§
    9²¸jª  ã                  ó²   — d dl mZ d dlmZ ddlmZmZ ddlmZm	Z	 ddl
mZ ddlmZ  e¦   «         rd d	lZerdd
lmZ ddlmZ dd„Z G d„ de¦  «        Zd	S )é    )Úannotations)ÚTYPE_CHECKINGé   )Úis_kernels_availableÚis_torch_available)ÚKERNELS_MAX_VERSIONÚKERNELS_MIN_VERSIONé   )ÚHfQuantizer)Úget_module_from_nameN)ÚPreTrainedModel)ÚNVFP4ConfigÚreturnútorch.device | Nonec                óÐ   — t          | t          ¦  «        rt          j        d| ¦  «        S 	 t          j        | ¦  «        } n# t          t
          f$ r Y d S w xY w| j        dk    r| nd S )NÚcuda)Ú
isinstanceÚintÚtorchÚdeviceÚ	TypeErrorÚRuntimeErrorÚtype)r   s    úw/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/transformers/quantizers/quantizer_nvfp4.pyÚ_as_cuda_devicer       sy   € Ý�&�#ÑÔð ,ÝŒ|˜F FÑ+Ô+Ð+ðÝ”˜fÑ%Ô%ˆˆøÝ•|Ð$ð ð ð Øˆtˆtðøøøà”[ FÒ*Ð*ˆ6ˆ6°Ð4s   ¬A ÁAÁAc                  óž   ‡ — e Zd ZU dZdZded<   d„ Zd„ Zdd„Zdˆ fd„Z	dd„Z
d„ Zedd„¦   «         Zedd„¦   «         Zedd„¦   «         Zˆ xZS )ÚNVFP4HfQuantizerzTQuantize eligible linear weights to NVFP4 while loading a full-precision checkpoint.Fr   Úquantization_configc           	     óÄ  — | j         rt          d¦  «        ‚t          ¦   «         s(t          dt          › dt
          › dt          › d�¦  «        ‚t          j                             ¦   «         st          d¦  «        ‚|€t          d¦  «        ‚t          |t          ¦  «        r!t          |                     ¦   «         ¦  «        n|h}d„ |D ¦   «         }d |v rt          d	¦  «        ‚t          |¦  «        d
k    rt          d¦  «        ‚|                     ¦   «         | _        t          j                             | j        ¦  «        \  }}|dk     rt          d|› |› d�¦  «        ‚d S )Nz=Loading pre-quantized NVFP4 checkpoints is not supported yet.zQNVFP4 quantization requires the `kernels` package. Install a compatible version (z <= version < z*), for example with `pip install kernels==z`.z'NVFP4 quantization requires a CUDA GPU.z0NVFP4 quantization requires a CUDA `device_map`.c                ó,   — h | ]}t          |¦  «        ’ŒS © )r   )Ú.0r   s     r   ú	<setcomp>z8NVFP4HfQuantizer.validate_environment.<locals>.<setcomp>?   s    € ÐFÐFÐF°F�¨Ñ/Ô/ÐFÐFÐFó    z8NVFP4 quantization does not support CPU or disk offload.r
   z�NVFP4 quantization currently supports one CUDA device. Tensor parallelism and multi-device `device_map` configurations are not supported yet.é
   zENVFP4 requires Blackwell (sm100+) block-scaled tensor cores; found smú.)Úpre_quantizedÚ
ValueErrorr   ÚImportErrorr	   r   r   r   Úis_availabler   r   ÚdictÚsetÚvaluesÚlenÚpopÚquantization_deviceÚget_device_capability)ÚselfÚ
device_mapÚkwargsÚdevicesÚcuda_devicesÚmajorÚminors          r   Úvalidate_environmentz%NVFP4HfQuantizer.validate_environment0   sŸ  € ØÔð 	^ÝÐ\Ñ]Ô]Ð]Ý#Ñ%Ô%ð 	ÝðRÝ1DðRð RÝTgðRð Rå:MðRð Rð Rñô ð õ
