+
    „°ñjš[  ã                  ó4  € R t ^ RIHt ^ RIt^ RIt^ RIt^ RIt^ RIt^ RI	t
^ RIt
^ RIt
^ RIHt ^ RIHtHt ^€t^@tRtRt^tR t ! R R4      tR	 tR
 R ltR R ltR R ltR R ltR R ltR t ! R R]P@                  4      t! ! R R]4      t"R R lt#R# )zøBlock-sparse attention on comfy_kitchen's sparse attention kernels (Sol-Attn adaptive
threshold, SLA-style top-k, or FastVideo's VSA). Generic models go through the
attention override; MiniMax-H3 gets the chunked qkv producer through block patches.)ÚannotationsN)ÚMiniMaxH3Model)ÚComfyExtensionÚioi   c           
     óL  € \        4       p\        P                  ! RT ;'       g    R4       Ft  pRV9   dQ   R VP                  R4       4       w  r4VP	                  \        \        W44      \        W44      ^,           4      4       KZ  VP                  \        V4      4       Kv  	  V# )z$'0, 1, 47-49' -> {0, 1, 47, 48, 49}.z\d+\s*-\s*\d+|\d+Ú Ú-c              3  ó8   "  € T F  p\        V4      x € K  	  R # 5i©N)Úint)Ú.0Úxs   & ÚGC:\Ai_Coding\apps\studio\ComfyUI\comfy_extras\nodes_sparse_attention.pyÚ	<genexpr>Ú#parse_block_list.<locals>.<genexpr>   s   é € Ð4¡O˜q”C˜—F�F£Oùs   ‚)
ÚsetÚreÚfindallÚsplitÚupdateÚrangeÚminÚmaxÚaddr   )ÚtextÚblocksÚpartÚaÚbs   &    r   Úparse_block_listr      sv   € ä‹U€FÜ—
’
Ð/°·°¸Ö<ˆØ�$Œ;Ù4 D§J¡J¨s¤OÓ4‰DˆAØ�M‰Mœ%¤ A£	¬3¨q«9°q­=Ó9Ö:à�J‰J”s˜4“yÖ!ñ =ð €Mó    c                  óD   € ] tR t^&tRtR tR tR tR tR t	R t
R tR	tR
# )ÚSparseAttnPatchzuOptions plus runtime state for one patched model; an ON_CLEANUP callback
resets the state when the sampling run ends.c                	ó¼   € Wn         W n        W�n        W0n        W@n        WPn        W`n        Wpn        W€n        W n	        \        4       V n        V P                  4        R # r
   )ÚtauÚ
topk_ratioÚextra_tokensÚvsaÚsigma_startÚ	sigma_endÚ
min_tokensÚdense_blocksÚsink_conditioningÚverboser   Ú	installedÚreset)Úselfr$   r%   r'   r(   r)   r*   r+   r,   r&   r-   s   &&&&&&&&&&&r   Ú__init__ÚSparseAttnPatch.__init__*   sJ   € àŒØ$ŒØ(ÔØŒØ&ÔØ"ŒØ$ŒØ(ÔØ!2ÔØŒÜ›ˆŒØ�
‰
Žr    c                	óN   € / V n         / V n        R V n        \        4       V n        R # r
   )ÚpooledÚ	vsa_plansÚvsa_roper   Ú_logged©r0   s   &r   r/   ÚSparseAttnPatch.reset9   s    € ØˆŒØˆŒØˆŒÜ“uˆŽr    c                	óº   € V P                   '       dI   WP                  9  d7   V P                  P                  V4       \        P                  ! R V 24       R# R# R# )zBlockSparseAttention: N)r-   r7   r   ÚloggingÚinfo)r0   ÚkeyÚmessages   &&&r   Úlog_onceÚSparseAttnPatch.log_once?   sA   € Ø�<�<ˆ<˜C§|¡|Ô3Ø�L‰L×Ñ˜SÔ!Ü�LŠLÐ1°'°Ð;Ö<ñ 4‰<r    c                ód  € VP                  R4      pVe:   \        V^ ,          4      pWPP                  8”  g   WPP                  8  d   RVR R2# W P                  8  d   V RV P                   2# V P
                  '       d/   Vf   V P                  RR4       R# W0P
                  9   d   R	V R
2# R# )z=Why this call stays dense regardless of its tensors, or None.ÚsigmasNzsigma z.3gz outside the start/end windowz tokens < min_tokens Úno_block_indexz>this model does not report block indices; dense_blocks ignoredzblock z in dense_blocks)ÚgetÚfloatr(   r)   r*   r+   r?   )r0   Útransformer_optionsÚtokensÚblock_indexrB   Úsigmas   &&&&  r   Údense_reasonÚSparseAttnPatch.dense_reasonD   s¶   € à$×(Ñ(¨Ó2ˆØÒÜ˜& �)Ó$ˆEØ×'Ñ'Ô'¨5·>±>Ô+AØ  c˜{Ð*GÐHÐHØ—O‘OÔ#Ø�XÐ2°4·?±?Ð2CÐDÐDØ××ÐØÒ"Ø—‘Ð.Ð0pÔqñ ð × 1Ñ 1Ô1Ø ˜}Ð,<Ð=Ð=Ùr    c                óÔ  € VP                  R4      pV P                  R8X  g   Ve   VP                  V8w  d   R# \        R VP                   4       R4      pVe   V^ ,          ^ 8:  d   R# ^ V^ ,          \
