Ë
    ÜÓ^j}w  ã                   óÈ  — 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mZ ddlmZmZ ddlmZ ddlmZ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/m0Z0 ddl1m2Z2m3Z3 ddl4m5Z5m6Z6 ddl7m8Z8 e G d„ dejr                  «      «       Z: G d„ dejr                  «      Z; G d„ dejr                  «      Z< ed«       G d„ dejr                  «      «       Z= G d „ d!ejr                  «      Z>d"„ Z? ed#«      dFd$„«       Z@d%ej‚                  d&eBd'ej‚                  fd(„ZC	 dGd)ejr                  d*ej‚                  d+ej‚                  d,ej‚                  d-ej‚                  dz  d.eDd/eDd0e,e.   fd1„ZE ee@«       G d2„ d3ejr                  «      «       ZF G d4„ d5e!«      ZGe/ G d6„ d7e*«      «       ZHe/ G d8„ d9eH«      «       ZI	 	 	 dHd:ej‚                  eJej‚                     z  dz  d;eBdz  d-ej‚                  dz  d'ej‚                  eBz  fd<„ZKe/ G d=„ d>eHe«      «       ZL G d?„ d@eeH«      ZM G dA„ dBe eH«      ZN G dC„ dDeeH«      ZOg dE¢ZPy)Ié    )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_hubÚuse_kernelized_func)Úcreate_causal_maskÚ!create_sliding_window_causal_mask)ÚFlashAttentionKwargs)ÚGenericForQuestionAnsweringÚ GenericForSequenceClassificationÚGenericForTokenClassificationÚGradientCheckpointingLayer)ÚMoeCausalLMOutputWithPastÚMoeModelOutputWithPast)ÚROPE_INIT_FUNCTIONSÚdynamic_rope_update)ÚALL_ATTENTION_FUNCTIONSÚPreTrainedModel)ÚUnpack)ÚTransformersKwargsÚauto_docstringÚcan_return_tuple)Úmaybe_autocastÚmerge_with_config_defaults)ÚOutputRecorderÚcapture_outputsé   )ÚMixtralConfigc                   ó�   ‡ — e Zd ZdZdefˆ fd„Zdej                  dej                  dej                  dej                  fd„Zˆ xZ	S )	ÚMixtralExpertsz2Collection of expert weights stored as 3D tensors.Úconfigc                 óò  •— t         ‰| �  «        |j                  | _        |j                  | _        |j                  | _        t        j                  t        j                  | j                  d| j                  z  | j
                  «      «      | _        t        j                  t        j                  | j                  | j
                  | j                  «      «      | _        t        |j                     | _        y )Né   )ÚsuperÚ__init__Únum_local_expertsÚnum_expertsÚhidden_sizeÚ
hidden_dimÚintermediate_sizeÚintermediate_dimr   Ú	ParameterÚtorchÚemptyÚgate_up_projÚ	down_projr   Ú
hidden_actÚact_fn©Úselfr)   Ú	__class__s     €úw/var/www/ramen.bs-engineer-server.com/venv/lib/python3.12/site-packages/transformers/models/mixtral/modeling_mixtral.pyr-   zMixtralExperts.__init__A   s­   ø€ Ü‰ÑÔØ!×3Ñ3ˆÔØ ×,Ñ,ˆŒØ &× 8Ñ 8ˆ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�ó    Úhidden_statesÚtop_k_indexÚtop_k_weightsÚreturnc                 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_classesr+   r%   r   )éÿÿÿÿéþÿÿÿ©ÚdimrF   )r5   Ú
zeros_likeÚno_gradr   Ú
functionalÚone_hotr/   ÚpermuteÚgreaterÚsumÚnonzeroÚwhereÚlinearr7   Úchunkr:   r8   Ú
index_add_ÚtoÚdtype)r<   r@   rA   rB   Úfinal_hidden_statesÚexpert_maskÚ
expert_hitÚ
expert_idxÚ	top_k_posÚ	token_idxÚcurrent_stateÚgateÚupÚcurrent_hidden_statess                 r>   ÚforwardzMixtralExperts.forwardJ   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)
Ú__name__Ú
