+ "ݳjy Rt^RIt^RIHtHt^RIHtHtHt^RI t ^RI H t ^RI H t ^RIHtHtHt^RIHt^RIHtHtHtHtHtHtHtHtHtHt^R IH t .R-Ot!RRlt"RRlt#RRlt$RRlt%!RR ]] 4t&!RR ]&4t'!RR ]'4t(!RR ]'4t)!RR]'4t*]'PW]^^] PX] PZRR7t.]'PW]R.^] P^] P`RR7t1]'PW]^^R] PX^R7t2](PW]R 7t3](PW]R 7t4]3t5]4t6]'PW]R.^] P^] PnR^R!7t8]'PW]] Pr] PX^^^^R"7t:]'PW]] Pr^] Pv^R#7t<]'PW]^^] PX] PZRR7t=])PW]^^] PXR$7t>])PW]R.^] P^] P`R%7t?])PW]R.^] P^] PnR%7t@])PW]R/^] P^] P`R0R&7tA])PW]R/^] P^] PnR0R&7tB]*PW]]P] P]!4R'7tE]*PW]]P] P]!^4R'7tH]*PW]]P] P]!^4R'7tIR(tJR)tKR*tLR+tMR,tNR#)1z6Implements modules used to perform fake quantization.N)ABCabstractmethod)AnyOptionalTuple)Node)Module)PerAxisPerGroup PerTensor)AffineQuantizedMinMaxObserver) AffineQuantizedObserverBaseFixedQParamsObserverHistogramObserver MappingTypeMovingAverageMinMaxObserver%MovingAveragePerChannelMinMaxObserverZeroPointDomain _with_args)default_fixed_qparams_range_0to1_observer,default_fixed_qparams_range_neg1to1_observer)_fake_quantize_affineFakeQuantizeBase FakeQuantizeFixedQParamsFakeQuantizeFusedMovingAvgObsFakeQuantizeAffineFakeQuantizec(V^8dQhRRR\/#qschemez torch.qschemereturnbool)formats"h/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/quantization/pt2e/fake_quantize.py __annotate__r&Gs_cfV\P\P\P39#N)torchper_channel_symmetricper_channel_affine per_channel_affine_float_qparamsr s&r%_is_per_channelr/Gs/  ##    .. r'c(V^8dQhRRR\/#rr")r$s"r%r&r&OsLLOLLr'cHV\P\P39#r))r*per_tensor_symmetricper_tensor_affiner.s&r%_is_per_tensorr4Os u1153J3JK KKr'c(V^8dQhRRR\/#rr")r$s"r%r&r&SsPPPTPr'cHV\P\P39#r))r*r2r+r.s&r%_is_symmetric_quantr7Ss u1153N3NO OOr'c(V^8dQhRRR\/#rr")r$s"r%r&r&Ws4r'c*V\P39#r))r*r-r.s&r%_is_float_qparamsr:Ws  .. r'c~aa]tRt^]toRtV3RlV3Rllt]R4t]R4t] PPRV3RlRll4t ] PPR4t ] PPRV3R lR ll4t] PPR 4t]R 4tV3R ltRtVtV;t#)raBase fake quantize module. Base fake quantize module Any fake quantize implementation should derive from this class. Concrete fake quantize module should follow the same API. In forward, they will update the statistics of the observed Tensor and fake quantize the input. They should also provide a `calculate_qparams` function that computes the quantization parameters given the collected statistics. c<V^8dQhRR/#)rr!N)r$ __classdict__s"r%r&FakeQuantizeBase.__annotate__msWW$Wr'c <\SV`4VPR\P!^.\P R74VPR\P!^.\P R74R#)z,Set fake_quant_enabled and observer_enabled.fake_quant_enableddtypeobserver_enabledN)super__init__register_bufferr*tensoruint8)self __class__s&r%rFFakeQuantizeBase.