+
    ƒ°ñjÓ  ã                  óT  € R t ^ RIHt ^ RIt^ RIHt ^ RIHu Ht ^ RI	H
t
 ^RIHtHt  ! R R]P                  4      t ! R R]P                  4      t ! R	 R
]P                  4      t ! R R]P                  4      t ! R R]P                  4      t ! R R]P                  4      tR# )z6Camera-token encoder and decoder for Depth Anything 3.)ÚannotationsN)Úoptimized_attention_for_device)Úaffine_inverseÚextri_intri_to_pose_encodingc                  óF   a € ] tR t^tRtR	RRRRRR/V 3R llltR tRtV ;t# )
Ú_MlpzIStandard 2-layer MLP with GELU. Matches upstream ``utils.attention.Mlp``.NÚdeviceÚdtypeÚ
operationsc               	ó¸   <€ \         SV `  4        T;'       g    TpT;'       g    TpVP                  WR WER7      V n        VP                  W#R WER7      V n        R# )T©Úbiasr   r	   N)ÚsuperÚ__init__ÚLinearÚfc1Úfc2)ÚselfÚin_featuresÚhidden_featuresÚout_featuresr   r	   r
   Ú	__class__s   &&&&$$$€ÚEC:\Ai_Coding\apps\studio\ComfyUI\comfy\ldm\depth_anything_3\camera.pyr   Ú_Mlp.__init__   s\   ø€ Ü‰ÑÔØ#×2Ð2 {ˆØ)×8Ð8¨[ˆØ×$Ñ$ [ÈÐU[Ð$ÓiˆŒØ×$Ñ$ _ÈÐV\Ð$ÓjˆŽó    c                	ój   € V P                  \        P                  ! V P                  V4      4      4      # ©N)r   ÚFÚgelur   ©r   Úxs   &&r   ÚforwardÚ_Mlp.forward   s"   € Ø�x‰xœŸš˜tŸx™x¨›{Ó+Ó,Ð,r   )r   r   )NN©	Ú__name__Ú
__module__Ú__qualname__Ú__firstlineno__Ú__doc__r   r!   Ú__static_attributes__Ú__classcell__©r   s   @r   r   r      s7   ø† ÙSñkÐW[ð kÐcgð kÐtx÷ kð k÷-ð -r   r   c                  ó>   a € ] tR t^ tRtRRRR/V 3R lltR tRtV ;t# )Ú_LayerScalez;Per-channel learnable scaling. Matches upstream LayerScale.r   Nr	   c               	ó†   <€ \         SV `  4        \        P                  ! \        P
                  ! WVR 7      4      V n        R# ©©r   r	   N)r   r   ÚnnÚ	ParameterÚtorchÚemptyÚgamma)r   Údimr   r	   r   s   &&$$€r   r   Ú_LayerScale.__init__#   s)   ø€ Ü‰ÑÔÜ—\’\¤%§+¢+¨cÈÔ"NÓOˆŽ
r   c                	óp   € WP                   P                  VP                  VP                  R 7      ,          # ))r	   r   )r5   Útor	   r   r   s   &&r   r!   Ú_LayerScale.forward'   s$   € Ø—:‘:—=‘= q§w¡w°q·x±x�=Ó@Õ@Ð@r   )r5   r#   r+   s   @r   r-   r-       s(   ø† ÙEðP dð P°$÷ P÷Að Ar   r-   c                  óB   a € ] tR t^+tRtRRRRRR/V 3R lltR tRtV ;t# )	Ú
_Attentionz­Self-attention with fused QKV projection. Mirrors upstream utils.attention.Attention;
Layout matches the HF safetensors (attn.qkv.{weight,bias} and attn.proj.{weight,bias}).r   Nr	   r
