformQuantizationObserverBase], optional): Custom observer for activation quantization. If not specified, the default observer is determined by `QUANT_CONFIG_DICT`. )rErFrGrHrIrJN)rBr)rbrDrErFrGrHrIrJs r>set_default_quant_config%QnnQuantizer.set_default_quant_configs&*%2  3"7%'% !r=cZXlURRUR5 gru)rinitialize_default_strategy_opsr~)rbrs r> set_recipeQnnQuantizer.set_recipe1s  44T5G5GHr=rcXlg)zv Set specific quant config from a callback function. If a node fits more than one callback, only apply the first one. N)r)rbrs r>set_submodule_qconfig_list'QnnQuantizer.set_submodule_qconfig_list5s '=#r=)r{rrzrqrrrrrrryrsrr~)FFFNFN)0r1r2r3r4r5r kHtpBackendrSM8750rnrwpropertyrrrrrrrrSfxNoder r'rr rrrr rrrxrrrr rrr rr)rrrBrr< __classcell__)rs@r>r(r( s'V-E,P,P!,!3!3 ) >kk* Ik Ik Ikd MK MD M k d %>ehhmm%>AS8T%>N+HHMM+ $,- .+,)588==) A(0(: A  A(x}(( &8J#7&D&"3z?"-c5j1A-d-!#     >I=&*5=1H+I&J= ==r=r(c^U4SjnU$) An example of nn_module_stack { 'L__self__': ('', 'executorch.backends.qualcomm.tests.models.SubModules'), 'L__self___add': ('add', 'executorch.backends.qualcomm.tests.models.Add') } c>URRS5=n(a!UR5H up#TU;dM g gNnn_module_stackTF)metarvalues)rr_ type_namemodule_type_strs r> predicate/get_submodule_type_predicate..predicateHsA"iimm,=> >? > / 6 6 8 "i/!9r=r0)rrs` r>get_submodule_type_predicater? r=c^U4SjnU$)rc>URRS5=n(aUR5H nTU;dM g gr)rrr})rrrmodule_name_strs r>r/get_submodule_name_predicate..predicate[s?"iimm,=> >? >',,."d*/r=r0)rrs` r>get_submodule_name_predicaterRrr=)F dataclassesrenumrr functoolsrtypingrrr r r r r rS5executorch.backends.qualcomm._passes.qnn_pass_managerr=executorch.backends.qualcomm.quantizer.backend_opinfo_adapterrr6executorch.backends.qualcomm.quantizer.registry_loaderr1executorch.backends.qualcomm.quantizer.validatorsr4executorch.backends.qualcomm.serialization.qc_schemarrr torch._opsrtorch.fxrtorchao.quantization.pt2er#torchao.quantization.pt2e.quantizerrrrrrrrr r!r"r#r$r%r&r' get_default_16bit_qnn_ptq_config__all__r)r6uint16int16r7int8r8int4r9r:r;uint8rMrBr(rrr0r=r>r s " GGG P T " E9$$=     E"! ,ll %5!  ,ll $5!  ,ll $'  ,ll  *ll  *% 01 # %  ,kk #4   ,ll #&  ,ll  *ll  )$01 "_Tn \\ \~q=9q=h &r=