Ë
    ÜÓ^j’|  ã                   ób  — d dl mZ d dlmZ d dlZd dlmc mZ d dlmZ ddl	m
Z ddlmZ ddlmZmZ dd	lmZ dd
lmZmZmZ ddlmZ ddlmZ ddlmZmZmZ ddlm Z m!Z! ddl"m#Z#m$Z$ ddl%m&Z&m'Z' ddl(m)Z) ddl*m+Z+m,Z,m-Z- ddl.m/Z/m0Z0m1Z1 ddl2m3Z3 ddl4m5Z5  ed«       G d„ dejl                  «      «       Z7 G d„ dejl                  «      Z8 G d„ dejl                  «      Z9 G d„ dejl                  «      Z:e G d „ d!ejl                  «      «       Z; G d"„ d#ejl                  «      Z<d$„ Z= ed%«      dHd&„«       Z>d'ej~                  d(e@d)ej~                  fd*„ZA	 dId+ejl                  d,ej~                  d-ej~                  d.ej~                  d/ej~                  dz  d0eBd1eBd2e)e+   fd3„ZCdJd4„ZDd5ej~                  d6eBd7e@d)ej~                  fd8„ZE G d9„ d:ejl                  «      ZF G d;„ d<e«      ZG G d=„ d>e'«      ZHe, G d?„ d@eH«      «       ZIe, G dA„ dBeHe«      «       ZJ G dC„ dDeeH«      ZK G dE„ dFeeH«      ZLg dG¢ZMy)Ké    )ÚCallable)ÚOptionalN)Únné   )Úinitialization)ÚACT2FN)ÚCacheÚDynamicCache)ÚGenerationMixin)Úuse_experts_implementationÚuse_kernel_forward_from_hubÚuse_kernel_func_from_hub)Úcreate_causal_mask)ÚFlashAttentionKwargs)Ú GenericForSequenceClassificationÚGenericForTokenClassificationÚ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is_flash_attention_requestedÚmaybe_autocastÚmerge_with_config_defaults)Úcapture_outputsé   )ÚMistral4ConfigÚRMSNormc                   óh   ‡ — e Zd Zddeddfˆ fd„Zdej                  dej                  fd„Zd„ Zˆ xZ	S )	ÚMistral4RMSNormÚepsÚreturnNc                 óŠ   •— t         ‰| �  «        t        j                  t	        j
                  |«      «      | _        || _        y)z>
        Mistral4RMSNorm is equivalent to T5LayerNorm
        N)ÚsuperÚ__init__r   Ú	ParameterÚtorchÚonesÚweightÚvariance_epsilon)ÚselfÚhidden_sizer'   Ú	__class__s      €úy/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/mistral4/modeling_mistral4.pyr+   zMistral4RMSNorm.__init__3   s1   ø€ ô 	‰ÑÔÜ—l‘l¤5§:¡:¨kÓ#:Ó;ˆŒØ #ˆÕó    Úhidden_statesc                 ó"  — |j                   }|j                  t        j                  «      }|j	                  d«      j                  dd¬«      }|t        j                  || j                  z   «      z  }| j                  |j                  |«      z  S )Né   éÿÿÿÿT)Úkeepdim)	ÚdtypeÚtor-   Úfloat32ÚpowÚmeanÚrsqrtr0   r/   )r1   r6   Úinput_dtypeÚvariances       r4   ÚforwardzMistral4RMSNorm.forward;   sy   € Ø#×)Ñ)ˆØ%×(Ñ(¬¯©Ó7ˆØ ×$Ñ$ QÓ'×,Ñ,¨R¸Ð,Ó>ˆØ%¬¯©°H¸t×?TÑ?TÑ4TÓ(UÑUˆØ�{‰{˜]×-Ñ-¨kÓ:Ñ:Ð:r5   c                 ó^   — t        | j                  j                  «      › d| j                  › �S )Nz, eps=)Útupler/   Úshaper0   )r1   s    r4   Ú
extra_reprzMistral4RMSNorm.extra_reprB   s*   € Ü˜Ÿ™×)Ñ)Ó*Ð+¨6°$×2GÑ2GÐ1HÐIÐIr5   )g�íµ ÷Æ°>)
Ú__name__Ú
__module__Ú__qualname__Úfloatr+   r-   ÚTensorrC   rG   Ú__classcell__©r3   s   @r4   r&   r&   1   s7   ø„ ñ$¨ð $¸$õ $ð; U§\¡\ð ;°e·l±ló ;öJr5   r&   c                   óÄ   ‡ — e Zd ZU ej                  ed<   ddefˆ fd„Ze	 	 	 ddedz  de	d   de
dz  ded	ef   fd
„«       Z ej                  «       ed„ «       «       Zˆ xZS )ÚMistral4RotaryEmbeddingÚinv_freqNÚconfigc                 ó®  •— t         ‰| �  «        |j                  | _        |j                  | _        || _        | j
                  j                  d   | _        | j                  }| j                  dk7  rt        | j                     } || j
                  |«      \  }| _
        | j                  d|d¬«       | j                  d|j                  «       d¬«       y )NÚ	rope_typeÚdefaultrQ   F)Ú
persistentÚoriginal_inv_freq)r*   r+   Úmax_position_embeddingsÚmax_seq_len_cachedÚoriginal_max_seq_lenrR   Úrope_parametersrT   Úcompute_default_rope_parametersr   Úattention_scalingÚregister_bufferÚclone)r1   rR   ÚdeviceÚrope_init_fnrQ   r3   s        €r4   r+   z Mistral4RotaryEmbedding.__init__I   s¶   ø€ Ü‰ÑÔØ"(×"@Ñ"@ˆÔØ$*×$BÑ$BˆÔ!àˆŒàŸ™×4Ñ4°[ÑAˆŒØ!%×!EÑ!EˆØ�>‰>˜YÒ&Ü.¨t¯~©~Ñ>ˆLÙ+7¸¿¹ÀVÓ+LÑ(ˆ�$Ô(à×Ñ˜Z¨¸eÐÔDØ×ÑÐ0°(·.±.Ó2BÈuÐÕUr5   r`   ztorch.deviceÚseq_lenr(   ztorch.Tensorc                 ó  — | j                   d   }t        | dd«      xs | j                  | j                  z  }d}d|t	        j