__module__Ú__qualname__Ú__doc__r&   r-   r5   ÚTensorrb   Ú__classcell__©r=   s   @r>   r(   r(   =   sM   ø„ á<ð0˜}õ 0ð#à—|‘|ð#ð —\‘\ð#ð —|‘|ð	#ð
 
�‰÷#r?   r(   c                   ó$   ‡ — e Zd Zˆ fd„Zd„ Zˆ xZS )ÚMixtralTopKRouterc                 ó  •— t         ‰| �  «        |j                  | _        |j                  | _        |j                  | _        t        j                  t        j                  | j
                  | j                  «      «      | _        y ©N)r,   r-   Únum_experts_per_tokÚtop_kr.   r/   r0   r1   r   r4   r5   r6   Úweightr;   s     €r>   r-   zMixtralTopKRouter.__init__f   s[   ø€ Ü‰ÑÔØ×/Ñ/ˆŒ
Ø!×3Ñ3ˆÔØ ×,Ñ,ˆŒÜ—l‘l¤5§;¡;¨t×/?Ñ/?ÀÇÁÓ#QÓRˆ�r?   c                 óp  — |j                  d| j                  «      }t        j                  || j                  «      }t
        j                  j                  j                  |j                  «       d¬«      }t        j                  || j                  d¬«      \  }}||j                  dd¬«      z  }|}|||fS )NrF   rH   T)rI   Úkeepdim)Úreshaper1   ÚFrS   rp   r5   r   rL   ÚsoftmaxÚfloatÚtopkro   rP   )r<   r@   Úrouter_logitsÚrouter_probsÚrouter_top_valueÚrouter_indicesÚrouter_scoress          r>   rb   zMixtralTopKRouter.forwardm   sŸ   € Ø%×-Ñ-¨b°$·/±/ÓBˆÜŸ™ °·±Ó<ˆÜ—x‘x×*Ñ*×2Ñ2°=×3FÑ3FÓ3HÈbÐ2ÓQˆÜ+0¯:©:°lÀDÇJÁJÐTVÔ+WÑ(Ð˜.ØÐ,×0Ñ0°RÀÐ0ÓFÑFÐØ(ˆØ˜m¨^Ð;Ð;r?   )rc   rd   re   r-   rb   rh   ri   s   @r>   rk   rk   e   s   ø„ ôSö<r?   rk   c                   ót   ‡ — e Zd Zˆ fd„Zdej
                  deej
                  ej
                  f   fd„Zˆ xZS )ÚMixtralSparseMoeBlockc                 ó¦   •— t         ‰| �  «        |j                  | _        |j                  | _        t        |«      | _        t        |«      | _	        y rm   )
r,   r-   rn   ro   Úrouter_jitter_noiseÚjitter_noiserk   r_   r(   Úexpertsr;   s     €r>   r-   zMixtralSparseMoeBlock.__init__x   sA   ø€ Ü‰ÑÔØ×/Ñ/ˆŒ
Ø"×6Ñ6ˆÔÜ% fÓ-ˆŒ	Ü% fÓ-ˆ�r?   r@   rC   c                 ó”  — |j                   \  }}}| j                  rQ| j                  dkD  rB|t        j                  |«      j                  d| j                  z
  d| j                  z   «      z  }|j                  d|j                   d   «      }| j                  |«      \  }}}| j                  |||«      }|j                  |||«      }|S )Nr   ç      ð?rF   )
ÚshapeÚtrainingr�   r5   Ú
empty_likeÚuniform_Úviewr_   r‚   rs   )r<   r@   Ú
batch_sizeÚsequence_lengthr1   Ú_rB   rA   s           r>   rb   zMixtralSparseMoeBlock.forward   sÀ   € Ø2?×2EÑ2EÑ/ˆ
�O ZØ�=Š=˜T×.Ñ.°Ò2ØœU×-Ñ-¨mÓ<×EÑEÀcÈD×L]ÑL]ÑF]Ð_bÐei×evÑevÑ_vÓwÑwˆMØ%×*Ñ*¨2¨}×/BÑ/BÀ2Ñ/FÓGˆØ(,¯	©	°-Ó(@Ñ%ˆˆ=˜+ØŸ™ ]°KÀÓOˆØ%×-Ñ-¨j¸/È:ÓVˆØÐr?   )	rc   rd   re   r-   r5   rg   Útuplerb   rh   ri   s   @r>   r~   r~   w   s1   ø„ ô.ð U§\¡\ð °e¸E¿L¹LÈ%Ï,É,Ð<VÑ6W÷ r?   r~   Ú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MixtralRMSNormÚepsrC   Nc                 óŠ   •— t         ‰| �  «        t        j                  t	        j
                  |«      «      | _        || _        y)z=
        MixtralRMSNorm is equivalent to T5LayerNorm
