+ "ݳj^RIt^RIHtHt^RIHtHt^RIHtH t H t ^RI t ^RI H t ^RI Ht^RIHtHt.ROtRt!RR ]4t]!RRR7!RR ]44t]!RRR7!RR ]44t] ]]]3,]3,t]P2R8dR]n]!RRR7!RR ]44t]!RRR7!RR ]44t]!RR44t!RR]4tR#)N)ABCabstractmethod) dataclassfield)CallableOptionalUnion)Tensor)Node)ObserverOrFakeQuantize!ObserverOrFakeQuantizeConstructor QuantizerQuantizationSpecBaseQuantizationSpecFixedQParamsQuantizationSpecSharedQuantizationSpecDerivedQuantizationSpecQuantizationAnnotationquantization_annotationc]tRt^%tRtRtR#)rzBase class for different types of quantization specs that allows users to specify how to quantize a Tensor (input/output of a Node) in the model N)__name__ __module__ __qualname____firstlineno____doc____static_attributes__rn/home/luca/screen-checking/.venv/lib/python3.14/site-packages/torchao/quantization/pt2e/quantizer/quantizer.pyrr%srT)eqfrozencJa]tRt^+toRtRtRtRtRtRt Rt V3Rlt Rt Vt R#)rzQuantization spec for common operators that allows user to specify how to quantize a Tensor, this includes dtype, quant_min, quant_max etc. NFcVPePVPeBVPVP8d'\RVP RVP R24hVPeVP^8d \R4hR#R#)Nz quant_min z must be <= quant_max .zCh_axis is < 0.) quant_min quant_max ValueErrorch_axis)selfs&r __post_init__QuantizationSpec.__post_init__=s~ NN &*/T^^,,B4>>BRRST  << # q(8./ /)9 #rc<V^8dQh/S[P;R&S[;R&S[S[,;R&S[S[,;R&S[S[P ,;R&S[S[,;R&S[;R&#)dtypeobserver_or_fake_quant_ctrr%r&qschemer( is_dynamic)torchr.r rintr0bool)format __classdict__s"r __annotate__QuantizationSpec.__annotate__+sz ;; !BA}#}#emm $+c]! !rr)rrrrrr%r&r0r(r1r*__annotate_func__r__classdictcell__r6s@rrr+s1 $I#I'+G!GJ0%rc<a]tRt^OtoRtRtRtRtV3RltRt Vt R#)rNFc<V^8dQh/S[P;R&S[;R&S[;R&S[S[,;R&S[S[,;R&S[S[P ,;R&S[;R&#)r-r.scale zero_pointr%r&r0r1)r2r.floatr3rr0r4)r5r6s"rr7)FixedQParamsQuantizationSpec.__annotate__Osq ;; LO  }#  }# emm $+rr) rrrrr%r&r0r1r9rr:r;s@rrrOs" $I#I'+GJrz-torchao.quantization.pt2e.quantizer.quantizerc0a]tRt^etoRtV3RltRtVtR#)rz_ Quantization spec for the Tensors whose quantization parameters are shared with other Tensors c&<V^8dQh/S[;R&#)r- edge_or_node) EdgeOrNode)r5r6s"rr7#SharedQuantizationSpec.__annotate__esrrN)rrrrrr9rr:r;s@rrresrcDa]tRt^otoRtRtRtRtRtRt V3Rlt Rt Vt R#)rz^Quantization spec for the Tensors whose quantization parameters are derived from other TensorsNFc0<V^8dQh/S[S[,;R&S[S[S[,.S[S[S[3,3,;R&S[P;R&S[S[ ,;R&S[S[ ,;R&S[S[P,;R&S[S[ ,;R&S[ ;R&#) r- derived_fromderive_qparams_fnr.r%r&r0r(r1) listrErr tupler r2r.rr3r0r4)r5r6s"rr7$DerivedQuantizationSpec.__annotate__osz""   &ffn@U UVV  ;; }#}#emm $+c]!rr) rrrrrr%r&r0r(r1r9rr:r;s@rrros*h $I#I'+G!GJrcNa]tRt^}toRt]!]R7tRtRt Rt V3Rlt Rt Vt R#)rzHow are input arguemnt or output should be quantized, expressed as QuantizationSpec, this corresponds to how a Tensor in the operator Graph is observed (PTQ) or fake quantized (QAT) )default_factoryNTFc<V^8dQh/S[S[S[S[,3,;R&S[S[,;R&S[;R&S[;R&#)r-input_qspec_map output_qspecallow_implicit_sharing _annotated)dictr rrr4)r5r6s"rr7#QuantizationAnnotation.__annotate__}sQ$)= >>?/07&!'',-rr)rrrrrrrUrQrRrSrTr9rr:r;s@rrr}s9 CHCO 48L $(J-rc|a]tRt^toV3RlRlt]V3RlRl4t]V3RlRl4tV3RlRltR t Vt R #) rcv<V^8dQhRS[PPRS[PP/#r-modelreturnr2fx GraphModule)r5r6s"rr7Quantizer.__annotate__s1  XX))     rc V#)aAllows for user defined transforms to run before annotating the graph. This allows quantizer to allow quantizing part of the model that are otherwise not quantizable. For example quantizer can a) decompose a compound operator like scaled dot product attention, into bmm and softmax if quantizer knows how to quantize bmm/softmax but not sdpa or b) transform scalars to tensor to allow quantizing scalares. Note: this is an optional method rr)rZs&&rtransform_for_annotation"Quantizer.transform_for_annotations  rcv<V^8dQhRS[PPRS[PP/#rYr\)r5r6s"rr7r_s/  ehh22 uxx7K7K rcR#Nrras&&rannotateQuantizer.annotate rcL<V^8dQhRS[PPRR/#)r-rZr[Nr\)r5r6s"rr7r_s%  ehh22 t rcR#rfrras&&rvalidateQuantizer.validaterirch<V^8dQhRS[PPRS[S[S[3,RR/#)r-rZedge_or_node_to_obs_or_fqr[N)r2r]r^rUrEr )r5r6s"rr7r_s>xx##$( 4J(J#K  rc R#)aA callback that will be called after the observers or fake quants are created for each sharing group, but before they are inserted into the graph. The callback can be used to make final quantization adjustments, such as enforcing specific scale and zero point on model input or output. Args: * `model`: the graph module being prepared. * `edge_or_node_to_obs_or_fq`: a dictionary mapping each annotated edge and node to the corresponding observer or fake quant object. Note that multiple edges and/or nodes can map to the same observer / fake quant instance if they were annotated with SharedQuantizationSpec. This dictionary can be modified by the callback. Nr)r)rZros&&&rprepare_obs_or_fq_callback$Quantizer.prepare_obs_or_fq_callbacks$ rrN) rrrrrbrrgrlrqrr:r;s@rrrsA       r)rrrrrErrr))sysabcrr dataclassesrrtypingrrr r2r torch.fxr torchao.quantization.pt2er r __all__Q_ANNOTATION_KEYrrrrLrE version_inforrrrrrrrr~s& #(,,  -3  d4  0+ 0! 0F d4 #7! 5t$d* + gKJ d4 1! d4  2 !   0,,r