                  d|dt        j                  ¬«      j                  |t        j                  ¬«      |z  z  z  }||fS )	a¨  
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        Ú
rope_thetaÚhead_dimNg      ð?r   r8   ©r;   )r`   r;   )	r[   Úgetattrr2   Únum_attention_headsr-   ÚarangeÚint64r<   rK   )rR   r`   rb   ÚbaseÚdimÚattention_factorrQ   s          r4   r\   z7Mistral4RotaryEmbedding.compute_default_rope_parametersY   s�   € ð& ×%Ñ% lÑ3ˆÜ�f˜j¨$Ó/Òc°6×3EÑ3EÈ×IcÑIcÑ3cˆàÐð Ø”U—\‘\ ! S¨!´5·;±;Ô?×BÑBÈ&ÔX]×XcÑXcÐBÓdÐgjÑjÑkñ
ˆð Ð)Ð)Ð)r5   c                 óN  — | j                   d d d …d f   j                  «       j                  |j                  d   dd«      j	                  |j
                  «      }|d d …d d d …f   j                  «       }t        |j
                  j                  t        «      r/|j
                  j                  dk7  r|j
                  j                  nd}t        |d¬«      5  |j                  «       |j                  «       z  j                  dd«      }t        j                  ||fd¬	«      }|j                  «       | j                  z  }|j                  «       | j                  z  }	d d d «       j	                  |j                   ¬
«      	j	                  |j                   ¬
«      fS # 1 sw Y   ŒAxY w)Nr   r9   r"   ÚmpsÚcpuF)Údevice_typeÚenabledr8   ©rl   rf   )rQ   rK   ÚexpandrF   r<   r`   Ú
isinstanceÚtypeÚstrr   Ú	transposer-   ÚcatÚcosr]   Úsinr;   )
r1   ÚxÚposition_idsÚinv_freq_expandedÚposition_ids_expandedrq   ÚfreqsÚembrz   r{   s
             r4   rC   zMistral4RotaryEmbedding.forwardw   sR  € ð !ŸM™M¨$²°4¨-Ñ8×>Ñ>Ó@×GÑGÈ×HZÑHZÐ[\ÑH]Ð_aÐcdÓe×hÑhÐij×iqÑiqÓrÐØ ,ªQ°²a¨ZÑ 8× >Ñ >Ó @Ðä'1°!·(±(·-±-ÄÔ'EÈ!Ï(É(Ï-É-Ð[`ÒJ`�a—h‘h—m’mÐfkˆÜ¨¸UÔCñ 	5Ø&×,Ñ,Ó.Ð1F×1LÑ1LÓ1NÑN×YÑYÐZ[Ð]^Ó_ˆEÜ—)‘)˜U E˜N°Ô3ˆCØ—'‘'“)˜d×4Ñ4Ñ4ˆCØ—'‘'“)˜d×4Ñ4Ñ4ˆC÷		5ð �v‰v˜AŸG™GˆvÓ$ c§f¡f°1·7±7 fÓ&;Ð;Ð;÷	5ð 	5ús   ÃBFÆF$©N)NNN)rH   rI   rJ   r-   rL   Ú__annotations__r#   r+   Ústaticmethodr   ÚintrE   rK   r\   Úno_gradr   rC   rM   rN   s   @r4   rP   rP   F   s˜   ø… Ø�l‰lÓñV˜~õ Vð  à(,Ø+/Ø"ñ*Ø Ñ%ð*à˜Ñ(ð*ð �t‘ð*ð 
ˆ~˜uÐ$Ñ	%ò	*ó ð*ð: €U‡]�]ƒ_Øñ<ó ó ô<r5   rP   c                   ó&   ‡ — e Zd Zdˆ fd„	Zd„ Zˆ xZS )ÚMistral4MLPc                 óÒ  •— t         ‰| �  «        || _        |j                  | _        |€|j                  n|| _        t        j                  | j                  | j                  d¬«      | _        t        j                  | j                  | j                  d¬«      | _        t        j                  | j                  | j                  d¬«      | _	        t        |j                     | _        y ©NF©Úbias)r*   r+   rR   r2   Úintermediate_sizer   ÚLinearÚ	gate_projÚup_projÚ	down_projr   Ú
hidden_actÚact_fn)r1   rR   r�   r3   s      €r4   r+   zMistral4MLP.__init__ˆ   s®   ø€ Ü‰ÑÔØˆŒØ!×-Ñ-ˆÔØ=NÐ=V ×!9Ò!9Ð\mˆÔÜŸ™ 4×#3Ñ#3°T×5KÑ5KÐRWÔXˆŒÜ—y‘y ×!1Ñ!1°4×3IÑ3IÐPUÔVˆŒÜŸ™ 4×#9Ñ#9¸4×;KÑ;KÐRWÔXˆŒÜ˜V×.Ñ.Ñ/ˆ�r5   c                 óˆ   — | j                  | j                  | j                  |«      «      | j                  |«      z  «      }|S r‚   )r‘   r“   r�   r�   )r1   r|   r‘   s      r4   rC   zMistral4MLP.forward’   s6   € Ø—N‘N 4§;¡;¨t¯~©~¸aÓ/@Ó#AÀDÇLÁLÐQRÃOÑ#SÓTˆ	ØÐr5   r‚   ©rH   rI   rJ   r+   rC   rM   rN   s   @r4   rˆ   rˆ   ‡   s   ø„ õ0ör5   rˆ   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚMistral4TopkRouterc                 ó”  •— t         ‰| �  «        |j                  | _        |j                  | _        |j                  | _        t        j                  t        j                  | j
                  | j                  «      «      | _        |j                  | _        |j                  | _        |j                   | _        |j"                  | _        y r‚   )r*   r+   Únum_experts_per_tokÚtop_kÚnum_local_expertsÚnum_expertsr2   Ú
hidden_dimr   r,   r-   Úzerosr/   Úrouted_scaling_factorÚn_groupÚ	num_groupÚ
topk_groupÚnorm_topk_prob©r1   rR   r3   s     €r4   r+   zMistral4TopkRouter.__init__˜   s�   ø€ Ü‰ÑÔØ×/Ñ/ˆŒ