        N)r,   r-   r   r4   r5   Úonesrp   Úvariance_epsilon)r<   r0   r‘   r=   s      €r>   r-   zMixtralRMSNorm.__init__Œ   s1   ø€ ô 	‰ÑÔÜ—l‘l¤5§:¡:¨kÓ#:Ó;ˆŒØ #ˆÕr?   r@   c                 ó"  — |j                   }|j                  t        j                  «      }|j	                  d«      j                  dd¬«      }|t        j                  || j                  z   «      z  }| j                  |j                  |«      z  S )Nr+   rF   T)rr   )	rW   rV   r5   Úfloat32ÚpowÚmeanÚrsqrtr”   rp   )r<   r@   Úinput_dtypeÚvariances       r>   rb   zMixtralRMSNorm.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:r?   c                 ó^   — t        | j                  j                  «      › d| j                  › �S )Nz, eps=)r�   rp   r…   r”   )r<   s    r>   Ú
extra_reprzMixtralRMSNorm.extra_repr›   s*   € Ü˜Ÿ™×)Ñ)Ó*Ð+¨6°$×2GÑ2GÐ1HÐIÐIr?   )g�íµ ÷Æ°>)
rc   rd   re   rv   r-   r5   rg   rb   r�   rh   ri   s   @r>   r�   r�   Š   s7   ø„ ñ$¨ð $¸$õ $ð; U§\¡\ð ;°e·l±ló ;öJr?   r�   c                   óÄ   ‡ — e Zd ZU ej                  ed<   ddefˆ fd„Ze	 	 	 ddedz  de	d   de
dz  ded	ef   fd
„«       Z ej                  «       ed„ «       «       Zˆ xZS )ÚMixtralRotaryEmbeddingÚinv_freqNr)   c                 ó®  •— t         ‰| �  «        |j                  | _        |j                  | _        || _        | j
                  j                  d   | _        | j                  }| j                  dk7  rt        | j                     } || j
                  |«      \  }| _
        | j                  d|d¬«       | j                  d|j                  «       d¬«       y )NÚ	rope_typeÚdefaultr    F)Ú
persistentÚoriginal_inv_freq)r,   r-   Úmax_position_embeddingsÚmax_seq_len_cachedÚoriginal_max_seq_lenr)   Úrope_parametersr¢   Úcompute_default_rope_parametersr   Úattention_scalingÚregister_bufferÚclone)r<   r)   ÚdeviceÚrope_init_fnr    r=   s        €r>   r-   zMixtralRotaryEmbedding.__init__¢   s¶   ø€ Ü‰ÑÔØ"(×"@Ñ"@ˆÔØ$*×$BÑ$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�$Ô(à×Ñ˜Z¨¸eÐÔDØ×ÑÐ0°(·.±.Ó2BÈuÐÕUr?   r®   ztorch.deviceÚseq_lenrC   ztorch.Tensorc                 ó  — | j                   d   }t        | dd«      xs | j                  | j                  z  }d}d|t	        j
                  d|dt        j                  ¬«      j                  |t        j                  ¬«      |z  z  z  }||fS )	a¨  
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        Ú
rope_thetaÚhead_dimNr„   r   r+   ©rW   )r®   rW   )	r©   Úgetattrr0   Únum_attention_headsr5   ÚarangeÚint64rV   rv   )r)   r®   r°   ÚbaserI   Úattention_factorr    s          r>   rª   z6MixtralRotaryEmbedding.compute_default_rope_parameters²   s�   € ð& ×%Ñ% lÑ3ˆÜ�f˜j¨$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)r?   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   rF   r%   ÚmpsÚcpuF)Údevice_typeÚenabledr+   rH   r´   )r    rv   Úexpandr…   rV   r®   Ú
isinstanceÚtypeÚstrr!   Ú	transposer5   ÚcatÚcosr«   ÚsinrW   )
r<   ÚxÚposition_idsÚinv_freq_expandedÚposition_ids_expandedr¾   ÚfreqsÚembrÆ   rÇ   s
             r>   rb   zMixtralRotaryEmbedding.forwardÐ   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$rm   )NNN)rc   rd   re   r5   rg   Ú__annotations__r&   r-   Ústaticmethodr   Úintr�   rv   rª   rK   r   rb   rh   ri   s   @r>   rŸ   rŸ   Ÿ   s˜   ø… Ø�l‰lÓñV˜}õ Vð  à'+Ø+/Ø"ñ*Ø Ñ$ð*à˜Ñ(ð*ð �t‘ð*ð 
ˆ~˜uÐ$Ñ	%ò	*ó ð*ð: €U‡]�]ƒ_Øñ<ó ó ô<r?   rŸ   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..NrF   r+   rH   )r…   r5   rÅ   )rÈ   Úx1Úx2s      r>   Úrotate_halfrÔ   à   sZ   € à	
ˆ3Ð"�!—'‘'˜"‘+ Ñ"Ð"Ð"Ñ	#€BØ	
ˆ3�—‘˜‘˜qÑ Ñ"Ð"Ñ	#€BÜ�9‰9�r�c˜2�Y BÔ'Ð'r?   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.