        ,           ^,
          \
        ,          3pV P                  R8w  d   VR3# \        R VP                   4       R4      pVf   WU3# WV^ ,          \
        ,          V^,          33# )	z«MiniMax-H3 conditioning rows as (exact-KV blocks, dense-query blocks):
the packed prefix stays exact for every query, and optionally the
target-audio query rows run dense.Úminimax_h3_layoutÚoffNc              3  ó>   "  € T F  w  rq3R 8X  g   K  W3x € K  	  R# 5i)ÚvideoN© ©r   r   r   Úkinds   &   r   r   Ú(SparseAttnPatch.sinks.<locals>.<genexpr>[   ó   é € ÐR©¡  tÈ'Á/”f�q•f«ùó   ‚’Úexact_kv_and_rowsc              3  ó>   "  € T F  w  rq3R 8X  g   K  W3x € K  	  R# 5i)ÚaudioNrQ   rR   s   &   r   r   rT   a   rU   rV   ©é    r[   )rZ   rZ   )rD   r,   Úseq_lenÚnextÚsegmentsÚ
BLOCK_SIZE)r0   rF   rG   ÚlayoutrP   r   rY   s   &&&    r   ÚsinksÚSparseAttnPatch.sinksT   sÚ   € ð %×(Ñ(Ð)<Ó=ˆØ×!Ñ! UÔ*¨fªnÀÇÁÐRXÔ@XØ!Ð!ÜÑR¨v¯ªÓRÐTXÓYˆØŠ=˜E !�H¨œMØ!Ð!Ø�e˜A•h¤Õ+¨aÕ/´JÕ>Ð?ˆØ×!Ñ!Ð%8Ô8Ø˜6�>Ð!ÜÑR¨v¯ªÓRÐTXÓYˆØŠ=Ø�>Ð!Ø˜a�¤JÕ.°°qµ	Ð:Ð:Ð:r    c                ó.  € \        VP                  4      \        VP                  4      \        V4      3pV P                  P                  V4      pVe   V# VP                  w  rVrxp	\        V4      \        V4      ^,          \        V4      ^,          3p
. ^ rËVP                   EF9  w  rÞpWí,
          pVR8w  dŸ   V\        ,           ^,
          \        ,          p\        P                  ! V\        ,          3R\        P                  VR7      p\        P                  ! WÞVR7      VRV% VP                  VP                  V\        4      4       VV,          pK´  V
^ ,          V
^,          ,          V
^,          ,          V8w  d   \        RV RV
 24      h\        w  pppR \!        V
\        4       4       w  ppp\        P                  ! VVV3R\        P                  VR7      p\        P                  ! WÞVR7      P                  ! V
!  VRV
^ ,          1RV
^,          1RV
^,          13&   VP                  VV,          VVV,          VVV,          V4      P#                  ^ ^^^^^4      P%                  R\        4      p\        P&                  ! V^ 8  P)                  \        P*                  4      ^RR	7      pVP                  \        P,                  ! V^V4      4       EK<  	  \        P.                  ! V4      pVP%                  R4      pV^ 8¬  p\        P0                  ! VP2                  \        P                  VR7      p\        P4                  ! V4      P7                  4       VVV,          &   R
\        VP9                  4       4      RVRVRVRV^ 8¬  P;                  ^4      P)                  \        P<                  4      /p\?        V P                  4      \@        8¼  d,   V P                  \C        \E        V P                  4      4       KI  W@P                  V&   V# )z¢Padded tile order: prefix segments in their own zero-padded 64-row
tiles, video in 4x4x4 cubes. `src` maps padded row -> source row (-1 =
pad), `inv` the reverse.NrP   ©ÚdtypeÚdevice©rf   zVSA: video segment of z% rows does not match the latent grid c              3  ó`   "  € T F$  w  rW,           ^,