   c               	óÚ   <€ \         SV `  4        W,          ^ 8X  g   Q hW n        W,          V n        VP	                  W^,          RW4R7      V n        VP	                  WRW4R7      V n        R# )é    Tr   N)r   r   Ú	num_headsÚhead_dimr   ÚqkvÚproj)r   r6   r?   r   r	   r
   r   s   &&&$$$€r   r   Ú_Attention.__init__/   sc   ø€ Ü‰ÑÔØ� !Ô#Ð#Ð#Ø"ŒØÕ(ˆŒØ×$Ñ$ S°­'¸ÀVÐ$ÓYˆŒØ×%Ñ% c°TÀ&Ð%ÓVˆŽ	r   c                	ó  € VP                   w  r#pV P                  V4      P                  W#^V4      pVP                  ^4      w  rgp\	        VP
                  RR7      p	V	! WgW€P                  R7      p
V P                  V
4      # )é   T)Úsmall_input)Úheads)ÚshaperA   ÚreshapeÚunbindr   r   r?   rB   )r   r    ÚBÚNÚCrA   ÚqÚkÚvÚattn_fnÚouts   &&         r   r!   Ú_Attention.forward7   sj   € Ø—'‘'‰ˆˆaØ�h‰h�q‹k×!Ñ! !¨¨1Ó-ˆØ—*‘*˜Q“-‰ˆˆaÜ0°·±ÀtÔLˆÙ�a˜A§^¡^Ô4ˆØ�y‰y˜‹~Ðr   )r@   r?   rB   rA   r#   r+   s   @r   r<   r<   +   s1   ø† ñ_ðW°ð W¸Tð WÈd÷ W÷ð r   r<   c                  óF   a € ] tR t^@tRtR	RRRRRR/V 3R llltR tRtV ;t# )
Ú_BlockzpPre-norm transformer block with LayerScale. Used by :class:CameraEnc. Layout follows upstream utils.block.Block.r   Nr	   r
   c               	óº  <€ \         SV `  4        VP                  WVR 7      V n        \	        WWVVR7      V n        V'       d   \        WVR 7      M\        P                  ! 4       V n	        VP                  WVR 7      V n
        \        V\        W,          4      WVVR7      V n        V'       d   \        WVR 7      V n        R# \        P                  ! 4       V n        R# )r0   )r   r	   r
   )r   r   r   r	   r
   N)r   r   Ú	LayerNormÚnorm1r<   Úattnr-   r1   ÚIdentityÚls1Únorm2r   ÚintÚmlpÚls2)	r   r6   r?   Ú	mlp_ratioÚinit_valuesr   r	   r
   r   s	   &&&&&$$$€r   r   Ú_Block.__init__C   s«   ø€ Ü‰ÑÔØ×)Ñ)¨#ÀEÐ)ÓJˆŒ
Ü˜s°fÐV`ÔaˆŒ	ßCN”;˜s¸Õ?ÔTV×T_ÒT_ÓTaˆŒØ×)Ñ)¨#ÀEÐ)ÓJˆŒ
Ü C¼¸S½_Ó9MÐV\ð  wAô  BˆŒßCN”;˜s¸Ô?ˆŽÔTV×T_ÒT_ÓTaˆŽr   c                	óÚ   € WP                  V P                  V P                  V4      4      4      ,           pWP                  V P	                  V P                  V4      4      4      ,           pV# r   )r[   rY   rX   r_   r^   r\   r   s   &&r   r!   Ú_Block.forwardL   sK   € Ø—‘˜Ÿ™ 4§:¡:¨a£=Ó1Ó2Õ2ˆØ—‘˜Ÿ™ $§*¡*¨Q£-Ó0Ó1Õ1ˆØˆr   )rY   r[   r_   r^   rX   r\   )é   ç{®Gáz„?r#   r+   s   @r   rU   rU   @   s7   ø† ÙzñbÐPTð bÐ\`ð bÐmq÷ bð b÷ð r   rU   c                  óR   a € ] tR t^RtRtRRRRRRR/R V 3R lllltR R	 ltR
tV ;t# )Ú	CameraEnca•  Encode per-view (extrinsics, intrinsics) into a camera token.

