o
    °ƒðj2d  ã                   @   sÔ   d dl Z d dlmZ d dlZd dlZd dlZzd dlmZ W n e	y)   dZY nw ddl
mZ ddlmZ ddlmZmZmZmZmZmZmZmZmZmZmZmZmZ ddlmZ G d	d
„ d
ƒZG dd„ dƒZdS )é    N)ÚAny)Úto_array_extendedé   )Ú
TensorData)Ú	ONNXModel)ÚDEQUANT_OP_NAMEÚONNX_TYPE_TO_NP_TYPEÚQUANT_OP_NAMEÚTENSOR_NAME_QUANT_SUFFIXÚfind_by_nameÚget_opset_versionÚmodel_has_infer_metadataÚnormalize_axisÚpack_bytes_to_4bitÚquantize_dataÚquantize_nparrayÚ&save_and_reload_model_with_shape_inferÚtensor_proto_to_array)ÚTensorQuantOverridesHelperc                   @   sL   e Zd Zdeeef fdd„Zddd„Zdd„ Zd	d
„ Z	dd„ Z
dd„ ZdS )ÚQuantizationParamsÚdatac                 K   sÞ   i | _ | ¡ D ]e\}}t|tƒstdt|ƒ› d|›d�ƒ‚|dkr8t|tttjt	fƒs8tdt|ƒ› d|›d�ƒ‚|dkrOt|tƒsO|d urOtdt|ƒ› d�ƒ‚|dkrg|j
tjtjfvrgtd|j
› d|›�ƒ‚|| j |< qd S )	NzKeys must be strings not z for k=Ú.Úaxisz1Values must be numpy arrays, int, float, str not z'Axis value must be an int or None, not Úscalez5scale must a float32 or float16 numpy element but is )r   ÚitemsÚ
isinstanceÚstrÚ	TypeErrorÚtypeÚintÚnpÚndarrayÚfloatÚdtypeÚfloat32Úfloat16Ú
ValueError)Úselfr   ÚkÚv© r*   ú©/tmp/claude-1000/-var-www-html-academy/3f42efd1-c2b7-4d08-9b05-30d86bb35479/scratchpad/piper/venv/lib/python3.10/site-packages/onnxruntime/quantization/base_quantizer.pyÚ__init__(   s   
÷zQuantizationParams.__init__Nc                 C   s   | j  ||¡S ©N)r   Úget)r'   ÚkeyÚdefault_valuer*   r*   r+   r.   5   ó   zQuantizationParams.getc                 c   s   � | j E d H  d S r-   ©r   ©r'   r*   r*   r+   Ú__iter__8   s   €zQuantizationParams.__iter__c                 C   s
   | j | S r-   r2   )r'   r/   r*   r*   r+   Ú__getitem__;   ó   
zQuantizationParams.__getitem__c                 C   s   || j |< d S r-   r2   )r'   r/   Úvaluer*   r*   r+   Ú__setitem__>   r1   zQuantizationParams.__setitem__c                 C   s
   t | jƒS r-   )Úlenr   r3   r*   r*   r+   Ú__len__A   r6   zQuantizationParams.__len__r-   )Ú__name__Ú