          V,          V,          x € K&  	  R# 5i©é   NrQ   )r   ÚgÚcs   &  r   r   Ú+SparseAttnPatch.vsa_plan.<locals>.<genexpr>}   s'   é € ÐOÑ;N±4°1˜1�5 1�9¨Õ*¨Q×.Ò.Ó;Nùs   ‚,.T)ÚdimÚstableÚnÚn_prefixÚsrcÚinvÚ	block_lenéÿÿÿÿ)#ÚtupleÚ	signaturer^   Ústrr5   rD   r   r_   ÚtorchÚfullÚint64ÚarangeÚappendÚviewÚRuntimeErrorÚVSA_CUBEÚzipÚpermuteÚreshapeÚargsortÚtoÚint8ÚgatherÚcatÚemptyr\   ÚnonzeroÚflattenÚnumelÚsumÚint32ÚlenÚVSA_PLAN_CACHEr]   Úiter)r0   r`   rf   r=   ÚplanÚ	_text_lenÚlatent_tÚlatent_hÚlatent_wÚ_audio_tÚgridÚtilesrq   r   r   rS   rp   ÚmÚsegÚctÚchÚcwÚptÚphÚpwÚpaddedÚcubesÚorderrr   Úlivers   s   &&&                            r   Úvsa_planÚSparseAttnPatch.vsa_planf   s.  € ô �V×%Ñ%Ó&¬¨f¯o©oÓ(>ÄÀFÃÐLˆØ�~‰~×!Ñ! #Ó&ˆØÒØˆKØ<B×<LÑ<LÑ9ˆ	˜X°Ü�H“œs 8›}°Õ1´3°x³=ÀAÕ3EÐFˆØ˜aˆxØ Ÿ/�/‰JˆA�$Ø•ˆAØ�wŒØœ•^ aÕ'¬JÕ6�Ü—j’j !¤j¥.Ð!2°B¼e¿k¹kÐRXÔY�ÜŸ,š, q°FÔ;��B�Q�Ø—‘˜SŸX™X a¬Ó4Ô5Ø˜A•�ÙØ�A�w˜˜a�Õ  4¨¥7Õ*¨aÔ/Ü"Ð%;¸A¸3Ð>cÐdhÐciÐ#jÓkÐkÜ!‰JˆB��BÙO¼3¸tÄXÔ;NÓO‰JˆB��BÜ—Z’Z  R¨ ¨b¼¿¹ÈFÔSˆFÜ38·<²<ÀÈVÔ3T×3YÒ3YÐ[_Ñ3`ˆF�8�D˜•G�8˜X˜d 1�g˜X x¨¨Q­ xÐ/Ñ0Ø—[‘[  r¥¨2¨r°R­x¸¸RÀ2½XÀrÓJß‘g˜a  A q¨!¨QÓ/·±¸¼JÓ0Gð ä—M’M 5¨1¡9§.¡.´·±Ó"<À!ÈDÔQˆEØ�L‰LœŸš e¨Q°Ó6×7ñ% *ô& —	’	˜%Ó ˆØ�m‰m˜BÓˆØ�a‰xˆÜ�kŠk˜&Ÿ.™.´·±ÀFÔKˆÜŸš tÓ,×4Ñ4Ó6ˆˆC��I‰Ø”S˜Ÿ™›Ó% z°8¸UÀCÈÐPSØ˜e q™j×-Ñ-¨aÓ0×3Ñ3´E·K±KÓ@ðBˆä�$—.‘.Ó!¤^Ô3Ø—‘œt¤D¨¯©Ó$8Ó9Ò:Ø"�‰�sÑØˆr    c                	óP  € V P                   pVe)   V^ ,          ! 4       VJ d   V^,          VJ d
   V^,          # VP                  ^VR,          3\        VP                  R,          4      ,           4      pV^ ,          V^ VR,          3&   \        P
                  ! V4      W$3V n         V# )Nrp   :é   NNrs   )r6   Ú	new_zerosrv   ÚshapeÚweakrefÚref)r0   Ú
rope_freqsr’   Úhitr¢   s   &&&  r   Úvsa_rope_freqsÚSparseAttnPatch.vsa_rope_freqs�   s�   € Ø�m‰mˆØŠ?˜s 1žv›x¨:Ó5¸#¸a½&ÀD».Ø�q•6ˆMØ×%Ñ% q¨$¨s­) n´u¸Z×=MÑ=MÈbÕ=QÓ7RÕ&RÓSˆØ!+¨A¥ˆˆq�$�u•+ˆ~ÑÜ Ÿš ZÓ0°$Ð?ˆŒØˆr    )r7   r+   r&   r.   r*   r4   r)   r(   r,   r$   r%   r-   r'   r5   r6   N)Ú__name__Ú
__module__Ú__qualname__Ú__firstlineno__Ú__doc__r1   r/   r?   rJ   ra   r¦   r°   Ú__static_attributes__rQ   r    r   r"   r"   &   s+   † ñ4òòò=ò
ò ;ò$(öTr    r"   c                óˆ  € V P                   P                  R8w  d   R# \        P                  ! V P                   4      '       g   R# V P                  \
        P                  \
        P                  \
        P                  39  d   RV P                   R2# V\        8w  d   RV R\         2# V P                  VP                  8w  g   V P                  VP                  8w  d   R# VP                  V P                  8w  g   VP                  V P                  8w  d*   R	V P                   R
VP                   R
VP                   2# R# )zGWhy these tensors can't go through the kernel, or None. q/k/v are BTHD.Úcudaznot on CUDAú(no compiled sol_attn kernel for this GPUzdtype z (kernel takes bf16/fp16)z	head_dim z != z#cross-attention or GQA (kept dense)zmixed dtypes Ú/N)rf   ÚtypeÚckÚsol_attn_is_availablere   ry   Úbfloat16Úfloat16Úfloat32ÚHEAD_DIMr«   )ÚqÚkÚvÚdim_heads   &&&&r   Ú_ineligiblerÇ   š   sä   € à‡x�x‡}�}˜ÔÙÜ×#Ò# A§H¡H×-Ò-Ù9Ø‡w�w”u—~‘~¤u§}¡}´e·m±mÐDÔDØ˜Ÿ™�yÐ 9Ð:Ð:Ø”8ÔØ˜8˜* D¬¨
Ð3Ð3Ø‡w�w�!—'‘'Ô˜QŸW™W¨¯©Ô/Ù4Ø‡w�w�!—'‘'Ô˜QŸW™W¨¯©Ô/Ø˜qŸw™w˜i q¨¯©¨	°°1·7±7°)Ð<Ð<Ùr    c               ó   € V ^8„  d   QhRR/# ©r©   Úpatchr"   rQ   )Úformats   "r   Ú__annotate__rÌ   «   s   € ÷ +ñ + ?ñ +r    c                ó   a a€ RV V3R llpV# )zÍAttention override; declined calls run ``previous`` (the override that was
on the hook before this one) or ``func``. Dense-only in VSA mode: a
VSA-trained model must never see plain block-sparse attention.c	                óZ  <a aaaaaaaaa	aa€ S	P                  R 4      ;'       g    / p
VV VVV	VVVVVV3R lpSf   SP                  '       d   V! 4       # S'       d   SP                  ^,          MSP                  ^,          pSP                  W¬V
