§
    9²¸js  ã                   óÊ   — d Z ddlmZmZmZ ddlmZ ddlmZ  e¦   «         r"ddl	Z	ddl
mZ dd	lmZmZmZ dd
lmZmZmZmZmZ  ej        e¦  «        Z G d„ de¦  «        ZdS )zLKeeping GGUF weights in their blocks instead of unpacking them at load time.é   )Úis_torch_availableÚis_torch_mps_availableÚlogging)Ú
GgufConfigé   )ÚHfQuantizeré    N)Úget_gguf_kernel)Ú
GgufHeaderÚload_gguf_state_dictÚread_gguf_metadata)Úadd_gguf_load_opsÚget_gguf_conversion_mappingÚget_gguf_planÚis_gguf_arch_supportedÚreplace_with_gguf_modulesc                   óÒ   ‡ — e Zd ZU dZded<   ded<   dZdZˆ fd„Zd„ Zd	„ Z	d
e
fd„Zd„ Zde
fd„Zd„ Zd„ Zd„ Zde
defd„Zedefd„¦   «         Zedefd„¦   «         Zddefd„Zˆ xZS )ÚGgufHfQuantizerzGLoads a quantized GGUF checkpoint with its weights left in GGUF blocks.r   Úquantization_configr   ÚheaderFc                 óÈ   •—  t          ¦   «         j        |fi |¤Ž d| _        i | _        i | _        i | _        g | _        g | _        d | _        d | _	        d | _
        d| _        d S )NTF)ÚsuperÚ__init__Úpre_quantizedÚpacked_modulesÚinput_permutationsÚ	quantizedÚnamesÚmappingr   ÚkernelÚdtypeÚ	supported)Úselfr   ÚkwargsÚ	__class__s      €úv/root/private/project/pallet_detection_v2/.venv/lib/python3.11/site-packages/transformers/quantizers/quantizer_gguf.pyr   zGgufHfQuantizer.__init__.   sr   ø€ Ø�‰ŒÔÐ,Ð7Ð7°Ð7Ð7Ð7Ø!ˆÔØ ˆÔØ"$ˆÔØˆŒØˆŒ
ØˆŒØˆŒØˆŒØˆŒ
ð ˆŒˆˆó    c                 óì   — | j         sd S | j        j        rd S | j        j        sd| j        _        d S t          ¦   «         | _        | j        s(d| j        _        t                               d¦  «         d S d S )NTzXNo GGUF matmul kernel is available for this device. We will dequantize the entire model.)	r"   r   Ú
dequantizer   Úhas_quantized_weightsr
   r    ÚloggerÚwarning)r#   Úargsr$   s      r&   Úvalidate_environmentz$GgufHfQuantizer.validate_environment=   s‡   € ØŒ~ð 	ØˆFØÔ#Ô.ð 	ØˆFØŒ{Ô0ð 	ð 37ˆDÔ$Ô/ØˆFÝ%Ñ'Ô'ˆŒØŒ{ð 	Ø26ˆDÔ$Ô/Ý�NŠNÐuÑvÔvÐvØˆFð	ð 	r'   c           	      ó&  ‡— dd„  t          ‰dˆfd„¦  «        ¦   «         D ¦   «         v rt          d¦  «        ‚| j        j        r‰S ‰€Ht	          ¦   «         r:dt          j        d¦  «        iŠt                               d	‰d         › d
�¦  «         ‰S )zHDefault to the backend the blocks are computed on, rather than the host.Údiskc                 ó,   — h | ]}t          |¦  «        ’ŒS © )Ústr)Ú.0Úplaces     r&   ú	<setcomp>z4GgufHfQuantizer.update_device_map.<locals>.<setcomp>S   s   € ÐdÐdÐd U•c˜%‘j”jÐdÐdÐdr'   Úvaluesc                  ó
   •— ‰ gS ©Nr2   )Ú