__module__Ú__qualname__Údictr   r   r,   r.   r4   r5   r8   r:   r*   r*   r*   r+   r   '   s    
r   c                   @   s€   e Zd Z	ddd„Zdejjdefdd„Zdd	„ Z	d
d„ Z
dd„ Zdd„ Zdd„ Zddd„Zddd„Z		d dd„Zdd„ ZdS )!ÚBaseQuantizerNc                 C   sº  t |ƒst|ƒ}dd„ |jjD ƒ| _| j dd„ |jjD ƒ¡ | j dd„ |jjD ƒ¡ t|ƒ| _	t
|ƒ| _|| _|| _|
r@|
ni | _d| jv oL| jd | _d | _d| jv oZ| jd | _| j dd ¡| _| j dd	¡| _| j d
¡| _t|d|ƒ| _t|d|ƒ| _	 |d ur tdd„ | ¡ D ƒƒr tddd„ | ¡ D ƒ› d�ƒ‚|| _|| _|| _|	| _t | j di ¡ƒ| _!dd„ | j	 "¡ D ƒ| _#| j! $| j#| j %¡ |¡\}}|sÕt&|ƒ‚| j! '¡ | _(d S )Nc                 S   ó   i | ]}|j |“qS r*   ©Úname)Ú.0Úvir*   r*   r+   Ú
<dictcomp>U   ó    z*BaseQuantizer.__init__.<locals>.<dictcomp>c                 S   r@   r*   rA   )rC   Úotr*   r*   r+   rE   V   rF   c                 S   r@   r*   rA   )rC   Úitr*   r*   r+   rE   W   rF   ÚEnableSubgraphÚForceQuantizeNoInputCheckÚWeightSymmetricÚActivationSymmetricFÚMinimumRealRangeÚtensor_typec                 s   s   � | ]	}t |tƒ V  qd S r-   )r   r   )rC   Útr*   r*   r+   Ú	<genexpr>{   s   € z)BaseQuantizer.__init__.<locals>.<genexpr>z(tensors_range contains unexpected types c                 S   s   h | ]}t |ƒ’qS r*   )r   )rC   r)   r*   r*   r+   Ú	<setcomp>}   rF   z)BaseQuantizer.__init__.<locals>.<setcomp>z, not TensorData.ÚTensorQuantOverridesc                 S   r@   r*   rA   )rC   Úinitzerr*   r*   r+   rE   ‡   rF   ))r   r   ÚgraphÚ
value_infoÚvalue_infosÚupdateÚoutputÚinputr   Úmodelr   Úopset_versionÚper_channelÚreduce_rangeÚextra_optionsÚenable_subgraph_quantizationÚparentÚforce_quantize_no_input_checkr.   Ú_is_weight_symmetricÚis_activation_symmetricÚmin_real_rangeÚgetattrÚactivation_qTypeÚweight_qTypeÚanyÚvaluesr   Útensors_rangeÚnodes_to_quantizeÚnodes_to_excludeÚop_types_to_quantizer   Útensor_quant_overridesÚinitializerÚinitializersÚis_validÚkeysr&   Úget_quant_typesÚtensor_quant_override_qtypes)r'   rZ   r\   r]   rg   rf   rj   rk   rl   rm   r^   Úoverrides_validÚoverrides_errr*   r*   r+   r,   F   sJ   

