+
    ªåûjŸO  ã                  óþ  a € 0 t $ R t^ RIHt ^ RIt^ RIt^ RIt^ RIHtH	t	 ^ RIH
t
 ^ RIHt ^ RIHt ^ RIt^ RIt]P"                  ! ]4      t/ tR R ltR	 R
 lt]! ^R7      R R l4       t]
! RR7       ! R R4      4       t ! R R]4      t ! R R]P4                  4      tR tR tR tR tR'R lt R t!R t"R t#R t$R t%]PL                  PN                  PP                  PR                  ]]PL                  PN                  PT                  PR                  ]]PL                  PN                  PV                  PR                  ]!]PL                  PN                  PX                  PZ                  ] ]PL                  PN                  PX                  P\                  ] ]PL                  PN                  P^                  PR                  ]"]PL                  PN                  P`                  PR                  ]#]PL                  PN                  Pb                  PR                  ]$]PL                  PN                  Pd                  PR                  ]%]PL                  PN                  Pf                  PR                  R  /
t4/ t5R!]6R"&   R# R$ lt7R% R& lt8R# )(z@Base classes for quantized tensors with typed layout parameters.)ÚannotationsN)ÚABCÚabstractmethod)Ú	dataclass)Ú	lru_cache)ÚAnyc               ó   € V ^8„  d   QhRR/# ©é   ÚnameÚstr© )Úformats   "ÚUC:\Ai_Coding\apps\studio\ComfyUI\.venv\Lib\site-packages\comfy_kitchen\tensor\base.pyÚ__annotate__r      s   € ÷ ñ  ñ ó    c                ó   € V\         V &   R # ©N©ÚLAYOUTS)r   Úclss   &&r   Úregister_layout_classr      s   € Ø„GˆDƒMr   c               ó   € V ^8„  d   QhRR/# r	   r   )r   s   "r   r   r      s   € ÷ ñ ˜3ñ r   c                ó   € \         V ,          # r   r   )r   s   &r   Úget_layout_classr      s   € Ü�4�=Ðr   )Úmaxsizec               ó   € V ^8„  d   QhRR/# )r
   Úreturnútuple[int, int] | Noner   )r   s   "r   r   r      s   € ÷ .ñ .Ð3ñ .r   c                 óŠ   € \         P                  P                  4       '       g   R# \         P                  P                  4       # )z1Get CUDA compute capability (SM version), cached.N)ÚtorchÚcudaÚis_availableÚget_device_capabilityr   r   r   Úget_cuda_capabilityr$      s-   € ô �:‰:×"Ñ"×$Ò$ÙÜ�:‰:×+Ñ+Ó-Ð-r   T)Úfrozenc                  óz   € ] tR t^(t$ RtR]R&   R]R&   R]R&   R tR	 tR
 R ltR R lt	R R lt
RR R lltRtR# )ÚBaseLayoutParamszÀBase dataclass for layout parameters with common functionality.

Subclasses should define additional fields and override _tensor_fields()
if they have additional tensor fields beyond 'scale'.
útorch.TensorÚscaleútorch.dtypeÚ
orig_dtypeútuple[int, ...]Ú
orig_shapec                	ó&   € V P                  4        R # r   )Ú_validate_tensor_fields©Úselfs   &r   Ú__post_init__ÚBaseLayoutParams.__post_init__3   s   € Ø×$Ñ$Ö&r   c                	ó   € R # r   r   r0   s   &r   r/   Ú(BaseLayoutParams._validate_tensor_fields6   s   € ár   c               ó   € V ^8„  d   QhRR/# )r
   r   z	list[str]r   )r   s   "r   r   ÚBaseLayoutParams.__annotate__:   s   € ÷ ñ  	ñ r   c                ó   € R.# )zBReturn list of field names that are tensors. Override in subclass.r)   r   r0   s   &r   Ú_tensor_fieldsÚBaseLayoutParams._tensor_fields:   s
   € àˆyÐr   c               ó    € V ^8„  d   QhRRRR/# )r
   Údeviceztorch.devicer   r'   r   )r   s   "r   r   r7   >   s   € ÷ $ñ $ ð $Ð1Añ $r   c                ó  € \         P                  ! V 4       Uu/ uF"  q"P                  \        WP                  4      bK$  	  ppV P	                  4        F  pW4,          P                  VR7      W4&   K  	  \        V 4      ! R/ VB # u upi )z/Move all tensor fields to the specified device.©r<   r   )ÚdataclassesÚfieldsr   Úgetattrr9   ÚtoÚtype)r1   r<   ÚfÚkwargsÚfields   &&   r   Ú	to_deviceÚBaseLayoutParams.to_device>   su   € ä9D×9KÒ9KÈDÔ9QÓRÑ9Q°A—&‘&œ' $¯©Ó/Ò/Ñ9QˆÐRØ×(Ñ(Ö*ˆEØ"�M×,Ñ,°FÐ,Ó;ˆF‹Mñ +ä�DŒzÑ#˜FÑ#Ð#ùò Ss   ™(Bc               ó   € V ^8„  d   QhRR/# )r
   r   r'   r   )r   s   "r   r   r7   E   s   € ÷ $ñ $Ð'ñ $r   c                ó  € \         P                  ! V 4       Uu/ uF"  qP                  \        WP                  4      bK$  	  ppV P	                  4        F  pW#,          P                  4       W#&   K  	  \        V 4      ! R/ VB # u upi )zClone all tensor fields.r   )r?   r@   r   rA   r9   ÚclonerC   )r1   rD   rE   rF   s   &   r   rK   ÚBaseLayoutParams.cloneE   sp   € ä9D×9KÒ9KÈDÔ9QÓRÑ9Q°A—&‘&œ' $¯©Ó/Ò/Ñ9QˆÐRØ×(Ñ(Ö*ˆEØ"�M×/Ñ/Ó1ˆF‹Mñ +ä�DŒzÑ#˜FÑ#Ð#ùò Ss   ™(Bc               ó$   € V ^8„  d   QhRRRRRR/# )r
   Úsrcr'   Únon_blockingÚboolr   ÚNoner   )r   s   "r   r   r7   L   s"   € ÷ >ñ >Ð-ð >¸Tð >Èdñ >r   c                ó6  € \         P                  ! V 4       F~  p\        WP                  4      pVP                  V P	                  4       9   d(   \        WP                  4      P                  WBR7       K^  \        P                  WP                  V4       K€  	  R# )z½Copy tensor fields in-place from src, reusing existing memory.

