+
    ªåûj”  ã                  ó.  € R t ^ RIHt ^ RIt^ RIHt ^ RIt^ RIt^ RI	H
t
 ^RIHtHtHtHtHt ]P"                  ! ]4      t ! R R]4      tRR R	 lltR
 t]! ]P.                  P0                  P2                  P4                  ]4      R 4       t]! ]P.                  P0                  P8                  P4                  ]4      R 4       t]! ]P.                  P0                  P<                  P4                  ]4      R 4       t]P.                  P0                  P@                  P4                  ]P.                  P0                  PB                  P4                  ]P.                  P0                  PD                  P4                  3 F  t#]! ]#]4      ! ]! ]#4      4       K  	  R# )z0FP8 (E4M3) quantization layout for tensor cores.)ÚannotationsN)Ú	dataclass)Úscaled_mm_v2)ÚBaseLayoutParamsÚQuantizedLayoutÚQuantizedTensorÚdequantize_argsÚregister_layout_opc                  óÀ   € ] tR t^tRtRt]! RR7       ! R R]4      4       t]	R]
P                  3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TensorCoreFP8Layoutaˆ  FP8 E4M3 quantization with per-tensor scaling.

Example:
    >>> x = torch.randn(128, 256, device="cuda", dtype=torch.bfloat16)
    >>> qt = QuantizedTensor.from_float(x, TensorCoreFP8Layout)
    >>> qt.shape
    torch.Size([128, 256])
    >>> dq = qt.dequantize()
    >>> torch.allclose(dq, x, rtol=0.1)
    True

Note:
    Requires SM >= 8.9 (Ada Lovelace) for hardware-accelerated matmul.
T)Úfrozenc                  ó   € ] tR t^)tR tRtR# )ÚTensorCoreFP8Layout.Paramsc           	     	óÞ   € \        V P                  \        P                  4      '       dC   \        P                  V R V P                  P                  \        P                  RR7      4       R# R# )ÚscaleT)ÚdtypeÚnon_blockingN)Ú
isinstancer   ÚtorchÚTensorÚobjectÚ__setattr__ÚtoÚfloat32)Úselfs   &ÚTC:\Ai_Coding\apps\studio\ComfyUI\.venv\Lib\site-packages\comfy_kitchen\tensor\fp8.pyÚ_validate_tensor_fieldsÚ2TensorCoreFP8Layout.Params._validate_tensor_fields+   sD   € Ü˜$Ÿ*™*¤e§l¡l×3Ò3Ü×"Ñ" 4¨°$·*±*·-±-ÄeÇmÁmÐbf°-Ó2gÖhñ 4ó    © N)Ú__name__Ú
__module__Ú__qualname__Ú__firstlineno__r   Ú__static_attributes__r   r   r   ÚParamsr   )   s   † ö	ir   r%   Nc               ó(   € V ^8„  d   QhRRRRRRRR/# )	é   Útensorútorch.Tensorr   z!torch.Tensor | float | str | Noner   ztorch.dtypeÚreturnztuple[torch.Tensor, Params]r   )Úformats   "r   Ú__annotate__Ú TensorCoreFP8Layout.__annotate__0   s8   € ÷ \ñ \àð\ð 1ð\ð ð	\ð 
%ñ\r   c                	ó  € VP                   p\        VP                  4      pVe   VR8X  dJ   \        P                  ! VP                  4       4      \        P                  ! V4      P                  ,          p\        V\        P                  4      '       g   \        P                  ! V4      pVP                  VP                  \        P                  R7      p\        P                  ! WV4      pWpP!                  W%VR7      3# )NÚrecalculate)Údevicer   ©r   Ú
orig_dtypeÚ
orig_shape)r   ÚtupleÚshaper   ÚamaxÚabsÚfinfoÚmaxr   r   r(   r   r0   r   ÚckÚquantize_per_tensor_fp8r%   )Úclsr(   r   r   Úkwargsr2   r3   Úqdatas   &&&&,   r   ÚquantizeÚTensorCoreFP8Layout.quantize/   s±   € ð —\‘\ˆ
Ü˜6Ÿ<™<Ó(ˆ
àŠ=˜E ]Ô2Ü—J’J˜vŸz™z›|Ó,¬u¯{ª{¸5Ó/A×/EÑ/EÕEˆEä˜%¤§¡×.Ò.Ü—L’L Ó'ˆEØ—‘ §¡´U·]±]�ÓCˆä×*Ò*¨6¸%Ó@ˆà—j‘j uÐPZ�jÓ[Ð[Ð[r   c               ó$   € V ^8„  d   QhRRRRRR/# )r'   r>   r)   Úparamsr%   r*   r   )r+   s   "r   r,   r-   F   s&   € ÷ Tñ T˜|ð T°Vð TÀñ Tr   c                	óX   € \         P                  ! WP                  VP                  4      # ©N)r:   Údequantize_per_tensor_fp8r   r2   ©r<   r>   rB   s   &&&r   Ú