Ø!×3Ñ3ˆÔØ ×,Ñ,ˆŒÜ—l‘l¤5§;¡;¨t×/?Ñ/?ÀÇÁÓ#QÓRˆŒØ%+×%AÑ%AˆÔ"ØŸ™ˆŒØ ×+Ñ+ˆŒØ$×3Ñ3ˆÕr5   c                 óÊ  — |j                  d| j                  «      }t        j                  || j                  «      }|j                  d«      }|j                  d| j                  | j                  | j                  z  «      j                  dd¬«      d   j                  d¬«      }t        j                  || j                  dd¬«      d   }t        j                  |«      }|j                  d|d«       |j                  d«      j                  d| j                  | j                  | j                  z  «      j!                  d| j                  «      }|j#                  |j%                  «        d«      }t        j                  || j&                  dd¬«      d   }	|j)                  d|	«      }
| j*                  r|
j                  dd	¬
«      dz   }|
|z  }
|
| j,                  z  }
||
|	fS )Nr9   r8   rs   r   F)Úkrl   Úsortedr"   ç        T)rl   r:   g#B’¡œÇ;)Úviewr�   ÚFÚlinearr/   Úsoftmaxr¡   rœ   ÚtopkÚsumr-   r¢   Ú
zeros_likeÚscatter_Ú	unsqueezert   ÚreshapeÚmasked_fillÚboolrš   Úgatherr£   rŸ   )r1   r6   Úrouter_logitsÚscoresÚgroup_scoresÚ	group_idxÚ
group_maskÚ
score_maskÚscores_for_choiceÚtopk_indicesÚtopk_weightsÚdenominators               r4   rC   zMistral4TopkRouter.forward£   s¬  € Ø%×*Ñ*¨2¨t¯©Ó?ˆÜŸ™ °·±Ó<ˆØ×&Ñ& rÓ*ˆà�K‰K˜˜DŸN™N¨D×,<Ñ,<ÀÇÁÑ,NÓO×TÑTÐUVÐ\^ÐTÓ_Ð`aÑb×fÑfÐkmÐfÓnð 	ô —J‘J˜|¨t¯©ÀBÈuÔUÐVWÑXˆ	Ü×%Ñ% lÓ3ˆ
Ø×Ñ˜A˜y¨!Ô,à× Ñ  Ó$ß‰V�B˜Ÿ™¨×(8Ñ(8¸D¿N¹NÑ(JÓKß‰W�R˜×)Ñ)Ó*ð 	ð
 #×.Ñ.°
·±Ó0AÐ/AÀ3ÓGÐÜ—z‘zÐ"3°t·z±zÀrÐRWÔXÐYZÑ[ˆØ—}‘} Q¨Ó5ˆØ×ÒØ&×*Ñ*¨r¸4Ð*Ó@À5ÑHˆKØ˜KÑ'ˆLØ# d×&@Ñ&@Ñ@ˆØ˜l¨LÐ8Ð8r5   r•   rN   s   @r4   r—   r—   —   s   ø„ ô	4ö9r5   r—   c                   óŠ   ‡ — e Zd ZdZˆ fd„Zdej                  dej                  dej                  dej                  fd„Zˆ xZS )ÚMistral4Expertsz2Collection of expert weights stored as 3D tensors.c                 óò  •— t         ‰| �  «        |j                  | _        |j                  | _        |j                  | _        t        j                  t        j                  | j                  d| j                  z  | j
                  «      «      | _        t        j                  t        j                  | j                  | j
                  | j                  «      «      | _        t        |j                     | _        y )Nr8   )r*   r+   r›   rœ   r2   r�   Úmoe_intermediate_sizeÚintermediate_dimr   r,   r-   ÚemptyÚgate_up_projr‘   r   r’   r“   r¤   s     €r4   r+   zMistral4Experts.__init__À   s­   ø€ Ü‰ÑÔØ!×3Ñ3ˆÔØ ×,Ñ,ˆŒØ &× <Ñ <ˆÔÜŸL™L¬¯©°T×5EÑ5EÀqÈ4×K`ÑK`ÑG`Ðbf×bqÑbqÓ)rÓsˆÔÜŸ™¤e§k¡k°$×2BÑ2BÀDÇOÁOÐUY×UjÑUjÓ&kÓlˆŒÜ˜V×.Ñ.Ñ/ˆ�r5   r6   Útop_k_indexÚtop_k_weightsr(   c                 óf  — t        j                  |«      }t        j                  «       5  t         j                  j                  j                  || j                  ¬«      }|j                  ddd«      }t        j                  |j                  d¬«      d«      j                  «       }d d d «       D ]ð  }|d   }|| j                  k(  rŒt        j                  |   «      \  }}	||	   }
t        j                  j                  |
| j                  |   «      j                  dd¬«      \  }}| j                  |«      |z  }t        j                  j                  || j                   |   «      }|||	|d f   z  }|j#                  d|	|j%                  |j&                  «      «       Œò |S # 1 sw Y   �ŒxY w)N)Únum_classesr8   r"   r   )r9   éþÿÿÿrs   r9   )r-   r¯   r†   r   Ú
functionalÚone_hotrœ   ÚpermuteÚgreaterr®   ÚnonzeroÚwherer«   rÆ   Úchunkr“   r‘   Ú
index_add_r<   r;   )r1   r6   rÇ   rÈ   Úfinal_hidden_statesÚexpert_maskÚ
expert_hitÚ
expert_idxÚ	top_k_posÚ	token_idxÚcurrent_stateÚgateÚupÚcurrent_hidden_statess                 r4   rC   zMistral4Experts.forwardÉ   s™  € ô $×.Ñ.¨}Ó=ÐÜ�]‰]‹_ñ 	SÜŸ(™(×-Ñ-×5Ñ5°kÈt×O_ÑO_Ð5Ó`ˆKØ%×-Ñ-¨a°°AÓ6ˆKÜŸ™ {§¡¸8 Ó'DÀaÓH×PÑPÓRˆJ÷	Sð
 %ò 
	nˆJØ# A™ˆJØ˜T×-Ñ-Ò-ØÜ#(§;¡;¨{¸:Ñ/FÓ#GÑ ˆI�yØ)¨)Ñ4ˆMÜ—}‘}×+Ñ+¨M¸4×;LÑ;LÈZÑ;XÓY×_Ñ_Ð`aÐgiÐ_Ój‰HˆD�"Ø$(§K¡K°Ó$5¸Ñ$:Ð!Ü$&§M¡M×$8Ñ$8Ð9NÐPT×P^ÑP^Ð_iÑPjÓ$kÐ!Ø$9¸MÈ)ÐU^Ð`dÐJdÑ<eÑ$eÐ!Ø×*Ñ*¨1¨iÐ9N×9QÑ9QÐRe×RkÑRkÓ9lÕmð
	nð #Ð"÷#	Sñ 	Sús   ªA=F&Æ&F0)	rH   rI   rJ   Ú__doc__r+   r-   rL   rC   rM   rN   s   @r4   rÁ   rÁ   ¼   sF   ø„ á<ô0ð#à—|‘|ð#ð —\‘\ð#ð —|‘|ð	#ð
 
�‰÷#r5   rÁ   c                   ó`   ‡ — e Zd ZdZdefˆ fd„Zdej                  dej                  fd„Zˆ xZ	S )ÚMistral4MoEz:
    A mixed expert module containing shared experts.