__init__msT  15<<5;;3WX /qc1UVr'cR#r)r=)rJxs&&r%forwardFakeQuantizeBase.forwardv r'c R#r)r=)rJkwargss&,r%calculate_qparams"FakeQuantizeBase.calculate_qparamszrQr'c$<V^8dQhRS[RR/#renabledr!Nr")r$r>s"r%r&r?s9999r'c8V'd^M^VP^&R#N)rArJrXs&&r%enable_fake_quant"FakeQuantizeBase.enable_fake_quant~s*1Qq"r'c(VPR4R#FN)r]rJs&r%disable_fake_quant#FakeQuantizeBase.disable_fake_quants u%r'c$<V^8dQhRS[RR/#rWr")r$r>s"r%r&r?s77t7t7r'c8V'd^M^VP^&R#rZ)rDr\s&&r%enable_observer FakeQuantizeBase.enable_observers(/1Qa r'c(VPR4R#r`)rfras&r%disable_observer!FakeQuantizeBase.disable_observers U#r'c ,\V3/VBpRVnV#)z'torchao.quantization.pt2e.fake_quantize)r __module__)clsrSfake_quant_constructors&, r% with_argsFakeQuantizeBase.with_argss$!+C!:6!:-V)%%r'cZ<V^8dQh/S[P;R&S[P;R&#)rrArDr*Tensor)r$r>s"r%r&r?]s' $ll"r'r=)T)__name__rl __qualname____firstlineno____doc__rFrrOrTr*jitexportr]rbrfri classmethodro__annotate_func____static_attributes____classdictcell__ __classcell__rKr>s@@r%rr]s WW     YY999 YY&& YY777 YY$$&&er'caa]tRt^toRt]RRR3V3Rllt]PPR4t Rt ]PPR4t V3Rlt V3R ltV3R ltR tVtV;t#) raSimulate the quantize and dequantize operations in training time. The output of this module is given by:: x_out = ( clamp(round(x/scale + zero_point), quant_min, quant_max) - zero_point ) * scale * :attr:`is_dynamic` indicates whether the fake quantie is a placeholder for dynamic quantization operators (choose_qparams -> q -> dq) or static quantization operators (q -> dq) * :attr:`scale` defines the scale factor used for quantization. * :attr:`zero_point` specifies the quantized value to which 0 in floating point maps to * :attr:`fake_quant_enabled` controls the application of fake quantization on tensors, note that statistics can still be updated. * :attr:`observer_enabled` controls statistics collection on tensors * :attr:`dtype` specifies the quantized dtype that is being emulated with fake-quantization, allowable values are torch.qint8 and torch.quint8. Args: observer (module): Module for observing statistics on input tensors and calculating scale and zero-point. observer_kwargs (optional): Arguments for the observer module Attributes: activation_post_process (Module): User provided module that collects statistics on the input tensor and provides a method to calculate scale and zero-point. NFc <\SV`4VeVeW#8:gQR4hVPR\P4p\ VR4'd)\ \ VR/4R/4PRV4p\P!V4PV8:gQR4hV\P!V4P8:gQR4hVPRVRV/4WER &V!R/VBVn VPPVn VPPVn VPPVn\VPP 4'd\P"pM\P$pVP'R \P(!R .\P"R 74VP'R \P(!^.VR 74VPP*VnVPP Vn\ VPR4'dVPP,MRVn\/VP 4'g=\1VP 4'g"QR\3VP 4,4h\/VP 4VnR#)Nz1quant_min must be less than or equal to quant_maxrCpkeywordszquant_min out of boundzquant_max out of bound quant_min quant_max is_dynamicscale?rB zero_pointch_axisr=zYOnly per channel and per tensor quantization are supported in fake quantize got qscheme: )rErFgetr*quint8hasattrgetattriinfominmaxupdateactivation_post_processrrrr:r floatintrGrHrCrr/r4stris_per_channel) rJobserverrrrobserver_kwargsrCzero_point_dtyperKs &&&&&, r%rFFakeQuantize.