ÿÿÿÿzBaseQuantizer.__init__Úweight_quant_typeÚreturnc                 C   s0   | j d ur| j S |tjjtjjtjjtjjfv S r-   )rb   ÚonnxÚTensorProtoÚINT4ÚINT8ÚINT16ÚFLOAT8E4M3FN)r'   rw   r*   r*   r+   Úis_weight_symmetric�   s   
üz!BaseQuantizer.is_weight_symmetricc                 C   s   t ‚r-   )ÚNotImplementedErrorr3   r*   r*   r+   Úquantize_modelš   s   zBaseQuantizer.quantize_modelc                 C   s   t || j ¡ ƒ}|d uS r-   )r   rZ   ro   )r'   Ú
input_namero   r*   r*   r+   Úis_input_a_initializer�   s   z$BaseQuantizer.is_input_a_initializerc                 C   s   | j S r-   )r\   r3   r*   r*   r+   Úis_per_channel¡   s   zBaseQuantizer.is_per_channelc                 C   sN   t || j ¡ ƒ}|d ur|jtjjtjjfv S | jr| j	d u r!dS | j	 
|¡S )NF)r   rZ   ro   Ú	data_typery   rz   ÚFLOATÚFLOAT16r_   r`   Úis_valid_quantize_weight)r'   Úweight_nameÚweightr*   r*   r+   rˆ   ¤   s   z&BaseQuantizer.is_valid_quantize_weightc                 C   sh   | j d urt| j ƒdkr|j| j vrdS |j| jvrdS |jttfv r%dS | jd ur2|j| jv r2dS dS )Nr   FT)rk   r9   rB   Úop_typerm   r   r	   rl   )r'   Únoder*   r*   r+   Úshould_quantize_node¬   s   
z"BaseQuantizer.should_quantize_nodeç      ð?c                 C   sŽ  t || j ¡ ƒ}t|ƒ}|t }| jtjjkr`t	 
|¡}|jt	jkr'tjj}	n|jt	jkr2tjj}	n	td|j› d�ƒ‚| t	j¡}
t	jdg|
jd�}| d¡}tj |
|¡}| j |g¡ d}nz|| | }t	j
|t	jd�t	j
|t	jd� }
|
 ¡ }
t	 t	 t	j¡j¡}t	 t	 t	j¡j¡}t	 |
|k ¡sœt	 |
|k¡r¥t d|› d�¡ t	  |
||¡ t	j¡}
t	j
|
t	jd� |j!¡}tj ||¡}| j |g¡ t	j
||jd� d¡}d	}| j}	|d
 }tj ||¡}| j |g¡ | jtjjkr÷| j}ntjj"}|d }| jtjjk�rtj# $|| jdgdg¡}n$|j%dk�r.t	j&|j't	jd� d¡}tj ||¡}n
tj# $||g dg¡}| j |g¡ ||||||	fS )z]
        Quantized the bias. Zero Point == 0 and Scale == Input_Scale * Weight_Scale
        zEOnly float16 or float32 are supported with float 8 but bias dtype is r   r   ©r#   éÿÿÿÿÚCastzQuantized bias `z<` exceeds the range of a int32. The bias scale is too small.ÚDequantizeLinearÚ_scaleÚ_zero_pointç        r   )(r   rZ   ro   r   r
   rg   ry   rz   r~   r    Úasarrayr#   r%   r‡   r$   r†   r   ÚastypeÚarrayÚreshapeÚnumpy_helperÚ
from_arrayÚinitializer_extendÚfloat64ÚroundÚiinfoÚint32ÚminÚmaxrh   ÚloggingÚwarningÚclipÚdimsÚINT32ÚhelperÚmake_tensorÚsizeÚzerosÚshape)r'   Ú	bias_nameÚinput_scaleÚweight_scaleÚbetaÚbias_initializerÚ	bias_dataÚquantized_bias_namer   Ú
node_qtypeÚquantized_dataÚ
bias_scaleÚbias_scale_dataÚpacked_bias_initializerÚ	node_typeÚ	int32_minÚ	int32_maxÚbias_np_dataÚquantized_bias_scale_nameÚpacked_bias_scale_initializerrN   Úquantized_bias_zp_nameÚpacked_bias_zp_initializerÚbias_zp_datar*   r*   r+   Úquantize_bias_static_impl¿   sj   



