+
    œ°³jäf  ã                   óà  € ^ RI t ^ RIHt ^ RIt^ RIHt ^RIHt ^RIH	t	 ^RI
HtHt ^RIHt ^RIHt ^R	IHt ^R
IHtHt ^RIHtHt ^RIHtHt ^RIHtHt ^RIH t  ^RI!H"t"H#t#H$t$ ^RI%H&t& ^RI'H(t(H)t) ^RI*H+t+ ^RI,H-t-  ! R R]P\                  4      t/]! R4       ! R R]P`                  4      4       t1 ! R R]P`                  4      t2 ! R R]P`                  4      t3R R lt4R2R R  llt5R3R! lt6R" t7R# t8]! R$4      R4R% l4       t9 ! R& R']P`                  4      t: ! R( R)]4      t;]# ! R* R+]4      4       t<]# ! R, R-]<4      4       t=]# ! R. R/]<]4      4       t> ! R0 R1]]<4      t?. R5Ot@R# )6é    N)ÚCallable)Únn)Úinitialization)ÚACT2FN)ÚCacheÚDynamicCache)ÚGenerationMixin)Úuse_kernel_forward_from_hub)Úcreate_causal_mask)Ú GenericForSequenceClassificationÚGradientCheckpointingLayer)ÚBaseModelOutputWithPastÚCausalLMOutputWithPast)ÚROPE_INIT_FUNCTIONSÚdynamic_rope_update)ÚALL_ATTENTION_FUNCTIONSÚPreTrainedModel)ÚUnpack)ÚTransformersKwargsÚauto_docstringÚcan_return_tuple)Údeprecate_kwarg)Úmaybe_autocastÚmerge_with_config_defaults)Úcapture_outputs©ÚMiniCPM3Configc                   ó^   a a€ ] tR t^-t oRtRV3R lV 3R llltV3R lV 3R lltRtVtV ;t	# )ÚMiniCPM3ScaledWordEmbeddingzT
This module overrides nn.Embeddings' forward by multiplying with embeddings scale.
c                ó2   <€ V ^8„  d   QhRS[ RS[ RS[ RS[/# )é   Únum_embeddingsÚembedding_dimÚpadding_idxÚembed_scale)ÚintÚfloat)ÚformatÚ__classdict__s   "€Úo/home/luca/screen-checking/.venv/lib/python3.14/site-packages/transformers/models/minicpm3/modeling_minicpm3.pyÚ__annotate__Ú(MiniCPM3ScaledWordEmbedding.__annotate__2   s0   ø€ ÷ Rñ R¡sð R¹3ð RÉSð RÑ_dñ Ró    c                ó–   <€ \         SV `  WV4       W@n        \        P                  ! \
        P                  ! V4      R R7      V n        R# )F©Ú
persistentN)ÚsuperÚ__init__Úscalar_embed_scaler   ÚBufferÚtorchÚtensorr%   )Úselfr"   r#   r$   r%   Ú	__class__s   &&&&&€r*   r2   Ú$MiniCPM3ScaledWordEmbedding.__init__2   s4   ø€ Ü‰Ñ˜¸ÔDØ"-ÔÜŸ9š9¤U§\¢\°+Ó%>È5ÔQˆÖr-   c                ó4   <€ V ^8„  d   QhRS[ P                  /# )r!   Ú	input_ids©r5   ÚTensor)r(   r)   s   "€r*   r+   r,   7   s   ø€ ÷ Sñ S¡§¡ñ Sr-   c                óŠ   <€ \         SV `  V4      V P                  P                  V P                  P
                  4      ,          # ©N)r1   Úforwardr%   ÚtoÚweightÚdtype)r7   r;   r8   s   &&€r*   r@   Ú#MiniCPM3ScaledWordEmbedding.forward7   s2   ø€ Ü‰w‰˜yÓ)¨D×,<Ñ,<×,?Ñ,?ÀÇÁ×@QÑ@QÓ,RÕRÐRr-   )r%   r3   )ç      ð?)
Ú__name__Ú
__module__Ú__qualname__Ú__firstlineno__Ú__doc__r2   r@   Ú__static_attributes__Ú__classdictcell__Ú__classcell__©r8   r)   s   @@r*   r   r   -   s(   ù‡ € ñ÷Rõ R÷
S÷ Só Sr-   r   ÚRMSNormc                   óZ   a a€ ] tR t^;t oRV3R lV 3R llltV3R lR ltR tRtVtV ;t	# )ÚMiniCPM3RMSNormc                ó$   <€ V ^8„  d   QhRS[ RR/# )r!   ÚepsÚreturnN)r'   )r(   r)   s   "€r*   r+   ÚMiniCPM3RMSNorm.__annotate__=   s   ø€ ÷ $ñ $©ð $¸$ñ $r-   c                óŽ   <€ \         SV `  4        \        P                  ! \        P
                  ! V4      4      V n        W n        R# )z.