__init__sF   Y%:) C )$''>Ex%% #r :JKOOU;;u%))Y6 P8P P6 E 2 6 66 P8P P6  " "KK#S T(2 %'/'B/'B$55??55??66AA T99AA B B${{ $yy  WellC5 &LM \5<<CS+TU1177 33;; t33Y??  ( ( 0 0 t||,,t||0L0L  $,,  L .dll;r'c6VPP4#r)rrTras&r%rTFakeQuantize.calculate_qparams++==??r'c<VP^,^8XEd VPVP44VP4wr#VP VP P 4VP VPP 4r2VP PVP8wdKVP PVP4VPPVP4VP PV4VPPV4VP^,^8XdVP'dd\P!VVP VPVPVPP VPP"4pV#\P$!VVP VPVPP VPP"4pV#))rDrdetachrTtordevicershaperesize_copy_rArr* fake_quantize_per_channel_affinerrrfake_quantize_per_tensor_affinerJX_scale _zero_points&& r%rOFakeQuantize.forwards   #q (  ( ( 4"&"8"8": F $**++,t556 zz6<</ ""6<<0'' (9(9: JJ  V $ OO ! !+ .  " "1 % *"""::JJOOLL00::00::  99JJOO00::00:: r'cRVP RVP RVPP RVPP RVP RVP RVP RVP RVP 2#) fake_quant_enabled=, observer_enabled= , quant_min= , quant_max=, dtype= , qscheme=z , ch_axis=, scale= , zero_point=) rArDrrrrCr rrrras&r% extra_reprFakeQuantize.extra_reprs"$"9"9!::MdNcNcMde55??@ TMiMiMsMsLtuZZL 4<<. 4<<.QZZL doo-> @ r'c|<\SV`WV4VPWR,&VPWR,&R#rrN)rE_save_to_state_dictrr)rJ destinationprefix keep_varsrKs&&&&r%r FakeQuantize._save_to_state_dict s6 #KC(, W$%-1__ \)*r'c H<RR.pVEFp W),p W9dW,p V R8Xd'VPPV P4M.V R8XgQhVPPV P4\P P 4'dKV R8XdVPPV 4KV R8XgQhVPPV 4KKV'gKVPV 4EK \S V`)VVVVVVV4R#r) rrrrr*rx is_scriptingrappendrE_load_from_state_dict) rJ state_dictrlocal_metadatastrict missing_keysunexpected_keys error_msgs local_statenamekeyvalrKs &&&&&&&& r%r"FakeQuantize._load_from_state_dict's - D-C  o 7?JJ&&syy1<///OO++CII699))++w ((-#|333--c2 , ##C(- . %        r'cZ<V^8dQh/S[P;R&S[P;R&#rrrrr)r$r>s"r%r&FakeQuantize.__annotate__s+H <<IJ Kr')rrrCrrr rr)rtrlrurvrwrrFr*rxryrTrOrrrr{r|r}r~rs@@r%rrsu!P- 1R>[>[=\ ] I&1177 66AAdll++  ?$,,  +r'c2VPVP3#r))rrras&r%rT*FixedQParamsFakeQuantize.calculate_qparamsmszz4??**r'c RVP RVP RVP RVP RVP RVP P RVP P RVP 2#) z:Define a string representation of the object's attributes.rrrrrrrr) rArDrrrCrrrr ras&r%r#FixedQParamsFakeQuantize.extra_reprqs"$"9"9!::MdNcNcMdeZZL doo->?ZZL T-I-I-S-S,TU55??@ 4<<. Z r')rrr)rtrlrurvrwrFr*rxryrTrr|r}r~rs@@r%rrVsP   YY++ YY   r'caa]tRtRtoRt]^^3V3RlV3Rlllt]PPV3RlRl4t ]PPV3RlRl4t V3R lR lt R t VtV;t#) ri|aDefine a fused module to observe the tensor. Fused module that is used to observe the input tensor (compute min/max), compute scale/zero_point and fake_quantize the tensor. This module uses calculation similar MovingAverageMinMaxObserver for the inputs, to compute the min/max values in order to compute the scale/zero_point. The qscheme input in the observer is used to differentiate between symmetric/affine quantization scheme. The output of this module is given by x_out = (clamp(round(x/scale + zero_point), quant_min, quant_max)-zero_point)*scale Similar to :class:`~torchao.quantization.pt2e.FakeQuantize`, and accepts the same attributes as the base class. c 6<V^8dQhRS[RS[RS[RS[RR/#)rrrrrr!N)rr)r$r>s"r%r&*FusedMovingAvgObsFakeQuantize.