 
ÿúz'BaseQuantizer.quantize_bias_static_implFc                 C   s~  |j t }|j d }|j d }t|ƒ}| jj|j i d�}	d|	v r%|	d j}d|	v r{d|	v r{tj|	d t| d�}
t |	d ¡}t	|| 
¡ ||
ƒ}t|
tjƒsWJ dt|
ƒ› �ƒ‚|
jtjkrc|
jtjkskJ d	|
j› �ƒ‚t|tjƒszJ dt|ƒ› �ƒ‚na|| jkr…|  |¡n| j}t| 
¡ ||	 d
|¡|	 d| jo™|¡| j|	 d¡|	 d¡d�\}
}}t|
tjƒs¹J dt|
ƒ› �ƒ‚|
jtjkrÅ|
jtjksÍJ d	|
j› �ƒ‚t|tjƒsÜJ dt|ƒ› �ƒ‚|j}tj ||g | d¡ ¡ ¡}tj ||g |
 d¡ ¡ ¡}| j ||g¡ |�sº| jtj j!k�rkt  ¡ }| j|_|j" #|j"¡ ||_ | 
¡  $¡  %¡ |_&t'du�rjt'|ƒ}|j(|j(k�sB| %¡ | %¡ k�rjt)d|j(› d| %¡ dd… › d| %¡ dd… › d|j(› dt*|ƒdd… › d�ƒ‚nH|tj j+tj j,fv �r�|jtj-tj.fv�rˆt)d|› d�ƒ‚t/t0| %¡ ƒƒ}tjj|||j"|dd�}ntj1|tj 2|¡d� |j"¡}tj3 4||¡}| j |g¡ |||fS )aš  
        :param weight: TensorProto initializer
        :param qType: type to quantize to
        :param keep_float_weight: Whether to quantize the weight. In some cases, we only want to qunatize scale and zero point.
                                  If keep_float_weight is False, quantize the weight, or don't quantize the weight.
        :return: quantized weight name, zero point name, scale name
        r”   r“   ©Údefault_valÚ
quant_typer   Ú
zero_pointr�   úUnexpected type úUnexpected dtype Ú	symmetricr]   ÚrminÚrmax©r]   rd   Úrmin_overrideÚrmax_override)r�   NzThe initializer of shape z! could not be created, expecting é
   z, got z and shape=z
raw=éÈ   r   úQuantized weights for ú. must be 8-bit before packing as 4-bit values.T©Úraw)5rB   r
   r   rn   Úget_per_tensor_overridesrN   r    r˜   r   r   Úflattenr   r!   r   r#   r$   r%   rg   r   rc   r   r.   r]   rd   r…   ry   r¨   r©   r™   ÚtolistrZ   rœ   rz   r~   r¦   ÚextendÚcopyÚtobytesÚraw_datar   r¬   ÚRuntimeErrorr   r{   ÚUINT4Úint8Úuint8Úbytesr   r–   Útensor_dtype_to_np_dtyperš   r›   )r'   rŠ   ÚqTyper]   Úkeep_float_weightÚq_weight_nameÚzp_nameÚ
scale_nameÚweight_dataÚquant_overridesrÆ   r   Úq_weight_datarÉ   Úscale_dtypeÚscale_initializerÚzero_initializerÚq_weight_initializerÚcheckÚpacked_datar*   r*   r+   Úquantize_initializer_impl  sŒ   
	



