
    qi~                     L   d dl mZ d dlmZ d dlZd dl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 ddlmZm Z  ddl!m"Z"m#Z# ddl$m%Z% ddl&m'Z'm(Z( ddl)m*Z*m+Z+m,Z, ddl-m.Z. ddl/m0Z0  ed       G d dejb                               Z2 G d dejb                        Z3 G d dejb                        Z4 G d dejb                        Z5 G d  d!ejb                        Z6d" Z7 ed#      d@d$       Z8d%ejr                  d&e:d'ejr                  fd(Z;	 dAd)ejb                  d*ejr                  d+ejr                  d,ejr                  d-ejr                  dz  d.e<d/e<d0e%e'   fd1Z= ee8       G d2 d3ejb                               Z> G d4 d5e      Z?e( G d6 d7e#             Z@e( G d8 d9e@             ZA	 	 	 dBd:ejr                  eBejr                     z  dz  d;e:dz  d-ejr                  dz  d'ejr                  e:z  fd<ZCe( G d= d>e@e             ZDg d?ZEy)C    )Callable)OptionalN)nn)
functional   )initialization)ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hubuse_kernelized_func)create_causal_mask)GradientCheckpointingLayer)MoeCausalLMOutputWithPastMoeModelOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstring)can_return_tuplemaybe_autocastmerge_with_config_defaults)capture_outputs   )GraniteMoeConfigRMSNormc                   h     e Zd Zddeddf fdZdej                  dej                  fdZd Z xZ	S )	GraniteMoeRMSNormepsreturnNc                     t         |           t        j                  t	        j
                  |            | _        || _        y)z@
        GraniteMoeRMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	Parametertorchonesweightvariance_epsilon)selfhidden_sizer$   	__class__s      d/opt/pipecat/venv/lib/python3.12/site-packages/transformers/models/granitemoe/modeling_granitemoe.pyr(   zGraniteMoeRMSNorm.__init__0   s1     	ll5::k#:; #    hidden_statesc                 "   |j                   }|j                  t        j                        }|j	                  d      j                  dd      }|t        j                  || j                  z         z  }| j                  |j                  |      z  S )N   T)keepdim)	dtypetor*   float32powmeanrsqrtr-   r,   )r.   r3   input_dtypevariances       r1   forwardzGraniteMoeRMSNorm.forward8   sy    #))%((7 $$Q',,R,>%Ht?T?T4T(UU{{]--k:::r2   c                 ^    t        | j                  j                         d| j                   S )Nz, eps=)tupler,   shaper-   )r.   s    r1   
extra_reprzGraniteMoeRMSNorm.extra_repr?   s*    ))*+6$2G2G1HIIr2   )gư>)
__name__
__module____qualname__floatr(   r*   Tensorr@   rD   __classcell__r0   s   @r1   r#   r#   .   s7    $ $$ $;U\\ ;ell ;Jr2   r#   c                        e Zd ZU ej                  ed<   ddef 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 xZS )GraniteMoeRotaryEmbeddinginv_freqNconfigc                    t         |           |j                  | _        |j                  | _        || _        | j
                  j                  d   | _        | j                  }| j                  dk7  rt        | j                     } || j
                  |      \  }| _
        | j                  d|d       | j                  d|j                         d       y )N	rope_typedefaultrN   F)
persistentoriginal_inv_freq)r'   r(   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrO   rope_parametersrQ   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r.   rO   devicerope_init_fnrN   r0   s        r1   r(   z"GraniteMoeRotaryEmbedding.__init__F   s    "("@"@$*$B$B!44[A!%!E!E>>Y&.t~~>L+7V+L($(ZeD0(..2BuUr2   r]   ztorch.deviceseq_lenr%   ztorch.Tensorc                    | 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).
        