P                  R4      4      pVe"   SP	                  RWÍ3RV RV 24       V! 4       # S'       d#   SP                  w  o poR SSS3 4       w  pppM.SP                  w  opoSS,          oVVV3R lSSS3 4       w  ppp\        VVVS4      pVeI   SP	                  R\        VP                  4      V3R	\        VP                  4       R
V 24       V! 4       # SP                  W¬4      w  ppSP                  \        P                  8X  d   R VVV3 4       w  ppp\        P                  ! VVVSP                  S	P                  R4      \        V4      \        V4      SP                  SP                   R7	      P#                  SP                  4      pSP	                  R\        VP                  4      3R\        VP                  4       RV RV 24       S'       d   VP%                  ^^4      # VP'                  SRSS,          4      # )rF   c                 óf   <€ S	SSS3p \        RR SRSRSRS
/SB pSf	   S! V / VB # S! S.V O5/ VB # )ÚmaskÚattn_precisionÚskip_reshapeÚskip_output_reshaperQ   )Údict)ÚargsÚkwrÑ   ÚfuncÚheadsrÄ   ÚkwargsrÐ   ÚpreviousrÃ   rÓ   rÒ   rÅ   s     €€€€€€€€€€€r   ÚdenseÚ8make_attention_override.<locals>.override.<locals>.dense³   sp   ø€ Ø�q˜!˜UÐ#ˆDÜñ I˜4ð I°ð IÈ\ð IØ*=ðIØAGñIˆBà(0Ò(8‘4˜Ð$ Ñ$ÐY¹hÀtÐ>YÈdÒ>YÐVXÑ>YÐYr    rH   rÛ   údense (ú
 tokens): c              3  óD   "  € T F  qP                  ^^4      x € K  	  R# 5iri   )Ú	transpose©r   Úts   & r   r   Ú<make_attention_override.<locals>.override.<locals>.<genexpr>Â   s   é € Ð?±Y°Ÿ+™+ a¨×+Ð+³Yùs   ‚ c              3  óJ   <"  € T F  qP                  SRSS4      x € K  	  R# 5i)rj   Nru   )r~   )r   râ   r   rÆ   rØ   s   & €€€r   r   rã   Æ   s!   øé € ÐLÁ)¸QŸ&™&  B¨¨x×8Ð8Ã)ùs   ƒ #Ú
ineligiblezdense z: c              3  ó^   "  € T F#  qP                  \        P                  4      x € K%  	  R # 5ir
   )r…   ry   r¿   rá   s   & r   r   rã   Í   s   é € ÐE¹°1Ÿ$™$œuŸ~™~×.Ð.»ùs   ‚+-Úscale)r$   rç   Úsink_blocksÚsink_qr%   Ú	token_augÚsparsezsparse z, sinks r»   ru   )rD   r'   r«   rJ   r?   rÇ   rv   ra   re   ry   rÁ   r½   Úsol_attnr$   Úlistr%   r&   r…   rà   rƒ   )r×   rÃ   rÄ   rÅ   rØ   rÐ   rÑ   rÒ   rÓ   rÙ   rF   rÛ   rG   ÚreasonÚ_ÚqsÚksÚvsÚsinkré   Úoutr   rÆ   rÊ   rÚ   s   fffffffffl           @@€€r   ÚoverrideÚ)make_attention_override.<locals>.override¯   sN  ÿü€ à$Ÿj™jÐ)>Ó?×EÐEÀ2Ð÷	Z÷ 	Zð Ò˜uŸyŸy˜yÙ“7ˆNß+�—‘˜–°·±¸µˆØ×#Ñ#Ð$7ÐAT×AXÑAXÐYfÓAgÓhˆØÒØ�N‰N˜G VÐ4¸À¸xÀzÐRXÐQYÐ6ZÔ[Ù“7ˆNßØ !§¡ÑˆAˆq�!�XÙ?°a¸¸A±YÓ?‰JˆB�‘BàŸW™W‰NˆAˆq�(Ø˜ÕˆHÞLÀ1ÀaÈÁ)ÓL‰JˆB��BÜ˜R  R¨Ó2ˆØÒØ�N‰N˜L¬%°·±«/¸6ÐBÀfÌUÐSU×S[ÑS[Ë_ÐL]Ð]_Ð`fÐ_gÐDhÔiÙ“7ˆNØ—{‘{Ð#6Ó?‰ˆˆfØ�7‰7”e—m‘mÔ#ÙE¸¸RÀ¹ÓE‰JˆB��BÜ�kŠk˜"˜b "¨%¯)©)¸6¿:¹:ÀgÓ;NÜ&*¨4£j¼¸f»ÐRW×RbÑRbØ$)×$6Ñ$6ô8ç8:¹¸1¿7¹7»ð 	ð 	�‰˜¤%¨¯©£/Ð2°g¼eÀBÇHÁH»oÐ=NÈhÐW[ÐV\Ð\]Ð^dÐ]eÐ4fÔgßØ—=‘=  AÓ&Ð&Ø�{‰{˜1˜b %¨(Õ"2Ó3Ð3r    )NNFFrQ   )rÊ   rÚ   rõ   s   ff r   Úmake_attention_overrider÷   «   s   ù€ ÷%4ð %4ðN €Or    c               ó   € V ^8„  d   QhRR/# rÉ   rQ   )rË   s   "r   rÌ   rÌ   Ù   s   € ÷ 	Cñ 	C˜Oñ 	Cr    c                ó    € VP                  R4      pW P                  9   d   R# \        W4      pV P                  P                  V4       W1R&   R# )zßPut this patch's override on top of whatever attention override is on the
hook. Runs at patch time and again from ON_PREPARE_STATE each step, so a node