Args:
    src: Source params to copy from (must be same type).
    non_blocking: If True, use non-blocking copy for tensors.
©rO   N)r?   r@   rA   r   r9   Úcopy_ÚobjectÚ__setattr__)r1   rN   rO   rF   Úsrc_vals   &&&  r   Ú	copy_fromÚBaseLayoutParams.copy_fromL   sj   € ô !×'Ò'¨Ö-ˆEÜ˜c§:¡:Ó.ˆGØ�z‰z˜T×0Ñ0Ó2Ô2Ü˜Ÿj™jÓ)×/Ñ/°Ð/ÖSä×"Ñ" 4¯©°WÖ=ó .r   r   N©F)Ú__name__Ú
__module__Ú__qualname__Ú__firstlineno__Ú__doc__Ú__annotations__r2   r/   r9   rG   rK   rX   Ú__static_attributes__r   r   r   r'   r'   (   s>   ‡ ñð
 ÓØÓØÓò'òõõ$õ$÷>ó >r   r'   c                  ó&  € ] tR t^[t$ RtR]R&   RtR]R&   ]]R R l4       4       t	]]R	 R
 l4       4       t
]]R R l4       4       t]]R R l4       4       t]R R l4       t]RR R ll4       t]R R l4       t]R R l4       t]R R l4       tRtR# )ÚQuantizedLayoutzNBase class for quantization layouts. Subclasses define inner Params dataclass.z	type[Any]ÚParamsNr   ÚMIN_SM_VERSIONc               ó    € V ^8„  d   QhRRRR/# )r
   Útensorr(   r   ztuple[torch.Tensor, Any]r   )r   s   "r   r   ÚQuantizedLayout.__annotate__e   s   € ÷ "ñ "˜lð "Ð9Qñ "r   c                	ó   € \         hr   ©ÚNotImplementedError)r   rg   rE   s   &&,r   ÚquantizeÚQuantizedLayout.quantizec   ó
   € ô "Ð!r   c               ó$   € V ^8„  d   QhRRRRRR/# )r
   Úqdatar(   Úparamsr   r   r   )r   s   "r   r   rh   j   s!   € ÷ "ñ "˜|ð "°Sð "¸\ñ "r   c                	ó   € \         hr   rj   ©r   rp   rq   s   &&&r   Ú
dequantizeÚQuantizedLayout.dequantizeh   rn   r   c               ó    € V ^8„  d   QhRRRR/# )r
   ÚqtensorÚQuantizedTensorr   ztuple[torch.Tensor, ...]r   )r   s   "r   r   rh   o   s   € ÷ "ñ "¨ð "Ð<Tñ "r   c                	ó   € \         hr   rj   ©r   rw   s   &&r   Úget_plain_tensorsÚ!QuantizedLayout.get_plain_tensorsm   rn   r   c               ó$   € V ^8„  d   QhRRRRRR/# )r
   rp   r(   rq   r   r   údict[str, torch.Tensor]r   )r   s   "r   r   rh   t   s"   € ÷ "ñ " |ð "¸Sð "ÐE\ñ "r   c                	ó   € \         hr   rj   rs   s   &&&r   Ústate_dict_tensorsÚ"QuantizedLayout.state_dict_tensorsr   rn   r   c               ó    € V ^8„  d   QhRRRR/# )r
   rw   rx   r   údict[str, Any]r   )r   s   "r   r   rh   x   s   € ÷ ñ ¨ð ¸Nñ r   c                ó   € / # )zDReturn layout-specific kwargs for requantizing a dequantized tensor.r   rz   s   &&r   Úrequantize_kwargsÚ!QuantizedLayout.requantize_kwargsw   s	   € ð ˆ	r   c               ó    € V ^8„  d   QhRRRR/# )r
   Úweightrx   Úxr(   r   )r   s   "r   r   rh   }   s   € ÷ ñ  oð ¸,ñ r   c	                ó   € \         # )zOReturn the activated linear result, or NotImplemented without modifying inputs.)ÚNotImplemented)	r   rˆ   r‰   ÚbiasÚ	input_actÚinput_act_weightÚinput_act_epsÚresidualÚresidual_scales	   &&&&&&&&&r   Úlinear_input_actÚ QuantizedLayout.linear_input_act|   s
   € ô
 Ðr   c               ó$   € V ^8„  d   QhRRRRRR/# )r
   rˆ   rx   Ústorager(   r   r   )r   s   "r   r   rh   „   s"   € ÷ ñ  ð Àð ÐRañ r   c                ó   € V# )z>Optionally optimize resident storage for small-M linear calls.r   )r   rˆ   r•   s   &&&r   Úoptimize_for_gemvÚ!QuantizedLayout.optimize_for_gemvƒ   s	   € ð ˆr   c               ó   € V ^8„  d   QhRR/# ©r
   r   rP   r   )r   s   "r   r   rh   ‰   s   € ÷ )ñ ) Tñ )r   c                ó^   € V P                   f   R# \        4       pVf   R# WP                   8¬  # )z@Check if fast quantized matmul is supported on current hardware.TF)re   r$   ©r   Úcaps   & r   Úsupports_fast_matmulÚ$QuantizedLayout.supports_fast_matmulˆ   s2   € ð ×ÑÒ%ÙÜ!Ó#ˆØŠ;ÙØ×(Ñ(Ñ(Ð(r   c               ó   € V ^8„  d   QhRR/# )r
   r   rƒ   r   )r   s   "r   r   rh   “   s   € ÷ 
ñ 
 ñ 
r   c           	     ón   € \        4       pRV P                  RV P                  RVRV P                  4       /# )z6Return hardware/software requirements for this layout.ÚlayoutÚmin_sm_versionÚcurrent_sm_versionÚfast_matmul_supported)r$   r[   re   rž   rœ   s   & r   Úget_requirementsÚ QuantizedLayout.get_requirements’   s=   € ô "Ó#ˆà�c—l‘lØ˜c×0Ñ0Ø  #Ø# S×%=Ñ%=Ó%?ð	
ð 	
r   r   ©NNNg        NN)r[   r\   r]   r^   r_   r`   re   Úclassmethodr   rl   rt   r{   r€   r…   r’   r—   rž   r¦   ra   r   r   r   rc   rc   [   sê   ‡ ÙXàÓð .2€NÐ*Ó1àØô"ó ó ð"ð Øô"ó ó ð"ð Øô"ó ó ð"ð Øô"ó ó ð"ð ôó ðð öó ðð ôó ðð ô)ó ð)ð ô
ó ô
r   rc   c                  ó˜  € ] tR t^žtRt]R R l4       tR R ltR R lt]	R R	 l4       t
]	R
 R l4       t]	R R l4       t]	R R l4       t]	R R l4       t]	R R l4       t]	R R l4       t]R R l4       tR-R R lltR tR tR tR R ltR.R  R! lltR" R# ltR/R$ R% lltR& R' ltR( t]R) 4       t]R0R* l4       t]R+ 4       tR,t R# )1rx   zÛ
Quantized tensor with typed layout parameters.