rope_thetahead_dimNg      ?r   r5   r8   )r]   r8   )	rX   getattrr/   num_attention_headsr*   arangeint64r9   rH   )rO   r]   r_   basedimattention_factorrN   s          r1   rY   z9GraniteMoeRotaryEmbedding.compute_default_rope_parametersV   s    & %%l3fj$/c63E3EIcIc3c U\\!S!5;;?BB&X]XcXcBdgjjk
 )))r2   c                 N   | j                   d d d d f   j                         j                  |j                  d   dd      j	                  |j
                        }|d d d d d f   j                         }t        |j
                  j                  t              r/|j
                  j                  dk7  r|j
                  j                  nd}t        |d      5  |j                         |j                         z  j                  dd      }t        j                  ||fd	      }|j                         | j                  z  }|j                         | j                  z  }	d d d        j	                  |j                   
      	j	                  |j                   
      fS # 1 sw Y   AxY w)Nr   r6   r   mpscpuF)device_typeenabledr5   ri   rc   )rN   rH   expandrC   r9   r]   
isinstancetypestrr   	transposer*   catcosrZ   sinr8   )
r.   xposition_idsinv_freq_expandedposition_ids_expandedrn   freqsembrw   rx   s
             r1   r@   z!GraniteMoeRotaryEmbedding.forwardt   sR    !MM$4-8>>@GGHZHZ[\H]_acdehhijiqiqr ,QaZ 8 > > @'1!((--'E!((--[`J`ahhmmfkUC 	5&,,.1F1L1L1NNYYZ[]^_E))UEN3C'')d444C'')d444C		5 vvAGGv$cff177f&;;;	5 	5s   BFF$)N)NNN)rE   rF   rG   r*   rI   __annotations__r    r(   staticmethodr   intrB   rH   rY   no_gradr   r@   rJ   rK   s   @r1   rM   rM   C   s    llV/ V  *.+/"* 4'*(* t* 
~u$	%	* *: U]]_<  <r2   rM   c                   6     e Zd Zdedededdf fdZd Z xZS )GraniteMoeParallelExpertsnum_experts
input_sizeoutput_sizer%   Nc                     t         |           t        j                  t	        j
                  |||            | _        || _        || _        || _	        y)a  
        Initialize the GraniteMoeParallelExperts module.
        The experts weights are stored in [num_experts, output_size, input_size] format. Such that it's compatible with
        many MoE libraries, such as [Megablock](https://github.com/databricks/megablocks) and
        [ScatterMoE](https://github.com/shawntan/scattermoe), as well as the
        [MoE kernel](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/layers/fused_moe/fused_moe.py)
        used in vllm.

        Args:
            num_experts (int):
                Number of experts.
            input_size (int):
                Size of the input.
            output_size (int):
                Size of the output.
        N)
r'   r(   r   r)   r*   emptyr,   r   r   r   )r.   r   r   r   r0   s       r1   r(   z"GraniteMoeParallelExperts.__init__   sD    " 	ll5;;{K#TU&$&r2   c                     |j                  |d      }g }t        | j                        D ]7  }|j                  t	        j
                  ||   | j                  |                9 t        j                  |d      }|S )a  
        Forward pass of the GraniteMoeParallelExperts module.

        Args:
            inputs (Tensor):
                Input tensor.
            expert_size:
                Expert size information.

        Returns:
            Tensor: Output tensor.
        r   rp   )	splitranger   appendFlinearr,   r*   rv   )r.   inputsexpert_size
input_listoutput_listiresultss          r1   r@   z!GraniteMoeParallelExperts.forward   sq     \\+1\5
t''( 	HAqxx
1t{{1~FG	H))KQ/r2   rE   rF   rG   r   r(   r@   rJ   rK   s   @r1   r   r      s)    'C 'S 's 't '.r2   r   c                   2     e Zd Zdededef fdZd Z xZS )GraniteMoeTopKGatingr   r   top_kc                     t         |           || _        || _        || _        t        j                  ||d      | _        y)a  
        Initialize the top-k gating mechanism.

