o p iI@sdZddlmZddlZddlZddlmZmZddlmZddl m Z m Z ddl m Z dd d Zdd dZGddde ZGddde ZGddde ZGddde ejdZdS)zEDefine some layers used to export quantization model with ONNX style.) annotationsN)_C_ops _legacy_C_ops) unique_name)in_dynamic_mode in_pir_mode)Layere4m3cCs|dkrdgt|j}|||<||}|d}|dkr2t|d|ddd|jS|dkrHt|d |d d d |jSt d ) Nrfloat32r @Z float8_e4m3fne5m2 Z float8_e5m2only support e4m3 or e5m2 now) lenshapenumelreshapeastypepaddlecastclipdtypeNotImplementedError)inputscaleaxistyperinpr#]/home/app/PaddleOCR-VL/.venv_paddleocr/lib/python3.10/site-packages/paddle/nn/quant/format.pyfake_fp8_quants$   r%cCsz|dkrdgt|j}|||<||}|dkr(|dd||jS|dkr9|dd||jStd) Nrr r r rrrr)rrrrrrr)rrr r!rr#r#r$fake_fp8_dequant0s  r&cs0eZdZfddZddZeddZZS)LinearQuanterDequantercst||_||_dSN)super__init___quanter _dequanter)selfquanterZ dequanter __class__r#r$r*?s  zLinearQuanterDequanter.__init__cCs0|}|jdur ||}|jdur||}|Sr()r+r,)r-routr#r#r$forwardDs     zLinearQuanterDequanter.forwardcCs"|dusJtt|t|Sr()r' LinearQuanter from_quanterLinearDequanterr.r#r#r$r4Ls z#LinearQuanterDequanter.from_quanter__name__ __module__ __qualname__r*r2 staticmethodr4 __classcell__r#r#r/r$r'>s  r'c:eZdZ    d fdd ZddZedd ZZS) r3Ncttj|dd}tjjtjjdtj j ddd}|j |j |dd|_|j|tjddd|_tjddd|_|durE|n|}tj|dd}tjjtjjd tj j ddd}|j |j |dd|_|j||durwd n||_||_||_t|jtr|jd d kr|jd dkrt|jdkrd|_d|_n2|jd dkr|jd dkrt|jdkrd|_d|_ntdd |jd >d |_d |jd |_t|jtr|jd |jd d |_dSdSNr )rzquant_dequant.scaleg?F)name initializerZ trainable)rattrrzquant_dequant.zero_pointr rr rrrrrzCurrently, only float8_e4m3 and float8_e5m2 formats are supported. Please set quant_bits to (4,3) or (5,2) for the corresponding format.r)r*rZ to_tensorZ frameworkZ ParamAttrutilsrgeneratennrCConstantZcreate_parameterr_scales set_valuein_accumin_state _zero_point _quant_axis _bit_length _group_size isinstancetupler_qmin_qmaxrr-scales zero_point quant_axis bit_lengthZ group_sizeZ scale_attrZzp_attrr/r#r$r*V\         zLinearQuanter.__init__c Cstr|jdkrt||j|jddS|jdkr!t||j|jddSt|jjdkrt|jdkrFt t | d|j|j|j |j}n(t |j|jd}t |j|jd}t t | d||j||j |j}| |jSt| d|j|jd |jd |jd |j d |j |jStrd |_t jjjd|jtdt jjdd d}t||j|j|j|j |j|j|j |jdd d \}}}}|S|j!"|j} |j!j#d||j|jdd| i|j|j|j |jdd| S)Nrr r!rrr rr r]r^qminqmaxT quant_outrErrrBrC stop_gradientFquantize_linearXZScaleZ ZeroPointYr]r^rarbr!ZinputsZoutputsattrs)$rrYr%rNrSrrrRsumrrroundrrXrepeat_interleaverUrrrfrTrrepircorecreate_persistable_valuerrKrLrCrMrrPrQ_helper"create_variable_for_type_inference append_op) r-rZ quant_weightnew_snew_zprc out_state out_accum out_scaler1r#r#r$r2s           zLinearQuanter.forwardcC t||||dSN)r\r]r^)r3r[ zero_pointsr]r^r6r#r#r$r4 zLinearQuanter.from_quanterNNr>r?r7r#r#r/r$r3Us<Zr3cr=) r5Nr>r?cr@rArIrZr/r#r$r*r_zLinearDequanter.