Properties:
    shape: Original (unpadded) shape
    storage_shape: Actual shape of quantized data (may be padded)
    is_padded: True if storage_shape != original shape
c               ó$   € V ^8„  d   QhRRRRRR/# ©r
   rp   r(   Ú
layout_clsr   rq   r   r   )r   s   "r   r   ÚQuantizedTensor.__annotate__©   s(   € ÷ 
ñ 
àð
ð ð
ð ñ	
r   c                	ó†   € \         P                  P                  V VP                  VP                  VP
                  R R7      # )F)r<   ÚdtypeÚrequires_grad)r    ÚTensorÚ_make_wrapper_subclassr-   r<   r+   )r   rp   r­   rq   s   &&&&r   Ú__new__ÚQuantizedTensor.__new__¨   s?   € ô �|‰|×2Ñ2ØØ×ÑØ—<‘<Ø×#Ñ#Øð 3ó 
ð 	
r   c               ó$   € V ^8„  d   QhRRRRRR/# r¬   r   )r   s   "r   r   r®   ·   s(   € ÷ 	ñ 	àð	ð ð	ð ñ		r   c                	óZ   € \        V\        4      '       g   Q hWn        W n        W0n        R # r   )Ú
isinstancer   Ú_qdataÚ_layout_clsÚ_params)r1   rp   r­   rq   s   &&&&r   Ú__init__ÚQuantizedTensor.__init__·   s'   € ô ˜*¤c×*Ò*Ð*Ð*ØŒØ%ÔØŽr   c               ó   € V ^8„  d   QhRR/# )r
   r   r   r   )r   s   "r   r   r®   Â   s   € ÷ 
ñ 
˜#ñ 
r   c           	     	ó–   € R \        V P                  4       RV P                   RV P                   RV P                  P
                   R2	# )zQuantizedTensor(shape=z, storage_shape=z	, layout=z, dtype=Ú))ÚtupleÚshapeÚstorage_shaperº   r»   r+   r0   s   &r   Ú__repr__ÚQuantizedTensor.__repr__Â   sV   € à$¤U¨4¯:©:Ó%6Ð$7ð 8Ø!×/Ñ/Ð0ð 1Ø×&Ñ&Ð'ð (Ø—\‘\×,Ñ,Ð-¨Qð0ð	
r   c               ó   € V ^8„  d   QhRR/# ©r
   r   r,   r   )r   s   "r   r   r®   Í   s   € ÷ (ñ (˜ñ (r   c                	ó@   € \        V P                  P                  4      # r   )rÁ   r¹   rÂ   r0   s   &r   rÃ   ÚQuantizedTensor.storage_shapeÌ   s   € ä�T—[‘[×&Ñ&Ó'Ð'r   c               ó   € V ^8„  d   QhRR/# )r
   r   r*   r   )r   s   "r   r   r®   Ñ   s   € ÷ !ñ !˜{ñ !r   c                ó.   € V P                   P                  # )zKThe dtype of the underlying quantized storage (e.g., float8_e4m3fn, uint8).)r¹   r°   r0   s   &r   Ústorage_dtypeÚQuantizedTensor.storage_dtypeÐ   s   € ð �{‰{× Ñ Ð r   c               ó   € V ^8„  d   QhRR/# )r
   r   Úintr   )r   s   "r   r   r®   Ö   s   € ÷ "ñ "˜ñ "r   c                ó.   € V P                   P                  # )z÷Return the actual storage size in bytes.