        Args:
            input_size (`int`):
                Size of the input.
            num_experts (`int`):
                Number of experts.
            top_k (`int`):
                Number of top experts to select.
        FbiasN)r'   r(   r   r   r   r   Linearlayer)r.   r   r   r   r0   s       r1   r(   zGraniteMoeTopKGating.__init__   s:     	&$
YYz;UC
r2   c                    | j                  |      j                         }|j                  | j                  d      \  }}t	        j
                  |d      j                  |      }t	        j                  |j                  d      | j                  g|j                  |j                        }|j                  d|d      }|j                         j                  d      }|j                         }|j!                         }	|	j#                  d      \  }
}|j%                  | j                  d      }|j!                         }||   }|||||fS )Nr   rp   r   r8   r]   trunc)rounding_mode)r   rH   topkr   r*   softmaxtype_aszerossizer   r8   r]   scatterlongsumtolistflattensortdiv)r.   r3   logitstop_k_logitstop_k_indicestop_k_gatesr   gatesr   top_k_experts_index_sorted_expertsbatch_indexbatch_gatess                 r1   r@   zGraniteMoeTopKGating.forward   s.   M*002&,kk$**!k&D#mmmLa8@@O a $"2"23;;L;LU`UgUg
 a2jjl&&q) "((* &--/"/"4"4Q"7*..tzz.Q "))+!"67#[+{FRRr2   r   rK   s   @r1   r   r      s'    D3 DS D D(Sr2   r   c                   .     e Zd ZdZdef fdZd Z xZS )GraniteMoeMoEz
    A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts.

    Args:
        config:
            Configuration object with model hyperparameters.
    rO   c                    t         |           |j                  | _        |j                  | _        t
        |j                     | _        t        |j                  | j                  | j                  dz        | _
        t        |j                  | j                  | j                        | _        t        | j                  |j                  |j                        | _        y )Nr5   )r   r   r   )r'   r(   r/   r   intermediate_sizer	   
hidden_act
activationr   num_local_expertsinput_linearoutput_linearr   num_experts_per_tokrouterr.   rO   r0   s     r1   r(   zGraniteMoeMoE.__init__   s     ,,!33 !2!235f6N6NPTP_P_aeaqaqtuauv6v7O7OQUQaQacgcrcrs*00,,
r2   c                    |j                         \  }}}|j                  d|      }| j                  |      \  }}}}}||   }	| j                  |	|      }
|
j	                  dd      }| j                  |d         |d   z  }
| j                  |
|      }||d d d f   z  }t        j                  ||z  | j                  f|j                  |j                        }|j                  d||      }|j                  ||| j                        }|S )Nr6   r5   rp   r   r   r   )r   reshaper   r   chunkr   r   r*   r   r   r8   r]   	index_addview)r.   layer_inputbszlengthemb_sizer   r   r   r   expert_inputsr3   chunked_hidden_statesexpert_outputsr   layer_outputs                  r1   r@   zGraniteMoeMoE.forward   s    + 0 0 2VX!))"h76:kk+6N3;[!#K0))-E - 3 3A2 3 >(=a(@ADYZ[D\\++M;G'+ag*>>S6\4??;>CWCW`n`u`uvq+~F#((fdooFr2   )rE   rF   rG   __doc__r    r(   r@   rJ   rK   s   @r1   r   r      s    
/ 
r2   r   c                     | 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..Nr6   r5   rp   )rC   r*   rv   )ry   x1x2s      r1   rotate_halfr     sZ    	
3"!''"+"""	#B	
3q ""	#B99rc2YB''r2   rotary_pos_embc                     |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.
    )	unsqueezer   )qkrw   rx   unsqueeze_dimq_embedk_embeds          r1   apply_rotary_pos_embr     sY    & --
&C
--
&C3w;q>C/0G3w;q>C/0GGr2   r3   n_repr%   c                     | j                   \  }}}}|dk(  r| S | dddddddddf   j                  |||||      } | j                  |||z  ||      S )z
    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)rC   rq   r   )r3   r   batchnum_key_value_headsslenrb   s         r1   	repeat_kvr   .  so    
 2?1D1D.Ehz!!Qa"23::5BUW\^bdlmM  (;e(CT8TTr2   modulequerykeyvalueattention_maskscalingdropoutkwargsc                    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 )Nr5   r   r6   )ri   r8   )ptrainingr   )r   num_key_value_groupsr*   matmulru   r   r   r   r:   r9   r8   r   r   
contiguous)r   r   r   r   r   r   r   r   
key_statesvalue_statesattn_weightsattn_outputs               r1   eager_attention_forwardr   :  s     3 ; ;<JUF$?$?@L<<z';';Aq'ABWLL!#n4==((2U]](SVVW\WbWbcL==((6??([L,,|\:K''1-88:K$$r2   c                       e Zd ZdZdedef fdZ	 	 	 	 ddej                  de	ej                  ej                  f   dz  dej                  dz  d	e
dz  d
ej                  dz  dee   de	ej                  ej                  f   fdZ xZS )GraniteMoeAttentionz=Multi-headed attention from 'Attention Is All You Need' paperrO   	layer_idxc                 ^   t         |           || _        || _        t	        |d|j
                  |j                  z        | _        |j                  |j                  z  | _	        |j                  | _        |j                  | _        d| _        t        j                  |j
                  |j                  | j                  z  |j                         | _        t        j                  |j
                  |j                  | j                  z  |j                         | _        t        j                  |j
                  |j                  | j                  z  |j                         | _        t        j                  |j                  | j                  z  |j
                  |j                         | _        y )Nrb   Tr   )r'   r(   rO   r   rd   r/   re   rb   r   r   attention_multiplierr   attention_dropout	is_causalr   r   attention_biasq_projk_projv_projo_projr.   rO   r   r0   s      r1   r(   zGraniteMoeAttention.__init__W  sJ   "
F4F4F&JdJd4de$*$>$>&B\B\$\!22!'!9!9ii : :T]] JQWQfQf
 ii : :T]] JQWQfQf
 ii : :T]] JQWQfQf
 ii&&68J8JQWQfQf
r2   Nr3   position_embeddingsr   past_key_valuescache_positionr   r%   c                    |j                   d d }g |d| j                  }| j                  |      j                  |      j	                  dd      }	| j                  |      j                  |      j	                  dd      }
| j                  |      j                  |      j	                  dd      }|\  }}t        |	|
||      \  }	}
|'|||d}|j                  |
|| j                  |      \  }
}t        j                  | j                  j                  t              } || |	|
||f| j                  sdn| j                   | j"                  d|\  }} |j$                  g |d j'                         }| j)                  |      }||fS )Nr6   r   r5   )rx   rw   r          )r   r   )rC   rb   r  r   ru   r  r  r   updater   r   get_interfacerO   _attn_implementationr   r   r   r   r   r   r  )r.   r3   r  r   r  r  r   input_shapehidden_shapequery_statesr   r   rw   rx   cache_kwargsattention_interfacer   r   s                     r1   r@   zGraniteMoeAttention.forwardn  s    $))#2.88b8$--8{{=166|DNNqRST[[/44\BLLQPQR
{{=166|DNNqRST&S#7jRUWZ#[ j&#&snUL'6'='=j,X\XfXfht'u$J(?(M(MKK,,.E)
 %8	%
  $}}C$2H2HLL	%
 	%
!\ *k));;;;FFHkk+.L((r2   NNNN)rE   rF   rG   r   r    r   r(   r*   rI   rB   r
   