device_maps   €r&   ú<lambda>z3GgufHfQuantizer.update_device_map.<locals>.<lambda>S   s   ø€ ÐU_ÐT`€ r'   zxOne or more modules is configured to be mapped to disk. Disk offload is not supported for models loaded from GGUF files.NÚ Úmpsz8No `device_map` was passed; loading the GGUF weights on ú.)	ÚgetattrÚRuntimeErrorr   r)   r   ÚtorchÚdevicer+   Úinfo)r#   r:   s    `r&   Úupdate_device_mapz!GgufHfQuantizer.update_device_mapN   s¼   ø€ ð
 ÐdÐdÐ.a­g°jÀ(ÐL`ÐL`ÐL`ÐL`Ñ.aÔ.aÑ.cÔ.cÐdÑdÔdÐdÐdÝð5ñô ð ð Ô#Ô.ð 	ØÐØÐÕ"8Ñ":Ô":ÐØ�eœl¨5Ñ1Ô1Ð2ˆJÝ�KŠKÐdÐS]Ð^`ÔSaÐdÐdÐdÑeÔeÐeØÐr'   Ú	gguf_filec                 óÊ   — t          |¦  «        \  }}t          |d         ¦  «        | _        | j        r/t          j        |¦  «        | _        |                      ¦   «          dS dS )zJParse the file's metadata, once `from_pretrained` knows where the file is.zgeneral.architectureN)r   r   r"   r   Ú	from_filer   r.   )r#   rE   ÚmetadataÚ_s       r&   Úread_headerzGgufHfQuantizer.read_header_   sf   € å(¨Ñ3Ô3‰ˆ�!Ý/°Ð9OÔ0PÑQÔQˆŒØŒ>ð 	(Ý$Ô.¨yÑ9Ô9ˆDŒKØ×%Ò%Ñ'Ô'Ð'Ð'Ð'ð	(ð 	(r'   c                 ó¬   — |€J| j         st          j        ¦   «         | _        | j        S | j        j        �| j        j        nt          j        }|| _        |S )z5Settle the dtype the model is loaded in, and keep it.)r"   rA   Úget_default_dtyper!   r   Úfloat32)r#   r!   s     r&   Úupdate_dtypezGgufHfQuantizer.update_dtypeg   sP   € àˆ=Ø”>ð "Ý"Ô4Ñ6Ô6�”
Ø”zÐ!Ø)-¬Ô):Ð)F�D”KÔ%Ð%ÍEÌMˆEØˆŒ
Øˆr'   Úcheckpoint_filec                 ó|   — | j         rt          | j        ¦  «        S ddlm}  ||d|| j        ¬¦  «        }|d         S )z6The file's tensors, quantized ones kept as raw blocks.r   )Úload_gguf_checkpointT)Úreturn_tensorsÚmodel_to_loadÚtorch_dtypeÚtensors)r"   r   r   Úmodeling_gguf_pytorch_utilsrQ   r!   )r#   rO   ÚmodelrQ   Úlegacys        r&   Úget_state_dictzGgufHfQuantizer.get_state_dictq   s_   € àŒ>ð 	5Ý'¨¬Ñ4Ô4Ð4àFÐFÐFÐFÐFÐFà%Ð%Ø¨DÀÐSWÔS]ð
ñ 
ô 
ˆð �iÔ Ð r'   c                 ó  — | j         sdS t          | j        j        |j        ¦  «        | _        t          | j        | j        ¦  «        \  | _        }| _        | _	        | j
        j        rdS t          ||| j        | j        ¦  «        | _        dS )z9Swap in `GgufLinear` wherever the weight can stay packed.N)r"   r   r   ÚarchitectureÚconfigr   r   r   r   r   r   r)   r   r    r!   r   )r#   rW   r$   Úpackables       r&   Ú$_process_model_before_weight_loadingz4GgufHfQuantizer._process_model_before_weight_loading}   s…   € àŒ~ð 	ØˆFÝ2°4´;Ô3KÈUÌ\ÑZÔZˆŒÝHUÐVZÔVaÐcgÔcoÑHpÔHpÑEˆŒ˜ $Ô"9¸4¼:ØÔ#Ô.ð 	ØˆFÝ7¸¸xÈÌÐVZÔV`ÑaÔaˆÔÐÐr'   c                 ó  — | j         s|S ddlm} | j                             ¦   «         D ]E\  }}| j                             |¦  «        }|�$|                     |j        j	        ¦  «        |_