Note: This returns the true memory footprint of the quantized data,
not the logical size based on orig_dtype. For example, an FP8 quantized
tensor with logical dtype bfloat16 returns the FP8 storage size.
)r¹   Únbytesr0   s   &r   rÑ   ÚQuantizedTensor.nbytesÕ   s   € ð �{‰{×!Ñ!Ð!r   c               ó   € V ^8„  d   QhRR/# rÇ   r   )r   s   "r   r   r®   à   s   € ÷ "ñ "˜oñ "r   c                	óž   € \        V P                  4      p\        VR 4      '       d   VP                  V P                  4      # V P                  # )Úget_logical_shape_from_storage)r   rº   ÚhasattrrÕ   rÃ   )r1   r­   s   & r   Úpadded_shapeÚQuantizedTensor.padded_shapeß   sC   € ä% d×&6Ñ&6Ó7ˆ
Ü�:Ð?×@Ò@Ø×<Ñ<¸T×=OÑ=OÓPÐPØ×!Ñ!Ð!r   c               ó   € V ^8„  d   QhRR/# rš   r   )r   s   "r   r   r®   ç   s   € ÷ <ñ <˜4ñ <r   c                	óH   € V P                   P                  V P                  8g  # r   )r»   r-   r×   r0   s   &r   Ú	is_paddedÚQuantizedTensor.is_paddedæ   s   € à�|‰|×&Ñ&¨$×*;Ñ*;Ñ;Ð;r   c               ó   € V ^8„  d   QhRR/# )r
   r   útype[QuantizedLayout]r   )r   s   "r   r   r®   ë   s   € ÷ 2ñ 2Ð1ñ 2r   c                	ó,   € \        V P                  4      # r   )r   rº   r0   s   &r   r­   ÚQuantizedTensor.layout_clsê   s   € ä × 0Ñ 0Ó1Ð1r   c               ó   € V ^8„  d   QhRR/# )r
   r   r   r   )r   s   "r   r   r®   ï   s   € ÷ ñ ˜ñ r   c                	ó   € V P                   # r   )r»   r0   s   &r   rq   ÚQuantizedTensor.paramsî   s   € à�|‰|Ðr   c               ó$   € V ^8„  d   QhRRRRRR/# )r
   rg   r(   r­   r   r   rx   r   )r   s   "r   r   r®   õ   s(   € ÷ .ñ .àð.ð ð.ð
 
ñ.r   c                	óP   € \        V4      P                  ! V3/ VB w  rEV ! WBV4      # r   )r   rl   )r   rg   r­   rE   rp   rq   s   &&&,  r   Ú
from_floatÚQuantizedTensor.from_floatô   s-   € ô )¨Ó4×=Ò=¸fÑOÈÑO‰ˆÙ�5 fÓ-Ð-r   Nc               ó(   € V ^8„  d   QhRRRRRRRR/# )	r
   rp   ztorch.Tensor | Nonerq   z
Any | NoneÚclone_paramsrP   r   rx   r   )r   s   "r   r   r®   þ   s2   € ÷ 
ñ 
à"ð
ð ð
ð ð	
ð
 
ñ
r   c                ó´   € Vf/   V'       d   V P                   P                  4       MV P                   p\        Ve   TMV P                  V P                  V4      # )ao  Create a copy with optionally modified qdata/params.

Args:
    qdata: New quantized data tensor. If None, uses self._qdata.
    params: New parameters. If None, uses self._params (cloned if clone_params=True).
    clone_params: If True and params is None, clone self._params. Set to False
        when you know params don't need cloning (e.g., they're already new).
)r»   rK   rx   r¹   rº   )r1   rp   rq   ré   s   &&&&r   Ú
_copy_withÚQuantizedTensor._copy_withþ   sJ   € ð Š>ß-9�T—\‘\×'Ñ'Ô)¸t¿|¹|ˆFÜØÒ&‰E¨D¯K©KØ×ÑØó
ð 	
r   c                	ó6   € V P                   P                  4       # r   )r¹   Údata_ptrr0   s   &r   rî   ÚQuantizedTensor.data_ptr  s   € Ø�{‰{×#Ñ#Ó%Ð%r   c                	ó6   € V P                   P                  4       # r   )r¹   Ú	is_pinnedr0   s   &r   rñ   ÚQuantizedTensor.is_pinned  s   € Ø�{‰{×$Ñ$Ó&Ð&r   c                	ó6   € V P                   P                  4       # r   )r¹   r•   r0   s   &r   r•   ÚQuantizedTensor.storage  s   € Ø�{‰{×"Ñ"Ó$Ð$r   c               ó    € V ^8„  d   QhRRRR/# )r
   r•   r(   r   rx   r   )r   s   "r   r   r®     s   € ÷ @ñ @¨ð @¸/ñ @r   c                ó8   € V P                   P                  W4      # )a#  Optimize a resident weight within its caller-owned byte buffer.