LongTensorr   r   r@   rJ   rK   s   @r1   r   r   S  s    G
/ 
C 
4 IM.2(,26))||)) #5<<#=>E)) t+	))
 )) ((4/)) +,)) 
u||U\\)	*))r2   r   c                        e Zd Zdedef fdZ	 	 	 	 ddej                  dej                  dz  dedz  dej                  dz  d	e
ej                  ej                  f   dz  d
ej                  fdZ xZS )GraniteMoeDecoderLayerrO   r   c                 B   t         |           |j                  | _        t        ||      | _        t        |j                  |j                        | _        t        |j                  |j                        | _        t        |      | _
        |j                  | _        y )N)rO   r   r$   )r'   r(   r/   r   	self_attnr#   rms_norm_epsinput_layernormpost_attention_layernormr   block_sparse_moeresidual_multiplierr  s      r1   r(   zGraniteMoeDecoderLayer.__init__  s|    !--,FiP01C1CI\I\](9&:L:LRXReRe(f% -f 5#)#=#= r2   Nr3   r   r  r  r  r%   c           	          |}| j                  |      } | j                  d|||||d|\  }}||| j                  z  z   }|}| j                  |      }| j	                  |      }||| j                  z  z   }|S )N)r3   r   r  r  r   )r  r  r  r  r  )	r.   r3   r   r  r  r  r   residualr   s	            r1   r@   zGraniteMoeDecoderLayer.forward  s     !,,];)4>> 
')+) 3
 