__init__c Cstr||jdkrt||j|jddS|jdkr!t||j|jddSt|jjdkr`|jdkr<| d|j|j}nt |j|j d}t |j|j d}| d||j|}| |j St| d|j|jd |jd |jd |jd |j |j Strd |_t jjjd|jtdt jjdd d}t||j|j|j|j|j|j|j|jdd d \}}}}|S|j |j } |jj!d||j|jdd| i|j|j|j|jdd| S)Nrr r`rrr rr r]r^rarbTrcrErdFdequantize_linearrgrirjrk)"rrYr&rNrSrrrRrmrrrorUrrrrTrXrrerprqrrrrKrLrCrMrrPrQrsrtru) r-rZquant_dequant_weightrvrwZ dequant_outrxryrzr1r#r#r$r23s            zLinearDequanter.forwardcCr{r|)r5r[r}r]r^r6r#r#r$r4r~zLinearDequanter.from_quanterrr7r#r#r/r$r5s<Vr5cs\eZdZdZfddZejdddZejdd d Zdd d Z ddZ dddZ Z S)ConvertibleQuantedLayeraAbstract class to help convert quantized layer to inference model. It defines some functions to convert quantizers and observers to quantize or dequantize operators that maintain the quantization parameters used during inference. Examples: .. code-block:: python >>> # Given codes in ./customized_quanter.py >>> class CustomizedQuantedLayer(ConvertibleQuantedLayer): ... def __init__(self): ... super().__init__() ... self.weight_a = paddle.create_parameter(shape=[1], dtype='float32') ... self.weight_b = paddle.create_parameter(shape=[1], dtype='float32') ... self.quanter_for_weight_a = None ... self.activation_weight = None ... ... def forward(self, input): ... qweight_a = self.quanter_for_weight_a(self.weight_a) ... weight_b = self.weight_b ... qinput = self.activation_weight(input) ... # compute with qweight_a, weight_b and qinput. ... return qweight * qinput + weight_b ... ... def weights_to_quanters(self): ... return [('weight_a', 'quanter_for_weight_a')] ... ... def activation_quanters(self): ... return ['activation_weight'] cstd|_dS)NF)r)r* convertedr-r/r#r$r*s  z ConvertibleQuantedLayer.__init__returnlist[tuple[str, str]]cCdS)uGet the name pairs of weights to be quantized and their corresponding quantizers. In the convert function of this abstract class, it will call the ‘weights_to_quanters’ function and do something as follows: For each pair, the quantizer will be converted to a quantize operator and a dequantize operator. Then, the weight will be quantized by the quantize operator. Finally, the quantize operator will be removed and the weights will be stored in integer data type. Returns: A list of name pairs. Each pair contains two names. The first is name of weight to be quantized and the second is name of corresponding quanter. Nr#rr#r#r$weights_to_quanterss z+ConvertibleQuantedLayer.weights_to_quanters list[str]cCr)zGet the names of quanters used to quantize activations. All the quanters or observers returned by this function will be converted to quantize and dequantize operators for deployment. Returns: A list of quanter names. Nr#rr#r#r$activation_quanterssz+ConvertibleQuantedLayer.activation_quantersr'cCsHt||sdSt||}|durdSt|}t|||||j|<|S)z9Convert quanter to an instance of LinearQuanterDequanter.N)hasattrgetattrr'r4setattr _sub_layers)r- quanter_namer.r#r#r$_convert_quanter_to_qdqs     z/ConvertibleQuantedLayer._convert_quanter_to_qdqcCs t||}||}||dS)z%Quantize the weight by given quanter.N)rrO)r- weight_namer.weightZqweightr#r#r$_quant_weightss z&ConvertibleQuantedLayer._quant_weightsFcCsz|jrJd|D] \}}||}|dur+|dur+|||jd|_d|jd<q |D]}||q0d|_dS)z2Convert current layer to onnx style for inference.z(The model should be converted only once.NFr+T)rrrrr+rr)r-Z remain_weightrrZqdqr#r#r$_converts     z ConvertibleQuantedLayer._convert)rr)rr)rr')F) r8r9r:__doc__r*abcabstractmethodrrrrrr<r#r#r/r$rs     r) metaclass)r r )r __future__rrrrrZ paddle.baserZpaddle.frameworkrrZ layer.layersr r%r&r'r3r5ABCMetarr#r#r#r$s      "