applied later cannot silently replace it; idempotent once it is on top.Úoptimized_attention_overrideN)rD   r.   r÷   r   )rÊ   rF   Úcurrentrõ   s   &&  r   Úinstall_overriderü   Ù   sG   € ð "×%Ñ%Ð&DÓE€GØ—/‘/Ô!ÙÜ& uÓ6€HØ	‡O�O×Ñ˜Ô!Ø:BÐ6Ó7r    c               ó   € V ^8„  d   QhRR/# rÉ   rQ   )rË   s   "r   rÌ   rÌ   å   s   € ÷ ñ Àñ r    c                ó*  € VP                   ^ ,          pVeP   VP                  \        P                  8w  g1   VP                  P
                  R8w  g   V P                  \        8w  d   R# VP                  W6V4      pVf)   \        P                  ! VP                  4      '       g   RpVe   VP                  RWg3RV RV 24       R# VP                  '       d;   VP                  R4      pVe   VP                  V8w  d   VP                  RR	4       R# R
# )zOWhether this H3 block call takes the sparse producer (decided before any work).r¹   Frº   rÛ   rÝ   rÞ   rM   Ú	no_layoutz)no H3 layout for this call; running denseT)r«   re   ry   r¿   rf   r¼   Úhead_dimrÂ   rJ   r½   r¾   r?   r'   rD   r\   )	Úattnr   r®   rF   rÊ   rH   Ún_tokensrî   r`   s	   &&&&&&   r   Úh3_eligibler  å   sà   € à�w‰w�q�z€HØÒ˜QŸW™W¬¯©Ô6¸!¿(¹(¿-¹-È6Ô:QÐUY×UbÑUbÔfnÔUnÙØ×ÑÐ 3¸{ÓK€FØ‚~œb×6Ò6°q·x±x×@Ò@Ø;ˆØÒØ�‰˜ Ð2°g¸h¸ZÀzÐRXÐQYÐ4ZÔ[ÙØ‡y‡y€yØ$×(Ñ(Ð)<Ó=ˆØŠ>˜VŸ^™^¨xÔ7Ø�N‰N˜;Ð(SÔTÙÙr    c               ó   € V ^8„  d   QhRR/# rÉ   rQ   )rË   s   "r   rÌ   rÌ   ø   s   € ÷ ;ñ ;Èñ ;r    c                óœ  a aaaaaaa€ SP                   ^ ,          pS P                  S P                  uoo\        P                  P                  S P                  P                  SP                  R7      p\        P                  P                  S P                  P                  SP                  R7      p/ RRuoooYbuop	\        P                  P                  4       ;_uu_ 4        VP                  '       d?   VP                  VR,          SP                  4      oSR,          oVP                  VS4      p	VS\        VP!                  RR4      4      3p
VP"                  P!                  V
4      pVRJ pV'       dg   \$        P&                  ! SS3\$        P(                  SP                  R7      \$        P&                  ! SS3\$        P(                  SP                  R7      3pRRR4       VP                  '       dY   ^ SR,          3;rÞRR	R
SR
,          /oS P*                  oSe/   SP-                  SSS,          4      P/                  ^SSS4      SR&   MVP1                  W64      w  rÞV VVVVVVV3R lp\2        P4                  ! VSSW—V33RX'       d   RMX^ ,          RV'       d   RMX^,          RVP6                  RVP8                  RVP:                  R\=        V4      R\=        V4      RS P                  P>                  /SB w  pppX^ ,          PA                  V4       V^,          PA                  V4       W´P"                  X
&   Se   RS RV^,           R2MRV RV 2pVPC                  RS3RV RV 24       VP/                  SSS,          4      pSe   VSR,          ,          pS PE                  V4      #   + '       g   i     ELã; i)z�H3 attention through the chunked producer: qkv projected in 4K-token
slices straight into the kernel's int8 carriers, full Q/K/V never built.rg   NrM   rp   Úuuidsrd   rq   ÚtailFrt   Úcoarse_gatec               3  óú  <"  € \        ^ S\        4       Fà  p S	f#   SP                  S
W \        ,            4      x € K)  S	R,          W \        ,            pS
VP                  ^ 4      ,          V^ 8¬  P	                  ^4      P                  S