q !=43K3K#KK 55mD--m< =43K3K#KKr2   r  )rE   rF   rG   r    r   r(   r*   rI   r
   r  rB   r@   rJ   rK   s   @r1   r  r    s    >/ >C > /3(,26HL|| t+ 	
 ((4/ #5<<#=>E 
r2   r  c                        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Z ej$                          fd       Z xZS )	GraniteMoePreTrainedModelrO   modelTr  r  F)r3   
attentionsc                     t         |   |       t        |t              r7t	        j
                  |j                  d| j                  j                         y y )Nr
  )r<   std)	r'   _init_weightsrr   r   initnormal_r,   rO   initializer_range)r.   r   r0   s     r1   r(  z'GraniteMoePreTrainedModel._init_weights  s>    f%f78LLSdkk6S6ST 9r2   )rE   rF   rG   r    r   base_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_supports_flex_attn_can_compile_fullgraph_supports_attention_backendr  r   _can_record_outputsr*   r   r(  rJ   rK   s   @r1   r#  r#    sp    &*#12#4"5N""&/)
 U]]_U Ur2   r#  c                       e Zd Zdef 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j                  dz  dee   defd                     Z xZS )GraniteMoeModelrO   c           	      .   t         |   |       |j                  | _        |j                  | _        t        j                  |j                  |j                  | j                        | _        t        j                  t        |j                        D cg c]  }t        ||       c}      | _        t        |j                  |j                        | _        t#        |      | _        d| _        |j(                  | _        | j+                          y c c}w )Nr  rO   F)r'   r(   pad_token_idpadding_idx
vocab_sizer   	Embeddingr/   embed_tokens
ModuleListr   num_hidden_layersr  layersr#   r  normrM   
rotary_embgradient_checkpointingembedding_multiplier	post_initr  s      r1   r(   zGraniteMoeModel.__init__  s     !.. ++LL):):F<N<NPTP`P`ammHMfNfNfHgh9#FI6h
 &f&8&8f>Q>QR	36B&+#$*$?$?! 	 is   DN	input_idsr   rz   r  inputs_embeds	use_cacher  r   r%   c                 b   |d u |d uz  rt        d      |r|t        | j                        }|| j                  |      }|F||j	                         nd}	t        j                  |	|	|j                  d   z   |j                        }||j                  d      }t        | j                  |||||      }
|| j                  z  }|}| j                  ||      }| j                  d | j                  j                   D ]  } ||f||
||||d|} | j                  |      }t!        ||      S )	Nz:You must specify exactly one of input_ids or inputs_embedsr9  r   r   )r]   )rO   rH  r   r  r  rz   )r  r   rz   r  rI  r  )last_hidden_stater  )
ValueErrorr   rO   r>  get_seq_lengthr*   rf   rC   r]   r   r   rE  rC  rA  r@  rB  r   )r.   rG  r   rz   r  rH  rI  r  r   past_seen_tokenscausal_maskr3   r  decoder_layers                 r1   r@   zGraniteMoeModel.forward  sp    -t";<YZZ0*$++>O  --i8M!CRC^==?de"\\ "2]5H5H5K"KTaThThN )33A6L(;;'))+%
 &(A(AA% #oom\J![[)H4;;+H+HI 
	M)	$7*) /#-	 	M
	 		-0%++
 	
r2   )NNNNNNN)rE   rF   rG   r    r(   r   r   r   r*   r  rI   r
   FloatTensorboolr   r   r   r@   rJ   rK   s   @r1   r7  r7    s    / "   .2.204(,26!%26;
##d*;
 t+;
 &&-	;

 ;
 ((4/;
 $;;
 ((4/;
 +,;
 
 ;
    ;
r2   r7  gate_logitsr   c                    | t        | t              syt        | t              rC| d   j                  }t        j                  | D cg c]  }|j                  |       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 )a  
    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   rp   r6   )rr   rB   r]   r*   rv   r9   r   r   r   r   one_hotr<   rH   rC   rq   r   r   r   )rS  r   r   r   compute_device
layer_gateconcatenated_gate_logitsrouting_weightsr   selected_expertsexpert_masktokens_per_expertrouter_prob_per_expert
batch_sizesequence_lengthr@  expert_attention_mask router_per_expert_attention_maskoverall_losss                      r1   load_balancing_loss_funcrc  +  s9   : *[%"@+u%$Q..#(99^i-jPZjmmN.K-jpq#r hh))112JPR1SO**_eDA((%%--.>LK!JJ{'8'8':B "'O!C&4&:&:#
O4::1=*B^_ 4AtT12V&
OUKXYWR,R	 	 "IIk&7&7&9<Q&QWXY\a\e\e!q]
 