May overwrite this weight's storage; use the returned tensor afterwards.
Other weights/metadata in the buffer remain intact. Both GEMV and GEMM must
remain supported. Unsupported layouts return self without modifying storage.
)r­   r—   )r1   r•   s   &&r   r—   Ú!QuantizedTensor.optimize_for_gemv  s   € ð �‰×0Ñ0°Ó?Ð?r   c               ó   € V ^8„  d   QhRR/# )r
   r‰   r(   r   )r   s   "r   r   r®   (  s   € ÷ 

ñ 

 ,ñ 

r   c           
     ó>   € V P                   P                  WW#WEWg4      # )zëCompute linear(act(x), self, bias), optionally followed by residual addcmul.

Return NotImplemented when the layout cannot handle this operation; the caller
owns the eager fallback. Norm weights must already be prepared on x's device.
)r­   r’   )r1   r‰   rŒ   r�   rŽ   r�   r�   r‘   s   &&&&&&&&r   r’   Ú QuantizedTensor.linear_input_act(  s%   € ð �‰×/Ñ/Ø�TÐ&6Àxó
ð 	
r   c               ó   € V ^8„  d   QhRR/# )r
   r   r(   r   )r   s   "r   r   r®   4  s   € ÷ ñ ˜Lñ r   c                	óF  € V P                   P                  4       '       g   V P                   P                  4       MV P                   p\        V P                  R R4      pV'       d¼   V P                  P
                  ^,          V P                  P
                  ^ ,          3pV P                  P                  WP                  4      pVP                  R,          V8w  d5   \        ;QJ d    . R V 4       F  NK  	  5M! R V 4       4      pWE,          pVP                  4       # V P                  P                  WP                  4      pV P                  P
                  pVP                  V8w  d5   \        ;QJ d    . R V 4       F  NK  	  5M! R V 4       4      pWE,          # V# )Ú
transposedF:Nr
   Nc              3  ó:   "  € T F  p\        ^ V4      x € K  	  R# 5i©é    N©Úslice©Ú.0Úss   & r   Ú	<genexpr>Ú-QuantizedTensor.dequantize.<locals>.<genexpr>@  s   é € ÐC±N¨qœu Q¨Ÿ{˜{³Nùó   ‚c              3  ó:   "  € T F  p\        ^ V4      x € K  	  R# 5irÿ   r  r  s   & r   r  r  G  s   é € Ð5±¨1œ5  AŸ;˜;³ùr  )r¹   Úis_contiguousÚ
contiguousrA   r»   r-   r­   rt   rÂ   rÁ   Út)r1   rp   Úis_transposedÚphysical_shapeÚfullÚslicesÚorigs   &      r   rt   ÚQuantizedTensor.dequantize4  s'  € ð 15·±×0IÑ0I×0KÒ0K�—‘×&Ñ&Ô(ÐQU×Q\ÑQ\ˆô   §¡¨l¸EÓBˆçØ"Ÿl™l×5Ñ5°aÕ8¸$¿,¹,×:QÑ:QÐRSÕ:TÐUˆNØ—?‘?×-Ñ-¨e·\±\ÓBˆDØ�z‰z˜"�~ Ô/ßœÑC±NÓCŸ™ÑC±NÓCÓC�Ø•|�Ø—6‘6“8ˆOà�‰×)Ñ)¨%·±Ó>ˆØ�|‰|×&Ñ&ˆØ�:‰:˜Ôß”UÑ5±Ó5—U‘UÑ5±Ó5Ó5ˆFØ•<ÐØˆr   c               ó    € V ^8„  d   QhRRRR/# )r
   Úprefixr   r   r~   r   )r   s   "r   r   r®   K  s   € ÷ Sñ S ð SÐ.Eñ Sr   c                	óÀ   € V P                   P                  V P                  V P                  4      pVP	                  4        UUu/ uF  w  r4V V 2VbK  	  upp# u uppi r   )r­   r€   r¹   r»   Úitems)r1   r  ÚtensorsÚsuffixrg   s   &&   r   Ú
state_dictÚQuantizedTensor.state_dictK  sN   € Ø—/‘/×4Ñ4°T·[±[À$Ç,Á,ÓOˆØBIÇ-Á-Ä/ÔRÁ/±°�6�(˜6˜(Ð# VÒ+Á/ÒRÐRùÓRs   ÁAc               ó    € V ^8„  d   QhRRRR/# )r
   rg   r(   r   rx   r   )r   s   "r   r   r®   O  s   € ÷ Jñ J¨Lð JÀñ Jr   c                ó¦   € V P                   P                  V 4      pVP                  V4       \        V 4      P                  ! WP
                  3/ VB # )zŽQuantize a float tensor using this tensor's layout-specific options.