MiniCPM3RMSNorm is equivalent to T5LayerNorm
N)r1   r2   r   Ú	Parameterr5   ÚonesrB   Úvariance_epsilon)r7   Úhidden_sizerS   r8   s   &&&€r*   r2   ÚMiniCPM3RMSNorm.__init__=   s/   ø€ ô 	‰ÑÔÜ—l’l¤5§:¢:¨kÓ#:Ó;ˆŒØ #Ör-   c                óN   <€ V ^8„  d   QhRS[ P                  RS[ P                  /# )r!   Úhidden_statesrT   r<   )r(   r)   s   "€r*   r+   rU   E   s#   ø€ ÷ ;ñ ;¡U§\¡\ð ;±e·l±lñ ;r-   c                ó:  € VP                   pVP                  \        P                  4      pVP	                  ^4      P                  RRR7      pV\        P                  ! W0P                  ,           4      ,          pV P                  VP                  V4      ,          # )r!   T)Úkeepdiméÿÿÿÿ)	rC   rA   r5   Úfloat32ÚpowÚmeanÚrsqrtrY   rB   )r7   r]   Úinput_dtypeÚvariances   &&  r*   r@   ÚMiniCPM3RMSNorm.forwardE   sw   € Ø#×)Ñ)ˆØ%×(Ñ(¬¯©Ó7ˆØ ×$Ñ$ QÓ'×,Ñ,¨R¸Ð,Ó>ˆØ%¬¯ª°H×?TÑ?TÕ4TÓ(UÕUˆØ�{‰{˜]×-Ñ-¨kÓ:Õ:Ð:r-   c                ó^   € \        V P                  P                  4       R V P                   2# )z, eps=)ÚtuplerB   ÚshaperY   )r7   s   &r*   Ú
extra_reprÚMiniCPM3RMSNorm.extra_reprL   s*   € Ü˜Ÿ™×)Ñ)Ó*Ð+¨6°$×2GÑ2GÐ1HÐIÐIr-   )rY   rB   )g�íµ ÷Æ°>)
rF   rG   rH   rI   r2   r@   rk   rK   rL   rM   rN   s   @@r*   rQ   rQ   ;   s%   ù‡ € ÷$õ $÷;ð ;÷Jò Jr-   rQ   c                   óÎ   a a€ ] tR t^Pt o]! RRR7      R
V3R lV 3R lll4       t]]! RRR7      R
V3R lR ll4       4       t]P                  ! 4       ]
R 4       4       tR	tVtV ;t# )ÚMiniCPM3RotaryEmbeddingÚdevicez5.18)Úversionc                ó    <€ V ^8„  d   QhRS[ /# ©r!   Úconfigr   )r(   r)   s   "€r*   r+   Ú$MiniCPM3RotaryEmbedding.__annotate__R   s   ø€ ÷ Oñ O™~ñ Or-   c                óä  <€ \         SV `  4        VP                  V n        VP                  V n        Wn        V P
                  P                  R ,          V n        V P                  pV P                  R8w  d   \        V P                  ,          pV! V P
                  V4      w  q@n
        \        P                  ! VRR7      V n        \        P                  ! VP                  4       RR7      V n        R# )Ú	rope_typeÚdefaultFr/   N)r1   r2   Úmax_position_embeddingsÚmax_seq_len_cachedÚoriginal_max_seq_lenrs   Úrope_parametersrv   Úcompute_default_rope_parametersr   Úattention_scalingr   r4   Úinv_freqÚcloneÚoriginal_inv_freq)r7   rs   ro   Úrope_init_fnr~   r8   s   &&&  €r*   r2   Ú MiniCPM3RotaryEmbedding.__init__Q   s¬   ø€ ä‰ÑÔØ"(×"@Ñ"@ˆÔØ$*×$BÑ$BˆÔ!àŒàŸ™×4Ñ4°[ÕAˆŒØ!%×!EÑ!EˆØ�>‰>˜YÔ&Ü.¨t¯~©~Õ>ˆLÙ+7¸¿¹ÀVÓ+LÑ(ˆÔ(äŸ	š	 (°uÔ=ˆŒÜ!#§¢¨8¯>©>Ó+;ÈÔ!NˆÖr-   c                óP   <€ V ^8„  d   QhRS[ RS[S[P                  S[3,          /# )r!   rs   rT   )r   ri   r5   r=   r'   )r(   r)   s   "€r*   r+   rt   d   s.   ø€ ÷ 5ñ 5±ð 5ÑZ_Ñ`e×`lÑ`lÑnsÐ`sÕZtñ 5r-   c           	     ó0  € V P                   R,          p\        V RR4      ;'       g    V P                  V P                  ,          pRpRV\        P
                  ! ^ V^\        P                  R7      V,          ,          ,          pVP                  V4      V3# )a~  
Computes the inverse frequencies according to the original RoPE implementation
Args:
    config ([`~transformers.PreTrainedConfig`]):
        The model configuration.
Returns:
    Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
    post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
Ú
rope_thetaÚhead_dimNrE   ©rC   )r{   ÚgetattrrZ   Únum_attention_headsr5   Úaranger'   rA   )rs   ro   ÚkwargsÚbaseÚdimÚattention_factorr~   s   &&,    r*   r|   Ú7MiniCPM3RotaryEmbedding.compute_default_rope_parametersb   s~   € ð ×%Ñ% lÕ3ˆÜ�f˜j¨$Ó/×cÐc°6×3EÑ3EÈ×IcÑIcÕ3cˆàÐà˜$¤5§<¢<°°3¸ÄÇÁÔ#MÐPSÕ#SÕTÕUˆØ�{‰{˜6Ó"Ð$4Ð4Ð4r-   c                ó^  € V P                   R,          P                  VP                  ^ ,          R	^4      P                  \        P
                  VP                  R7      pVR