    rR   c                 óÂ   •— t         ‰| �  «        || _        t        |«      | _        t        |«      | _        t        ||j                  |j                  z  ¬«      | _
        y )N)rR   r�   )r*   r+   rR   rÁ   Úexpertsr—   rÛ   rˆ   rÃ   Ún_shared_expertsÚshared_expertsr¤   s     €r4   r+   zMistral4MoE.__init__é   sP   ø€ Ü‰ÑÔØˆŒÜ& vÓ.ˆŒÜ& vÓ.ˆŒ	Ü)Ø¨V×-IÑ-IÈF×LcÑLcÑ-cô
ˆÕr5   r6   r(   c                 óò   — |}|j                   }| j                  |«      \  }}}|j                  d|j                   d   «      } | j                  |||«      j                  |Ž }|| j	                  |«      z   }|S )Nr9   )rF   rÛ   r©   râ   rä   )r1   r6   Ú	residualsÚ
orig_shapeÚ_r¾   r½   s          r4   rC   zMistral4MoE.forwardò   s   € Ø!ˆ	Ø"×(Ñ(ˆ
Ø(,¯	©	°-Ó(@Ñ%ˆˆ<˜Ø%×*Ñ*¨2¨}×/BÑ/BÀ2Ñ/FÓGˆØT˜Ÿ™ ]°LÀ,ÓO×TÑTÐV`ÐaˆØ%¨×(;Ñ(;¸IÓ(FÑFˆØÐr5   )
rH   rI   rJ   rÞ   r#   r+   r-   rL   rC   rM   rN   s   @r4   rà   rà   ä   s/   ø„ ñð
˜~õ 
ð U§\¡\ð °e·l±l÷ r5   rà   c                 óš   — | dd| j                   d   dz  …f   }| d| j                   d   dz  d…f   }t        j                  | |fd¬«      S )z*Rotates half the hidden dims of the input..Nr9   r8   rs   )rF   r-   ry   )r|   Úx1Úx2s      r4   Úrotate_halfrì   ü   sZ   € à	
ˆ3Ð"�!—'‘'˜"‘+ Ñ"Ð"Ð"Ñ	#€BØ	
ˆ3�—‘˜‘˜qÑ Ñ"Ð"Ñ	#€BÜ�9‰9�r�c˜2�Y BÔ'Ð'r5   Úrotary_pos_embc                 óž   — |j                  |«      }|j                  |«      }| |z  t        | «      |z  z   }||z  t        |«      |z  z   }||fS )a…  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.
    )r±   rì   )Úqr¦   rz   r{   Úunsqueeze_dimÚq_embedÚk_embeds          r4   Úapply_rotary_pos_embró     sY   € ð& �-‰-˜Ó
&€CØ
�-‰-˜Ó
&€CØ�3‰wœ; q›>¨CÑ/Ñ0€GØ�3‰wœ; q›>¨CÑ/Ñ0€GØ�GÐÐr5   r6   Ún_repr(   c                 óª   — | j                   \  }}}}|dk(  r| S | dd…dd…ddd…dd…f   j                  |||||«      } | j                  |||z  ||«      S )zÔ
    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"   N)rF   rt   r²   )r6   rô   ÚbatchÚnum_key_value_headsÚslenre   s         r4   Ú	repeat_kvrù     so   € ð
 2?×1DÑ1DÑ.€EÐ  hØ�‚zØÐØ!¢!¢Q¨ªa²Ð"2Ñ3×:Ñ:¸5ÐBUÐW\Ð^bÐdlÓm€MØ× Ñ  Ð(;¸eÑ(CÀTÈ8ÓTÐTr5   ÚmoduleÚqueryÚkeyÚvalueÚattention_maskÚscalingÚdropoutÚkwargsc                 ó  — t        || j                  «      }t        || j                  «      }	t        j                  ||j	                  dd«      «      |z  }
|�|
|z   }
t
        j                  j                  |
dt        j                  ¬«      j                  |j                  «      }
t
        j                  j                  |
|| j                  ¬«      }
t        j                  |
|	«      }|j	                  dd«      j                  «       }||
fS )Nr8   r   r9   )rl   r;   )ÚpÚtrainingr"   )rù   Únum_key_value_groupsr-   Úmatmulrx   r   rÌ   r¬   r=   r<   r;   r   r  Ú
contiguous)rú   rû   rü   rý   rþ   rÿ   r   r  Ú
key_statesÚvalue_statesÚattn_weightsÚattn_outputs               r4   Úeager_attention_forwardr  )  sà   € ô ˜3 × ;Ñ ;Ó<€JÜ˜U F×$?Ñ$?Ó@€Lä—<‘<  z×';Ñ';¸A¸qÓ'AÓBÀWÑL€LØÐ!Ø# nÑ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$Ð$r5   c                 ó¦  — |dd|j                   d   dz  …f   j                  |«      }|dd|j                   d   dz  …f   j                  |«      }| dddd…f   | dddd…f   }}|dddd…f   |dddd…f   }	}t        j                  ||z  ||z  z
  ||z  ||z  z   gd¬«      }
t        j                  ||z  |	|z  z
  |	|z  ||z  z   gd¬«      }|
|fS )aû  
    Applies interleaved Rotary Position Embedding to the query and key tensors.

    DeepSeek lays the rotary dimensions out in interleaved pairs `(x0, x1), (x2, x3), ...`, each rotated by a
    single frequency. We compute that rotation directly on the even/odd slices instead of de-interleaving with a
    `view`/`transpose`/`reshape`; the output is bit-identical to the de-interleaved `rotate_half` formulation while
    avoiding the extra contiguous copy.

    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.
        position_ids (`torch.Tensor`):
            The position indices of the tokens corresponding to the query and key tensors. For example, this can be
            used to pass offsetted position ids when working with a KV-cache.
        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.