dequantizeÚTensorCoreFP8Layout.dequantizeE   s   € ä×+Ò+¨E·<±<À×ARÑARÓSÐSr   c               ó    € V ^8„  d   QhRRRR/# )r'   Úqtensorr   r*   z!tuple[torch.Tensor, torch.Tensor]r   )r+   s   "r   r,   r-   J   s   € ÷ 5ñ 5¨ð 5Ð<]ñ 5r   c                	óF   € VP                   VP                  P                  3# rD   )Ú_qdataÚ_paramsr   )r<   rJ   s   &&r   Úget_plain_tensorsÚ%TensorCoreFP8Layout.get_plain_tensorsI   s   € à�~‰~˜wŸ™×4Ñ4Ð4Ð4r   c               ó$   € V ^8„  d   QhRRRRRR/# )r'   r>   r)   rB   r%   r*   zdict[str, torch.Tensor]r   )r+   s   "r   r,   r-   N   s"   € ÷ 
ñ 
 |ð 
¸Vð 
ÐH_ñ 
r   c                ó"   € RVRVP                   /# )u7   Return key suffix â†’ tensor mapping for serialization.Ú Ú_scale)r   rF   s   &&&r   Ústate_dict_tensorsÚ&TensorCoreFP8Layout.state_dict_tensorsM   s   € ð �Ø�f—l‘lð
ð 	
r   r   )é   é	   )r    r!   r"   r#   Ú__doc__ÚMIN_SM_VERSIONr   r   r%   Úclassmethodr   Úfloat8_e4m3fnr?   rG   rN   rT   r$   r   r   r   r   r      s’   † ñð €Ná�dÔôiÐ!ó ió ðið
 ð 48Ø"×0Ñ0ö	\ó ð\ð* ôTó ðTð ô5ó ð5ð ô
ó ô
r   r   c               ó4   € V ^8„  d   QhRRRRRRRRRRRR	R
R/# )r'   Úinput_qdatar)   Úweight_qdataÚscale_aÚscale_bÚbiasztorch.Tensor | NoneÚ	out_dtypeztorch.dtype | Noner*   r   )r+   s   "r   r,   r,   X   sN   € ÷ ñ Øðàðð ðð ð	ð
 ðð "ðð ñr   c           	     ó@   € \        V P                  4       VVVVVR 7      # ))r_   r`   ra   rb   )r   Ú
contiguous)r]   r^   r_   r`   ra   rb   s   &&&&&&r   Ú_fp8_scaled_mmre   X   s,   € ô Ø×ÑÓ ØØØØØôð r   c                ó   a € V 3R lpV# )zâFactory for shape-changing operations (view, reshape, t, etc.).

These ops work directly on FP8 since it's not packed (1:1 element mapping).
The aten_op is applied to _qdata and the result is wrapped in a new QuantizedTensor.
c                óL  <€ V^ ,          p\        V\        4      '       g	   S! V/ VB # S! VP                  .VR,          O5/ VB p\        P	                  VP
                  P                  VP
                  P                  \        VP                  4      R7      p\        VRV4      # )é    :é   NNr1   r   )
r   r   rL   r   r%   rM   r   r2   r4   r5   )ÚqtÚargsr=   Úinput_tensorÚ	new_qdataÚ
new_paramsÚaten_ops   &&&   €r   ÚhandlerÚ(_make_fp8_shape_handler.<locals>.handlerp   sœ   ø€ Ø˜A•wˆÜ˜,¬×8Ò8Ù˜DÐ+ FÑ+Ð+ñ ˜L×/Ñ/ÐE°$°rµ(ÒE¸fÑEˆ	ô )×/Ñ/Ø×&Ñ&×,Ñ,Ø#×+Ñ+×6Ñ6Ü˜YŸ_™_Ó-ð 0ó 
ˆ
ô
 ˜yÐ*?ÀÓLÐLr   r   )ro   rp   s   f r   Ú_make_fp8_shape_handlerrr   j   s   ø€ õMð  €Nr   c                óÖ  € V^ ,          V^,          rC\        V4      ^8”  d
   V^,          MRp\        V\        4      '       d   \        V\        4      '       g4   \        P                  P
                  P                  ! \        W4V34      !  # \        P                  V4      w  rg\        P                  V4      w  r‰VP                  RVP                  P                  4      p
VP                  4       p \        WkWyWZ4      pVP                  \        P                   \        P"                  39   dS   \        P%                  Wy,          VP                  P                  \'        VP(                  4      R7      p\        VRV4      # V#   \*        \,        3 dW   p\.        P1                  RT R24       \        P                  P
                  P                  ! \        Y4T34      !  u Rp?# Rp?ii ; i)z�FP8 linear: output = input @ weight.T + bias

Uses torch._scaled_mm for hardware-accelerated FP8 matmul when both
input and weight are FP8 QuantizedTensors.
