+ "ݳj28^RIHt^RIt^RIHtHtHtHt^RIH t H t H t H t H t HtHtHtHt^RIHtHtHtHt^RIHtHtHtHt^RIHtHtHt!RR ]P@PB4t"!R R ]"4t#!R R ]"4t$!RR]"4t%!RR]%4t&R#))OptionalN)PerAxisPerGroupPerRowPerToken) _DTYPE_TO_BIT_WIDTH_DTYPE_TO_QVALUE_BOUNDS MappingType_choose_scale_float8_dequantize_affine_float8_fake_quantize_affine_quantize_affine_float8_Roundchoose_qparams_affine)_get_per_token_block_sizeget_block_sizeget_group_qparams_symmetricget_groupwise_affine_qparams)FakeQuantizeConfigBaseFloat8FakeQuantizeConfigInt4WeightFakeQuantizeConfigIntxFakeQuantizeConfig) _fake_quantize_per_channel_group_fake_quantize_per_token_log_deprecation_warningc^a]tRt^0toRtV3RlRlt]V3RlRl4tV3RltRt Vt R#) FakeQuantizerBasezX Generic module for applying fake quantization to a tensor, as specified in the config. c <V^8dQhRS[/#return)str)format __classdict__s"h/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/quantization/qat/fake_quantizer.py __annotate__FakeQuantizerBase.__annotate__7s11#1c (RVP,#)zU Return a human readable representation of this `FakeQuantizer` with config details. zFakeQuantizer(%s)configselfs&r%__repr__FakeQuantizerBase.__repr__7s#T[[00r(c$<V^8dQhRS[RR/#)r r+r!rr)r#r$s"r%r&r'>s??2?7J?r(c\V\4'd \V4#\V\4'd \ V4#\V\ 4'd \ V4#\RV 24h)zUnknown config type: ) isinstancerIntxFakeQuantizerrInt4WeightFakeQuantizerrFloat8FakeQuantizer ValueErrorr*s&r% from_configFakeQuantizerBase.from_config=s` f4 5 5$V, ,  < = =*62 2  8 9 9&v. .4VH=> >r(c&<V^8dQh/S[;R&#r r+r1)r#r$s"r%r&r'0s #" r(N) __name__ __module__ __qualname____firstlineno____doc__r. staticmethodr8__annotate_func____static_attributes____classdictcell__)r$s@r%rr0s/ 11 ??r(rcTaa]tRt^ItoRtV3RlV3RlltV3RlRltRtVtV;t #)r6z_ Generic module for applying float8 fake quantization to a tensor, as specified in the config. c <V^8dQhRS[/#r;)r)r#r$s"r%r& Float8FakeQuantizer.__annotate__NsUU7Ur(c|<\SV`4WnRVn\P P R4R#)Tz,torchao.quantization.qat.Float8FakeQuantizerNsuper__init__r+enabledtorch_C_log_api_usage_oncer-r+ __class__s&&r%rLFloat8FakeQuantizer.__init__Ns-    $$%STr(cN<V^8dQhRS[PRS[P/#r xr!rNTensor)r#r$s"r%r&rHTs#%,,r(cVP'gV#VPp\VPVPP 4p\ VVVPPVPPVPPR7p\WVPP4p\WTV4pV#)) hp_value_lb hp_value_ub) rMdtypershaper+ granularityr rZr[rr )r-rVoriginal_dtype block_sizescaleqdqs&& r%forwardFloat8FakeQuantizer.forwardTs|||H#AGGT[[-D-DE $  KK   // //   $Adkk.?.? @ &q @ r(r+rM) r=r>r?r@rArLrdrDrE __classcell__rRr$s@@r%r6r6Is%UU r(r6cxaa]tRt^etoRtV3RlV3RlltV3RlRltV3RlRltV3RlR ltR t Vt V;t #) r5z Generic module for applying int4 fake quantization to a weight tensor, targeting the following MSLK kernels: torch.ops.mslk.f8i4bf16_shuffled torch.ops.mslk.bf16i4bf16_shuffled torch.ops.mslk.bf16i4bf16_rowwise c <V^8dQhRS[/#r;)r)r#r$s"r%r&$Int4WeightFakeQuantizer.__annotate__nsYY;Yr(c|<\SV`4WnRVn\P P R4R#)Tz0torchao.quantization.qat.Int4WeightFakeQuantizerNrJrQs&&r%rL Int4WeightFakeQuantizer.