    )Ú	unsqueezerÔ   )ÚqÚkrÆ   rÇ   Úunsqueeze_dimÚq_embedÚk_embeds          r>   Úapply_rotary_pos_embrÝ   ç   sY   € ð& �-‰-˜Ó
&€CØ
�-‰-˜Ó
&€CØ�3‰wœ; q›>¨CÑ/Ñ0€GØ�3‰wœ; q›>¨CÑ/Ñ0€GØ�GÐÐr?   r@   Ún_reprC   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)r…   rÀ   rs   )r@   rÞ   ÚbatchÚnum_key_value_headsÚslenr³   s         r>   Ú	repeat_kvrã     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Tr?   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 )Nr+   r   rF   )rI   rW   )Úpr†   r%   )rã   Únum_key_value_groupsr5   ÚmatmulrÄ   r   rL   ru   r–   rV   rW   rê   r†   Ú
contiguous)rä   rå   ræ   rç   rè   ré   rê   rë   Ú
key_statesÚvalue_statesÚattn_weightsÚattn_outputs               r>   Úeager_attention_forwardrõ     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$Ð$r?   c                   óô   ‡ — 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
dz  d
ee   de	ej                  ej                  dz  f   fd„Zˆ xZS )ÚMixtralAttentionz=Multi-headed attention from 'Attention Is All You Need' paperr)   Ú	layer_idxc                 ó  •— t         ‰| �  «        || _        || _        t	        |dd «      xs |j
                  |j                  z  | _        |j                  |j                  z  | _	        | j                  dz  | _
        |j                  | _        d| _        t        j                  |j
                  |j                  | j                  z  d¬«      | _        t        j                  |j
                  |j                  | j                  z  d¬«      | _        t        j                  |j
                  |j                  | j                  z  d¬«      | _        t        j                  |j                  | j                  z  |j
                  d¬«      | _        y )Nr³   g      à¿TF©Úbias)r,   r-   r)   rø   rµ   r0   r¶   r³   rá   rî   ré   Úattention_dropoutÚ	is_causalr   ÚLinearÚq_projÚk_projÚv_projÚo_proj©r<   r)   rø   r=   s      €r>   r-   zMixtralAttention.__init__*  s2  ø€ Ü‰ÑÔØˆŒØ"ˆŒÜ ¨
°DÓ9Òm¸V×=OÑ=OÐSY×SmÑSmÑ=mˆŒØ$*×$>Ñ$>À&×B\ÑB\Ñ$\ˆÔ!Ø—}‘} dÑ*ˆŒØ!'×!9Ñ!9ˆÔØˆŒÜ—i‘i × 2Ñ 2°F×4NÑ4NÐQU×Q^ÑQ^Ñ4^ÐejÔkˆŒÜ—i‘i × 2Ñ 2°F×4NÑ4NÐQU×Q^ÑQ^Ñ4^ÐejÔkˆŒÜ—i‘i × 2Ñ 2°F×4NÑ4NÐQU×Q^ÑQ^Ñ4^ÐejÔkˆŒÜ—i‘i × :Ñ :¸T¿]¹]Ñ JÈF×L^ÑL^ÐejÔkˆ�r?   Nr@   Úposition_embeddingsrè   Úpast_key_valuesrë   rC   c           
      ó6  — |j                   d d }g |¢d‘| j                  ‘­}| j                  |«      j                  |«      j	                  dd«      }| j                  |«      j                  |«      j	                  dd«      }	| j                  |«      j                  |«      j	                  dd«      }
|\  }}t        ||	||«      \  }}	|� |j                  |	|
| j                  «      \  }	}
t        j                  | j                  j                  t        «      } || ||	|
|f| j                  sdn| j                   | j"                  t%        | j                  dd «      dœ|¤Ž\  }} |j&                  g |¢d‘­Ž j)                  «       }| j+                  |«      }||fS )NrF   r%   r+   ç        Úsliding_window)rê   ré   r  )r…   r³   rÿ   r‰   rÄ   r   r  rÝ   Úupdaterø   r   Úget_interfacer)   Ú_attn_implementationrõ   r†   rü   ré   rµ   rs   rð   r  )r<   r@   r  rè   r  rë   Úinput_shapeÚhidden_shapeÚquery_statesrñ   rò   rÆ   rÇ   Úattention_interfacerô   ró   s                   r>   rb   zMixtralAttention.forward8  s¢  € ð $×)Ñ)¨#¨2Ð.ˆØ8˜Ð8 bÐ8¨$¯-©-Ñ8ˆà—{‘{ =Ó1×6Ñ6°|ÓD×NÑNÈqÐRSÓTˆØ—[‘[ Ó/×4Ñ4°\ÓB×LÑLÈQÐPQÓRˆ