ÿ 
ù

ÿ
 
ÿÿÿþÿ€
ÿÿ
z'BaseQuantizer.quantize_initializer_implTc           &      C   s„  t || j ¡ ƒ}|d u rtd|ƒ‚t|ƒ}t|jƒ}t||ƒ\}	}
|	s0td|› d|› d|› �ƒ‚|
}|j| }| jj	|d|igd�}t|ƒ}|dkrY||krYtd|› d	|› d
�ƒ‚t|d d |ƒ\}}|rj||kr|td|› d|› d|d d › d�ƒ‚d|d v r‰|d d j
}|d  d|  |¡¡}|d  d| jo�|¡}g }g }g }t|jƒ}t|ƒ}d||< t|ƒD ]ñ}| ||¡}||k rÄ|nd}|| }d|v �r5d|v �r5tj|d t| d�}t |d ¡}t|| ¡ ||ƒ}t|tjƒsþJ dt|ƒ› �ƒ‚|jtjk�r|jtjk�sJ d|j› �ƒ‚t|tjƒ�s$J dt|ƒ› �ƒ‚t|tjƒ�s4J dt|ƒ› �ƒ‚n]t| ¡ |||| j| d¡| d¡d�\}}}t|tjƒ�s\J dt|ƒ› �ƒ‚|jtjk�rj|jtjk�srJ d|j› �ƒ‚t|tjƒ�s‚J dt|ƒ› �ƒ‚t|tjƒ�s’J dt|ƒ› �ƒ‚| |¡ | |¡ | t |¡  |¡¡ q¶t !||¡}|t" }|d }|d } |j#| g}!t$j% &| |j'|!t (|¡ )¡ ¡}"t$j% &|||!t (|¡ )¡ ¡}#| j *|"|#g¡ |�s=|t$j+j,t$j+j-fv �r |jtj.tj/fv�rt0d|› d�ƒ‚t1t2| 3¡ ƒƒ}$t$j%j&||||$dd �}%| j *|%g¡ ntj|t$j% 4|¡d�  |j#¡}t$j5 6||¡}%| j *|%g¡ ||| fS )!Nz{} is not an initializerzWeight z# has a per-channel axis with value z  that is out-of-bounds for rank r   rÃ   r   z.Per-channel tensor quantization overrides for z must have either 1 or z& elements in the list of dictionaries.r   z"Tensor quantization overrides for z& specify an unexpected axis. Expected z
, but got r   rÅ   rÉ   r]   r   rÆ   r�   rÇ   rÈ   rÊ   rË   rÌ   r”   r“   rÑ   rÒ   TrÓ   )7r   rZ   ro   r&   r   r9   r¬   r   rn   Úget_per_channel_overridesrN   r.   r   r]   ÚlistÚrangeÚtaker    r˜   r   r   rÖ   r   r!   r   r#   r$   r%   r   rd   Úappendr–   r™   Úconcatenater
   r¦   ry   r¨   r©   r…   Úhstackr×   rœ   rz   r{   rÝ   rÞ   rß   rÜ   rà   r   rÚ   rá   rš   r›   )&r'   r‰   rg   Úchannel_axisr]   rã   ro   ÚweightsÚweights_rankÚis_axis_validÚ	axis_normÚchannel_countÚquant_overrides_for_channelsÚnum_channel_overridesÚis_axis_override_validÚaxis_overriderÉ   Úzero_point_listÚ
scale_listÚquantized_per_channel_data_listÚweights_shapeÚreshape_dimsÚiÚper_channel_dataÚchannel_override_indexÚchannel_quant_overridesrÆ   r   Úquantized_per_channel_dataÚquantized_weightsrä   rå   ræ   Úzero_scale_shaperë   rì   rï   rí   r*   r*   r+   Ú quantize_weight_per_channel_impln  sâ   	

ÿÿ

ÿÿÿÿ
ÿÿ
ÿ
ÿ ÿù 

ÿ ÿ

ÿÿ
ÿ
ÿ
þý
z.BaseQuantizer.quantize_weight_per_channel_implc                 C   sþ   | j d u rd S | j ¡ D ]p}|jdv r_|  |¡sqt| j ¡ |jd  ƒdkr(q|jd | j vs8|jd | j vr9q| j |jd  }t	|t
ƒsVtdt|ƒ› d|jd ›d�ƒ‚|| j |jd < q|jdkr||  |¡sjqt
t d¡t d	¡d
�| j |jd < qd S )N)ÚClipÚRelur   r   rÇ   z for r   ÚSoftmaxr•   rŽ   )ÚlowestÚhighest)rj   rZ   Únodesr‹   r�   r9   Úinput_name_to_nodesrY   rX   r   r   r   r   r    r$   )r'   rŒ   Útdr*   r*   r+   Úadjust_tensor_rangesü  s(   


 
 

$€ïz"BaseQuantizer.adjust_tensor_rangesr-   )rŽ   )FF)TF)r;   r<   r=   r,   ry   rz   ÚDataTypeÚboolr   r�   rƒ   r„   rˆ   r�   rÂ   rð   r  r  r*   r*   r*   r+   r?   E   s     
õJ


T`
ú r?   ) r£   Útypingr   Únumpyr    ry   Úonnx.numpy_helperÚonnx.reference.op_runr   ÚImportErrorÚ	calibrater   Ú
onnx_modelr   Úquant_utilsr   r   r	   r
   r   r   r   r   r   r   r   r   r   rn   r   r   r?   r*   r*   r*   r+   Ú<module>   s    þ<