g^(hddlmZmZddlmZmZddlmZmZm Z ddl Z ddl m Z ddl m Z ddlmZddlmZgd ZGd d eZed d GddeZed d GddeZ e eeefefZde_ed d GddeZed d GddeZeGddZGddeZy))ABCabstractmethod) dataclassfield)CallableOptionalUnionN)Tensor)ObserverOrFakeQuantize)"_ObserverOrFakeQuantizeConstructor)Node) QuantizerQuantizationSpecBaseQuantizationSpecFixedQParamsQuantizationSpec EdgeOrNodeSharedQuantizationSpecDerivedQuantizationSpecQuantizationAnnotationceZdZdZy)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____doc__g/var/www/skyplay_api_hub/venv/lib/python3.12/site-packages/torch/ao/quantization/quantizer/quantizer.pyrrsrrT)eqfrozenceZdZUdZej ed<eed<dZe e ed<dZ e e ed<dZ e ejed<dZ e e ed<d Zeed <d Zy) rzQuantization spec for common operators that allows user to specify how to quantize a Tensor, this includes dtype, quant_min, quant_max etc. dtypeobserver_or_fake_quant_ctrN quant_min quant_maxqschemech_axisF is_dynamicc|jK|j?|j|jkDr&td|jd|jd|j|jdkr tdyy)Nz quant_min z must be <= quant_max .rzCh_axis is < 0.)r#r$ ValueErrorr&)selfs r __post_init__zQuantizationSpec.__post_init__1s~ NN &*/T^^,,B4>>BRRST  << # q(8./ /)9 #r)rrrrtorchr!__annotations__r r#rintr$r%r&r'boolr,rrrrrsh ;; !CB#Ix}##Ix}#'+GXemm $+!GXc]!J0rrceZdZUejed<eed<eed<dZe eed<dZ e eed<dZ e ejed<dZ e ed <y) rr!scale zero_pointNr#r$r%Fr')rrrr-r!r.floatr/r#rr$r%r'r0rrrrrCsO ;; LO#Ix}##Ix}#'+GXemm $+Jrrz)torch.ao.quantization.quantizer.quantizerceZdZUdZeed<y)rzg Quantization spec for the Tensors whose quantization parameters are shared with other Tensors edge_or_nodeN)rrrrrr.rrrrrXs rrceZdZUdZeeed<eeege e e ffed<e jed<dZ eeed<dZeeed<dZee j"ed<dZeeed <d Zeed <y) rz^Quantization spec for the Tensors whose quantization parameters are derived from other Tensors derived_fromderive_qparams_fnr!Nr#r$r%r&Fr')rrrrlistrr.rr tupler r-r!r#rr/r$r%r&r'r0rrrrrbshz""&ffn@U UVV ;;#Ix}##Ix}#'+GXemm $+!GXc]!JrrcneZdZUdZeeZeeee fe d<dZ ee e d<dZ e e d<dZe e d <y) rzHow are input arguemnt or output should be quantized, expressed as QuantizationSpec, this corresponds to how a Tensor in the operator Graph is observed (PTQ) or fake quantized (QAT) )default_factoryinput_qspec_mapN output_qspecTallow_implicit_sharingF _annotated)rrrrrdictr>r rrr.r?r@r0rArrrrrps[ CHCOT$)= >>? 48L(/07 $(D'Jrrc^eZdZdejj dejj fdZedejj dejj fdZedejj ddfdZ dejj de e e fddfdZ y) rmodelreturnc|S)aAllows for user defined transforms to run before annotating the graph. This allows quantizer to allow quantizing part of the model that are otherwise not quantizable. For example quantizer can a) decompose a compound operator like scaled dot product attention, into bmm and softmax if quantizer knows how to quantize bmm/softmax but not sdpa or b) transform scalars to tensor to allow quantizing scalares. Note: this is an optional method rr+rDs rtransform_for_annotationz"Quantizer.transform_for_annotations  rcyNrrGs rannotatezQuantizer.annotate rNcyrJrrGs rvalidatezQuantizer.validaterLredge_or_node_to_obs_or_fqcy)aA callback that will be called after the observers or fake quants are created for each sharing group, but before they are inserted into the graph. The callback can be used to make final quantization adjustments, such as enforcing specific scale and zero point on model input or output. Args: * `model`: the graph module being prepared. * `edge_or_node_to_obs_or_fq`: a dictionary mapping each annotated edge and node to the corresponding observer or fake quant object. Note that multiple edges and/or nodes can map to the same observer / fake quant instance if they were annotated with SharedQuantizationSpec. This dictionary can be modified by the callback. Nr)r+rDrOs rprepare_obs_or_fq_callbackz$Quantizer.prepare_obs_or_fq_callbacks$ r)rrrr-fx GraphModulerHrrKrNrBrr rQrrrrrs XX))       ehh22 uxx7K7K   ehh22 t  xx##$( 4J(J#K  rr)abcrr dataclassesrrtypingrrr r-r torch.ao.quantizationr torch.ao.quantization.qconfigr torch.fxr __all__rrrr;rrrrrrrrrr[s #(,, 8L 3  d4  0+ 0! 0F d4 #7! 5t$d* + C  d4 1! d4  2 !   0,,r