Maps a 9-D pose-encoding vector through a small MLP up to the backbone's
``embed_dim``, then runs ``trunk_depth`` transformer blocks. The output
has shape ``(B, S, embed_dim)`` and is injected at block ``alt_start``
of the DINOv2 backbone in place of the cls token.

Parameters mirror the upstream ``cam_enc.py`` so HF weights load directly.
r   Nr	   r
   c               ó4   € V ^8„  d   QhRRRRRRRRRRRRRR	/# )
é   Údim_outr]   Údim_inÚtrunk_depthÚ
target_dimr?   r`   ra   Úfloat© )Úformats   "r   Ú__annotate__ÚCameraEnc.__annotate__]   sP   € ÷ 
ñ 
àð
ð ð
ð ð	
ð
 ð
ð ð
ð ð
ð ñ
r   c               	ó^  <€ \         SV `  4        W@n        W0n        \        P
                  ! \        V4       Uu. uF  p\        WVVW‰V
R 7      NK  	  up!  V n        V
P                  WV	R7      V n
        V
P                  WV	R7      V n        \        VV^,          VW‰V
R7      V n        R# u upi ))r?   r`   ra   r   r	   r
   r0   )r   r   r   r   r	   r
   N)r   r   rn   rm   r1   Ú
SequentialÚrangerU   ÚtrunkrW   Ú
token_normÚ
trunk_normr   Úpose_branch)r   rk   rl   rm   rn   r?   r`   ra   r   r	   r
   Ú_kwargsÚ_r   s   &&&&&&&&$$$, €r   r   ÚCameraEnc.__init__]   s·   ø€ ô 	‰ÑÔØ$ŒØ&ÔÜ—]’]ô ˜;Ô'ó	%
ñ (�ô �7¸9Ø*Ø ¸*÷Fñ (ñ	%
ñ ˆŒ
ð %×.Ñ.¨wÈUÐ.ÓSˆŒØ$×.Ñ.¨wÈUÐ.ÓSˆŒÜØØ# q�LØ Ø°:ô	
ˆÖùò%
s   ¹B*c               ó$   € V ^8„  d   QhRRRRRR/# )rj   Ú
extrinsicsútorch.TensorÚ
intrinsicsÚreturnrp   )rq   s   "r   rr   rs   |   s"   € ÷ 	ñ 	 ,ð 	¸Lð 	Ø".ñ	r   c                ó*  € \        V4      p\        WBV4      pV P                  VP                  V P                  P                  P
                  P                  4      4      pV P                  V4      pV P                  V4      pV P                  V4      pV# )z9Encode camera parameters into ``(B, S, dim_out)`` tokens.)
r   r   rz   r9   r   Úweightr	   rx   rw   ry   )r   r   r�   Úimage_size_hwÚc2wsÚpose_encodingÚtokenss   &&&&   r   r!   ÚCameraEnc.forward|   s|   € ô ˜jÓ)ˆÜ4°TÀ}ÓUˆØ×!Ñ! -×"2Ñ"2°4×3CÑ3C×3GÑ3G×3NÑ3N×3TÑ3TÓ"UÓVˆØ—‘ Ó(ˆØ—‘˜FÓ#ˆØ—‘ Ó(ˆØˆr   )rz   rn   rx   rw   rm   ry   )i   é	   re   rŠ   é   re   rf   r#   r+   s   @r   rh   rh   R   s:   ø† ññ
ð ð
ð  ð
ð -1÷
ó 
÷>	ó 	r   rh   c                  óV   a € ] tR t^ˆtRtRRRRRRR/R V 3R lllltRR R	 lltR
tV ;t# )Ú	CameraDecaš  Decode the final cam token into a 9-D pose encoding.

Output layout: ``[T(3), quat_xyzw(4), fov_h, fov_w]``. The translation is
always predicted by the network; the quaternion and FoV can either be
predicted or supplied via ``camera_encoding`` (used at training time
when GT cameras are available -- not exercised at inference here).