Caller-supplied kwargs override preserved layout options such as scale.
)r­   r…   ÚupdaterC   ræ   rº   )r1   rg   rE   Úqkwargss   &&, r   Úrequantize_from_floatÚ%QuantizedTensor.requantize_from_floatO  sD   € ð
 —/‘/×3Ñ3°DÓ9ˆØ�‰�vÔÜ�D‹z×$Ò$ V×-=Ñ-=ÑIÀÑIÐIr   c           	     	óØ  € R .p/ p/ p\         P                  ! V P                  4       F—  p\        V P                  VP                  4      p\        V\        P                  4      '       dG   RVP                   2p\        P                  WV4       VP                  V4       WbVP                  &   K‰  WSVP                  &   K™  	  VRV P                  R\        V P                  4      RVRV/3# )r¹   Ú_param_r­   Úparams_classÚtensor_fieldsÚnon_tensor_fields)r?   r@   r»   rA   r   r¸   r    r²   rU   rV   Úappendrº   rC   )r1   Úinner_tensorsr$  r%  rF   ÚvalueÚ	attr_names   &      r   Ú__tensor_flatten__Ú"QuantizedTensor.__tensor_flatten__Z  sÐ   € Ø!˜
ˆØˆØÐä ×'Ò'¨¯©Ö5ˆEÜ˜DŸL™L¨%¯*©*Ó5ˆEÜ˜%¤§¡×.Ò.Ø% e§j¡j \Ð2�	Ü×"Ñ" 4°EÔ:Ø×$Ñ$ YÔ/Ø,5˜eŸj™jÓ)à05 %§*¡*Ó-ñ 6ð Ø˜$×*Ñ*ØœD §¡Ó.Ø˜]ØÐ!2ð	
ð 
ð 	
r   c                	óÎ   € \        VR ,          4      pVR,          P                  4        F  w  rVW,          WE&   K  	  VR,          ! R/ VB p\        V R,          VR,          V4      # )r%  r$  r#  r¹   r­   r   )Údictr  rx   )r'  ÚctxÚ
outer_sizeÚouter_strideÚparams_kwargsÚ
field_namer)  rq   s   &&&&    r   Ú__tensor_unflatten__Ú$QuantizedTensor.__tensor_unflatten__p  sf   € ä˜SÐ!4Õ5Ó6ˆØ%(¨Õ%9×%?Ñ%?Ö%AÑ!ˆJØ(5Õ(@ˆMÓ%ñ &Bð �^Ö$Ñ5 }Ñ5ˆÜ˜}¨XÕ6¸¸LÕ8IÈ6ÓRÐRr   c                	ó²  € T;'       g    / pV'       d
   V^ ,          MRp\         P                  V4      pVe
   V! WSV4      # \        V4      pV'       dC   V\        9   d8   \        V,          pVP                   F  p	W˜9   g   K  W‰,          ! WSV4      u # 	  \
        P                  RT RV'       d   VP                  MR R24       V P                  WV4      # )r   NzUnhandled op z for Úunknownz, dequantizing)	Ú_DISPATCH_TABLEÚgetÚ_get_layout_from_argsÚ_LAYOUT_DISPATCH_TABLEÚ__mro__ÚloggerÚdebugr[   Ú_dequant_and_fallback)
r   ÚfuncÚtypesÚargsrE   ÚqtÚhandlerr­   Úop_handlersÚ
parent_clss
   &&&&&     r   Ú__torch_dispatch__Ú"QuantizedTensor.__torch_dispatch__{  sÃ   € à—�˜2ˆßˆT�!ŽW $ˆô "×%Ñ% dÓ+ˆØÒÙ˜2 VÓ,Ð,ô +¨4Ó0ˆ
ß˜$Ô"8Ô8Ü0°Õ6ˆKØ(×0Ô0�
ØÖ,Ø&Ö2°2¸VÓDÒDñ 1ô
 	�‰�} T F¨%Çz°
×0CÒ0CÐW`Ð/aÐaoÐpÔqØ×(Ñ(¨°VÓ<Ð<r   c                	ó6   € V! \        V4      / \        V4      B # r   ©Údequantize_args)r   r?  rA  rE   s   &&&&r   r>  Ú%QuantizedTensor._dequant_and_fallback‘  s   € á”_ TÓ*ÐF¬o¸fÓ.EÑFÐFr   )rº   r»   r¹   )NNTr¨   )Ú )r   N)!r[   r\   r]   r^   r_   Ústaticmethodr´   r¼   rÄ   ÚpropertyrÃ   rÌ   rÑ   r×   rÛ   r­   rq   r©   ræ   rë   rî   rñ   r•   r—   r’   rt   r  r  r*  r3  rF  r>  ra   r   r   r   rx   rx   ž   s:  † ñð ô
ó ð
õ	õ
ð ô(ó ð(ð ô!ó ð!ð ô"ó ð"ð ô"ó ð"ð ô<ó ð<ð ô2ó ð2ð ôó ðð
 ô.ó ð.÷
ò0&ò'ò%õ@÷

õ÷.SõJò
ð, ñSó ðSð ó=ó ð=ð* ñGó ôGr   rx   c                óN  € \        V \        4      '       d   V P                  4       # \        V \        4      '       d/   V P	                  4        UUu/ uF  w  rV\        V4      bK  	  upp# \        V \        \        34      '       d   \        V 4      ! R V  4       4      # V # u uppi )z¾Recursively dequantize QuantizedTensors in args/kwargs.