__init__ns-    $$%WXr(cN<V^8dQhRS[PRS[P/#r wr!rW)r#r$s"r%r&rkts'YYY%,,Yr(cVVP'gV#VPP\P8XdVP V4#VPP\P 8XdVPV4#\RVPP 24h)zUnknown activation dtype ) rMr+activation_dtyperN float8_e4m3fn_fp8_activations_forwardbfloat16_bf16_activations_forwardr7)r-rps&&r%rdInt4WeightFakeQuantizer.forwardts~|||H ;; ' '5+>+> >003 3 [[ ) )U^^ ;11!4 489U9U8VWX Xr(cN<V^8dQhRS[PRS[P/#rorW)r#r$s"r%r&rk~s#''%,,'5<<'r(c  <VP4^8XgQhVPP\P8XgQh\ VP \44p\VV\PRR7p\W\P4p\WCVP4pRp^pVPVP ^,RVPP4p\P!\P!V4RRR7p\P !W, VR7p \P"!V 4p ^VPP3p \%VV V V \P&R ^R7p V P)VP4#) z Apply int4 fake quantization to the weight tensor where the input activations are expected to be rowwise fp8, using the following as a reference: https://github.com/meta-pytorch/MSLK/blob/main/mslk/quantize/shuffle.py g-q=)rZư>Fdimkeepdimmin) quant_dtype quant_min quant_maxi)r|r+rrrNrsrr]rr rr r\view group_sizeamaxabsclamp zeros_liker int8to) r-rpper_row_block_size fp8_scalew_fp8epsmslk_scale_quant_max w_fp8_groupedmax_absra zero_pointper_group_block_sizefqs && r%rt0Int4WeightFakeQuantizer._fp8_activations_forward~sD uuw!||{{++u/B/BBBB,AGGVX>(      (e6I6IJ)%AGGD  5;;q>2t{{7M7MN **UYY}52uM G:D%%e, !4;;#9#9: "     uuQWW~r(cN<V^8dQhRS[PRS[P/#rorW)r#r$s"r%r&rks#,,5<<,ELL,r(c VP4^8XgQhVPP\P8XgQhRp^^rC^pVP \P 4PVP^,RVPP4p\P!VRRR7p\P!VRRR7p\P!Wx, VR7V, p WV,,p \P!Wh, V , 4PW44p W, p W,V ,p V PVP4P VP4#)z Apply int4 fake quantization to the weight tensor where the input activations are expected to be bf16, using the following as a reference: https://github.com/meta-pytorch/MSLK/blob/main/mslk/quantize/shuffle.py rzTr{r~r)r|r+rrrNrurfloat32rr]rraminrrapplyr\) r-rprqminqmaxmslk_symmetric_qmax w_groupedmax_valmin_valrarrs && r%rv1Int4WeightFakeQuantizer._bf16_activations_forwards uuw!||{{++u~~===dDD',,QWWQZT[[=S=ST **YB=**YB= G-37$>'::: \\9.%7 8 > >t J  % Z* $wwqww""177++r(rf) r=r>r?r@rArLrdrtrvrDrErgrhs@@r%r5r5es<YY YY''R,,,r(r5caa]tRt^toRtV3RlV3RlltV3RlRltV3RlRltV3RlR ltV3R lR lt V3R lR lt Rt Vt V;t #)r4z` Generic module for applying integer fake quantization to a tensor, as specified in the config. c <V^8dQhRS[/#r;)r)r#r$s"r%r&IntxFakeQuantizer.__annotate__s " "5 "r(c<\SV`4\PP R4WnRVnRVnRVnRVn RVn R#)z*torchao.quantization.qat.IntxFakeQuantizerTNg& .>F) rKrLrNrOrPr+rMrar _scale_eps _initializedrQs&&r%rLIntxFakeQuantizer.__init__sL  $$%QR  -1 26!r(cN<V^8dQhRS[PRS[P/#rUrW)r#r$s"r%r&rs'SSS%,,Sr(c VP'gV#VPP'd:VP'g(VPeVP f \ R4h\VPP\4'dVPV4#\VPP\\34'dVPV4#\ RVPP,4h)z Apply fake quantization to the tensor based on the bit-width, granularity, symmetry, and other properties specified in the config. zScales and zero points must be initialized for range learning. Please call `torchao.quantization.qat.initialize_fake_quantizers` before initializing the optimizer and