Nrb   r1   r   zFP8 _scaled_mm failed: z , falling back to dequantization)Úlenr   r   r   ÚnnÚ
functionalÚlinearr   r   rN   ÚgetrM   r2   Útre   r   r[   Úfloat8_e5m2r%   r4   r5   ÚRuntimeErrorÚ	TypeErrorÚloggerÚwarning)rj   rk   r=   rl   Úweightra   r]   r_   r^   r`   rb   Úweight_tÚoutputÚoutput_paramsÚes   &&&            r   Ú_handle_fp8_linearr„   …   s�  € ð   �7 D¨¥G�&Ü˜$“i !”mˆ4�Ž7¨€Dô �|¤_×5Ò5¼*ÀVÌ_×:]Ò:]Ü�x‰x×"Ñ"×)Ò)¬?¸LÐRVÐ;WÓ+XÑYÐYä.×@Ñ@ÀÓNÑ€KÜ/×AÑAÀ&ÓIÑ€LØ—
‘
˜;¨×(<Ñ(<×(GÑ(GÓH€Ið �~‰~Ó€HðZÜ °wÈÓYˆð �<‰<œE×/Ñ/´×1BÑ1BÐCÔCÜ/×6Ñ6ØÕ'Ø'×/Ñ/×:Ñ:Ü  §¡Ó.ð 7ó ˆMô
 # 6Ð+@À-ÓPÐPØˆøäœ)Ð$ô ZÜ�‰Ð0°°Ð3SÐTÔUÜ�x‰x×"Ñ"×)Ò)¬?¸LÐRVÐ;WÓ+XÑYÕYûðZús&   Ã1BF Å?F ÆG(ÆAG#ÇG(Ç#G(c                óä  € V^ ,          V^,          rC\        V\        4      '       d   \        V\        4      '       g   \        P                  ! \	        V4      !  # \
        P                  V4      w  rV\
        P                  V4      w  rxVP                  RVP                  P                  4      p	 \        WWWhV	R7      #   \        \        3 d!    \        P                  ! \	        T4      !  u # i ; i)z)FP8 matrix multiplication: output = a @ brb   )rb   )r   r   r   Úmmr   r   rN   rx   rM   r2   re   r{   r|   )
rj   rk   r=   ÚaÚbÚa_qdatar_   Úb_qdatar`   rb   s
   &&&       r   Ú_handle_fp8_mmr‹   ¬   s¼   € ð ��7�D˜•G€qä�qœ/×*Ò*¬z¸!¼_×/MÒ/MÜ�xŠxœ¨Ó.Ñ/Ð/ä*×<Ñ<¸QÓ?Ñ€GÜ*×<Ñ<¸QÓ?Ñ€GØ—
‘
˜;¨¯	©	×(<Ñ(<Ó=€Ið0Ü˜g°ÈIÔVÐVøÜœ)Ð$ô 0Ü�xŠxœ¨Ó.Ñ/Ò/ð0ús   Â0B> Â>.C/Ã.C/c                óô  € V^ ,          V^,          V^,          rTp\        V\        4      '       d   \        V\        4      '       g   \        P                  ! \	        V4      !  # \
        P                  V4      w  rg\
        P                  V4      w  r‰VP                  RVP                  P                  4      p
 \        WhWyW:4      #   \        \        3 d!    \        P                  ! \	        T4      !  u # i ; i)z)FP8 addmm: output = bias + input @ weightrb   )r   r   r   Úaddmmr   r   rN   rx   rM   r2   re   r{   r|   )rj   rk   r=   ra   rl   r   r]   r_   r^   r`   rb   s   &&&        r   Ú_handle_fp8_addmmrŽ   ¾   sÆ   € ð "& a¥¨$¨q­'°4¸µ7˜€Dä�|¤_×5Ò5¼*ÀVÌ_×:]Ò:]Ü�{Š{œO¨DÓ1Ñ2Ð2ä.×@Ñ@ÀÓNÑ€KÜ/×AÑAÀ&ÓIÑ€LØ—
‘
˜;¨×(<Ñ(<×(GÑ(GÓH€Ið3Ü˜k¸È4Ó[Ð[øÜœ)Ð$ô 3Ü�{Š{œO¨DÓ1Ñ2Ò2ð3ús   Â9C Ã.C7Ã6C7)NN)$rX   Ú
__future__r   ÚloggingÚdataclassesr   r   Úcomfy_kitchenr:   Úcomfy_kitchen.scaled_mm_v2r   Úbaser   r   r   r   r	   Ú	getLoggerr    r}   r   re   rr   ÚopsÚatenrw   Údefaultr„   r†   r‹   r�   rŽ   ÚviewÚreshapery   Ú_aten_opr   r   r   Ú<module>rœ      sN  ðÙ 6Ý "ã Ý !ã ã Ý 3÷õ ð 
×	Ò	˜8Ó	$€ô<
˜/ô <
÷Bò$ñ6 �E—I‘I—N‘N×)Ñ)×1Ñ1Ð3FÓGñ#Zó Hð#ZñL �E—I‘I—N‘N×%Ñ%×-Ñ-Ð/BÓCñ0ó Dð0ñ" �E—I‘I—N‘N×(Ñ(×0Ñ0Ð2EÓFñ3ó Gð3ð* 
‡I�I‡N�N×Ñ×ÑØ	‡I�I‡N�N×Ñ×"Ñ"Ø	‡I�I‡N�N×Ñ×Ñó€Hñ
 �xÐ!4Ô5Ñ6MÈhÓ6WÖXór   