,          P                  4       p\        VP                  P                  \        4      '       d2   VP                  P                  R8w  d   VP                  P                  MRp\        VRR7      ;_uu_ 4        W4,          P                  ^^4      p\        P                  ! Wf3R	R7      pVP                  4       V P                  ,          pVP                  4       V P                  ,          p	R R R 4       XP                  VP                   R7      X	P                  VP                   R7      3#   + '       g   i     LH; i)N)rC   ro   ÚmpsÚcpuF)Údevice_typeÚenabled©r�   r‡   )NºNNNNr`   )r–   Nr–   )r~   Úexpandrj   rA   r5   r'   ro   Ú
isinstanceÚtypeÚstrr   Ú	transposeÚcatÚcosr}   ÚsinrC   )
r7   ÚxÚposition_idsÚinv_freq_expandedÚposition_ids_expandedr“   ÚfreqsÚembr�   rž   s
   &&&       r*   r@   ÚMiniCPM3RotaryEmbedding.forwardv   s2  € ð �M‰M˜-Õ(×/Ñ/°×0BÑ0BÀ1Õ0EÀrÈ1ÓM×PÑPÔW\×WbÑWbÐkl×ksÑksÐPÓtð 	ð !-¨ZÕ 8× >Ñ >Ó @Ðä'1°!·(±(·-±-Ä×'EÒ'EÈ!Ï(É(Ï-É-Ð[`ÔJ`�a—h‘h—m’mÐfkˆä¨¸U×CÖCØ&Õ>×IÑIÈ!ÈQÓOˆEÜ—)’)˜U˜N°Ô3ˆCØ—'‘'“)˜d×4Ñ4Õ4ˆCØ—'‘'“)˜d×4Ñ4Õ4ˆC÷	 Dð �v‰v˜AŸG™GˆvÓ$ c§f¡f°1·7±7 fÓ&;Ð;Ð;÷ D×Cús   Ã(A4FÆF,	)r}   rs   r~   ry   r€   rz   rv   r?   )rF   rG   rH   rI   r   r2   Ústaticmethodr|   r5   Úno_gradr   r@   rK   rL   rM   rN   s   @@r*   rn   rn   P   su   ù‡ € Ù�X vÔ.÷Oô Oó /ðOð  Ù�X vÔ.÷5ñ 5ó /ó ð5ð$ ‡]‚]ƒ_Øñ<ó ó ÷<ð <r-   rn   c                   ó8   a a€ ] tR t^‰t oV 3R ltR tRtVtV ;t# )ÚMiniCPM3MLPc                ó  <€ \         SV `  4        Wn        VP                  V n        VP                  V n        \
        P                  ! V P                  V P                  VP                  R 7      V n        \
        P                  ! V P                  V P                  VP                  R 7      V n	        \
        P                  ! V P                  V P                  VP                  R 7      V n
        \        VP                  ,          V n        R# )©ÚbiasN)r1   r2   rs   rZ   Úintermediate_sizer   ÚLinearÚmlp_biasÚ	gate_projÚup_projÚ	down_projr   Ú
hidden_actÚact_fn©r7   rs   r8   s   &&€r*   r2   ÚMiniCPM3MLP.__init__Š   s¶   ø€ Ü‰ÑÔØŒØ!×-Ñ-ˆÔØ!'×!9Ñ!9ˆÔÜŸš 4×#3Ñ#3°T×5KÑ5KÐRX×RaÑRaÔbˆŒÜ—y’y ×!1Ñ!1°4×3IÑ3IÐPV×P_ÑP_Ô`ˆŒÜŸš 4×#9Ñ#9¸4×;KÑ;KÐRX×RaÑRaÔbˆŒÜ˜V×.Ñ.Õ/ˆŽr-   c                ó�   € V P                  V P                  V P                  V4      4      V P                  V4      ,          4      pV# r?   )r²   r´   r°   r±   )r7   rŸ   r²   s   && r*   r@   ÚMiniCPM3MLP.forward”   s6   € Ø—N‘N 4§;¡;¨t¯~©~¸aÓ/@Ó#AÀDÇLÁLÐQRÃOÕ#SÓTˆ	ØÐr-   )r´   rs   r²   r°   rZ   r­   r±   ©	rF   rG   rH   rI   r2   r@   rK   rL   rM   rN   s   @@r*   r©   r©   ‰   s   ù‡ € õ0÷ò r-   r©   c                ód   € V ^8„  d   QhR\         P                  R\        R\         P                  /# )r!   r]   Ún_reprT   )r5   r=   r&   )r(   s   "r*   r+   r+   ™   s.   € ÷ 	Uñ 	UœUŸ\™\ð 	U´#ð 	U¼%¿,¹,ñ 	Ur-   c                ó˜   € V P                   w  r#rEV^8X  d   V # V R,          P                  W#WV4      p V P                  W#V,          WE4      # )zÈ
This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
)r–   r–   Nr–   r–   )rj   r—   Úreshape)r]   r»   ÚbatchÚnum_key_value_headsÚslenr†   s   &&    r*   Ú	repeat_kvrÁ   ™   sU   € ð
 2?×1DÑ1DÑ.€E Ø�„zØÐØ!Ð"2Õ3×:Ñ:¸5ÐW\ÐdlÓm€MØ× Ñ  ¸eÕ(CÀTÓTÐTr-   c                ó   € V ^8„  d   QhR\         P                  R\        P                  R\        P                  R\        P                  R\        P                  R,          R\        R\        R	\
        \        ,          /# )
r!   ÚmoduleÚqueryÚkeyÚvalueÚattention_maskNÚscalingÚdropoutr‹   )r   ÚModuler5   r=   r'   r   r   )r(   s   "r*   r+   r+   ¥   sw   € ÷ %ñ %Ü�I‰Ið%ä�<‰<ð%ô 
�‰ð%ô �<‰<ð	%ô
 —L‘L 4Õ'ð%ô ð%ô ð%ô Ô'Õ(ñ%r-   c                 ó"  € \        W P                  4      p\        W0P                  4      p	\        P                  ! WP	                  ^^4      4      V,          p
Ve	   W¤,           p
\
        P                  P                  V