Parameters mirror the upstream ``cam_dec.py`` so HF weights load directly.
r   Nr	   r
   c               ó   € V ^8„  d   QhRR/# )rj   rl   r]   rp   )rq   s   "r   rr   ÚCameraDec.__annotate__“   s   € ÷ 
ñ 
˜sñ 
r   c               	óÚ  <€ \         SV `  4        Tp\        P                  ! VP	                  WfW#R 7      \        P
                  ! 4       VP	                  WfW#R 7      \        P
                  ! 4       4      V n        VP	                  V^W#R 7      V n        VP	                  V^W#R 7      V n        \        P                  ! VP	                  V^W#R 7      \        P
                  ! 4       4      V n	        R# r/   )
r   r   r1   ru   r   ÚReLUÚbackboneÚfc_tÚfc_qvecÚfc_fov)r   rl   r   r	   r
   r{   Údr   s   &&$$$, €r   r   ÚCameraDec.__init__“   s½   ø€ ä‰ÑÔØˆÜŸšØ×Ñ˜a¨6ÐÓ?Ü�GŠG‹IØ×Ñ˜a¨6ÐÓ?Ü�GŠG‹Ió	
ˆŒð ×%Ñ% a¨°6Ð%ÓGˆŒ	Ø!×(Ñ(¨¨A°fÐ(ÓJˆŒÜ—m’mØ×Ñ˜a ¨6ÐÓ?Ü�GŠG‹Ió
ˆŽr   c               ó$   € V ^8„  d   QhRRRRRR/# )rj   Úfeatr€   Úcamera_encodingz'torch.Tensor | None'r‚   rp   )rq   s   "r   rr   r�   ¤   s#   € ÷ =ñ =˜Lð =Ø!6ð=ØBNñ=r   c                ó  € VP                   R,          w  r4VP                  W4,          R4      pV P                  V4      pV P                  VP	                  4       4      P                  W4^4      pVf`   V P                  VP	                  4       4      P                  W4^4      pV P                  VP	                  4       4      P                  W4^4      pMVR,          pVRRR13,          p\        P                  ! WVV.RR7      # )zADecode ``(B, N, dim_in)`` cam tokens into ``(B, N, 9)`` pose enc.:Nrj   NN.)r6   éÿÿÿÿ).:rE   é   Néþÿÿÿ)	rH   rI   r’   r“   ro   r”   r•   r3   Úcat)r   r™   rš   rK   rL   Úout_tÚout_qvecÚout_fovs   &&&     r   r!   ÚCameraDec.forward¤   sÌ   € ð �z‰z˜"�~‰ˆØ�|‰|˜A�E 2Ó&ˆØ�}‰}˜TÓ"ˆØ—	‘	˜$Ÿ*™*›,Ó'×/Ñ/°°aÓ8ˆØÒ"Ø—|‘| D§J¡J£LÓ1×9Ñ9¸!ÀÓBˆHØ—k‘k $§*¡*£,Ó/×7Ñ7¸¸aÓ@‰Gà& xÕ0ˆHØ% c¨2©3 hÕ/ˆGÜ�yŠy˜%¨7Ð3¸Ô<Ð<r   )r’   r•   r”   r“   )i   r   r#   r+   s   @r   r�   r�   ˆ   s4   ø† ññ
Øð
Ø'+ð
Ø8<÷
ó 
÷"=õ =r   r�   )r(   Ú
__future__r   r3   Útorch.nnr1   Útorch.nn.functionalÚ
functionalr   Úcomfy.ldm.modules.attentionr   Ú	transformr   r   ÚModuler   r-   r<   rU   rh   r�   rp   r   r   Ú<module>r«      s€   ðÙ <å "ã Ý ß Ð å Fß Cô-ˆ2�9‰9ô -ôA�"—)‘)ô Aô�—‘ô ô*ˆR�Y‰Yô ô$3�—	‘	ô 3ôl)=�—	‘	ö )=r   