+ "ݳj-6^RIt^RIt^RIHt^RIHt^RIHtHt^RI t ^RI H t ^RI H t ^RIHt^RIHt^RIHt]P(!]4t]P/]P04]P2!]P44t]P8!R 4t]P=]4]P?]4!R R ] ]4t!RR R llt"!RR] PF4t$/3RRllt%!RR4t&]!RR]44t']!]'4RRl4t(R#)N) dataclass)Enum)OptionalUnion)quantize_per_channel_group) AOBaseConfig)'_choose_qparams_and_quantize_affine_hqq) register_quantize_module_handlerz4%(asctime)s - %(name)s - %(levelname)s - %(message)sc&]tRt^'tRtRtRtRtR#)UIntxChooseQParamsAlgorithmz[Variant of quantization algorithm to calculate scale and zero_point for UIntx quantization.min_maxhqqN)__name__ __module__ __qualname____firstlineno____doc__MIN_MAXHQQ__static_attributes__r_/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/experimental/quant_api.pyr r 'seG Crr c\V^8dQhR\PR\R\R\/#)vals group_sizenbithas_weight_zeros)torchTensorintbool)formats"r __annotate__r%9s0'2'2 ,,'2$''2/2'2FJ'2rc V^8dV^8:gQhV'd*^V^, ,)p^V^, ,^, pM^p^V,^, pVPwrxVPRV4p\P!V^R7wr\P!V^R7wrW, We, , p V'g\P !V 4p M#V\P !W, 4, p VPWx4pV PVR4p V PVR4p \TT T TTV'd\PM\PVR7pV'gRp WV 3#))axis)inputscales zero_points quant_min quant_maxdtyperN) shapereshaper minmax zeros_likeroundrint8uint8)rrrrsignedqminqmaxnkvmins_vmaxs group_scales group_zeros group_qvalss&&&&& r _quantizerC9s6 19" " tax!dQh1$T Q ::DA <<J 'DyyA&HEyyA&HEMdk2L &&|4 U[[)=>> << D''2.L%%a,K,"ejj K  k 11rctaa]tRt^ctoRV3RlV3Rlllt]P 3V3RlRlltRtRt Vt V;t #)UIntxWeightOnlyQuantizedLinearcD<V^8dQhRS[S[P,/#)rbias)rr r!)r$ __classdict__s"rr%+UIntxWeightOnlyQuantizedLinear.__annotate__ds 2 2u||$ 2rc<\SV`4WnW nVe \P !VRR7VnR#VPRR4R#)NF requires_gradrG)super__init___pack_weights_op _linear_opnn ParameterrGregister_parameter)selfpack_weight_op linear_oprG __class__s&&&&rrN'UIntxWeightOnlyQuantizedLinear.__init__dsC  .#   T?DI  # #FD 1rc <V^8dQhRS[/#)ruintx_choose_qparams_algorithm)r )r$rHs"rr%rIrs*P*P )D *Prc W nW0nV\P8XdU\ VVV^VP VP RRR7wrVrxVP\P4pV)V,pMTV\P8Xd2\WPVPRRR7wrVpV)V,pM\RV 24h\P!VRR7Vn\P!VRR7VnVP#VP%44PVP R7p \P!V RR7VnR#) r'FT)nbitsrr( compute_dtypedeviceverbose raw_output)rr8z,Unsupported uintx_choose_qparams_algorithm: rK)r^N)rrr rr r.r^tor r7rrC ValueErrorrQrR weight_scales weight_zerosrOcpupacked_weights) rTweightsrrrZ weight_qvalsrcrdr>rfs &&&&& rquantize_and_pack_weights8UIntxWeightOnlyQuantizedLinear.quantize_and_pack_weightsrs7 $ )-H-L-L L7)")--">>!#  9L(??5;;7L)==8L +/J/R/R R8A$))dSX9 5L)==8L>?]>^_  \\-uMLLUK..|/?/?/ABEE&&F !ll>OrcVP4^8gQhVP4^8Xd`VPVVPVPVPVP 4pVP eW P ,pV#VP^RpVPR,pVPP^,pVPVPRV4VPVPVPVP 4P!.VOVN5!pVP eW P ,pV#)rr/) dimrPrfrrcrdrGr0r1)rTxoutput lead_shaper<r;s&& rforward&UIntxWeightOnlyQuantizedLinear.forwards#uuw!|| 557a<__##""!! Fyy$))+MWWQr] GGBK    $ $Q ' IIb!     OO        ' "  " ! " 99 ii'F r)rPrOrGrrrfrcrd)N) rrrrrNr rrirpr__classdictcell__ __classcell__)rWrHs@@rrErEcs3 2 2&GbFiFi *P*PXrrEc8V^8dQhR\P/#)rmodulerQModule)r$s"rr%r%sNNbiiNrc VR,pVR,p\V\P4'dQhV^8dV^8:gQhVP4FwrE\V\P4'g\ WQ4K2VP P 4'g\RV R24h\\\PPRV R24\\PPRV R24VPR7p\WV4VPVP W24K R #) rrzFUIntxWeightOnlyQuantizedLinear requires contiguous weights for layer 'zF'. Please call .contiguous() on the weight tensor before quantization. _pack_weight_bit_linear_fp_act_ bit_weightrUrVrGN) isinstancerQLinearnamed_children)_replace_linear_with_quantized_linear_mpsweight is_contiguousrbrEgetattrr opstorchaorGsetattrri)rukwargsrrnamechildqlinears&& rrrs %J &>D&")),,, , 19" ",,. %++ 5e D<<--// \]a\bcZZ5&uyy'8'8M$s:ST!II%%j'IZZ G F' *  - -ellD M#/rcXa]tRt^toRRRRRRRR/V3RlRlltV3RlR ltR tVtR#) UIntxWeightOnlyLinearQuantizerr^N precisionbitwidth groupsizec <V^8dQhRS[S[,RS[S[P,RS[S[,RS[S[,/#)rr^rrr)rstrr r.r")r$rHs"rr%+UIntxWeightOnlyLinearQuantizer.