    .Nr9   r8   r   r"   rs   )rF   r±   r-   ry   )rï   r¦   rz   r{   r}   rð   Úq1Úq2Úk1Úk2rñ   rò   s               r4   Úapply_rotary_pos_emb_interleaver  B  s  € ð8 ˆcÐ'�S—Y‘Y˜r‘] aÑ'Ð'Ð'Ñ
(×
2Ñ
2°=Ó
A€CØ
ˆcÐ'�S—Y‘Y˜r‘] aÑ'Ð'Ð'Ñ
(×
2Ñ
2°=Ó
A€Càˆs�A�D�q�Dˆy‰\˜1˜S ! $ Q $˜Y™<ˆ€BØˆs�A�D�q�Dˆy‰\˜1˜S ! $ Q $˜Y™<ˆ€Bä�i‰i˜˜c™ B¨¡HÑ,¨b°3©h¸¸c¹Ñ.AÐBÈÔK€GÜ�i‰i˜˜c™ B¨¡HÑ,¨b°3©h¸¸c¹Ñ.AÐBÈÔK€GØ�GÐÐr5   Úpositions_idsÚbetarX   c           	      ó„   — d|t        j                  dt        j                  | |z  «      z   «      z  z   }|d d …d d d …d f   S ©Nr"   )r-   ÚlogÚfloor)r  r  rX   rÿ   s       r4   Úget_llama_4_attn_scaler  i  sB   € Ø�$œŸ™ 1¤u§{¡{°=ÐCZÑ3ZÓ'[Ñ#[Ó\Ñ\Ñ\€GØ’1�dšA˜tÐ#Ñ$Ð$r5   c                   ó.  ‡ — e Zd ZdZdedefˆ fd„Z	 ddej                  de	ej                  ej                  f   dej                  dz  d	ej                  d
e
dz  dee   de	ej                  ej                  dz  e	ej                     dz  f   fd„Zˆ xZS )ÚMistral4Attentionz=Multi-headed attention from 'Attention Is All You Need' paperrR   Ú	layer_idxc                 óf  •— t         ‰| �  «        || _        || _        |j                  |j
                  z  | _        |j                  | _        |j                  | _        |j                  | _	        |j                  | _
        |j                  | _        |j                  | _        |j                  | _        |j                  | _        d| _        | j                  €?t!        j"                  |j$                  | j                  | j                  z  d¬«      | _        n“t!        j"                  |j$                  |j                  |j(                  ¬«      | _        t-        |j                  «      | _        t!        j"                  |j                  | j                  | j                  z  d¬«      | _        t!        j"                  |j$                  | j                  | j                  z   |j(                  ¬«      | _        t-        | j                  «      | _        t!        j"                  | j                  | j                  | j                  | j                  z   z  d¬«      | _        t!        j"                  | j                  | j                  z  |j$                  |j(                  ¬«      | _        | j                  dz  | _        y )NTFr‹   g      à¿)r*   r+   rR   r  rh   r÷   r  Úattention_dropoutÚ	num_headsÚq_lora_rankÚqk_rope_head_dimÚkv_lora_rankÚ
v_head_dimÚqk_nope_head_dimÚqk_head_dimÚ	is_causalr   rŽ   r2   Úq_projÚattention_biasÚq_a_projr&   Úq_a_layernormÚq_b_projÚkv_a_proj_with_mqaÚkv_a_layernormÚ	kv_b_projÚo_projrÿ   ©r1   rR   r  r3   s      €r4   r+   zMistral4Attention.__init__q  s  ø€ Ü‰ÑÔØˆŒØ"ˆŒØ$*×$>Ñ$>À&×B\ÑB\Ñ$\ˆÔ!Ø!'×!9Ñ!9ˆ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!×-Ñ-ˆÔàˆŒØ×ÑÐ#ÜŸ)™) F×$6Ñ$6¸¿¹È×IYÑIYÑ8YÐ`eÔfˆD�KäŸI™I f×&8Ñ&8¸&×:LÑ:LÐSY×ShÑShÔiˆDŒMÜ!0°×1CÑ1CÓ!DˆDÔÜŸI™I f×&8Ñ&8¸$¿.¹.È4×K[ÑK[Ñ:[ÐbgÔhˆDŒMä"$§)¡)Ø×ÑØ×Ñ × 5Ñ 5Ñ5Ø×&Ñ&ô#
ˆÔô
 .¨d×.?Ñ.?Ó@ˆÔÜŸ™Ø×ÑØ�N‰N˜d×3Ñ3°d·o±oÑEÑFØô
ˆŒô —i‘iØ�N‰N˜TŸ_™_Ñ,Ø×ÑØ×&Ñ&ô
ˆŒð ×'Ñ'¨DÑ1ˆ�r5   Nr6   Úposition_embeddingsrþ   r}   Úpast_key_valuesr  r(   c                 ó   — |j                   d d \  }}||d| j                  f}	||d| j                  | j                  z   f}
| j                  €| j                  |«      }n/| j                  | j                  | j                  |«      «      «      }|j                  |	«      j                  dd«      }t        j                  || j                  | j                  gd¬«      \  }}| j                  |«      }t        j                  || j                  | j                  gd¬«      \  }}| j!                  | j#                  |«      «      j                  |