beginning training.zUnknown granularity '%s')rMr+range_learningrrarr7r3r^r_per_token_forwardrr_per_channel_or_group_forward)r-rVs&&r%rdIntxFakeQuantizer.forwards |||H KK & & &%%%#t'>L  dkk--x 8 8**1- -  //'81D E E55a8 87$++:Q:QQR Rr(cN<V^8dQhRS[PRS[P/#rUrW)r#r$s"r%r&rs'TTELLTU\\Tr(c  4VPP'd \R4h\VPP,wr#VP 4'd\ V\P\V4VPPVVVPPVPPVPPR7 wVn VnVP4\!WPVPW#4#)z4 Perform per token fake quantization on the tensor. z(Symmetric per token is not supported yet) mapping_typer` target_dtyperrr scale_dtypezero_point_dtype)r+ is_symmetricNotImplementedErrorr r\_should_compute_qparamsrr ASYMMETRICrrscale_precisionzero_point_precisionrar(_maybe_update_qparams_for_range_learningr)r-rVrrs&& r%r$IntxFakeQuantizer._per_token_forwards ;; # # #%&PQ Q,T[[->->?   ' ' ) )*?(334Q7![[..KKOO KK77!%!A!A + 'DJ  9 9 ;'::tSSr(cN<V^8dQhRS[PRS[P/#rUrW)r#r$s"r%r&rs#5 5 u||5  5 r(c  VPPpVPPpVPPpVPPpVPP p\ V\4'd,VP^8XgQhVP4R,pM5\ V\4'dVPpM\RV,4hVP4'd\VPP,pV'd3\!VVVVVPP"R7wVnVnM1\)VVVVVPP"R7wVnVnVP&P+V4VnVP-4\.VPP,wr\1VVP$VP&V V VV4#)z Perform per channel or per group fake quantization on the tensor. We express per channel using per group where the group size is the size of the last dimension of the tensor. zUnexpected granularity '%s')rr)r+r^rrzero_point_domainrr3raxissizerrr7rrr\rrrarrrrr r) r-rVr^rrrrr bit_widthrrs && r%r/IntxFakeQuantizer._per_channel_or_group_forwards kk-- ++55#{{?? KK99{{//  k7 + +##q( (("J  X . .$//J:[HI I  ' ' ) )+DKK,=,=>I0K#  1-T_1M#  1-T_#oo001EFDO  9 9 ;,T[[->->? / JJ OO      r(c <V^8dQhRS[/#r)bool)r#r$s"r%r&r7sWWWr(c VPP;'g%VPRJ;'gVPRJ#)z? Return whether we need to compute new scales and zero points. N)r+ is_dynamicrarr,s&r%r)IntxFakeQuantizer._should_compute_qparams7s7{{%%VVt);VVtRV?VVr(c<V^8dQhRR/#)r r!Nr<)r#r$s"r%r&r=sQQ$Qr(c *VPP'di\VP\P P 4'g5\VP\P P 4'dR#VPVPr!\VPP,wr4\P!WPR7p\P P VRR7VnVPP'dVPP4R#\P!V4p\P!W#V4p\P P VRR7VnR#)z If range learning is enabled, turn scales and zero points into trainable parameters. This function is idempotent and should only be called once. Nr~T) requires_grad)r+rr3rarNnn Parameterrr r\rrrzero_rr)r-rarrrs& r%r:IntxFakeQuantizer._maybe_update_qparams_for_range_learning=s  ***$**ehh&8&899$//588+=+=>>  JJz,T[[->->?  E7XX''T'B ;; # # # OO ! ! #j1JZtr?r@rArLrdrrrrrDrErgrhs@@r%r4r4sU " "SS4TT,5 5 nWW QQQr(r4cBaa]tRtRtoRtV3RlV3RlltRtVtV;t#) FakeQuantizeriVzW (Deprecated) Please use :class:`~torchao.quantization.qat.IntxFakeQuantizer` instead. c <V^8dQhRS[/#r;r1)r#r$s"r%r&FakeQuantizer.__annotate__[s''5'r(c<<\SV`V4\V4R#)N)rKrLrrQs&&r%rLFakeQuantizer.__init__[s   &r(r<) r=r>r?r@rArLrDrErgrhs@@r%rrVs'''r(r)'typingrrN torchao.quantization.granularityrrrr%torchao.quantization.quant_primitivesrr r r r r rrrtorchao.quantization.utilsrrrrfake_quantize_configrrrrutilsrrrrModulerr6r5r4rr<r(r%rs     ??2+8V,/V,rTQ)TQp'%'r(