P                  4      ,          pSeA   S! V4      SR,          P                  SSS,          4      W VP                  ^ ,          ,           % SP                  V4      x € Kâ  	  R# 5i)r[   Nrr   r  )	r   ÚPRODUCER_CHUNKÚqkv_projÚ	clamp_minÚ	unsqueezer…   re   r~   r«   )ÚiÚidxÚxcr  ÚextraÚgater   rØ   rp   r’   r   s      €€€€€€€€r   ÚchunksÚ#h3_sparse_attention.<locals>.chunks  sË   øé € Ü�q˜!œ^Ö,ˆAØŠ|Ø—m‘m A a¬NÕ(:Ð$;Ó<Ò<ÙØ�u•+˜a¤NÕ 2Ð3ˆCØ�3—=‘= Ó#Õ$¨¨q©×';Ñ';¸AÓ'>×'AÑ'AÀ!Ç'Á'Ó'JÕJˆBØÒÙTXÐY[ÓT\��mÕ$×)Ñ)¨!¨U°XÕ-=Ó>¸qÀRÇXÁXÈaÅ[ÅÐQØ—-‘- Ó#Ô#ó -ùs   ƒC8C;ÚkmeanÚvscaler$   r%   rê   rè   ré   Úrope_epszVSA tiles (z padded rows, z prefix tiles)zsinks r»   Úproducerzsparse producer path: z	 tokens, rs   rQ   )#r«   rØ   r   ÚcomfyÚmodel_managementÚcast_toÚq_normÚweightrf   Úk_normÚmodel_prefetchÚpause_malloc_graphr'   r¦   r°   rv   rD   r4   ry   r‰   rÁ   Úto_gate_compressÚ	new_emptyr~   ra   r½   Úsol_attn_chunkedr$   r%   r&   rí   ÚepsÚcopy_r?   Úout_proj)r  r   r®   rF   rÊ   rH   r  ÚqwrÖ   Úfreqsr=   r4   Úfirstró   ré   r  rô   r  r  Úmoder  r  r   rØ   rp   r’   s   ff&&&&              @@@@@@r   Úh3_sparse_attentionr+  ø   sg  ÿ€ ð �w‰w�q�z€HØ—j‘j $§-¡-€O€Eˆ8Ü	×	Ñ	×	'Ñ	'¨¯©×(:Ñ(:À1Ç8Á8Ð	'Ó	L€BÜ	×	Ñ	×	'Ñ	'¨¯©×(:Ñ(:À1Ç8Á8Ð	'Ó	L€BØ˜D $Ð€Eˆ4�Ø€H€A€uÜ	×	Ñ	×	0Ñ	0×	2Õ	2Ø�9�9ˆ9Ø—>‘>Ð"5Ð6IÕ"JÈAÏHÉHÓUˆDØ�S•	ˆAØ×(Ñ(¨°TÓ:ˆEà˜AœuÐ%8×%<Ñ%<¸WÀbÓ%IÓJÐKˆØ—‘×!Ñ! #Ó&ˆØ˜$�ˆßä—’˜U HÐ-´U·]±]È1Ï8É8ÔTÜ—’˜U HÐ-´U·]±]È1Ï8É8ÔTðˆF÷ 
3ð ‡y‡y€yØ˜D Õ,Ð-Ð-ˆØ˜ ¨T°+Õ->Ð?ˆØ×$Ñ$ˆØÒØ#$§;¡;¨q°%¸(Õ2BÓ#C×#HÑ#HÈÈAÈuÐV^Ó#_ˆE�-Ñ øà—{‘{Ð#6ÓA‰ˆ÷	$ô 	$ô ×,Ò,Ø��5˜% b ñ+ç‰d ¨¥ð+÷ ‰t &¨¥)ð+ð �I‰Ið	+ð #(×"2Ñ"2ð	+ð ?D×>PÑ>Pð	+ô
 ˜“Jð+ô
 (,¨F£|ð+ð —‘—‘ð+ð %*ñ+Ñ€Cˆ�ð ˆ1…I‡O�O�EÔØ
ˆ1…I‡O�O�FÔØ‡L�L�ÑØEIÒEUˆ[˜˜˜>¨$¨q­'¨°.ÑAÐ]cÐdhÐciÐijÐkqÐjrÐ[s€DØ	‡N�N�J �?Ð&<¸X¸JÀiÐPTÈvÐ$VÔWØ
�(‰(�1�e˜hÕ&Ó
'€CØÒØ�$�u•+ÕˆØ�=‰=˜ÓÐ÷e 
3×	2Ð	2ús   Ã$C<N:Î:O	c               ó   € V ^8„  d   QhRR/# rÉ   rQ   )rË   s   "r   rÌ   rÌ   6  s   € ÷ 
ñ 
°?ñ 
r    c                ó8   a aaa€ R/ 3V VV3R lloVV VV3R lpV# )zBRuns the block with its attention swapped for the sparse producer.Nc                ó6   <€ \        SP                  WVSS4      # r
   )r+  r  )Úhr®   rF   ÚblockrH   rÊ   s   &&&€€€r   Ú	attentionÚ&make_h3_block_patch.<locals>.attention8  s   ø€ Ü" 5§:¡:¨qÐ>QÐSXÐZeÓfÐfr    c                óœ   <€ \        SP                  V R ,          V R,          V R,          SS4      '       d	   / V CRS/Cp VR,          ! V 4      # )Úimgr®   rF   r1  Úoriginal_block)r  r  )rÕ   r  r1  r0  rH   rÊ   s   &&€€€€r   Úblock_patchÚ(make_h3_block_patch.<locals>.block_patch;  sO   ø€ Ü�u—z‘z 4¨¥;°°\Õ0BÀDÐI^ÕD_ÐafÐhs×tÒtØ3�dÐ3˜K¨Ñ3ˆDØÐ%Ö& tÓ,Ð,r    rQ   )r0  rH   rÊ   r6  r1  s   fff @r   Úmake_h3_block_patchr8  6  s&   û€ à $¸"÷ gñ g÷-ð -ð
 Ðr    c       
        óº  a€ V P                  R 4      pV'       d!   V	'       d   \        P                  ! R4       ^ p	\        WV\	        VP                  V4      4      \	        VP                  V4      4      WgW‰V
R7
      oV P                  4       p\        SVP                  R,          4       VP                  \        P                  P                  P                  RV3R l4       VP                  \        P                  P                  P                  RV3R l4       V P                  R4      p\        V\         4      '       d†   \#        VP$                  4       F#  w  rïVP'                  \)        WþS4      RR	V4       K%  	  V'       d@   VP$                  ^ ,          P*                  P,                  f   \        P.                  ! R
4       V# V'       d   \1        R4      hV# )Úmodel_samplingzDVSA: extra_tokens ignored (the trained sparse pattern is the target))