«      j                  dd«      }t        j                  || j                  | j                  gd¬«      \  }}|j                  |d|| j                  «      }|\  }}| j$                  j&                  rt)        ||||«      \  }}nt+        ||||«      \  }} |j,                  g |j                   d d ¢d‘­Ž }t        j.                  ||fd¬«      }t        j.                  ||fd¬«      }|t1        || j$                  j2                  j5                  d«      | j$                  j2                  j5                  d«      «      j7                  |j8                  «      z  }|� |j;                  ||| j<                  «      \  }}t?        | j$                  «      rH| j                  | j                  k7  r/tA        jB                  |d| j                  | j                  z
  g«      }tE        jF                  | j$                  jH                  tJ        «      } || ||||f| jL                  sdn| jN                  | jP                  d	œ|¤Ž\  }}t?        | j$                  «      r4| j                  | j                  k7  r|d d …d d …d d …d | j                  …f   }|jS                  ||d«      jU                  «       }| jW                  |«      }||fS )
Nr9   r"   r8   rs   Úllama_4_scaling_betaÚ original_max_position_embeddingsr   r¨   )r   rÿ   ),rF   r%  r$  r#  r   r'  r+  r*  r)  r©   rx   r-   Úsplitr!  r,  r"  r.  r-  rR   Úrope_interleaver  ró   rt   ry   r  r[   Úgetr<   r;   Úupdater  r   rª   Úpadr   Úget_interfaceÚ_attn_implementationr  r  r  rÿ   r²   r  r/  )r1   r6   r1  rþ   r}   r2  r  Ú
batch_sizeÚ
seq_lengthÚquery_shapeÚ	key_shapeÚq_statesÚq_passÚq_rotÚcompressed_kvÚk_passÚk_rotr	  rz   r{   Úquery_statesr  Úattention_interfacer  r
  s                            r4   rC   zMistral4Attention.forwardœ  sž  € ð "/×!4Ñ!4°S°bÐ!9Ñˆ
�JØ! :¨r°4×3CÑ3CÐDˆØ ¨R°×1FÑ1FÈÏÉÑ1XÐYˆ	à×ÑÐ#Ø—{‘{ =Ó1‰Hà—}‘} T×%7Ñ%7¸¿¹ÀmÓ8TÓ%UÓVˆHØ—=‘= Ó-×7Ñ7¸¸1Ó=ˆÜŸ™ H¨t×/DÑ/DÀd×F[ÑF[Ð.\ÐbdÔe‰ˆ�à×/Ñ/°Ó>ˆÜŸ™ M°D×4EÑ4EÀt×G\ÑG\Ð3]ÐceÔf‰ˆ�à—‘ × 3Ñ 3°FÓ ;Ó<×AÑAÀ)ÓL×VÑVÐWXÐZ[Ó\ˆÜ$Ÿ{™{¨6°D×4IÑ4IÈ4Ï?É?Ð3[ÐacÔdÑˆ�à—
‘
˜: q¨*°d×6KÑ6KÓLˆà&‰ˆˆSØ�;‰;×&Ò&Ü:¸5À%ÈÈcÓR‰LˆE‘5ä/°°u¸cÀ3ÓG‰LˆE�5Ø�—‘Ð4˜fŸl™l¨3¨BÐ/Ð4°Ò4ˆä—y‘y &¨% °bÔ9ˆÜ—Y‘Y ¨˜°BÔ7ˆ
à#Ô&<ØØ�K‰K×'Ñ'×+Ñ+Ð,BÓCØ�K‰K×'Ñ'×+Ñ+Ð,NÓOó'
÷ ‰"ˆ\×ÑÓ
 ñ	!ˆð Ð&Ø'6×'=Ñ'=¸jÈ,ÐX\×XfÑXfÓ'gÑ$ˆJ˜ä'¨¯©Ô4¸×9IÑ9IÈTÏ_É_Ò9\ÜŸ5™5 °°4×3CÑ3CÀdÇoÁoÑ3UÐ/VÓWˆLä(?×(MÑ(MØ�K‰K×,Ñ,Ô.Eó)
Ðñ %8ØØØØØð	%
ð  $Ÿ}š}‘C°$×2HÑ2HØ—L‘Lñ	%
ð ñ	%
Ñ!ˆ�\ô (¨¯©Ô4¸×9IÑ9IÈTÏ_É_Ò9\Ø%¢aªªAÐ/@°·±Ð/@Ð&@ÑAˆKà!×)Ñ)¨*°jÀ"ÓE×PÑPÓRˆØ—k‘k +Ó.ˆØ˜LÐ(Ð(r5   r‚   )rH   rI   rJ   rÞ   r#   r…   r+   r-   rL   rE   r	   r   r   rC   rM   rN   s   @r4   r  r  n  sÄ   ø„ ÙGð)2˜~ð )2¸#õ )2ðb )-ñF)à—|‘|ðF)ð # 5§<¡<°·±Ð#=Ñ>ðF)ð Ÿ™ tÑ+ð	F)ð
 —l‘lðF)ð  ™ðF)ð Ð-Ñ.ðF)ð 
ˆu�|‰|˜UŸ\™\¨DÑ0°%¸¿¹Ñ2EÈÑ2LÐLÑ	M÷F)r5   r  c                   ó  ‡ — e Zd Zdedefˆ fd„Z	 	 	 	 	 ddej                  dej                  dz  dej                  dz  de	dz  d	e
dz  d
eej                  ej                  f   dz  dee   dej                  fd„Zˆ xZS )ÚMistral4DecoderLayerrR   r  c                 ó`  •— t         ‰| �  «        |j                  | _        t        ||¬«      | _        ||j
                  k\  rt        |«      | _        nt        |«      | _        t        |j                  |j                  ¬«      | _        t        |j                  |j                  ¬«      | _        y )N)rR   r  ©r'   )r*   r+   r2   r  Ú	self_attnÚfirst_k_dense_replacerà   Úmlprˆ   r&   Úrms_norm_epsÚinput_layernormÚpost_attention_layernormr0  s      €r4   r+   zMistral4DecoderLayer.__init__æ  s‡   ø€ Ü‰ÑÔØ!×-Ñ-ˆÔä*°&ÀIÔNˆŒà˜×4Ñ4Ò4Ü" 6Ó*ˆD�Hä" 6Ó*ˆDŒHä.¨v×/AÑ/AÀv×GZÑGZÔ[ˆÔÜ(7¸×8JÑ8JÐPV×PcÑPcÔ(dˆÕ%r5   Nr6   rþ   r}   r2  Ú	use_cacher1  r  r(   c           
      óÀ   — |}| j                  |«      } | j                  d||||||dœ|¤Ž\  }}	||z   }|}| j                  |«      }| j                  |«      }||z   }|S )N)r6   rþ   r}   r2  rS  r1  © )rQ  rM  rR  rO  )
r1   r6   rþ   r}   r2  rS  r1  r  Úresidualrè   s
             r4   rC   zMistral4DecoderLayer.forwardô  s�   € ð !ˆØ×,Ñ,¨]Ó;ˆà)˜4Ÿ>™>ð 
