
    Uhg                     4   d dl mZmZmZmZ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mZ ddl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"m#Z# ddl$m%Z%  e"       rd dl&m'Z' ddl(m)Z)  e#jT                  e+      Z, G d dejZ                        Z. G d dejZ                        Z/ G d dejZ                        Z0 G d dejZ                        Z1 G d dejZ                        Z2d Z3d:dZ4dejj                  de6d ejj                  fd!Z7	 d;d"ejZ                  d#ejj                  d$ejj                  d%ejj                  d&eejj                     d'e8d(e8fd)Z9 G d* d+ejZ                        Z: G d, d-e      Z;e! G d. d/e             Z< G d0 d1ejZ                        Z=e! G d2 d3e<             Z>	 	 	 d<d4eejj                  eejj                     df   d5ee6   d&eejj                     d eejj                  e6f   fd6Z? G d7 d8e<e      Z@g d9ZAy)=    )CallableListOptionalTupleUnionN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)AttentionMaskConverter)GradientCheckpointingLayer)BaseModelOutputWithPastMoeCausalLMOutputWithPastMoeModelOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)auto_docstringis_torch_flex_attn_availablelogging   )GraniteMoeSharedConfig)	BlockMask)make_flex_block_causal_maskc                   `     e Zd ZdZdef fdZdej                  dej                  fdZ xZ	S )GraniteMoeSharedMLPz~
    MLP layer for shared experts

    Args:
        config:
            Configuration object with model hyperparameters.
    configc                 h   t         t        |           |j                  | _        |j
                  | _        t        |j                     | _        t        j                  | j                  | j                  dz  d      | _        t        j                  | j                  | j                  d      | _        y )N   Fbias)superr   __init__hidden_size
input_sizeshared_intermediate_sizer
   
hidden_act
activationr   Linearinput_linearoutput_linearselfr    	__class__s     /var/www/catia.catastroantioquia-mas.com/valormas/lib/python3.12/site-packages/transformers/models/granitemoeshared/modeling_granitemoeshared.pyr&   zGraniteMoeSharedMLP.__init__:   s    !413 ,,!:: !2!23IIdoot7G7G!7KRWXYYt'7'7uU    hidden_statesreturnc                     | j                  |      }|j                  dd      }| j                  |d         |d   z  }| j                  |      }|S )Nr"   dimr   r   )r-   chunkr+   r.   )r0   r4   chunked_hidden_statess      r2   forwardzGraniteMoeSharedMLP.forwardC   s^    ))-8 - 3 3A2 3 >(=a(@ADYZ[D\\**=9r3   )
__name__
__module____qualname____doc__r   r&   torchTensorr<   __classcell__r1   s   @r2   r   r   1   s2    V5 VU\\ ell r3   r   c                   ,     e Zd Zd fd	Zd Zd Z xZS )GraniteMoeSharedRMSNormc                     t         |           t        j                  t	        j
                  |            | _        || _        y)zF
        GraniteMoeSharedRMSNorm is equivalent to T5LayerNorm
        N)r%   r&   r   	ParameterrA   onesweightvariance_epsilon)r0   r'   epsr1   s      r2   r&   z GraniteMoeSharedRMSNorm.__init__L   s1     	ll5::k#:; #r3   c                 "   |j                   }|j                  t        j                        }|j	                  d      j                  dd      }|t        j                  || j                  z         z  }| j                  |j                  |      z  S )Nr"   r7   T)keepdim)	dtypetorA   float32powmeanrsqrtrK   rJ   )r0   r4   input_dtypevariances       r2   r<   zGraniteMoeSharedRMSNorm.forwardT   sy    #))%((7 $$Q',,R,>%Ht?T?T4T(UU{{]--k:::r3   c                 ^    t        | j                  j                         d| j                   S )Nz, eps=)tuplerJ   shaperK   r0   s    r2   
extra_reprz"GraniteMoeSharedRMSNorm.extra_repr[   s*    ))*+6$2G2G1HIIr3   )gư>)r=   r>   r?   r&   r<   r[   rC   rD   s   @r2   rF   rF   K   s    $;Jr3   rF   c                   6     e Zd Zdedededdf fdZd Z xZS )GraniteMoeSharedParallelExpertsnum_expertsr(   output_sizer5   Nc                     t         |           t        j                  t	        j
                  |||            | _        || _        || _        || _	        y)a  
        Initialize the GraniteMoeSharedParallelExperts 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   rH   rA   emptyrJ   r^   r(   r_   )r0   r^   r(   r_   r1   s       r2   r&   z(GraniteMoeSharedParallelExperts.__init__`   sD    " 	ll5;;{K#TU&$&r3   c                     |j                  |d      }g }t        | j                        D ]7  }|j                  t	        j
                  ||   | j                  |                9 t        j                  |d      }|S )a  
        Forward pass of the GraniteMoeSharedParallelExperts module.