__annotate__s=          r'c <\SV`!WV3/VB\VP\\ 34'gQR4hVP R\P!^.\PR74VP R\P!^.\PR74\VPP4Vn R#)zLFused observer+fake_quant module only works with MovingAverageMinMaxObserverrArBrDN) rErF isinstancerrrrGr*rHlongr7r is_symmetric_quant)rJrrrrrKs&&&&,r%rF&FusedMovingAvgObsFakeQuantize.__init__s iK?K  ( ( (*O P    [   15<<5::3VW /qc1TU"5  ( ( 0 0# r'c^<V^8dQhRS[S[PS[P3,/#rr!)tupler*rs)r$r>s"r%r&rs)@@5u||)C#D@r'c6VPP4#r)rras&r%rT/FusedMovingAvgObsFakeQuantize.calculate_qparamsrr'c <V^8dQhRS[/#rr)r$r>s"r%r&rs  C r'c,RVP RVP RVP RVP RVP RVP P RVP P RVP RVP P 2#) rrrrrrrrz, reduce_range=) rArDrrrCrrrr reduce_rangeras&r%r(FusedMovingAvgObsFakeQuantize.extra_reprs"$"9"9!::MdNcNcMdeZZL doo->htzzlS55??@ TMiMiMsMsLtu||nOD4P4P4]4]3^ ` r'cN<V^8dQhRS[PRS[P/#rrr!rr)r$r>s"r%r&rs#   %,, r'c\P!VVPVPVPP VPP VPVPVPPVPPVPPVPVPVP4 #r))r*fused_moving_avg_obs_fake_quantrDrArmin_valmax_valrraveraging_constantrrrrr)rJrs&&r%rO%FusedMovingAvgObsFakeQuantize.forwards44  ! !  # #  ( ( 0 0  ( ( 0 0 JJ OO  ( ( ; ;  ( ( 2 2  ( ( 2 2 LL     # #  r')r)rtrlrurvrwrrFr*rxryrTrrOr|r}r~rs@@r%rr|sl&4   ( YY@@ YY     r'ca]tRtRtoRt]R3V3RlRllt]PPV3RlRl4t V3RlR lt ]PPV3R lR l4t V3R lR lt V3RltRtVtR#)ria2Simulate quantize and dequantize with affine quantization in training time. This fake quantize module supports all granularities including per-tensor, per-axis, per-group, per-block, per-row, and per-token quantization through the affine quantization primitives. It works with AffineQuantizedObserverBase subclasses like AffineQuantizedMinMaxObserver and AffineQuantizedMovingAverageMinMaxObserver. The output of this module is given by:: x_out = fake_quantize_affine(x, block_size, scale, zero_point, ...) Args: observer (module): Module for observing statistics on input tensors and calculating scale and zero-point. Must be a subclass of AffineQuantizedObserverBase. is_dynamic (bool): Whether this is for dynamic quantization. **observer_kwargs: Arguments passed to the observer constructor. Attributes: activation_post_process (AffineQuantizedObserverBase): The observer instance. scale (Tensor): The quantization scale. zero_point (Tensor): The quantization zero point. target_dtype (torch.dtype): The target quantized dtype. quant_min (Optional[int]): Minimum quantization value. quant_max (Optional[int]): Maximum quantization value. zero_point_domain (ZeroPointDomain): Domain of the zero point. mapping_type (MappingType): Mapping type for quantization. is_dynamic (bool): Whether this is dynamic quantization. block_size (Optional[Tuple[int, ...]]): Block size for groupwise