R\        P                  R7      P                  VP                  4      p
\
        P                  P                  W¦V P                  R7      p
\        P                  ! W©4      pVP	                  ^^4      P                  4       pWº3# )r!   )r�   rC   )ÚpÚtrainingr`   )rÁ   Únum_key_value_groupsr5   Úmatmulr›   r   Ú
functionalÚsoftmaxra   rA   rC   rÉ   rÍ   Ú
contiguous)rÃ   rÄ   rÅ   rÆ   rÇ   rÈ   rÉ   r‹   Ú
key_statesÚvalue_statesÚattn_weightsÚattn_outputs   &&&&&&&,    r*   Úeager_attention_forwardr×   ¥   sÒ   € ô ˜3× ;Ñ ;Ó<€JÜ˜U×$?Ñ$?Ó@€Lä—<’< ×';Ñ';¸A¸qÓ'AÓBÀWÕL€LØÒ!Ø#Õ4ˆä—=‘=×(Ñ(¨¸2ÄUÇ]Á]Ð(ÓS×VÑVÐW\×WbÑWbÓc€LÜ—=‘=×(Ñ(¨È6Ï?É?Ð(Ó[€LÜ—,’,˜|Ó:€KØ×'Ñ'¨¨1Ó-×8Ñ8Ó:€KàÐ$Ð$r-   c                 ój   € V ^8:  d   R# RV,          \         P                  ! V 4      ,          R,           # )é   rE   gš™™™™™¹?)ÚmathÚlog)ÚscaleÚmscales   &&r*   Úyarn_get_mscalerÞ   ¾   s(   € Ø�„zÙØ��<œ$Ÿ(š( 5›/Õ)¨CÕ/Ð/r-   c                 ó®   € V P                  R R4      R8w  d>   V P                  R^ 4      pV R,          pV'       d   \        W24      pW,          V,          pV# )rv   rw   Úmscale_all_dimÚfactor)ÚgetrÞ   )r{   rÈ   rà   Úscaling_factorrÝ   s   &&   r*   Úyarn_apply_mscalerä   Ä   sS   € Ø×Ñ˜;¨	Ó2°iÔ?Ø(×,Ñ,Ð-=¸qÓAˆØ(¨Õ2ˆßÜ$ ^ÓDˆFØÕ&¨Õ/ˆGØ€Nr-   c                óÌ   € V RRV P                   R,          ^,          13,          pV RV P                   R,          ^,          R13,          p\        P                  ! V) V3RR7      # )z*Rotates half the hidden dims of the input..Nr•   r`   )rj   r5   rœ   )rŸ   Úx1Úx2s   &  r*   Úrotate_halfrè   Î   sZ   € à	
ˆ3Ð"�!—'‘'˜"•+ Õ"Ð"Ð"Õ	#€BØ	
ˆ3�—‘˜•˜qÕ Ñ"Ð"Õ	#€BÜ�9Š9�r�c˜2�Y BÔ'Ð'r-   Úrotary_pos_embc                óÈ   € VP                  V4      pVP                  V4      pW,          \        V 4      V,          ,           pW,          \        V4      V,          ,           pWV3# )aI  Applies Rotary Position Embedding to the query and key tensors.

Args:
    q (`torch.Tensor`): The query tensor.
    k (`torch.Tensor`): The key tensor.
    cos (`torch.Tensor`): The cosine part of the rotary embedding.
    sin (`torch.Tensor`): The sine part of the rotary embedding.
    unsqueeze_dim (`int`, *optional*, defaults to 1):
        The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
        sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
        that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
        k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
        cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
        the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
Returns:
    `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
)Ú	unsqueezerè   )ÚqÚkr�   rž   Úunsqueeze_dimÚq_embedÚk_embeds   &&&&&  r*   Úapply_rotary_pos_embrñ   Õ   sS   € ð& �-‰-˜Ó
&€CØ
�-‰-˜Ó
&€CØ�wœ; q›>¨CÕ/Õ0€GØ�wœ; q›>¨CÕ/Õ0€GØÐÐr-   c                   ój   a a€ ] tR t^ït oRtV3R lV 3R lltV3R lR ltR	V3R lR lltRtVt	V ;t
# )
ÚMiniCPM3Attentiona1  
Multi-head Latent Attention (MLA), structurally identical to `DeepseekV2Attention`.
The only difference is the rotary convention: MiniCPM3 keeps the original cos/sin RoPE
(`apply_rotary_pos_emb`) instead of DeepSeek-V2's complex rotary, so we inherit the
module construction and override only `forward`.
c                ó&   <€ V ^8„  d   QhRS[ RS[/# ©r!   rs   Ú	layer_idx©r   r&   )r(   r)   s   "€r*   r+   ÚMiniCPM3Attention.__annotate__÷   s   ø€ ÷ 5]ñ 5]™~ð 5]¹#ñ 5]r-   c                ó‚  <€ \         SV `  4        Wn        W n        VP                  V n        VP
                  V n        VP                  V n        VP                  V n        VP                  V n	        VP                  V n
        VP                  V n        VP                  V n        VP                  VP                  ,           V n        VP                  VP                  ,          V n        R V n        V P                  f?   \"        P$                  ! V P