 Œz×&Ò&Ñ(Ô(ð 	JÝÐHÑIÔIÐIàÐÝÐOÑPÔPÐPÝ.8¸ÅTÑ.JÔ.JÐ\•#�j×'Ò'Ñ)Ô)Ñ*Ô*Ð*ÐQ[ÐP\ˆØFÐF¸gÐFÑFÔFˆØ�<ÐÐÝÐWÑXÔXÐXÝˆ|ÑÔ Ò!Ð!ÝðEñô ð ð $0×#3Ò#3Ñ#5Ô#5ˆÔ å”z×7Ò7¸Ô8PÑQÔQ‰ˆˆuØ�2Š:ˆ:ÝÐvÐglÐvÐnsÐvÐvÐvÑwÔwÐwð ˆ:r$   c                óx   — t          |dd ¦  «        }t          |dd ¦  «        }|�|dk    rt          d¦  «        ‚|S )NÚdistributed_configÚtp_sizer
   ztNVFP4 quantization does not support tensor parallelism yet because its scale metadata does not have a sharding plan.)Úgetattrr(   )r2   Úconfigr;   r<   s       r   Úupdate_tp_planzNVFP4HfQuantizer.update_tp_planM   sS   € Ý$ VÐ-AÀ4ÑHÔHÐÝÐ,¨i¸Ñ>Ô>ˆØÐ 7¨Q¢; ;Ýð#ñô ð ð ˆr$   Úmodelr   Ú
param_nameÚstrr   Úboolc                ó`   — ddl m} t          ||¦  «        \  }}t          ||¦  «        o|dk    S )Nr   )ÚNVFP4LinearÚweight)Úintegrations.nvfp4rE   r   r   )r2   r@   rA   r4   rE   ÚmoduleÚtensor_names          r   Úparam_needs_quantizationz)NVFP4HfQuantizer.param_needs_quantizationW   sB   € Ø4Ð4Ð4Ð4Ð4Ð4å2°5¸*ÑEÔEÑˆ�Ý˜& +Ñ.Ô.ÐJ°;À(Ò3JÐJr$   Úparamútorch.TensorÚfloatc                óz   •— |                       ||¦  «        rdS t          ¦   «                              |||¦  «        S )Ng      à?)rJ   ÚsuperÚparam_element_size)r2   r@   rA   rK   Ú	__class__s       €r   rP   z#NVFP4HfQuantizer.param_element_size]   s<   ø€ Ø×(Ò(¨°
Ñ;Ô;ð 	à�3Ý‰wŒw×)Ò)¨%°¸UÑCÔCÐCr$   c                óŒ   — ddl m} |                      || j        j        |j        ¦  «        | _         ||| j        ¬¦  «         d S )Nr   )Úreplace_with_nvfp4_linear)Úmodules_to_not_convert)rG   rS   Úget_modules_to_not_convertr   rT   Ú_keep_in_fp32_modules)r2   r@   r4   rS   s       r   Ú$_process_model_before_weight_loadingz5NVFP4HfQuantizer._process_model_before_weight_loadingc   s]   € ØBÐBÐBÐBÐBÐBà&*×&EÒ&EØ�4Ô+ÔBÀEÔD_ñ'
ô '
ˆÔ#ð 	"Ð! %ÀÔ@[Ð\Ñ\Ô\Ð\Ð\Ð\r$   c                ó.   — ddl m}  || j        ¦  «        S )Nr   )ÚNVFP4Quantize)rG   rY   r0   )r2   rY   s     r   Úget_quantize_opsz!NVFP4HfQuantizer.get_quantize_opsk   s'   € Ø6Ð6Ð6Ð6Ð6Ð6àˆ}˜TÔ5Ñ6Ô6Ð6r$   c                ó   — dS ©NFr!   ©r2   s    r   Úis_serializablez NVFP4HfQuantizer.is_serializablep   ó   € àˆur$   c                ó   — dS r\   r!   r]   s    r   Úis_trainablezNVFP4HfQuantizer.is_trainablet   r_   r$   c                ó   — dS )NTr!   r]   s    r   Úis_compileablezNVFP4HfQuantizer.is_compileablex   s   € àˆtr$   )r@   r   rA   rB   r   rC   )r@   r   rA   rB   rK   rL   r   rM   )r@   r   )r   rC   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__Úrequires_calibrationÚ__annotations__r9   r?   rJ   rP   rW   rZ   Úpropertyr^   ra   rc   Ú__classcell__)rQ   s   @r   r   r   *   s  ø€ € € € € € Ø^Ð^à ÐØ$Ð$Ð$Ñ$ðxð xð xð:ð ð ðKð Kð Kð KðDð Dð Dð Dð Dð Dð]ð ]ð ]ð ]ð7ð 7ð 7ð
 ðð ð ñ „Xðð ðð ð ñ „Xðð ðð ð ñ „Xðð ð ð ð r$   r   )r   r   )Ú
__future__r   Útypingr   Úutilsr   r   Úutils.import_utilsr   r	   Úbaser   Úquantizers_utilsr   r   Úmodeling_utilsr   Úutils.quantization_configr   r   r   r!   r$   r   ú<module>rt      s  ðð #Ð "Ð "Ð "Ð "Ð "à  Ð  Ð  Ð  Ð  Ð  à <Ð <Ð <Ð <Ð <Ð <Ð <Ð <Ø IÐ IÐ IÐ IÐ IÐ IÐ IÐ IØ Ð Ð Ð Ð Ð Ø 2Ð 2Ð 2Ð 2Ð 2Ð 2ð ÐÑÔð Ø€L€L€Làð 8Ø0Ð0Ð0Ð0Ð0Ð0Ø7Ð7Ð7Ð7Ð7Ð7ð5ð 5ð 5ð 5ðPð Pð Pð Pð P�{ñ Pô Pð Pð Pð Pr$   