Ø—{‘{ =Ó1×6Ñ6°|ÓD×NÑNÈqÐRSÓTˆà&‰ˆˆSÜ#7¸ÀjÐRUÐWZÓ#[Ñ ˆ�jàÐ&Ø'6×'=Ñ'=¸jÈ,ÐX\×XfÑXfÓ'gÑ$ˆJ˜ä(?×(MÑ(MØ�K‰K×,Ñ,Ô.Eó)
Ðñ %8ØØØØØð
%
ð  $Ÿ}š}‘C°$×2HÑ2HØ—L‘LÜ" 4§;¡;Ð0@À$ÓGñ
%
ð ñ
%
Ñ!ˆ�\ð *�k×)Ñ)Ð;¨;Ð;¸Ò;×FÑFÓHˆØ—k‘k +Ó.ˆØ˜LÐ(Ð(r?   rm   )rc   rd   re   rf   r&   rÐ   r-   r5   rg   r�   r	   r   r   rb   rh   ri   s   @r>   r÷   r÷   &  s    ø„ áGðl˜}ð l¸õ lð& )-ñ')à—|‘|ð')ð # 5§<¡<°·±Ð#=Ñ>ð')ð Ÿ™ tÑ+ð	')ð
  ™ð')ð Ð-Ñ.ð')ð 
ˆu�|‰|˜UŸ\™\¨DÑ0Ð0Ñ	1÷')r?   r÷   c                   óö   ‡ — e Zd Zdedefˆ fd„Z	 	 	 	 ddej                  deej                  ej                  f   dz  dej                  dz  dej                  dz  d	e
dz  d
ee   dej                  fd„Zˆ xZS )ÚMixtralDecoderLayerr)   rø   c                 ó  •— t         ‰| �  «        |j                  | _        t        ||«      | _        t        |«      | _        t        |j                  |j                  ¬«      | _	        t        |j                  |j                  ¬«      | _
        y )N©r‘   )r,   r-   r0   r÷   Ú	self_attnr~   Úmlpr�   Úrms_norm_epsÚinput_layernormÚpost_attention_layernormr  s      €r>   r-   zMixtralDecoderLayer.__init__c  sl   ø€ Ü‰ÑÔØ!×-Ñ-ˆÔä)¨&°)Ó<ˆŒä(¨Ó0ˆŒÜ-¨f×.@Ñ.@Àf×FYÑFYÔZˆÔÜ(6°v×7IÑ7IÈv×ObÑObÔ(cˆÕ%r?   Nr@   r  rè   rÉ   r  rë   rC   c           	      ó¾   — |}| j                  |«      } | j                  d|||||dœ|¤Ž\  }}||z   }|}| j                  |«      }| j                  |«      }||z   }|S )N)r@   r  rè   rÉ   r  © )r  r  r  r  )	r<   r@   r  rè   rÉ   r  rë   ÚresidualrŒ   s	            r>   rb   zMixtralDecoderLayer.forwardm  s‹   € ð !ˆØ×,Ñ,¨]Ó;ˆØ)˜4Ÿ>™>ð 
Ø'Ø 3Ø)Ø%Ø+ñ
ð ñ
Ñˆ�qð ! =Ñ0ˆØ ˆØ×5Ñ5°mÓDˆØŸ™ Ó/ˆØ  =Ñ0ˆØÐr?   )NNNN)rc   rd   re   r&   rÐ   r-   r5   rg   r�   Ú
LongTensorr	   r   r   rb   rh   ri   s   @r>   r  r  b  s¯   ø„ ðd˜}ð d¸õ dð IMØ.2Ø04Ø(,ñà—|‘|ðð # 5§<¡<°·±Ð#=Ñ>ÀÑEðð Ÿ™ tÑ+ð	ð
 ×&Ñ&¨Ñ-ðð  ™ðð Ð+Ñ,ðð 
�‰÷r?   r  c                   ó–   ‡ — 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¬«      eedœZ ej(                  «       ˆ fd	„«       Zˆ xZS )
ÚMixtralPreTrainedModelr)   ÚmodelTr  r  r   )Úindex)rx   r@   Ú
attentionsc                 ó`  •— t         ‰| �  |«       | j                  j                  }t	        |t
        «      rEt        j                  |j                  d|¬«       t        j                  |j                  d|¬«       y t	        |t        «      r#t        j                  |j                  d|¬«       y y )Nr  )r˜   Ústd)r,   Ú_init_weightsr)   Úinitializer_rangerÁ   r(   ÚinitÚnormal_r7   r8   rk   rp   )r<   rä   r#  r=   s      €r>   r$  z$MixtralPreTrainedModel._init_weights›  sy   ø€ ä‰Ñ˜fÔ%Ø�k‰k×+Ñ+ˆÜ�fœnÔ-Ü�L‰L˜×,Ñ,°3¸CÕ@Ü�L‰L˜×)Ñ)°¸Ö=Ü˜Ô 1Ô2Ü�L‰L˜Ÿ™¨S°cÖ:ð 3r?   )rc   rd   re   r&   rÎ   Úbase_model_prefixÚsupports_gradient_checkpointingÚ_no_split_modulesÚ_skip_keys_device_placementÚ_supports_flash_attnÚ_supports_sdpaÚ_supports_flex_attnÚ_can_compile_fullgraphÚ_supports_attention_backendr#   rk   r  r÷   Ú_can_record_outputsr5   rK   r$  rh   ri   s   @r>   r  r  ˆ  sy   ø… àÓØÐØ&*Ð#Ø.Ð/ÐØ#4Ð"5ÐØÐØ€NØÐà!ÐØ"&Ðá'Ð(9ÀÔCØ,Ø&ñÐð €U‡]�]ƒ_ó;ó ô;r?   r  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MixtralModelr)   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 )Nr  ©r)   F)r,   r-   Úpad_token_idÚpadding_idxÚ
vocab_sizer   Ú	Embeddingr0   Úembed_tokensÚ
ModuleListÚrangeÚnum_hidden_layersr  Úlayersr�   r  ÚnormrŸ   Ú
