i- SSKrSSKrSSKJr SSKJr SSKJrJr SSK r SSK J r SSK J r SSKJr SSKJr SSKJr \R("\5r\R/\R05 \R2"\R45r\R8"S 5r\R=\5 \R?\5 "S S \ \5r!SS \ RDS \#S\#S\$4Sjjr%"SS\ RL5r'04S\ RL4Sjjr("SS5r)\"SS\55r*\"\*5S\ RRLS\*S\ RRL4Sj5r+g)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$\rSrSrSrSrSrSrg)UIntxChooseQParamsAlgorithm'z[Variant of quantization algorithm to calculate scale and zero_point for UIntx quantization.min_maxhqqN)__name__ __module__ __qualname____firstlineno____doc__MIN_MAXHQQ__static_attributes__r]/var/www/html/ai-image-ml/venv/lib/python3.13/site-packages/torchao/experimental/quant_api.pyr r 'seG Crr vals group_sizenbithas_weight_zerosc vUS:aUS::deU(aSUS- -*nSUS- -S- nO SnSU-S- nURupxURSU5n[R"USS9up[R"USS9upX- Xe- - n U(d[R "U 5n OU[R "X- 5- n URXx5nU RUS5n U RUS5n [UU U UUU(a[RO[RUS9nU(dSn XU 4$)Nr)axis)inputscales zero_points quant_min quant_maxdtyper) shapereshapetorchminmax zeros_likeroundrint8uint8)rrrrsignedqminqmaxnkvmins_vmaxs group_scales group_zeros group_qvalss r _quantizer>9s6 19" " tax!dQh1$T Q ::DA <<J 'DyyA&HEyyA&HEMdk2L &&|4 U[[)=>> << D''2.L%%a,K,"ejj K  k 11rc~^\rSrSrSS\\R 4U4Sjjjr\R4S\4Sjjr Sr Sr U=r $) UIntxWeightOnlyQuantizedLinearcbiasc>[TU]5 XlX lUb[R "USS9UlgURSS5 g)NF requires_gradrB)super__init___pack_weights_op _linear_opnn ParameterrBregister_parameter)selfpack_weight_op linear_oprB __class__s rrG'UIntxWeightOnlyQuantizedLinear.__init__dsC  .#   T?DI  # #FD 1ruintx_choose_qparams_algorithmc ~X lX0lU[R:XaM[ UUUSUR UR SSS9upVpxUR[R5nU*U-nOLU[R:Xa*[XRURSSS9upVnU*U-nO[SU35e[R"USS9Ul[R"USS9UlUR#UR%55RUR S9n [R"U SS9Ulg) Nr FT)nbitsrr# compute_dtypedeviceverbose raw_output)rr3z,Unsupported uintx_choose_qparams_algorithm: rD)rV)rrr rr r)rVtor,r2rr> ValueErrorrJrK weight_scales weight_zerosrHcpupacked_weights) rMweightsrrrR weight_qvalsr[r\r9r^s rquantize_and_pack_weights8UIntxWeightOnlyQuantizedLinear.quantize_and_pack_weightsrs7 $ )-H-L-L L7)")--">>!#  9L(??5;;7L)==8L +/J/R/R R8A$))dSX9 5L)==8L>?]>^_  \\-uMLLUK..|/?/?/ABEE&&F !ll>OrcpUR5S:deUR5S:XaZURUURURURUR 5nUR bX R -nU$URSSnURSnURRSnURURSU5URURURUR 5R"/UQUP76nUR bX R -nU$)Nrr") dimrIr^rr[r\rBr*r+)rMxoutput lead_shaper7r6s rforward&UIntxWeightOnlyQuantizedLinear.forwards#uuw!|| 557a<__##""!! Fyy$))+MWWQr] GGBK    $ $Q ' IIb!     