Useful for fallback implementations that need to convert quantized
tensors back to regular tensors before calling the underlying op.
c              3  ó8   "  € T F  p\        V4      x € K  	  R # 5ir   rI  )r  Úas   & r   r  Ú"dequantize_args.<locals>.<genexpr>£  s   é € Ð;±d°œ/¨!×,Ð,³dùs   ‚)	r¸   rx   rt   r-  r  rJ  ÚlistrÁ   rC   )rA  ÚkÚvs   &  r   rJ  rJ  ˜  sƒ   € ô �$œ×(Ò(Ø�‰Ó Ð Ü	�Dœ$×	Ò	Ø26·*±*´,Ô?±,©$¨!�”? 1Ó%Ò%±,Ò?Ð?Ü	�Dœ4¤˜-×	(Ò	(Ü�DŒzÑ;±dÓ;Ó;Ð;Ø€Kùó @s   ÁB!c                ó.  € VP                  R4      pVP                  R4      pV R,           F¤  p\        V\        P                  4      '       d   TpK'  \        V\        P                  4      '       d   TpKK  \        V\
        4      '       g   Kc  \        P                  ! \        4      ;_uu_ 4        \        P                  ! V4      pRRR4       K¦  	  \        V\
        4      '       d   \        P                  ! V4      pW#3#   + '       g   i     Kè  ; i)z.Extract device and dtype from .to() arguments.r<   r°   :é   NNN)	r8  r¸   r    r<   r°   r   Ú
contextlibÚsuppressÚ	Exception)rA  rE   r<   r°   Úargs   &&   r   Ú_parse_to_argsr\  ©  s¿   € à�Z‰Z˜Ó!€FØ�J‰J�wÓ€EØ�B�xˆxˆÜ�cœ5Ÿ<™<×(Ò(ØŠFÜ˜œUŸ[™[×)Ò)ØŠEÜ˜œS×!Ô!Ü×$Ò$¤Y×/Õ/ÜŸš cÓ*�÷ 0Ñ/ñ ô �&œ#×ÒÜ—’˜fÓ%ˆØˆ=Ð÷	 0×/Ð/ús   Â1DÄDc                óV   € V P                  V P                  P                  4       R 7      # ©)rp   )rë   r¹   Údetach©rB  rA  rE   s   &&&r   Ú_handle_detachra  º  s    € Ø�=‰=˜rŸy™y×/Ñ/Ó1ˆ=Ó2Ð2r   c                óV   € V P                  V P                  P                  4       R 7      # r^  )rë   r¹   rK   r`  s   &&&r   Ú_handle_clonerc  ¾  s   € Ø�=‰=˜rŸy™yŸ™Ó0ˆ=Ó1Ð1r   c                óT  € \        W4      w  rEVR J;'       d    W@P                  P                  8g  pVR J;'       d    WPP                  P                  8g  pV'       g   V'       g   V'       g   V # V'       d9   V P                  P                  VR7      pV P                  P                  V4      p	MHV'       d   V P                  P                  4       MV P                  pV P                  P                  4       p	V'       d   \        P                  ! W•R7      p	V P                  W‰RR7      # )Nr>   ©r+   F©rp   rq   ré   )r\  r¹   r<   r»   r+   rB   rG   rK   r?   Úreplacerë   )
rB  rA  rE   Ú
force_copyÚtarget_deviceÚtarget_dtypeÚneeds_deviceÚneeds_dtypeÚ	new_qdataÚ
new_paramss
   &&&&      r   Ú
_handle_toro  Â  sÔ   € Ü"0°Ó">Ñ€Mà ¨Ð,×RÐR°Ç)Á)×BRÑBRÑ1R€LØ dÐ*×TÐT¨|¿z¹z×?TÑ?TÑ/T€Kç§·JØˆ	çØ—I‘I—L‘L¨�LÓ6ˆ	Ø—Z‘Z×)Ñ)¨-Ó8‰
ç)3�B—I‘I—O‘OÔ%¸¿¹ˆ	Ø—Z‘Z×%Ñ%Ó'ˆ
çÜ ×(Ò(¨ÔMˆ
à�=‰=˜yÈ%ˆ=ÓPÐPr   c                ó   € \        WVR R7      # )T)rh  )ro  r`  s   &&&r   Ú_handle_to_copyrq  Ø  s   € Ü�b °4Ô8Ð8r   c                óš   € V P                   P                  4       '       d   V # V P                  V P                   P                  4       R 7      # r^  )r¹   r
  rë   r  r`  s   &&&r   Ú_handle_contiguousrs  Ü  s8   € Ø	‡y�y×Ñ× Ò Øˆ	Ø�=‰=˜rŸy™y×3Ñ3Ó5ˆ=Ó6Ð6r   c                ó6   € V P                   P                  4       # r   )r¹   r
  r`  s   &&&r   Ú_handle_is_contiguousru  â  s   € Ø�9‰9×"Ñ"Ó$Ð$r   c                ój  € V^ ,          V^,          rC\        V\        4      '       g#   \        R\        V4      P                   R24      hVP
                  VP
                  8w  d&   \        RVP
                   RVP
                   24      hVP                  P                  pVP                  R\        V4      ^8¬  4      pVP                  P                  VP                  VR7       VP                  P                  VP                  VR7       \        P                  ! VP                  VR7      Vn        V# )r   zCannot copy z to QuantizedTensorzLayout mismatch: z vs rO   rS   re  )r¸   rx   Ú	TypeErrorrC   r[   rº   r»   r+   r8  Úlenr¹   rT   rX   r?   rg  )rB  rA  rE   ÚdstrN   Údst_orig_dtyperO   s   &&&    r   Ú_handle_copy_r{  æ  së   € Ø�A�w˜˜Q�ˆÜ�cœ?×+Ò+Ü˜,¤t¨C£y×'9Ñ'9Ð&:Ð:MÐNÓOÐOØ
‡�˜#Ÿ/™/Ô)ÜÐ+¨C¯O©OÐ+<¸DÀÇÁÐ@QÐRÓSÐSà—[‘[×+Ñ+€NØ—:‘:˜n¬c°$«i¸1©nÓ=€Là‡J�J×Ñ�S—Z‘Z¨lÐÔ;Ø‡K�K×Ñ˜#Ÿ+™+°LÐÔAÜ×%Ò% c§k¡k¸nÔM€C„KØ€Jr   c                óF  € VP                  R R4      pVP                  R4      p\        P                  ! V P                  VR7      pV P
                  P                  4       pVe   VP                  V4      pVe   \        P                  ! WcR7      pV P                  WVRR7      # )r°   Nr<   r>   re  Frf  )Úpopr8  r    Ú
empty_liker¹   r»   rK   rG   r?   rg  rë   )rB  rA  rE   rj  ri  rm  rn  s   &&&    r   Ú_handle_empty_liker  ö  s‡   € Ø—:‘:˜g tÓ,€LØ—J‘J˜xÓ(€Mä× Ò  §¡°=ÔA€IØ—‘×!Ñ!Ó#€JàÒ Ø×)Ñ)¨-Ó8ˆ
ØÒÜ ×(Ò(¨ÔMˆ
à�=‰=˜yÈ%ˆ=ÓPÐPr   c                ó   € R # )Tr   r`  s   &&&r   Ú<lambda>r�    s   € ÑW[r   z+dict[Any, dict[type[QuantizedLayout], Any]]r:  c               ó    € V ^8„  d   QhRRRR/# )r
   Útorch_opr   r­   rÞ   r   )r   s   "r   r   r     s   € ÷ ñ  ð Ð2Gñ r   c                ó   a a€ VV 3R lpV# )zÁDecorator to register a layout-specific operation handler.