__annotate__sJ,',' ,'EKK( ,' 3- ,' C= ,'rcV'dVR8wd \R4hWnV'd@V\P\P\P 39d \ R4hW nVf^p\PRV R24V\^^49d \ R4hW0n Vf^p\PRV R24VR 9d \ R4hW@n R#) mpszHOnly device=mps is currently supported in UIntxWeightOnlyLinearQuantizerz[Only precisions float32, float16 & bfloat16 are supported in UIntxWeightOnlyLinearQuantizerNz&bitwidth not specified, defaulting to .zDOnly bitwidts 1 to 7 are supported in UIntxWeightOnlyLinearQuantizerz'groupsize not specified, defaulting to zQOnly groupsizes 32, 64, 128 & 256 are supported in UIntxWeightOnlyLinearQuantizer @) NotImplementedErrorr^r float32float16bfloat16rbrloggerwarningrangerr)rTr^rrrs&$$$$rrN'UIntxWeightOnlyLinearQuantizer.__init__s fo%Z !K  MM MM NN+  m 'N  H NNCH:QO P 5A; &V %M  I NNDYKqQ R . .c 'NrcN<V^8dQhRS[PRS[P/#)rmodelreturnrv)r$rHs"rr%rs#  bii BII rcVP'dVPVP4pVP'dVPVP4p\VRVPRVP /R7V#)rr)r)r^rarrrr)rTrs&&rquantize'UIntxWeightOnlyLinearQuantizer.quantizesa ;;;HHT[[)E >>>HHT^^,E1 dnn   r)rr^rr)rrrrrNrrrrrHs@rrrsB,'!%,',0 ,' #' ,' $( ,','\  rrcVa]tRtRtoRt^t^t]Pt Rt V3Rlt Rt Vt R#)UIntxWeightOnlyConfigia[ Configuration for applying uintx weight-only asymmetric per-group quantization to linear layers for MPS devices. Args: bitwidth (int): Number of bits for quantization, must be between 1 and 7 inclusive. Default is 4. group_size (int): Group size for quantization. Must be one of [32, 64, 128, 256]. Default is 128. uintx_choose_qparams_algorithm (Union[UIntxChooseQParamsAlgorithm, str]): Algorithm for choosing quantization parameters. Options: - "min_max" (default): Simple min-max scaling - "hqq": Half-Quadratic Quantization for better accuracy c6VP\^^49d\RVP 24hVPR9d\RVP 24h\ VP \ 4'd\VP 4VnR#R#)r'z0bitwidth must be between 1 and 7 inclusive, got z2group_size must be one of [32, 64, 128, 256], got Nr)rrrbrr~rZrr )rTs&r __post_init__#UIntxWeightOnlyConfig.__post_init__,s ==a +B4==/R  ??"4 4DT__DUV  d993 ? ?2M333D / @rcT<V^8dQh/S[;R&S[;R&S[S[S[3,;R&#)rrrrZ)r"rr r)r$rHs"rr%"UIntxWeightOnlyConfig.__annotate__s9"#$%&%**Es*J$K'r)rZN)rrrrrrrr rrZr__annotate_func__rrrrs@rrrs0 HJ#++# /rrcV^8dQhR\PPR\R\PP/#)rruconfigr)r rQrwr)r$s"rr%r%=s4 HHOO%: XX__rc VPpVPpVPpVPP 4'g \ R4h\ \\PPRV R24\\PPRV R24VPR7pVPVPW#V4V#)zUIntxWeightOnlyQuantizedLinear requires contiguous weights. Please call .contiguous() on the weight tensor before quantization.ryrzr{r|r}) rrrZrrrbrErr rrrGri)rurrrrZrs&& r _uintx_weight_only_mps_transformr<s ??D""J%+%J%J" == & & ( ( R  -uyy00M$s2KL%))++tfJ-OP [[G  %% t)G Nr)T))loggingsys dataclassesrenumrtypingrrr torch.nnrQ$torch.ao.quantization.fx._decomposedrtorchao.core.configr%torchao.quantization.quant_primitivesr %torchao.quantization.transform_moduler getLoggerrrsetLevelWARNING StreamHandlerstdouthandler Formatter formatter setFormatter addHandlerrr rCrwrErrrrrrrrs  !" -   8 $     +   T U Y' #t$'2TURYYUrIKN6;;| #L# #L""789r