Ø'Ø)Ø%Ø+ØØ 3ñ
ð ñ
Ñˆ�qð ! =Ñ0ˆð !ˆØ×5Ñ5°mÓDˆØŸ™ Ó/ˆØ  =Ñ0ˆØÐr5   )NNNFN)rH   rI   rJ   r#   r…   r+   r-   rL   Ú
LongTensorr	   r´   rE   r   r   rC   rM   rN   s   @r4   rJ  rJ  å  s¾   ø„ ðe˜~ð e¸#õ eð" /3Ø04Ø(,Ø!&ØHLñà—|‘|ðð Ÿ™ tÑ+ðð ×&Ñ&¨Ñ-ð	ð
  ™ðð ˜$‘;ðð # 5§<¡<°·±Ð#=Ñ>ÀÑEðð Ð+Ñ,ðð 
�‰÷r5   rJ  c                   óŒ   ‡ — e Zd ZU eed<   dZdZdgZdgZdZ	dZ
dZdZdZeedœZg Zg Z ej(                  «       ˆ fd„«       Zˆ xZS )ÚMistral4PreTrainedModelrR   ÚmodelTrJ  r2  )r6   Ú
attentionsc                 ó¬  •— t         ‰| �  |«       t        |t        «      r7t	        j
                  |j                  d| j                  j                  ¬«       y t        |t        «      rmt	        j
                  |j                  d| j                  j                  ¬«       t	        j
                  |j                  d| j                  j                  ¬«       y y )Nr¨   )r?   Ústd)r*   Ú_init_weightsru   r—   ÚinitÚnormal_r/   rR   Úinitializer_rangerÁ   rÆ   r‘   )r1   rú   r3   s     €r4   r^  z%Mistral4PreTrainedModel._init_weights'  sˆ   ø€ ä‰Ñ˜fÔ%Ü�fÔ0Ô1Ü�L‰L˜Ÿ™¨S°d·k±k×6SÑ6SÖTÜ˜¤Ô0Ü�L‰L˜×,Ñ,°3¸D¿K¹K×<YÑ<YÕZÜ�L‰L˜×)Ñ)°¸¿¹×9VÑ9VÖWð 1r5   )rH   rI   rJ   r#   rƒ   Úbase_model_prefixÚsupports_gradient_checkpointingÚ_no_split_modulesÚ_skip_keys_device_placementÚ_supports_flash_attnÚ_supports_sdpaÚ_supports_flex_attnÚ_can_compile_fullgraphÚ_supports_attention_backendrJ  r  Ú_can_record_outputsÚ_keep_in_fp32_modules_strictÚ"_keys_to_ignore_on_load_unexpectedr-   r†   r^  rM   rN   s   @r4   rY  rY    s|   ø… ØÓØÐØ&*Ð#Ø/Ð0ÐØ#4Ð"5ÐØÐØ€NØÐà!ÐØ"&Ðà-Ø'ñÐð $&Ð Ø)+Ð&à€U‡]�]ƒ_óXó ôXr5   rY  c                   óò   ‡ — e Zd Zdefˆ fd„Zeee	 	 	 	 	 	 ddej                  dz  dej                  dz  dej                  dz  dedz  dej                  dz  d	edz  d
ee   defd„«       «       «       Zˆ xZS )ÚMistral4ModelrR   c           	      ó  •— t         ‰| �  |«       |j                  | _        |j                  | _        t        j                  |j                  |j                  | j                  «      | _        t        j                  t        |j                  «      D �cg c]  }t        ||«      ‘Œ c}«      | _        t        |j                  |j                  ¬«      | _        t#        |¬«      | _        d| _        | j)                  «        y c c}w )NrL  ©rR   F)r*   r+   Úpad_token_idÚpadding_idxÚ
vocab_sizer   Ú	Embeddingr2   Úembed_tokensÚ
ModuleListÚrangeÚnum_hidden_layersrJ  Úlayersr&   rP  ÚnormrP   Ú
rotary_embÚgradient_checkpointingÚ	post_initr0  s      €r4   r+   zMistral4Model.__init__3  sÆ   ø€ Ü‰Ñ˜Ô Ø!×.Ñ.ˆÔØ ×+Ñ+ˆŒäŸL™L¨×):Ñ):¸F×<NÑ<NÐPT×P`ÑP`ÓaˆÔÜ—m‘mÜFKÈF×LdÑLdÓFeÖf¸Ô! &¨)Õ4Òfó
ˆŒô $ F×$6Ñ$6¸F×<OÑ<OÔPˆŒ	Ü1¸Ô@ˆŒØ&+ˆÔ#ð 	�‰Õùò gs   ÂDNÚ	input_idsrþ   r}   r2  Úinputs_embedsrS  r  r(   c           
      ó<  — |d u |d uz  rt        d«      ‚|€| j                  |«      }|r|€t        | j                  ¬«      }|€V|�|j	                  «       nd}t        j                  |j                  d   |j                  ¬«      |z   }|j                  d«      }t        | j                  ||||¬«      }	|}
| j                  |
|¬«      }| j                  d | j                  j                   D ]  } ||
f|	||||dœ|¤Ž}
Œ | j                  |
«      }
t        |
|¬	«      S )
Nz:You must specify exactly one of input_ids or inputs_embedsrq  r   r"   )r`   )rR   r€  rþ   r2  r}   )r}   )rþ   r1  r}   r2  rS  )Úlast_hidden_stater2  )Ú
ValueErrorrv  r
   rR   Úget_seq_lengthr-   ri   rF   r`   r±   r   r|  rz  ry  r{  r   )r1   r  rþ   r}   r2  r€  rS  r  Úpast_seen_tokensÚcausal_maskr6   r1  Údecoder_layers                r4   rC   zMistral4Model.forwardC  sL  € ð ˜Ð -°tÐ";Ò<ÜÐYÓZÐZàÐ Ø*.×*;Ñ*;¸IÓ*FˆMá˜Ð0Ü*°$·+±+Ô>ˆOàÐØCRÐC^˜×=Ñ=Ô?ÐdeÐÜ Ÿ<™<¨×(;Ñ(;¸AÑ(>À}×G[ÑG[Ô\Ð_oÑoˆLØ'×1Ñ1°!Ó4ˆLä(Ø—;‘;Ø'Ø)Ø+Ø%ô
ˆð &ˆØ"Ÿo™o¨mÈ,˜oÓWÐà!Ÿ[™[Ð)H¨4¯;©;×+HÑ+HÐIò 		ˆMÙ)Øðà*Ø$7Ø)Ø /Ø#ñð ñ‰Mð		ð Ÿ	™	 -Ó0ˆÜ&Ø+Ø+ô