                  V P                  V P                  ,          RR7      MRV n        V P                  e8   \"        P$                  ! V P
                  VP                  VP(                  R7      MRV n        V P                  e   \-        VP                  4      MRV n        V P                  e?   \"        P$                  ! VP                  V P                  V P                  ,          RR7      MRV n        \"        P$                  ! V P
                  VP                  VP                  ,           VP(                  R7      V n        \-        VP                  4      V n        \"        P$                  ! VP                  V P                  V P                  V P                  ,           ,          RR7      V n        \"        P$                  ! V P                  V P                  ,          V P
                  VP(                  R7      V n        \;        VP<                  V P                  R,          4      V n        R# )TNFr«   g      à¿) r1   r2   rs   rö   Úattention_dropoutrZ   r‰   Ú	num_headsÚq_lora_rankÚqk_rope_head_dimÚkv_lora_rankÚ
v_head_dimÚqk_nope_head_dimÚqk_head_dimr¿   rÎ   Ú	is_causalr   r®   Úq_projÚattention_biasÚq_a_projrQ   Úq_a_layernormÚq_b_projÚkv_a_proj_with_mqaÚkv_a_layernormÚ	kv_b_projÚo_projrä   r{   rÈ   ©r7   rs   rö   r8   s   &&&€r*   r2   ÚMiniCPM3Attention.__init__÷   s{  ø€ Ü‰ÑÔØŒØ"ŒØ!'×!9Ñ!9ˆÔØ!×-Ñ-ˆÔØ×3Ñ3ˆŒà!×-Ñ-ˆÔØ &× 7Ñ 7ˆÔØ"×/Ñ/ˆÔØ ×+Ñ+ˆŒØ &× 7Ñ 7ˆÔØ!×2Ñ2°V×5LÑ5LÕLˆÔØ$*×$>Ñ$>À&×B\ÑB\Õ$\ˆÔ!àˆŒð ×ÑÒ'ô �IŠI�d×&Ñ&¨¯©¸×9IÑ9IÕ(IÐPUÕVàð 	Œð ×ÑÒ+ô �IŠI�d×&Ñ&¨×(:Ñ(:À×AVÑAVÕWàð 	Œð
 EI×DTÑDTÒD`œ_¨V×-?Ñ-?Ô@ÐfjˆÔð ×ÑÒ+ô �IŠI�f×(Ñ(¨$¯.©.¸4×;KÑ;KÕ*KÐRWÕXàð 	Œô #%§)¢)Ø×ÑØ×Ñ &×"9Ñ"9Õ9Ø×&Ñ&ô#
ˆÔô
 .¨f×.AÑ.AÓBˆÔÜŸšØ×ÑØ�N‰N˜d×3Ñ3°d·o±oÕEÕFØô
ˆŒô —i’iØ�N‰N˜TŸ_™_Õ,Ø×ÑØ×&Ñ&ô
ˆŒô )¨×)?Ñ)?À×AQÑAQÐVZÕA[Ó\ˆŽr-   c                ó’   <€ V ^8„  d   QhRS[ P                  RS[ P                  RS[S[ P                  S[ P                  3,          /# )r!   Úkv_nopeÚk_rotrT   )r5   r=   ri   )r(   r)   s   "€r*   r+   rø   .  s?   ø€ ÷ (ñ (¡§¡ð (±e·l±lð (ÁuÉUÏ\É\Ñ[`×[gÑ[gÐMgÕGhñ (r-   c                ó‚  € VP                   w  r4rTW5RV P                  V P                  ,           3pV P                  V4      P	                  V4      P                  ^^4      p\        P                  ! WP                  V P                  .RR7      w  rxVP                  RVP                   ^,          RR4      pVP                  ! . VP                   RR OV P                  V P                  ,           N5!  p	V	RRV P                  13,          P                  V4       V	RV P                  R13,          P                  V4       W˜3# )as  Expands the compressed latents into key and value states. Args:
    - kv_nope: key + value without positional encoding, shape [batch_size, 1, seqlen, self.kv_lora_rank]
    - k_rot: shared key with positional encoding, shape [batch_size, 1, seqlen, self.qk_rope_head_dim]
Returns the key and value states, two tensors of shape [batch, num_heads, seq, (k or v)_head_dim].
r•   N.r`   )rj   r   rÿ   r
  Úviewr›   r5   Úsplitr—   Ú	new_emptyrý   Úcopy_)
r7   r  r  Ú
batch_sizeÚ_Ú
seq_lengthÚ	key_shapeÚk_noperÔ   rÓ   s
   &&&       r*   Ú	expand_kvÚMiniCPM3Attention.expand_kv.  s  € ð (/§}¡}Ñ$ˆ
�zØ¨R°×1FÑ1FÈÏÉÕ1XÐYˆ	à—.‘. Ó)×.Ñ.¨yÓ9×CÑCÀAÀqÓIˆÜ$Ÿ{š{¨7×5JÑ5JÈDÏOÉOÐ4\ÐbdÔeÑˆØ—‘˜R §¡¨a¥°"°bÓ9ˆð ×&Ò&Ðj¨¯©°c°rÐ(:Ðj¸D×<QÑ<QÐTX×TiÑTiÕ<iÓjˆ
Ø�3Ð/˜$×/Ñ/Ð/Ð/Õ0×6Ñ6°vÔ>Ø�3˜×-Ñ-Ñ/Ð/Õ0×6Ñ6°uÔ=ØÐ'Ð'r-   c                ó  <€ V ^8„  d   QhRS[ P                  RS[S[ P                  S[ P                  3,          RS[ P                  R,          RS[R,          RS[S[ P                  S[ P                  R,          3,          /# )r!   r]   Úposition_embeddingsrÇ   NÚpast_key_valuesrT   )r5   r=   ri   r   )r(   r)   s   "€r*   r+   rø   A  sw   ø€ ÷ 6)ñ 6)á—|‘|ð6)ñ #¡5§<¡<±·±Ð#=Õ>ð6)ñ Ÿ™ tÕ+ð	6)ñ
  �ð6)ñ 