rotary_embÚgradient_checkpointingÚ	post_initr  s      €r>   r-   zMixtralModel.__init__¨  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‘mÜEJÈ6×KcÑKcÓEdÖe¸	Ô  ¨Õ3Òeó
ˆŒô # 6×#5Ñ#5¸6×;NÑ;NÔOˆŒ	Ü0¸Ô?ˆŒØ&+ˆÔ#ð 	�‰Õùò fs   ÂDNÚ	input_idsrè   rÉ   r  Úinputs_embedsÚ	use_cacherë   rC   c           
      óz  — |d u |d uz  rt        d«      ‚|r|€t        | j                  ¬«      }|€| j                  |«      }|€V|�|j	                  «       nd}t        j                  |j                  d   |j                  ¬«      |z   }|j                  d«      }| j                  j                  €t        n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_embedsr5  r   r%   )r®   )r)   rD  rè   r  rÉ   )rÉ   )rè   rÉ   r  rE  r  )Úlast_hidden_stater  )Ú
ValueErrorr
   r)   r:  Úget_seq_lengthr5   r·   r…   r®   r×   r  r   r   r@  r>  r=  r?  r   )r<   rC  rè   rÉ   r  rD  rE  rë   Úpast_seen_tokensÚmask_functionÚcausal_maskr@   r  Údecoder_layers                 r>   rb   zMixtralModel.forward¸  sd  € ð ˜Ð -°tÐ";Ò<ÜÐYÓZÐZá˜Ð0Ü*°$·+±+Ô>ˆOàÐ Ø ×-Ñ-¨iÓ8ˆMàÐØCRÐC^˜×=Ñ=Ô?ÐdeÐÜ Ÿ<™<¨×(;Ñ(;¸AÑ(>À}×G[ÑG[Ô\Ð_oÑoˆLØ'×1Ñ1°!Ó4ˆLà.2¯k©k×.HÑ.HÐ.PÕ*ÔVwˆÙ#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)Ø /Ø#Ø$7ñð ñ‰Mð		ð Ÿ	™	 -Ó0ˆä%Ø+Ø+ô
ð 	
r?   )NNNNNN)rc   rd   re   r&   r-   r"   r$   r   r5   r  rg   r	   ÚFloatTensorÚboolr   r   r   rb   rh   ri   s   @r>   r3  r3  ¦  sÍ   ø„ ð˜}õ ð   ØØð .2Ø.2Ø04Ø(,Ø26Ø!%ñ4
à×#Ñ# dÑ*ð4
ð Ÿ™ tÑ+ð4
ð ×&Ñ&¨Ñ-ð	4
ð
  ™ð4
ð ×(Ñ(¨4Ñ/ð4
ð ˜$‘;ð4
ð Ð+Ñ,ð4
ð 
 ò4
ó ó ó  ô4
r?   r3  Úgate_logitsr/   c                 ó¾  — | �t        | t        «      syt        | t        «      rC| d   j                  }t        j                  | D �cg c]  }|j                  |«      ‘Œ c}d¬«      }t        j                  j                  j                  d¬«      }t        j                  ||d¬«      \  }}	t        j                  j                  j                  |	|«      }
|€>t        j                  |
j                  «       d¬«      }t        j                  |d¬«      }�n|j                  \  }}|j                  d   ||z  z  }|ddd…dd…ddf   j                  |||||f«      j                  d||«      j                  «      }t        j                   |
j                  «       |z  d¬«      t        j                   |d¬«      z  }|ddd…dd…df   j                  ||||f«      j                  d|«      j                  |«      }t        j                   ||z  d¬«      t        j                   |d¬«      z  }t        j                   ||j#                  d«      z  «      }||z  S c c}w )aÄ  
    Computes auxiliary load balancing loss as in Switch Transformer - implemented in Pytorch.

    See Switch Transformer (https://huggingface.co/papers/2101.03961) for more details. This function implements the loss
    function presented in equations (4) - (6) of the paper. It aims at penalizing cases where the routing between
    experts is too unbalanced.

    Args:
        gate_logits:
            Logits from the `gate`, should be a tuple of model.config.num_hidden_layers tensors of
            shape [batch_size X sequence_length, num_experts].
        num_experts:
            Number of experts
        top_k:
            The number of experts to route per-token, can be also interpreted as the `top-k` routing
            parameter.
        attention_mask (`torch.Tensor`, *optional*):
            The attention_mask used in forward function
            shape [batch_size X sequence_length] if not None.