OO        ' "  " ! " 99 ii'F r)rIrHrBrrr^r[r\)N)rrrrrr,TensorrGr rrarir __classcell__)rPs@rr@r@csO (, 2u||$ 2 2&GbFiFi *P )D *PXrr@modulec nUSnUSn[U[R5(aeUS:aUS::deUR5HupE[U[R5(d [ XQ5 M1UR R 5(d[SUS35e[[[RRSUS35[[RRS US 35URS 9n[XU5 URUR X25 M g) Nrrr zFUIntxWeightOnlyQuantizedLinear requires contiguous weights for layer 'zF'. Please call .contiguous() on the weight tensor before quantization. _pack_weight_bit_linear_fp_act_ bit_weightrNrOrB) isinstancerJLinearnamed_children)_replace_linear_with_quantized_linear_mpsweight is_contiguousrZr@getattrr,opstorchaorBsetattrra)rmkwargsrrnamechildqlinears rrxrxs %J &>D&")),,, , 19" ",,. %++ 5e D<<--// \]a\bcZZ5&uyy'8'8M$s:ST!II%%j'IZZ G F' *  - -ellD M#/rc \rSrSrSSSSS.S\\S\\RS\\S\\4Sjjr S \ RS \ R4S jr S r g) UIntxWeightOnlyLinearQuantizerN)rV precisionbitwidth groupsizerVrrrcU(aUS:wa [S5eXlU(a>U[R[R[R 4;a [ S5eX lUcSn[RSUS35 U[SS5;a [ S 5eX0l UcS n[RS US35 US ;a [ S 5eX@l g)NmpszHOnly device=mps is currently supported in UIntxWeightOnlyLinearQuantizerz[Only precisions float32, float16 & bfloat16 are supported in UIntxWeightOnlyLinearQuantizerz&bitwidth not specified, defaulting to .r r!zDOnly bitwidts 1 to 7 are supported in UIntxWeightOnlyLinearQuantizerz'groupsize not specified, defaulting to  @rzQOnly groupsizes 32, 64, 128 & 256 are supported in UIntxWeightOnlyLinearQuantizer) NotImplementedErrorrVr,float32float16bfloat16rZrloggerwarningrangerr)rMrVrrrs rrG'UIntxWeightOnlyLinearQuantizer.__init__s fo%Z !K  MM MM NN+  m 'N  H NNCH:QO P 5A; &V %M  I NNDYKqQ R . .c 'NrmodelreturncUR(aURUR5nUR(aURUR5n[UURUR S.S9 U$)N)rr)r)rVrYrrxrr)rMrs rquantize'UIntxWeightOnlyLinearQuantizer.quantizesY ;;HHT[[)E >>HHT^^,E1 "nn   r)rrVrr)rrrrrstrr,r)intrGrJModulerrrrrrrsn!%+/"&#' ,' ,'EKK( ,' 3- ,' C= ,'\ bii BII rrch\rSrSr%SrSr\\S'Sr\\S'\ Rr \ \ \ 4\S'SrS rg ) 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 rrrrrRc,UR[SS5;a[SUR35eURS;a[SUR35e[ UR [ 5(a[UR 5Ulgg)Nr r!z0bitwidth must be between 1 and 7 inclusive, got rz2group_size must be one of [32, 64, 128, 256], got )rrrZrrurRrr )rMs r __post_init__#UIntxWeightOnlyConfig.__post_init__,s ==a +B4==/R  ??"4 4DT__DUV  d993 ? ?2M333D / @r)rRN)rrrrrrr__annotations__rr rrRrrrrrrrrrsB HcJ#++#E*Es*J$K rrconfigrc URnURnURnURR 5(d [ S5e[ [[RRSUS35[[RRSUS35URS9nURURX#U5 U$)NzUIntxWeightOnlyQuantizedLinear requires contiguous weights. Please call .contiguous() on the weight tensor before quantization.rprqrrrsrt) rrrRryrzrZr@r{r,r|r}rBra)rmrrrrRrs r _uintx_weight_only_mps_transformr<s ??D""J%+%J%J" == & & ( ( R  -uyy00M$s2KL%))++tfJ-OP [[G  %% t)G Nr)T),loggingsys dataclassesrenumrtypingrrr,torch.nnrJ$torch.ao.quantization.fx._decomposedrtorchao.core.configr%torchao.quantization.quant_primitivesr %torchao.quantization.transform_moduler getLoggerrrsetLevelWARNING StreamHandlerstdouthandler Formatter formatter setFormatter addHandlerrr rkrboolr>rr@rxrrrrrrrsi !" -   8 $     +   T U Y' #t&TX'2 ,,'2$''2/2'2FJ'2TURYYUrIKNbiiN6;;| #L# #L""78 HHOO%: XX__9r