‰u�|‰|™UŸ\™\¨DÕ0Ð0Õ	1ñ6)r-   c                óœ  € VP                   R R w  rgWgRV P                  3pV P                  f   V P                  V4      p	M/V P	                  V P                  V P                  V4      4      4      p	V	P                  V4      P                  ^^4      p	\        P                  ! W�P                  V P                  .RR7      w  r«V P                  V4      p\        P                  ! WÀP                  V P                  .RR7      w  rÞV P                  V4      P                  V^WpP                  4      pVP                  V^WpP                  4      pVw  pp\!        W¾VV4      w  r¾Ve   VP#                  WþV P$                  4      w  rþ\        P&                  ! W«3RR7      pV P)                  Wþ4      w  pp\*        P,                  ! V P.                  P0                  \2        4      pT! V VVVV3RV P4                  '       g   RMV P6                  RV P8                  /VB w  ppVP;                  WgR4      P=                  4       pV P?                  V4      pVV3# )Nr•   rÉ   ç        rÈ   r`   ) rj   r  rü   r  r  r  r  r  r›   r5   r  r   rý   r  rþ   r	  rñ   Úupdaterö   rœ   r  r   Úget_interfacers   Ú_attn_implementationr×   rÍ   rú   rÈ   r½   rÒ   r  )r7   r]   r  rÇ   r  r‹   r  r  Úquery_shapeÚq_statesÚq_passÚq_rotÚcompressed_kvÚkv_passr  Úk_passr�   rž   Úquery_statesrÓ   rÔ   Úattention_interfacerÖ   rÕ   s   &&&&&,                  r*   r@   ÚMiniCPM3Attention.forwardA  s   € ð "/×!4Ñ!4°S°bÐ!9Ñˆ
Ø!¨r°4×3CÑ3CÐDˆà×ÑÒ#Ø—{‘{ =Ó1‰Hà—}‘} T×%7Ñ%7¸¿¹ÀmÓ8TÓ%UÓVˆHØ—=‘= Ó-×7Ñ7¸¸1Ó=ˆÜŸš H×/DÑ/DÀd×F[ÑF[Ð.\ÐbdÔe‰ˆà×/Ñ/°Ó>ˆÜŸš ]×5FÑ5FÈ×H]ÑH]Ð4^ÐdfÔg‰ˆà×$Ñ$ WÓ-×2Ñ2°:¸qÀ*×N_ÑN_Ó`ˆØ—
‘
˜: q¨*×6KÑ6KÓLˆà&‰ˆˆSô ,¨E¸#¸sÓC‰ˆð Ò&Ø+×2Ñ2°6À$Ç.Á.ÓQ‰MˆFä—y’y & °bÔ9ˆà#'§>¡>°&Ó#@Ñ ˆ
�Lä(?×(MÒ(MØ�K‰K×,Ñ,Ô.Eó)
Ðñ %8ØØØØØñ	%
ð  $Ÿ}Ÿ}˜}‘C°$×2HÑ2Hð	%
ð —L‘Lð	%
ð ñ	%
Ñ!ˆ�\ð "×)Ñ)¨*À"ÓE×PÑPÓRˆØ—k‘k +Ó.ˆØ˜LÐ(Ð(r-   )rú   rs   rZ   r  r	  r  r
  rþ   rö   rû   rÎ   r  r  r  r  rü   r  r  r   rý   rÈ   rÿ   )NN)rF   rG   rH   rI   rJ   r2   r  r@   rK   rL   rM   rN   s   @@r*   ró   ró   ï   s0   ù‡ € ñ÷5]ó 5]÷n(ð (÷&6)÷ 6)ò 6)r-   ró   c                   óT   a a€ ] tR tRt oV3R lV 3R lltRV3R lR lltRtVtV ;t# )ÚMiniCPM3DecoderLayeriz  c                ó&   <€ V ^8„  d   QhRS[ RS[/# rõ   r÷   )r(   r)   s   "€r*   r+   Ú!MiniCPM3DecoderLayer.__annotate__{  s   ø€ ÷ Wñ W™~ð W¹#ñ Wr-   c                óŒ  <€ \         SV `  4        VP                  V n        \        WR 7      V n        \        V4      V n        \        VP                  VP                  R7      V n	        \        VP                  VP                  R7      V n
        VP                  \        P                  ! VP                  4      ,          V n        R# ))rs   rö   ©rS   N)r1   r2   rZ   ró   Ú	self_attnr©   ÚmlprQ   Úrms_norm_epsÚinput_layernormÚpost_attention_layernormÚscale_depthrÚ   ÚsqrtÚnum_hidden_layersÚresidual_scaler  s   &&&€r*   r2   ÚMiniCPM3DecoderLayer.__init__{  s�   ø€ Ü‰ÑÔØ!×-Ñ-ˆÔä*°&ÔNˆŒä˜vÓ&ˆŒÜ.¨v×/AÑ/AÀv×GZÑGZÔ[ˆÔÜ(7¸×8JÑ8JÐPV×PcÑPcÔ(dˆÔ%ð %×0Ñ0´4·9²9¸V×=UÑ=UÓ3VÕVˆÖr-   c                ó.  <€ V ^8„  d   QhRS[ P                  RS[ P                  R,          RS[ P                  R,          RS[R,          RS[R,          RS[S[ P                  S[ P                  3,          R,          RS[S[,          R	S[ P                  /# )
r!   r]   rÇ   Nr    r  Ú	use_cacher  r‹   rT   )r5   r=   Ú
LongTensorr   Úboolri   r   r   )r(   r)   s   "€r*   r+   r2  ˆ  sœ   ø€ ÷ ñ á—|‘|ðñ Ÿ™ tÕ+ðñ ×&Ñ&¨Õ-ð	ñ
  �ðñ ˜$•;ðñ #¡5§<¡<±·±Ð#=Õ>ÀÕEðñ Ñ+Õ,ðñ 
�‰ñr-   c                ó  € TpV P                  V4      pV P                  ! RR VRVRVRVRVRV/VB w  rW�V P                  ,          ,           pTpV P                  V4      pV P	                  V4      pW�V P                  ,          ,           pV# )r]   rÇ   r    r  r@  r  © )r8  r5  r=  r9  r6  )
r7   r]   rÇ   r    r  r@  r  r‹   Úresidualr  s
   &&&&&&&,  r*   r@   ÚMiniCPM3DecoderLayer.forwardˆ  s½   € ð !ˆØ×,Ñ,¨]Ó;ˆØŸ>š>ñ 
Ø'ð
à)ð
ð &ð
ð ,ð	