    Returns:
        The auxiliary loss.
    Nr   rH   rF   )rÁ   r�   r®   r5   rÅ   rV   r   rL   ru   rw   rM   r˜   rv   r…   rÀ   rs   rP   r×   )rP  r/   ro   rè   Úcompute_deviceÚ
layer_gateÚconcatenated_gate_logitsÚrouting_weightsrŒ   Úselected_expertsrY   Útokens_per_expertÚrouter_prob_per_expertrŠ   r‹   r=  Úexpert_attention_maskÚ router_per_expert_attention_maskÚoverall_losss                      r>   Úload_balancing_loss_funcr\  ò  s9  € ð: Ð¤*¨[¼%Ô"@Øä�+œuÔ%Ø$ Q™×.Ñ.ˆÜ#(§9¡9Ð^iÖ-jÐPZ¨j¯m©m¸NÕ.KÒ-jÐpqÔ#rÐ ä—h‘h×)Ñ)×1Ñ1Ð2JÐPRÐ1ÓS€OäŸ*™* _°eÀÔDÑ€AÐä—(‘(×%Ñ%×-Ñ-Ð.>ÀÓL€KàÐä!ŸJ™J {×'8Ñ'8Ó':ÀÔBÐô "'§¡¨OÀÔ!CÒà&4×&:Ñ&:Ñ#ˆ
�OØ4×:Ñ:¸1Ñ=À*ÈÑB^Ñ_Ðð ˜4¢¢A t¨TÐ1Ñ2ß‰VÐ&¨
°OÀUÈKÐXÓYß‰W�R˜ Ó,ß‰R�Óð	 	ô "ŸI™I k×&7Ñ&7Ó&9Ð<QÑ&QÐWXÔYÔ\a×\eÑ\eØ! qô]
ñ 
Ðð ˜4¢¢A tÐ+Ñ,ß‰VÐ&¨
°OÀ[ÐQÓRß‰W�R˜Ó%ß‰R�Óð	 	)ô "'§¡¨?Ð=]Ñ+]ÐcdÔ!eÔhm×hqÑhqØ,°!ôi
ñ "
Ðô —9‘9Ð.Ð1G×1QÑ1QÐRSÓ1TÑTÓU€LØ˜+Ñ%Ð%ùò[ .ks   ÁIc                   óN  ‡ — 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dz  dee	j                  z  dee   defd„«       «       Zˆ xZS )ÚMixtralForCausalLMzlm_head.weightzmodel.embed_tokens.weightÚlm_headÚcolwise_gather_outputr@   Úlogitsc                 óN  •— t         ‰| �  |«       t        |«      | _        |j                  | _        t        j                  |j                  |j                  d¬«      | _        |j                  | _	        |j                  | _        |j                  | _        | j                  «        y )NFrú   )r,   r-   r3  r  r8  r   rþ   r0   r_  Úrouter_aux_loss_coefr.   r/   rn   rB  r;   s     €r>   r-   zMixtralForCausalLM.__init__J  s   ø€ Ü‰Ñ˜Ô Ü! &Ó)ˆŒ
Ø ×+Ñ+ˆŒÜ—y‘y ×!3Ñ!3°V×5FÑ5FÈUÔSˆŒØ$*×$?Ñ$?ˆÔ!Ø!×3Ñ3ˆÔØ#)×#=Ñ#=ˆÔ ð 	�‰Õr?   NrC  rè   rÉ   r  rD  ÚlabelsrE  Úoutput_router_logitsÚlogits_to_keeprë   rC   c
                 ój  — |�|n| j                   j                  } | j                  d|||||||dœ|
¤Ž}|j                  }t	        |	t
        «      rt        |	 d«      n|	}| j                  |dd…|dd…f   «      }d}|� | j                  ||| j                  fi |
¤Ž}d}|rYt        |j                  | j                  | j                  |«      }|�+|| j                  |j                  |j                   «      z  z  }t#        ||||j$                  |j&                  |j(                  |j                  ¬«      S )a~  
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
            config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
            (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.

        Example:

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

        >>> model = MixtralForCausalLM.from_pretrained("mistralai/Mixtral-8x7B-v0.1")
        >>> tokenizer = AutoTokenizer.from_pretrained("mistralai/Mixtral-8x7B-v0.1")

        >>> 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."