r$   r%   r'   r(   r)   r*   r+   r,   r&   r-   rF   Úblock_sparse_attentionc                ó*   <€ \        SVR ,          4      # )rF   )rü   )Úmodel_patcherÚtimestepÚmodel_optionsrÊ   s   &&&€r   Ú<lambda>Ú.apply_block_sparse_attention.<locals>.<lambda>R  s!   ø€ ÔK[Ð\aÐcpð  rGõ  dHô  LIr    c                ó$   <€ SP                  4       # r
   )r/   )r=  rÊ   s   &€r   r@  rA  T  s   ø€ ÈEÏKÉKÌMr    Údiffusion_modelÚditÚdouble_blockz_VSA: the model has no to_gate_compress layers; running the fine stage without the coarse branchz&VSA selection needs a MiniMax-H3 model)Úget_model_objectr;   r<   r"   rE   Úpercent_to_sigmaÚclonerü   r?  Úadd_callback_with_keyr  Úpatcher_extensionÚCallbacksMPÚON_PREPARE_STATEÚ
ON_CLEANUPÚ
isinstancer   Ú	enumerater   Úset_model_patch_replacer8  r  r!  ÚwarningÚ
ValueError)Úmodelr$   r%   r'   Ústart_percentÚend_percentr*   r+   r,   r&   r-   r:  rš   rC  r  r0  rÊ   s   &$$$$$$$$$$     @r   Úapply_block_sparse_attentionrV  C  s}  ø€ à×+Ñ+Ð,<Ó=€Nß
�|ä�ŠÐ[Ô\ØˆÜ ÀÜ(-¨n×.MÑ.MÈmÓ.\Ó(]Ü&+¨N×,KÑ,KÈKÓ,XÓ&YØ'1Ø.?Ðdkô	m€Eð
 	�‰‹€AÜ�U˜AŸO™OÐ,AÕBÔCØ×ÑœE×3Ñ3×?Ñ?×PÑPÐRjô IôJà×ÑœE×3Ñ3×?Ñ?×JÑJØ4Ô6Yô[ð ×,Ñ,Ð->Ó?€OÜ�/¤>×2Ò2Ü! /×"8Ñ"8Ö9‰HˆAØ×%Ñ%Ô&9¸%ÀEÓ&JÈEÐSaÐcdÖeñ :ç�?×)Ñ)¨!Õ,×1Ñ1×BÑBÒJÜ�OŠOÐ}Ô~ð €H÷ 
ÜÐAÓBÐBØ€Hr    c                  ó@   € ] tR tRt]R 4       t]RR R ll4       tRtR# )ÚBlockSparseAttentionia  c                	óŒ  € \         P                  ! R RRRR.R\         P                  P                  RRR7      \         P                  P                  R	R
\         P                  P                  R\         P                  P                  RRRRRRR7      .4      \         P                  P                  R\         P                  P                  RRRRRRR7      .4      \         P                  P                  R\         P                  P                  RRRRRRR7      .4      .RR7      \         P                  P                  RRRRR R!R7      \         P                  P                  R"RRRR R#R7      \         P                  P                  R$R%RR&R'7      \         P                  P                  R(R)^ R:R*RR+R,7      \         P                  P                  R-R.^ R.^@RR/R,7      \         P                  P                  R0. R;OR1RR2R37      \         P                  P                  R4R5RR6R'7      .	\         P                  P                  RR7R87      .R97      # )<rX  zModel Sparse Attentionzmodel/patchTzBlock Sparse Attentionas  Applies block-sparse attention to eligible model attention layers, reducing compute for long sequences. The speed gain grows with sequence length since short sequences are usually faster dense. Outside the start/end_percent, dense_blocks and under min_tokens, the model uses the dense model attention backend. Use the node Model Attention Backend to select that fallback.rS  zThe model to patch.)ÚtooltipÚ	selectionÚmethodúsol-attnr$   çÍÌÌÌÌÌô?ç        g      @gš™™™™™©?zrThreshold in score-distribution sigmas. Higher is sparser: 1.0 keeps ~16% of key blocks exact, 1.5 ~7%, 2.0 ~2.7%.)Údefaultr   r   ÚsteprZ  ÚslaÚkeep_percentg      $@g      à?g     ÀW@z¾Percent of key blocks each query block keeps exactly (sinks and the diagonal ride on top). The selection SLA-style LoRAs are distilled against; without such a LoRA higher is closer to dense.r'   z¤Percent of video cubes each query cube keeps; FastH3-VSA checkpoints are trained at 10. Uses the model's to_gate_compress layers for the coarse branch when present.aÏ  Method used to choose key blocks for full token-level attention. sol-attn: Sparsifying Online Attention uses a training-free adaptive threshold for each attention head and query block. sla: Sparse-Linear Attention keeps a fixed percentage of the highest-scoring key blocks; use only with model weights trained for this pattern. vsa: Video Sparse Attention (FastVideo) uses 3D video-cube tiling and a learned coarse attention branch; requires FastH3 model weights.)