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

        Returns:
            Tensor: Output tensor.
        r   r8   )	splitranger^   appendFlinearrJ   rA   cat)r0   inputsexpert_size
input_listoutput_listiresultss          r2   r<   z'GraniteMoeSharedParallelExperts.forwardw   sq     \\+1\5
t''( 	HAqxx
1t{{1~FG	H))KQ/r3   r=   r>   r?   intr&   r<   rC   rD   s   @r2   r]   r]   _   s)    'C 'S 's 't '.r3   r]   c                   2     e Zd Zdededef fdZd Z xZS )GraniteMoeSharedTopKGatingr(   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.
        Fr#   N)r%   r&   r^   r(   rs   r   r,   layer)r0   r(   r^   rs   r1   s       r2   r&   z#GraniteMoeSharedTopKGating.__init__   s:     	&$
YYz;UC
r3   c                    | j                  |      j                         }|j                  | j                  d      \  }}t	        j
                  |d      j                  |      }t	        j                  |j                  d      | j                  g|j                  |j                        }|j                  d|d      }|j                         j                  d      }|j                         }|j!                         }	|	j#                  d      \  }
}|j%                  | j                  d      }|j!                         }||   }|||||fS )Nr   r8   r   rO   devicetrunc)rounding_mode)ru   floattopkrs   rA   softmaxtype_aszerossizer^   rO   rx   scatterlongsumtolistflattensortdiv)r0   r4   logitstop_k_logitstop_k_indicestop_k_gatesr   gatesrj   top_k_experts_index_sorted_expertsbatch_indexbatch_gatess                 r2   r<   z"GraniteMoeSharedTopKGating.forward   s.   M*002&,kk$**!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Rr3   ro   rD   s   @r2   rr   rr      s'    D3 DS D D&Sr3   rr   c                   .     e Zd ZdZdef fdZd Z xZS )GraniteMoeSharedMoEz
    A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts.

    Args:
        config:
            Configuration object with model hyperparameters.
    r    c                    t         t        |           |j                  | _        |j
                  | _        t        |j                     | _        t        |j                  | j                  | j                  dz        | _        t        |j                  | j                  | j                        | _        t        | j                  |j                  |j                        | _        y )Nr"   )r(   r^   rs   )r%   r   r&   r'   r(   intermediate_sizer
   r*   r+   r]   num_local_expertsr-   r.   rr   num_experts_per_tokrouterr/   s     r2   r&   zGraniteMoeSharedMoE.__init__   s    !413 ,,!33 !2!23;$$doot7G7G!7K
 =$$d&6&6
 100,,
r3   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                        }||	fS )a  
        Forward pass of the mixture of experts layer.

        Args:
            layer_input (Tensor):
                Input tensor.