quantization. Example:: >>> # Per-tensor fake quantize >>> fq = AffineFakeQuantize( ... observer=AffineQuantizedMinMaxObserver, ... mapping_type=MappingType.ASYMMETRIC, ... target_dtype=torch.uint8, ... granularity=PerTensor(), ... ) >>> x = torch.randn(10, 20) >>> y = fq(x) >>> # Groupwise fake quantize >>> fq = AffineFakeQuantize( ... observer=AffineQuantizedMinMaxObserver, ... mapping_type=MappingType.SYMMETRIC, ... target_dtype=torch.int8, ... granularity=PerGroup(128), ... ) >>> x = torch.randn(10, 256) >>> y = fq(x) Fc <V^8dQhRS[/#)rrr")r$r>s"r%r&AffineFakeQuantize.__annotate__s""""""r'c \PV4W#R&V!R/VBVnVPPVnVPPVnVPP VnVPP VnVPPVnVPPVnW n VP \P8Xd\PM\PpVPR\P !R.\PR74VPR\P !^.VR74RVnRVnR#)rrrrBrNFr=)rrFr target_dtyperrzero_point_domain mapping_type granularityrrFLOATr*rrrGrH block_size _initialized)rJrrrrs&&&, r%rFAffineFakeQuantize.__init__s) !!$'(2 %DLE E $!88EE55??55??!%!=!=!O!O 88EE77CC$%%)>)>> KK  WellC5 &LM \5<<CS+TU59!r'c^<V^8dQhRS[S[PS[P3,/#r)rr*rs)r$r>s"r%r&r s)@@5u||)C#D@r'c 6VPP4#)z=Calculate the quantization parameters (scale and zero_point).rras&r%rT$AffineFakeQuantize.calculate_qparamss++==??r'cN<V^8dQhRS[PRS[P/#rrr)r$r>s"r%r&r$s#$$$%,,$r'c ^VP'gBVPVP44VPPVnRVnVP^,^8XEd<VPVP44VP 4wr#VP VPP4pVP VPP4pVPPVP8wdKVPPVP4VPPVP4VPPV4VPPV4VPPVnVP^,^8XdnVPfQR4h\VVPVPVPVPVP VP"VP$4pV#)Tzblock_size must be set before fake quantization. Ensure the model is run at least one forward pass with both observer and fake_quant disabled.)rrrrrDrTrrrrrrrrArrrrrrs&& r%rOAffineFakeQuantize.forward$s    ( ( 4"::EEDO $D    #q (  ( ( 4"&"8"8": FYYtzz001F%..)?)?@Kzz6<</ ""6<<0'' (9(9: JJ  V $ OO ! !+ ."::EEDO  " "1 % *??. > . & !!&& Ar'c <V^8dQhRS[/#rr)r$r>s"r%r&rKs   C  r'c$RVP RVP RVP RVP RVP RVP RVP RVP RVP R VP R VP 2#) rrrrz, target_dtype=z, granularity=z, mapping_type=z, zero_point_domain=rrz , block_size=) rArDrrrrrrrrrras&r%rAffineFakeQuantize.extra_reprJs"$"9"9!:; $ 5 567( T^^4DE --.nT=M=M?//* , r'cN<V^8dQhRS[PPRS[/#)rmodel observer_node)r*fx GraphModuler)r$r>s"r%r&rWs) C CUXX11 C$ Cr'c <VPPW4R#)a Converts the fake quantize node in the graph into its quantized representation. This method delegates to the underlying activation_post_process (observer)'s convert method, which handles the creation of quantize_affine/dequantize_affine operations in the graph. Args: model: graph module to convert the fake quantize node in observer_node: the fake quantize node to convert N)rconvert)rJrrs&&&r%rAffineFakeQuantize.convertWs $$,,UBr'cZ<V^8dQh/S[P;R&S[P;R&#rrr)r$r>s"r%r&rs+n <<op qr') rrrrrrrrrrN)rtrlrurvrwr rFr*rxryrTrOrrr{r|r})r>s@r%rrs{4v/ """"H YY@@$$L YY     C Cmr'T)rrrrCr rF)rrrrrCrr)rrrrCr rr)rr rCrrrr)rr rrCr)rrrrC)rrrrCr )rrrrCr eps)rrrrc\V\PP4'dXVPP P R^4^,p\P!RRV4pVR8H;'gVR8H#R#)zFReturn true if given mod is an instance of FakeQuantize script module..z\.