ð
  ð
ð !4ð
ð ñ
Ñˆð !°4×3FÑ3FÕ#FÕFˆà ˆØ×5Ñ5°mÓDˆØŸ™ Ó/ˆØ °4×3FÑ3FÕ#FÕFˆØÐr-   )rZ   r8  r6  r9  r=  r5  )NNNFNr¹   rN   s   @@r*   r0  r0  z  s    ù‡ € ÷Wó W÷÷ ò r-   r0  c                   óš   a a€ ] tR tRt oRtRtR.tR.tRtRt	Rt
RtRtR]R]/t]P"                  ! 4       V 3R l4       tV3R	 ltR
tVtV ;t# )ÚMiniCPM3PreTrainedModeli¦  ÚmodelTr0  r  r]   Ú
attentionsc                ó¬   <€ \         SV `  V4       \        V\        4      '       d.   \        P
                  ! VP                  VP                  4       R # R # r?   )r1   Ú_init_weightsr˜   r   ÚinitÚ	constant_r%   r3   )r7   rÃ   r8   s   &&€r*   rL  Ú%MiniCPM3PreTrainedModel._init_weights¸  s=   ø€ ä‰Ñ˜fÔ%Ü�fÔ9×:Ò:Ü�NŠN˜6×-Ñ-¨v×/HÑ/HÖIñ ;r-   c                ó&   <€ V ^8„  d   Qh/ S[ ;R&   # rr   r   )r(   r)   s   "€r*   r+   Ú$MiniCPM3PreTrainedModel.__annotate__¦  s   ø‡ ‚ áÑò r-   rD  )rF   rG   rH   rI   Úbase_model_prefixÚsupports_gradient_checkpointingÚ_no_split_modulesÚ_skip_keys_device_placementÚ_supports_flash_attnÚ_supports_sdpaÚ_supports_flex_attnÚ_can_compile_fullgraphÚ_supports_attention_backendr0  ró   Ú_can_record_outputsr5   r§   rL  Ú__annotate_func__rK   rL   rM   rN   s   @@r*   rH  rH  ¦  sy   ù‡ € ð  ÐØ&*Ð#Ø/Ð0ÐØ#4Ð"5ÐØÐØ€NØÐà!ÐØ"&ÐàÐ-ØÐ'ðÐð
 ‡]‚]ƒ_ôJó ðJ÷' … r-   rH  c                   ór   a a€ ] tR tRt oV3R lV 3R llt]]]RV3R lR ll4       4       4       tRt	Vt
V ;t# )ÚMiniCPM3Modeli¿  c                ó    <€ V ^8„  d   QhRS[ /# rr   r   )r(   r)   s   "€r*   r+   ÚMiniCPM3Model.__annotate__Á  s   ø€ ÷ ñ ™~ñ r-   c           	     ó  <€ \         SV `  V4       VP                  V n        VP                  V n        \        VP                  VP                  V P                  VP                  R 7      V n        \        P                  ! \        VP                  4       Uu. uF  p\        W4      NK  	  up4      V n        \        VP                  VP                   R7      V n        \%        VR7      V n        RV n        V P+                  4        R# u upi ))r%   r4  ©rs   FN)r1   r2   Úpad_token_idr$   Ú
vocab_sizer   rZ   Ú	scale_embÚembed_tokensr   Ú
ModuleListÚranger<  r0  ÚlayersrQ   r7  Únormrn   Ú
rotary_embÚgradient_checkpointingÚ	post_initr  s   && €r*   r2   ÚMiniCPM3Model.__init__Á  sÓ   ø€ Ü‰Ñ˜Ô Ø!×.Ñ.ˆÔØ ×+Ñ+ˆŒä7Ø×Ñ˜v×1Ñ1°4×3CÑ3CÐQW×QaÑQaô
ˆÔô —m’mÜFKÈF×LdÑLdÔFeÓfÑFe¸Ô! &Ö4ÑFeÑfó
ˆŒô $ F×$6Ñ$6¸F×<OÑ<OÔPˆŒ	Ü1¸Ô@ˆŒØ&+ˆÔ#ð 	�‰Öùò gs   ÂDc                óþ   <€ V ^8„  d   QhRS[ P                  R,          RS[ P                  R,          RS[ P                  R,          RS[R,          RS[ P                  R,          RS[R,          RS[S[,          R	S[/# )
r!   r;   NrÇ   r    r  Úinputs_embedsr@  r‹   rT   )	r5   rA  r=   r   ÚFloatTensorrB  r   r   r   )r(   r)   s   "€r*   r+   r`  Ö  s“   ø€ ÷ 2
ñ 2
á×#Ñ# dÕ*ð2
ñ Ÿ™ tÕ+ð2
ñ ×&Ñ&¨Õ-ð	2
ñ
  �ð2
ñ ×(Ñ(¨4Õ/ð2
ñ ˜$•;ð2
ñ Ñ+Õ,ð2
ñ 
!ñ2
r-   c                óˆ  € VR J VR J,          '       d   \        R4      hVf   V P                  V4      pV'       d   Vf   \        V P                  R7      pVfb   Ve   VP	                  4       M^ p\
        P                  ! VP                  ^,          VP                  R7      V,           pVP                  ^ 4      p\        V P                  VVVVR7      p	Tp
V P                  W£R7      pV P                  R V P                  P                    F  pV! V
3RV	RVRVR	VR
V/VB p
K  	  V P                  V
4      p
\        V
VR7      # )Nz:You must specify exactly one of input_ids or inputs_embedsrb  )ro   )rs   rp  rÇ   r  r    )r    rÇ   r  r    r  r@  )Úlast_hidden_stater  )Ú
ValueErrorrf  r   rs   Úget_seq_lengthr5   rŠ   rj   ro   rë   r   rk  ri  r<  rj  r   )r7   r;   rÇ   r    r  rp  r@  r‹   Úpast_seen_tokensÚcausal_maskr]   r  Údecoder_layers   &&&&&&&,     r*   r@   ÚMiniCPM3Model.forwardÓ  se  € ð ˜Ð -°tÐ";×<Ô<ÜÐYÓZÐZàÒ Ø*.×*;Ñ*;¸IÓ*FˆMç˜Ò0Ü*°$·+±+Ô>ˆOàÒØCRÒC^˜×=Ñ=Ô?ÐdeÐÜ Ÿ<š<¨×(;Ñ(;¸AÕ(>À}×G[ÑG[Ô\Ð_oÕoˆLØ'×1Ñ1°!Ó4ˆLä(Ø—;‘;Ø'Ø)Ø+Ø%ô
ˆð &ˆØ"Ÿo™o¨m˜oÓWÐà!Ÿ[™[Ð)H¨4¯;©;×+HÑ+HÓIˆMÙ)Øñà*ðð %8ðð *ð	ð