 4At+,V&
O[QRWR%R	 	) "'?=]+]cd!ehmhqhq,!i
 "
 99.1G1Q1QRS1TTUL+%%[ .ks   Ic                   d    e Zd ZddiZddiZddgdgfiZdef fdZee		 	 	 	 	 	 	 	 	 dd
e
j                  d	z  de
j                  d	z  de
j                  d	z  ded	z  de
j                  d	z  de
j                  d	z  ded	z  de
j                  d	z  dee
j                  z  deez  fd              Z xZS )GraniteMoeForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr3   r   rO   c                 p   t         |   |       t        |      | _        |j                  | _        t        j                  |j                  |j                  d      | _        |j                  | _	        |j                  | _        |j                  | _        |j                  | _        | j                          y )NFr   )r'   r(   r7  r$  r<  r   r   r/   rf  router_aux_loss_coefr   r   r   logits_scalingrF  r   s     r1   r(   zGraniteMoeForCausalLM.__init__  s     $V,
 ++yy!3!3V5F5FUS$*$?$?!!33#)#=#= $33 	r2   NrG  r   rz   r  rH  labelsoutput_router_logitsr  logits_to_keepr%   c
           
         ||n| j                   j                  } | j                  d||||||d|
}|j                  }t	        |	t
              rt        |	 d      n|	}| j                  |dd|ddf         }|| j                   j                  z  }d}|* | j                  ||fd| j                   j                  i|
}d}|rYt        |j                  | j                  | j                  |      }|+|| j                  |j!                  |j"                        z  z  }t%        ||||j&                  |j(                  |j*                  |j                        S )al  
        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, GraniteMoeForCausalLM

        >>> model = GraniteMoeForCausalLM.from_pretrained("ibm/PowerMoE-3b")
        >>> tokenizer = AutoTokenizer.from_pretrained("ibm/PowerMoE-3b")

        >>> 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)rG  r   rz   r  rH  r  r<  )lossaux_lossr   r  r3   r%  router_logitsr   )rO   rl  r$  rK  rr   r   slicerf  rj  loss_functionr<  rc  rq  r   r   ri  r9   r]   r   r  r3   r%  )r.   rG  r   rz   r  rH  rk  rl  r  rm  r   outputsr3   slice_indicesr   ro  rp  s                    r1   r@   zGraniteMoeForCausalLM.forward  s   L %9$D $++JjJj 	 $** 
)%+')
 
  118B>SV8W~ot4]kmA}a,?@A$++444%4%%  ;;11 	D /%%  ((	H !11HKK4LLL(#33!//))!//
 	
r2   )	NNNNNNNNr   )rE   rF   rG   _tied_weights_keys_tp_plan_pp_planr    r(   r   r   r*   r  rI   r
   rQ  rR  r   rB   r   r@   rJ   rK   s   @r1   re  re  }  s9   *,GH23H_-z:;H/   .2.204(,26*.,026-.S
##d*S
 t+S
 &&-	S

 S
 ((4/S
   4'S
 #TkS
 ((4/S
 ell*S
 
*	*S
  S
r2   re  )re  r7  r#  )r   )r
  )Nr5   N)Fcollections.abcr   typingr   r*   r   torch.nnr   r    r   r)  activationsr	   cache_utilsr
   r   
generationr   integrationsr   r   r   masking_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   utils.genericr   r   r   utils.output_capturingr   configuration_granitemoer    Moduler#   rM   r   r   r   r   r   rI   r   r   rH   r   r   r  r#  r7  rB   rc  re  __all__r   r2   r1   <module>r     sv  , %    $ & ! . ) f f / 9 Q K F & 7 Y Y 5 6 Y'J		 J (J(><		 ><B*		 *Z.S299 .Sb(BII (V( *+ ,2	UU\\ 	U# 	U%,, 	U& %II%<<% 
% <<	%
 LL4'% % % '(%2 )*C)")) C) +C)L"7 "J U U U. P
/ P
 P
j #
*.	O&ell 33d:O&tO& LL4'	O&
 \\CO&d g
5 g
 g
T Tr2   