        ŒF ||¦  «         d|_        |S )z`Fill in what needs the weights already in place: the input permutations, then the layer kernels.r   )Úkernelize_ggml_layersN)r"   Úintegrations.gguf.kernelsr`   r   Úitemsr   ÚgetÚtoÚweightrB   Úinput_permutationÚ_weight_conversions)r#   rW   r$   r`   Ú
param_nameÚmoduleÚpermutations          r&   Ú#_process_model_after_weight_loadingz3GgufHfQuantizer._process_model_after_weight_loading‡   sž   € àŒ~ð 	ØˆLØEÐEÐEÐEÐEÐEà"&Ô"5×";Ò";Ñ"=Ô"=ð 	Pð 	PÑˆJ˜ØÔ1×5Ò5°jÑAÔAˆKØÐ&Ø+6¯>ª>¸&¼-Ô:NÑ+OÔ+O�Ô(øØÐ˜eÑ$Ô$Ð$ð %)ˆÔ!Øˆr'   c                 ó¨   ‡ — ‰ j         s|S ˆ fd„‰ j                             ¦   «         D ¦   «         }t          ‰ j        |z   |‰ j        ‰ j        ¦  «        S )zUPrepend this file's conversions: the GGUF -> transformers mapping, and the unpacking.c                 ó.   •— i | ]\  }}|‰j         v¯||“ŒS r2   )r   )r4   ÚnameÚtr#   s      €r&   ú
<dictcomp>z=GgufHfQuantizer.update_weight_conversions.<locals>.<dictcomp>ž   s-   ø€ ÐfÐfÐf¡  qÀdÐRVÔReÐFeÐFe�T˜1ÐFeÐFeÐFer'   )r"   r   rb   r   r   r   r!   )r#   Úweight_conversionsÚ	to_unpacks   `  r&   Úupdate_weight_conversionsz)GgufHfQuantizer.update_weight_conversionsš   s\   ø€ àŒ~ð 	&Ø%Ð%ØfÐfÐfÐf¨D¬N×,@Ò,@Ñ,BÔ,BÐfÑfÔfˆ	Ý  ¤Ð0BÑ!BÀIÈtÌzÐ[_Ô[eÑfÔfÐfr'   rh   Úreturnc                 ó   — dS ©NFr2   )r#   rW   rh   r$   s       r&   Úparam_needs_quantizationz(GgufHfQuantizer.param_needs_quantization¡   ó   € Øˆur'   c                 ó   — dS rv   r2   ©r#   s    r&   Úis_trainablezGgufHfQuantizer.is_trainable¤   s   € àˆur'   c                 ó   — dS )NTr2   rz   s    r&   Úis_compileablezGgufHfQuantizer.is_compileable¨   s   € àˆtr'   Nc                 ó   — dS rv   r2   )r#   Úsafe_serializations     r&   Úis_serializablezGgufHfQuantizer.is_serializable¬   rx   r'   r9   )Ú__name__Ú
__module__Ú__qualname__Ú__doc__Ú__annotations__Úrequires_calibrationÚ requires_parameters_quantizationr   r.   rD   r3   rJ   rN   rY   r^   rk   rs   Úboolrw   Úpropertyr{   r}   r€   Ú__classcell__)r%   s   @r&   r   r   &   sŠ  ø€ € € € € € ØQÐQà%Ð%Ð%Ñ%ØÐÐÑØ ÐØ',Ð$ðð ð ð ð ðð ð ð"ð ð ð"( Sð (ð (ð (ð (ðð ð ð
!¨cð 
!ð 
!ð 
!ð 
!ðbð bð bðð ð ð&gð gð gð¸#ð ÈDð ð ð ð ð ð˜dð ð ð ñ „Xðð ð ð ð ð ñ „Xððð ¸$ð ð ð ð ð ð ð ð r'   r   )r„   Úutilsr   r   r   Úutils.quantization_configr   Úbaser   rA   ra   r
   Úintegrations.gguf.readerr   r   r   Úintegrations.gguf.utilsr   r   r   r   r   Ú
get_loggerr�   r+   r   r2   r'   r&   ú<module>r‘      s4  ðð SÐ Rà GÐ GÐ GÐ GÐ GÐ GÐ GÐ GÐ GÐ GØ 2Ð 2Ð 2Ð 2Ð 2Ð 2Ø Ð Ð Ð Ð Ð ð ÐÑÔð Ø€L€L€Là;Ð;Ð;Ð;Ð;Ð;Ø_Ð_Ð_Ð_Ð_Ð_Ð_Ð_Ð_Ð_ðð ð ð ð ð ð ð ð ð ð ð ð ð ð 
ˆÔ	˜HÑ	%Ô	%€ðGð Gð Gð Gð G�kñ Gô Gð Gð Gð Gr'   