Údisplay_nameÚoptionsrZ  rT  gš™™™™™É?g      ð?g{®Gáz„?zXPercentage point when sparse attention begins. Before this point, attention stays dense.rU  zZPercentage point when sparse attention ends. After this point, attention returns to dense.r+   r   z=Transformer blocks that always run dense, e.g. '0, 1, 47-49'.)r`  ÚadvancedrZ  r*   é 0  i   z'Sequences shorter than this stay dense.)r`  r   r   ra  rf  rZ  r&   é   z¤Extra top-scoring tokens each query block attends beyond its selected blocks. Closer to dense for more attention time; 256 recommended, 0 disables. Ignored for VSA.r,   rW   zÎMiniMax-H3 only. exact_kv: every query attends the packed text/audio/reference rows exactly (~3% cost). exact_kv_and_rows: additionally runs the target-audio query rows dense (keeps generated audio intact).)re  r`  rf  rZ  r-   FzOLogs whether each attention shape used sparse attention or why it stayed dense.z.The model with block-sparse attention applied.)rd  rZ  )Únode_idrd  ÚcategoryÚis_experimentalÚsearch_aliasesÚdescriptionÚinputsÚoutputsi   )Úexact_kvrW   rN   )r   ÚSchemaÚModelÚInputÚDynamicComboÚOptionÚFloatÚStringÚIntÚComboÚBooleanÚOutput)Úclss   &r   Údefine_schemaÚ"BlockSparseAttention.define_schemab  s[  € ä�yŠyØ*Ø1Ø"Ø Ø4Ð5ðXô
 —‘—‘˜wÐ0E�ÓFÜ—‘×%Ñ% kÀÜ—O‘O×*Ñ*¨:ÜŸ™Ÿ™ u°c¸sÈÐRVð0ið 'ó jð8ó ô
 —O‘O×*Ñ*¨5ÜŸ™Ÿ™ ~¸tÈÐRVÐ]`ð0sð 'ó tð3ó ô —O‘O×*Ñ*¨5ÜŸ™Ÿ™ ~¸tÈÐRVÐ]`ð0\ð 'ó ]ð3ó ðSð$eð% &ó fô, —‘—‘˜¸ÀÈ#ÐTXð (Bð ó Cä—‘—‘˜}°c¸sÈÐRVð (Dð ó Eä—	‘	—‘ ¸ÀTØ(gð  ó iä—‘—‘˜\°5¸aÀWÐSVÐaeØ%Nð ó Pä—‘—‘˜^°S¸aÀSÈrÐ\`ð&8ð ó 9ô —‘—‘Ð2Ò<dØ':ÀTð(kð ó lô
 —
‘
× Ñ  °EÀDØ)zð !ó |ðS+ôX —X‘X—_‘_°'ÐCs�_ÓtÐuôm7
ð 7	
r    c               ó   € V ^8„  d   QhRR/# )r©   Úreturnzio.NodeOutputrQ   )rË   s   "r   rÌ   Ú!BlockSparseAttention.__annotate__ž  s   € ÷ ,ñ ,ØYfñ,r    c
                	óÒ   € VR ,          p
\        TVP                  RR4      V
R8X  d   RMVR,          R,          V
R8H  VVV\        V4      VVV	R7      p\        P                  ! V4      # )	r[  r$   r^  r]  r_  rc  g      Y@r'   )
r$   r%   r'   rT  rU  r*   r+   r,   r&   r-   )rV  rD   r   r   Ú
NodeOutput)r|  rS  r[  rT  rU  r+   r*   r&   r,   r-   r*  Úpatched_models   &&&&&&&&&&  r   ÚexecuteÚBlockSparseAttention.execute�  sr   € ð ˜Õ%ˆÜ4ØØ—‘˜e SÓ)Ø" jÔ0‘s°iÀÕ6OÐRWÕ6WØ˜‘Ø'Ø#Ø!Ü)¨,Ó7Ø/Ø%Øô
ˆô �}Š}˜]Ó+Ð+r    rQ   N)r   rg  r[   rW   F)r²   r³   r´   rµ   Úclassmethodr}  r…  r·   rQ   r    r   rX  rX  a  s)   † Øñ8
ó ð8
ðt ö,ó ô,r    rX  c                  ó   € ] tR tRtR tRtR# )ÚBlockSparseAttentionExtensioni±  c              ƒ  	ó   "  € \         .# 5ir
   )rX  r8   s   &r   Úget_node_listÚ+BlockSparseAttentionExtension.get_node_list²  s   é € Ü$Ð%Ð%ùs   ‚
rQ   N)r²   r³   r´   rµ   r‹  r·   rQ   r    r   r‰  r‰  ±  s   † ö&r    r‰  c               ó   € V ^8„  d   QhRR/# )r©   r€  r‰  rQ   )rË   s   "r   rÌ   rÌ   ¶  s   € ÷ +ñ +Ð =ñ +r    c               ƒ  ó   "  € \        4       # 5ir
   )r‰  rQ   r    r   Úcomfy_entrypointr�  ¶  s   é € Ü(Ó*Ð*ùs   ‚)é   r�  r�  )$r¶   Ú
__future__r   r;   r   r¬   Úcomfy_kitchenr½   ry   Úcomfy.model_managementr  Úcomfy.model_prefetchÚcomfy.patcher_extensionÚcomfy.ldm.minimax.modelr   Úcomfy_api.latestr   r   rÂ   r_   r
  r€   r�   r   r"   rÇ   r÷   rü   r  r+  r8  rV  Ú	ComfyNoderX  r‰  r�  rQ   r    r   Ú<module>r™     s£   ðñWõ #ã Û 	Û ã Û ã Û Û Ý 2ß /à€Ø€
Ø€Ø€Ø€ò	÷qñ qòhõ"+õ\	Cõõ&;õ|
òô<M,˜2Ÿ<™<ô M,ô`& Nô &÷
+r    