ð 	
r5   )NNNNNN)rH   rI   rJ   r#   r+   r    r!   r   r-   rW  rL   r	   ÚFloatTensorr´   r   r   r   rC   rM   rN   s   @r4   ro  ro  1  sÍ   ø„ ð˜~õ ð   ØØð .2Ø.2Ø04Ø(,Ø26Ø!%ñ2
à×#Ñ# dÑ*ð2
ð Ÿ™ tÑ+ð2
ð ×&Ñ&¨Ñ-ð	2
ð
  ™ð2
ð ×(Ñ(¨4Ñ/ð2
ð ˜$‘;ð2
ð Ð+Ñ,ð2
ð 
!ò2
ó ó ó  ô2
r5   ro  c                   óB  ‡ — e Zd ZddiZddiZddgdgfiZˆ fd„Zee	 	 	 	 	 	 	 	 dd	e	j                  dz  d
e	j                  dz  de	j                  dz  dedz  de	j                  dz  de	j                  dz  dedz  dee	j                  z  dee   defd„«       «       Zˆ xZS )ÚMistral4ForCausalLMzlm_head.weightzmodel.embed_tokens.weightÚlm_headÚcolwise_gather_outputr6   Úlogitsc                 óè   •— t         ‰| �  |«       t        |«      | _        |j                  | _        t        j                  |j                  |j                  d¬«      | _        | j                  «        y rŠ   )
r*   r+   ro  rZ  rt  r   rŽ   r2   r‹  r~  r¤   s     €r4   r+   zMistral4ForCausalLM.__init__�  sU   ø€ Ü‰Ñ˜Ô Ü" 6Ó*ˆŒ
Ø ×+Ñ+ˆŒÜ—y‘y ×!3Ñ!3°V×5FÑ5FÈUÔSˆŒð 	�‰Õr5   Nr  rþ   r}   r2  r€  ÚlabelsrS  Úlogits_to_keepr  r(   c	           
      óx  —  | j                   d||||||dœ|	¤Ž}
|
j                  }t        |t        «      rt	        | d«      n|}| j                  |dd…|dd…f   «      }d}|�* | j                  d||| j                  j                  dœ|	¤Ž}t        |||
j                  |
j                  |
j                  ¬«      S )aá  
        Example:

        ```python
        >>> from transformers import AutoTokenizer, Mistral4ForCausalLM

        >>> model = Mistral4ForCausalLM.from_pretrained("meta-mistral4/Mistral4-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-mistral4/Mistral4-2-7b-hf")

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

        >>> # Generate
        >>> 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  rþ   r}   r2  r€  rS  N)r�  r�  rt  )Úlossr�  r2  r6   r[  rU  )rZ  r‚  ru   r…   Úslicer‹  Úloss_functionrR   rt  r   r2  r6   r[  )r1   r  rþ   r}   r2  r€  r�  rS  r�  r  Úoutputsr6   Úslice_indicesr�  r’  s                  r4   rC   zMistral4ForCausalLM.forwardŠ  sÙ   € ð> ,6¨4¯:©:ð ,
ØØ)Ø%Ø+Ø'Øñ,
ð ñ,
ˆð  ×1Ñ1ˆä8BÀ>ÔSVÔ8Wœ˜~˜o¨tÔ4Ð]kˆØ—‘˜mªA¨}ºaÐ,?Ñ@ÓAˆàˆØÐØ%�4×%Ñ%Ðp¨V¸FÈtÏ{É{×OeÑOeÑpÐioÑpˆDä%ØØØ#×3Ñ3Ø!×/Ñ/Ø×)Ñ)ô
ð 	
r5   )NNNNNNNr   )rH   rI   rJ   Ú_tied_weights_keysÚ_tp_planÚ_pp_planr+   r   r   r-   rW  rL   r	   rˆ  r´   r…   r   r   r   rC   rM   rN   s   @r4   rŠ  rŠ  {  s  ø„ à*Ð,GÐHÐØÐ2Ð3€HØ˜_Ð-°¨zÐ:Ð;€Hôð Øð .2Ø.2Ø04Ø(,Ø26Ø*.Ø!%Ø-.ñ6
à×#Ñ# dÑ*ð6
ð Ÿ™ tÑ+ð6
ð ×&Ñ&¨Ñ-ð	6
ð
  ™ð6
ð ×(Ñ(¨4Ñ/ð6
ð × Ñ  4Ñ'ð6
ð ˜$‘;ð6
ð ˜eŸl™lÑ*ð6
ð Ð+Ñ,ð6
ð 
 ò6
ó ó ô6
r5   rŠ  c                   ó   — e Zd Zy)Ú!Mistral4ForSequenceClassificationN©rH   rI   rJ   rU  r5   r4   r›  r›  Å  ó   „ Ør5   r›  c                   ó   — e Zd Zy)ÚMistral4ForTokenClassificationNrœ  rU  r5   r4   rŸ  rŸ  É  r�  r5   rŸ  )rY  ro  rŠ  r›  rŸ  )r"   )r¨   r  )NÚcollections.abcr   Útypingr   r-   Útorch.nn.functionalr   rÌ   rª   Ú r   r_  Úactivationsr   Úcache_utilsr	   r
   Ú
generationr   Úintegrationsr   r   r   Úmasking_utilsr   Úmodeling_flash_attention_utilsr   Úmodeling_layersr   r   r   Úmodeling_outputsr   r   Úmodeling_rope_utilsr   r   Úmodeling_utilsr   r   Úprocessing_utilsr   Úutilsr   r   r   Úutils.genericr   r   r    Úutils.output_capturingr!   Úconfiguration_mistral4r#   ÚModuler&   rP   rˆ   r—   rÁ   rà   rì   ró   rL   r…   rù   rK   r  r  r  r  rJ  rY  ro  rŠ  r›  rŸ  Ú__all__rU  r5   r4   ú<module>rµ     sj  ðõ( %Ý ã ß Ð Ý å &Ý !ß .Ý )ß mÑ mÝ /Ý B÷ñ ÷
 Pß Kß FÝ &ß IÑ Iß eÑ eÝ 5Ý 2ñ ˜YÓ'ôJ�b—i‘ió Jó (ðJô(><˜bŸi™iô ><ôB�"—)‘)ô ô "9˜Ÿ™ô "9ðJ ô$#�b—i‘ió $#ó ð$#ôN�"—)‘)ô ò0(ñ Ð*Ó+òó ,ðð2	U˜UŸ\™\ð 	U°#ð 	U¸%¿,¹,ó 	Uð& ñ%Ø�I‰Ið%à�<‰<ð%ð 
�‰ð%ð �<‰<ð	%ð
 —L‘L 4Ñ'ð%ð ð%ð ð%ð Ð'Ñ(ó%ó2$ðN%¨%¯,©,ð %¸eð %Ð^að %Ðfk×frÑfró %ô
t)˜Ÿ	™	ô t)ôn,Ð5ô ,ô^X˜oô Xð: ôF
Ð+ó F
ó ðF
ðR ôF
Ð1°?ó F
ó ðF
ôR	Ð(HÐJaô 	ô	Ð%BÐD[ô 	ò�r5   