Args:
    torch_op: PyTorch operation (e.g., torch.ops.aten.linear.default)
    layout_cls: Layout class (e.g., TensorCoreFP8Layout)
c                óP   <€ S\         9  d
   / \         S&   V \         S,          S&   V # r   )r:  )Úhandler_funcr­   rƒ  s   &€€r   Ú	decoratorÚ%register_layout_op.<locals>.decorator  s-   ø€ ØÔ1Ô1Ø/1Ô" 8Ñ,Ø7CÔ˜xÕ(¨Ñ4ØÐr   r   )rƒ  r­   r‡  s   ff r   Úregister_layout_opr‰    s   ù€ öð
 Ðr   c               ó   € V ^8„  d   QhRR/# )r
   r   ztype[QuantizedLayout] | Noner   )r   s   "r   r   r   %  s   € ÷ 	ñ 	Ð#?ñ 	r   c                ó  € V  F†  p\        V\        4      '       d   \        VP                  4      u # \        V\        \
        34      '       g   KN  V F2  p\        V\        4      '       g   K  \        VP                  4      u u # 	  Kˆ  	  R# )z.Extract layout class from operation arguments.N)r¸   rx   r   rº   rS  rÁ   )rA  r[  Úitems   &  r   r9  r9  %  sf   € ãˆÜ�cœ?×+Ò+Ü# C§O¡OÓ4Ò4Ü˜œd¤E˜]×+Ô+Û�Ü˜d¤O×4Ô4Ü+¨D×,<Ñ,<Ó=Ô=ó ñ	 ñ r   rZ   )9Ú__conditional_annotations__r_   Ú
__future__r   rX  r?   ÚloggingÚabcr   r   r   Ú	functoolsr   Útypingr   r    Útorch._dynamoÚ	getLoggerr[   r<  r   r   r   r$   r'   rc   r²   rx   rJ  r\  ra  rc  ro  rq  rs  ru  r{  r  ÚopsÚatenr_  ÚdefaultrK   Ú_to_copyrB   Údtype_layoutr°   r  r
  rT   r~  Ú!_has_compatible_shallow_copy_typer7  r:  r`   r‰  r9  )r�  s   @r   Ú<module>r›     s÷  øðÜ FÝ "ã Û Û ß #Ý !Ý Ý ã Û à	×	Ò	˜8Ó	$€à
€õõñ �1Ôô.ó ð.ñ �$Ô÷/>ð />ó ð/>ôd@
�cô @
ôFuG�e—l‘lô uGòtò"ò"3ò2ôQò,9ò7ò%òò Qð  
‡I�I‡N�N×Ñ×!Ñ! >Ø	‡I�I‡N�N×Ñ× Ñ  -Ø	‡I�I‡N�N×Ñ×#Ñ# _Ø	‡I�I‡N�N×Ñ×"Ñ" JØ	‡I�I‡N�N×Ñ×Ñ˜ZØ	‡I�I‡N�N×Ñ×%Ñ%Ð'9Ø	‡I�I‡N�N× Ñ ×(Ñ(Ð*?Ø	‡I�I‡N�N×Ñ× Ñ  -Ø	‡I�I‡N�N×Ñ×%Ñ%Ð'9Ø	‡I�I‡N�N×4Ñ4×<Ñ<Ñ>[ð€ð GIÐ ÐCÓ Hõ÷	r   