o
    hl                     @   s   d dl Z d dlmZmZ d dlmZ ddlmZ er ddlm	Z	 ddl
mZ dd	lmZmZmZmZ dd
lmZ e r?d dlZeeZG dd deZdS )    N)TYPE_CHECKINGOptional)version   )HfQuantizer   )PreTrainedModel)replace_with_aqlm_linear)is_accelerate_availableis_aqlm_availableis_torch_availablelogging)QuantizationConfigMixinc                       s   e Zd ZdZdZdgZdZdef fddZdd	 Z	dddZ
		dddZdddZedded fddZdddZ  ZS )AqlmHfQuantizerzS
    Quantizer of the AQLM method. Enables the loading of prequantized models.
    TaqlmNquantization_configc                    s   t  j|fi | || _d S N)super__init__r   )selfr   kwargs	__class__ z/var/www/html/construction_image-detection-poc/venv/lib/python3.10/site-packages/transformers/quantizers/quantizer_aqlm.pyr   -   s   
zAqlmHfQuantizer.__init__c                 O   s    t  stdt stdd S )NzGUsing `aqlm` quantization requires Accelerate: `pip install accelerate`zDUsing `aqlm` quantization requires AQLM: `pip install aqlm[gpu,cpu]`)r
   ImportErrorr   )r   argsr   r   r   r   validate_environment1   s
   z$AqlmHfQuantizer.validate_environmenttorch_dtypetorch.dtypereturnc                 C   s:   |d u rt j rt j}td |S t j}td |S )NzCUDA available. Assuming AQLM inference on GPU and loading the model in `torch.float16`. To overwrite it, set `torch_dtype` manually.zCUDA is unavailable. Assuming AQLM inference on CPU and loading the model in `torch.float32`. To overwrite it, set `torch_dtype` manually.)torchcudais_availablefloat16loggerinfofloat32)r   r   r   r   r   update_torch_dtype8   s   
z"AqlmHfQuantizer.update_torch_dtypemodelr   c                 K   s"   t || j| jjd | j|j_d S )N)r   linear_weights_not_to_quantize)r	   r   r*   configr   r)   r   r   r   r   $_process_model_before_weight_loadingF   s   z4AqlmHfQuantizer._process_model_before_weight_loadingc                 K   s   |S r   r   r,   r   r   r   #_process_model_after_weight_loadingR      z3AqlmHfQuantizer._process_model_after_weight_loadingc                 C   sB   t tj dt dk}|rdS tdtj d d dS )Nr   z1.0.2Tz$Currently installed `aqlm` version (zw) doesn't support training. If you wish to train a quantized model, please update `aqlm` with `pip install aqlm>=1.0.2`F)r   parse	importlibmetadatar%   warning)r   r)   aqlm_supports_trainingr   r   r   is_trainableU   s   zAqlmHfQuantizer.is_trainablec                 C   s   dS )NTr   )r   safe_serializationr   r   r   is_serializable`   r/   zAqlmHfQuantizer.is_serializable)r   r   r    r   )r)   r   r   )__name__
__module____qualname____doc__requires_calibrationrequired_packagesoptimum_quantizerr   r   r   r(   r-   r.   propertyr   r5   r7   __classcell__r   r   r   r   r   $   s    



r   )r1   typingr   r   	packagingr   baser   modeling_utilsr   integrationsr	   utilsr
   r   r   r   utils.quantization_configr   r!   
get_loggerr8   r%   r   r   r   r   r   <module>   s   