___torch_mangle_\d+z4torchao.quantization.pt2e.fake_quantize.FakeQuantizezEtorchao.quantization.pt2e.fake_quantize.FusedMovingAvgObsFakeQuantizeF) rr*rxRecursiveScriptModule_cqualified_namesplitresub)modsuffixrs& r%_is_fake_quant_script_moduler"&sx#uyy6677&&,,S!4Q7vv.F; J J W WVW r'cx\V\4'g\V4'dVP4R#R#)zDisable fake quantization for the module. Disable fake quantization for this module, if applicable. Example usage:: # model is any PyTorch model model.apply(torchao.quantization.pt2e.disable_fake_quant) N)rrr"rbr s&r%rbrb4s/#'((,H,M,M  -Nr'cx\V\4'g\V4'dVP4R#R#)zEnable fake quantization for the module. Enable fake quantization for this module, if applicable. Example usage:: # model is any PyTorch model model.apply(torchao.quantization.pt2e.enable_fake_quant) N)rrr"r]r$s&r%r]r]As/#'((,H,M,M -Nr'cx\V\4'g\V4'dVP4R#R#)zDisable observation for this module. Disable observation for this module, if applicable. Example usage:: # model is any PyTorch model model.apply(torchao.quantization.pt2e.disable_observer) N)rrr"rir$s&r%ririNs/#'((,H,M,M -Nr'cx\V\4'g\V4'dVP4R#R#)zEnable observation for this module. Enable observation for this module, if applicable. Example usage:: # model is any PyTorch model model.apply(torchao.quantization.pt2e.enable_observer) N)rrr"rfr$s&r%rfrf[s/#'((,H,M,M -Nr')rrrrrrbrir]rfdefault_fake_quantdefault_weight_fake_quantdefault_dynamic_fake_quant.default_fixed_qparams_range_neg1to1_fake_quant+default_fixed_qparams_range_0to1_fake_quant*default_symmetric_fixed_qparams_fake_quant'default_affine_fixed_qparams_fake_quant%default_per_channel_weight_fake_quantdefault_embedding_fake_quant!default_embedding_fake_quant_4bitdefault_histogram_fake_quantdefault_fused_act_fake_quantdefault_fused_wt_fake_quant'default_fused_per_channel_wt_fake_quant(fused_wt_fake_quant_range_neg_127_to_1274fused_per_channel_wt_fake_quant_range_neg_127_to_127default_affine_fake_quantdefault_groupwise_fake_quant"default_affine_per_axis_fake_quantiig0?)Orwrabcrrtypingrrrr*torch.fxrtorch.nnrtorchao.quantizationr r r .torchao.quantization.pt2e._affine_quantizationr "torchao.quantization.pt2e.observerr rrrrrrrrr%torchao.quantization.quant_primitivesr__all__r/r4r7r:rrrrrrorr3r(qint8r2r)r*r+r,r-r.r+r/r-r0quint4x2r1r2r3r4r5r6r7 ASYMMETRICrIr8 SYMMETRICint8r9r:r"rbr]rirfr=r'r%rIsJ= #'' ==    @LP 7&sF7&t| #| ~# |# LB LB JbCbCJ"++ ( ,,  # # ,)22 ( ++  & & 3 *33 ( ,, 42J1S1S 92T2./G.P.P 6/Q/+ 3++V'(4(>(> 2 ++  ' ' )?)% ,55 2  2 2 ,,  6  %1$:$: 2  2 2 .. %;%! ,55  ,,  # # 6  =FF ( ,, G  <EE ( ++  & & F+H*Q*Q 2 ++  ' ' +R+',I+R+R ( ++  & & ,S,( "++6kk++  ,5 /88 *'' 9  2;; *&& <   &8%A%A *&& &B&"   !    r'