 !0ðð $ðð ñŠMñ Jð Ÿ	™	 -Ó0ˆÜ&Ø+Ø+ô
ð 	
r-   )rf  rl  ri  rj  r$   rk  rd  )NNNNNN)rF   rG   rH   rI   r2   r   r   r   r@   rK   rL   rM   rN   s   @@r*   r^  r^  ¿  s=   ù‡ € ÷ó ð$  ØØ÷2
ñ 2
ó ó ó  ÷2
ð 2
r-   r^  c                   ó„   a a€ ] tR tRt oRR/tRR/tRR.R.3/tRR/tV 3R	 lt]	]
RV3R
 lR ll4       4       tRtVtV ;t# )ÚMiniCPM3ForCausalLMi  zlm_head.weightzmodel.embed_tokens.weightÚlm_headÚcolwise_gather_outputr]   ÚlogitsÚkeep_full_weightc                óì   <€ \         SV `  V4       \        V4      V n        VP                  V n        \
        P                  ! VP                  VP                  R R7      V n        V P                  4        R# )Fr«   N)
r1   r2   r^  rI  rd  r   r®   rZ   r|  rm  rµ   s   &&€r*   r2   ÚMiniCPM3ForCausalLM.__init__  sU   ø€ Ü‰Ñ˜Ô Ü" 6Ó*ˆŒ
Ø ×+Ñ+ˆŒÜ—y’y ×!3Ñ!3°V×5FÑ5FÈUÔSˆŒð 	�‰Ör-   c                óP  <€ V ^8„  d   QhRS[ P                  R,          RS[ P                  R,          RS[ P                  R,          RS[R,          RS[ P                  R,          RS[ P                  R,          RS[R,          R	S[S[ P                  ,          R
S[S[,          RS[	/
# )r!   r;   NrÇ   r    r  rp  Úlabelsr@  Úlogits_to_keepr‹   rT   )
r5   rA  r=   r   rq  rB  r&   r   r   r   )r(   r)   s   "€r*   r+   Ú MiniCPM3ForCausalLM.__annotate__  s»   ø€ ÷ 6
ñ 6
á×#Ñ# dÕ*ð6
ñ Ÿ™ tÕ+ð6
ñ ×&Ñ&¨Õ-ð	6
ñ
  �ð6
ñ ×(Ñ(¨4Õ/ð6
ñ × Ñ  4Õ'ð6
ñ ˜$•;ð6
ñ ™eŸl™lÕ*ð6
ñ Ñ+Õ,ð6
ñ 
 ñ6
r-   c	                óÎ  € V P                   ! RRVRVRVRVRVRV/V	B p
V
P                  pW°P                  P                  ,          p\	        V\
        4      '       d   \        V) R4      MTpV P                  VRVR3,          4      pRpVe-   V P                  ! RR	VR
VRV P                  P                  /V	B p\        VVV
P                  V
P                  V
P                  R7      # )a\  
Example:

```python
>>> from transformers import AutoTokenizer, MiniCPM3ForCausalLM

>>> model = MiniCPM3ForCausalLM.from_pretrained("openbmb/MiniCPM3-4B")
>>> tokenizer = AutoTokenizer.from_pretrained("openbmb/MiniCPM3-4B")

>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")

>>> generate_ids = model.generate(inputs.input_ids, max_length=30)
>>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
"Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you."
```r;   rÇ   r    r  rp  r@  Nr–   r~  rƒ  rd  )Úlossr~  r  r]   rJ  rD  )rI  rs  rs   Úlogits_scalingr˜   r&   Úslicer|  Úloss_functionrd  r   r  r]   rJ  )r7   r;   rÇ   r    r  rp  rƒ  r@  r„  r‹   Úoutputsr]   Úslice_indicesr~  r‡  s   &&&&&&&&&,     r*   r@   ÚMiniCPM3ForCausalLM.forward  s  € ð< ,0¯:ª:ñ ,
Øð,
à)ð,
ð &ð,
ð ,ð	,
ð
 (ð,
ð  ð,
ð ñ,
ˆð  ×1Ñ1ˆà%¯©×(BÑ(BÕBˆÜ8BÀ>ÔSV×8WÒ8Wœ˜~˜o¨tÔ4Ð]kˆØ—‘˜m¨A¨}¸aÐ,?Õ@ÓAˆàˆØÒØ×%Ò%Ñp¨VÐp¸FÐpÈtÏ{É{×OeÑOeÐpÐioÑpˆDä%ØØØ#×3Ñ3Ø!×/Ñ/Ø×)Ñ)ô
ð 	
r-   )r|  rI  rd  )NNNNNNNr   )rF   rG   rH   rI   Ú_tied_weights_keysÚ_tp_planÚ_pp_planÚ
_fsdp_planr2   r   r   r@   rK   rL   rM   rN   s   @@r*   r{  r{    sf   ù‡ € à*Ð,GÐHÐØÐ2Ð3€HØ˜_Ð-°¨zÐ:Ð;€HØÐ/Ð0€Jõð Ø÷6
ñ 6
ó ó ÷6
ð 6
r-   r{  c                   ó   € ] tR tRtRtR# )Ú!MiniCPM3ForSequenceClassificationiV  rD  N)rF   rG   rH   rI   rK   rD  r-   r*   r“  r“  V  s   † Ûr-   r“  )r!  )rÙ   rÙ   )rÙ   )rH  r^  r{  r“  )ArÚ   Úcollections.abcr   r5   r   Ú r   rM  Úactivationsr   Úcache_utilsr   r   Ú
generationr	   Úintegrationsr
   Úmasking_utilsr   Úmodeling_layersr   r   Úmodeling_outputsr   r   Úmodeling_rope_utilsr   r   Úmodeling_utilsr   r   Úprocessing_utilsr   Úutilsr   r   r   Úutils.deprecationr   Úutils.genericr   r   Úutils.output_capturingr   Úconfiguration_minicpm3r   Ú	Embeddingr   rÊ   rQ   rn   r©   rÁ   r×   rÞ   rä   rè   rñ   ró   r0  rH  r^  r{  r“  Ú__all__rD  r-   r*   Ú<module>r§     sm  ðó* Ý $ã Ý å &Ý !ß .Ý )Ý 7Ý /ß [ß Oß Kß FÝ &ß IÑ IÝ 0ß GÝ 5Ý 2ôS "§,¡,ô Sñ ˜YÓ'ôJ�b—i‘ió Jó (ðJô(6<˜bŸi™iô 6<ôr�"—)‘)ô õ 	U÷%ô20òò(ñ Ð-Ó.óó /ðô2H)˜Ÿ	™	ô H)ôV)Ð5ô )ðX ôJ˜oó Jó ðJð0 ôH
Ð+ó H
ó ðH
ðV ôG
Ð1°?ó G
ó ðG
ôT	Ð(HÐJaô 	ò s‚r-   