        ```N)rC  rè   rÉ   r  rD  rE  re  )ÚlossÚaux_lossra  r  r@   r!  rx   r  )r)   re  r  rG  rÁ   rÐ   Úslicer_  Úloss_functionr8  r\  rx   r/   rn   rc  rV   r®   r   r  r@   r!  )r<   rC  rè   rÉ   r  rD  rd  rE  re  rf  rë   Úoutputsr@   Úslice_indicesra  rh  ri  s                    r>   rb   zMixtralForCausalLM.forwardV  sU  € ðN %9Ð$DÑ È$Ï+É+×JjÑJjð 	ð
 +5¨$¯*©*ð 	+
ØØ)Ø%Ø+Ø'ØØ!5ñ	+
ð ñ	+
ˆð  ×1Ñ1ˆä8BÀ>ÔSVÔ8Wœ˜~˜o¨tÔ4Ð]kˆØ—‘˜mªA¨}ºaÐ,?Ñ@ÓAˆàˆØÐØ%�4×%Ñ% f¨f°d·o±oÑPÈÑPˆDàˆÙÜ/Ø×%Ñ%Ø× Ñ Ø×(Ñ(Øó	ˆHð Ð!Ø˜×1Ñ1°H·K±KÀÇÁÓ4LÑLÑL�ä(ØØØØ#×3Ñ3Ø!×/Ñ/Ø×)Ñ)Ø!×/Ñ/ô
ð 	
r?   )	NNNNNNNNr   )rc   rd   re   Ú_tied_weights_keysÚ_tp_planÚ_pp_planr-   r    r   r5   r  rg   r	   rN  rO  rÐ   r   r   r   rb   rh   ri   s   @r>   r^  r^  D  s5  ø„ à*Ð,GÐHÐØÐ2Ð3€HØ˜_Ð-°¨zÐ:Ð;€Hô
ð Øð .2Ø.2Ø04Ø(,Ø26Ø*.Ø!%Ø,0Ø-.ñP
à×#Ñ# dÑ*ðP
ð Ÿ™ tÑ+ðP
ð ×&Ñ&¨Ñ-ð	P
ð
  ™ðP
ð ×(Ñ(¨4Ñ/ðP
ð × Ñ  4Ñ'ðP
ð ˜$‘;ðP
ð # T™kðP
ð ˜eŸl™lÑ*ðP
ð Ð+Ñ,ðP
ð 
#òP
ó ó ôP
r?   r^  c                   ó   — e Zd Zy)Ú MixtralForSequenceClassificationN©rc   rd   re   r  r?   r>   rr  rr  «  ó   „ Ør?   rr  c                   ó   — e Zd Zy)ÚMixtralForTokenClassificationNrs  r  r?   r>   rv  rv  ¯  rt  r?   rv  c                   ó   — e Zd Zy)ÚMixtralForQuestionAnsweringNrs  r  r?   r>   rx  rx  ³  rt  r?   rx  )r^  rx  r3  r  rr  rv  )r%   )r  )Nr+   N)QÚcollections.abcr   Útypingr   r5   Útorch.nn.functionalr   rL   rt   Ú r   r&  Úactivationsr   Úcache_utilsr	   r
   Ú
generationr   Úintegrationsr   r   r   r   Úmasking_utilsr   r   Úmodeling_flash_attention_utilsr   Úmodeling_layersr   r   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"   Úutils.output_capturingr#   r$   Úconfiguration_mixtralr&   ÚModuler(   rk   r~   r�   rŸ   rÔ   rÝ   rg   rÐ   rã   rv   rõ   r÷   r  r  r3  r�   r\  r^  rr  rv  rx  Ú__all__r  r?   r>   ú<module>rŽ     sÁ  ðõ4 %Ý ã ß Ð Ý å &Ý !ß .Ý )÷ó ÷ SÝ B÷ó ÷ Rß Kß FÝ &ß IÑ Iß Gß EÝ 0ð ô$#�R—Y‘Yó $#ó ð$#ôN<˜Ÿ	™	ô <ô$˜BŸI™Iô ñ& ˜YÓ'ôJ�R—Y‘Yó Jó (ðJô(><˜RŸY™Yô ><òB(ñ Ð*Ó+òó ,ðð2	U˜UŸ\™\ð 	U°#ð 	U¸%¿,¹,ó 	Uð& ñ%Ø�I‰Ið%à�<‰<ð%ð 
�‰ð%ð �<‰<ð	%ð
 —L‘L 4Ñ'ð%ð ð%ð ð%ð Ð'Ñ(ó%ñ2 Ð)Ó*ô8)�r—y‘yó 8)ó +ð8)ôv#Ð4ô #ðL ô;˜_ó ;ó ð;ð: ôH
Ð)ó H
ó ðH
ðZ #Ø
Ø*.ñ	O&Ø—‘  e§l¡lÑ 3Ñ3°dÑ:ðO&à�t‘ðO&ð —L‘L 4Ñ'ð	O&ð
 ‡\�\�CÑóO&ðd ôc
Ð/°ó c
ó ðc
ôL	Ð'GÐI_ô 	ô	Ð$AÐCYô 	ô	Ð"=Ð?Uô 	ò�r?   