        Returns:
            Tensor:
                Output tensor.
            Tensor:
                Router logits.
        r7   r"   r8   r   r   Nrw   )r   reshaper   r-   r:   r+   r.   rA   r   r(   rO   rx   	index_addview)r0   layer_inputbszlengthemb_sizer   r   r   rj   router_logitsexpert_inputsr4   r;   expert_outputsr   layer_outputs                   r2   r<   zGraniteMoeSharedMoE.forward   s    !, 0 0 2VX!))"h7BF++kBZ?;[-#K0))-E - 3 3A2 3 >(=a(@ADYZ[D\\++M;G'+ag*>>S6\4??;>CWCW`n`u`uvq+~F#((fdooF]**r3   )r=   r>   r?   r@   r   r&   r<   rC   rD   s   @r2   r   r      s    
5 
&+r3   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..Nr7   r"   r8   )rY   rA   rh   )xx1x2s      r2   rotate_halfr      sZ    	
3"!''"+"""	#B	
3q ""	#B99rc2YB''r3   c                     |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.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        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kcossinposition_idsunsqueeze_dimq_embedk_embeds           r2   apply_rotary_pos_embr      sY    ( --
&C
--
&C3w;q>C/0G3w;q>C/0GGr3   r4   n_repr5   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)rY   expandr   )r4   r   batchnum_key_value_headsslenhead_dims         r2   	repeat_kvr     so    
 2?1D1D.Ehz!!Qa"23::5BUW\^bdlmM  (;e(CT8TTr3   modulequerykeyvalueattention_maskscalingdropoutc                 T   t        || j                        }t        || j                        }	t        j                  ||j	                  dd            |z  }
|#|d d d d d d d |j
                  d   f   }|
|z   }
t        j                  j                  |
dt        j                        j                  |j                        }
t        j                  j                  |
|| j                        }
t        j                  |
|	      }|j	                  dd      j                         }||
fS )Nr"   r	   r7   )r9   rO   )ptrainingr   )r   num_key_value_groupsrA   matmul	transposerY   r   
functionalr}   rQ   rP   rO   r   r   
contiguous)r   r   r   r   r   r   r   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputs                r2   eager_attention_forwardr   &  s    3 ; ;<JUF$?$?@L<<z';';Aq'ABWLL!$Q1.D
0@0@0D.D%DE#k1 ==((2U]](SVVW\WbWbcL==((6??([L,,|\:K''1-88:K$$r3   c                   d    e Zd ZdZddedee   f fdZ	 	 	 	 	 	 ddej                  deej                     deej                     dee   d	ed
eej                     deeej                  ej                  f      deej                  eej                     eeej                        f   fdZ xZS )GraniteMoeSharedAttentionz=Multi-headed attention from 'Attention Is All You Need' paperr    	layer_idxc                    t         |           || _        || _        |-t        j                  d| j                  j                   d       |j                  | _        |j                  | _	        |j                  | _        | j                  | j                  z  | _        |j                  | _        | j                  | j                  z  | _        d| _        |j                   | _        | j                  | j                  z  | j                  k7  r&t%        d| j                   d| 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                  |j*                        | _        y )NzInstantiating z without passing a `layer_idx` is not recommended and will lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` when creating this class.Tz?hidden_size must be divisible by num_heads (got `hidden_size`: z and `num_heads`: z).r#   )r%   r&   r    r   loggerwarning_oncer1   r=   attention_dropoutr'   num_attention_heads	num_headsr   r   r   	is_causalattention_multiplierr   
ValueErrorr   r,   attention_biasq_projk_projv_projo_projr0   r    r   r1   s      r2   r&   z"GraniteMoeSharedAttention.__init__F  s   " !8!8 9 :, , "(!9!9!--33((DNN:#)#=#= $(NNd6N6N$N!22MMDNN*t/?/??QRVRbRbQc$T^^$4B8 
 ii 0 0$..4==2PW]WlWlmii 0 0$2J2JT]]2Zagavavwii 0 0$2J2JT]]2Zagavavwii 0 0$2B2BI^I^_r3   r4   r   r   past_key_value	use_cachecache_positionposition_embeddingsr5   c                     |j                         \  }	}
}| j                  |      }| j                  |      }| j                  |      }|j	                  |	|
| j
                  | j                        j                  dd      }|j	                  |	|
| j                  | j                        j                  dd      }|j	                  |	|
| j                  | j                        j                  dd      }||nd\  }}|t        ||||      \  }}|'|||d}|j                  ||| j                  |      \  }}t        }| j                  j                  dk7  r^| j                  j                  dk(  r(|j                  dd      rt         j#                  d	       nt$        | j                  j                     } || ||||f| j&                  sd
n| j(                  | j*                  d|\  }}|j	                  |	|
d      }| j-                  |      }|||fS )Nr   r"   )NN)r   r   r   eagersdpaoutput_attentionsFz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.        )r   r   r7   )r   r   r   r   r   r   r   r   r   r   updater   r   r    _attn_implementationgetr   r   r   r   r   r   r   )r0   r4   r   r   r   r   r   r   r   r   q_lenr   query_statesr   r   r   r   cache_kwargsattention_interfacer   r   s                        r2   r<   z!GraniteMoeSharedAttention.forwardf  s
    &**,UA{{=1[[/
{{=1#((eT^^T]]S]]^_abc__S%1I1I4==Yccdeghi
#((eT5M5Mt}}]gghiklm*=*I&|S*';L*VY[^'_$L*%#&snUL'5'<'<ZW[WeWegs't$J(?;;++w6{{//69fjjI\^c>d##L
 '>dkk>^>^&_#$7	%
  $}}C$2H2HLL	%
 	%
!\ "&&sE26kk+.L.88r3   N)NNNFNN)r=   r>   r?   r@   r   r   rp   r&   rA   rB   
LongTensorr   boolr   r<   rC   rD   s   @r2   r   r   C  s    G`5 `(3- `F 2637*.59KO69||69 !.69 u//0	69
 !69 69 !!1!1269 &eELL%,,,F&GH69 
u||Xell3XeELL>Q5RR	S69r3   r   c                   r    e Zd Zdedef fdZ	 	 	 	 	 	 	 	 ddej                  deej                     deej                     dee
   dee   d	ee   d
eej                     dee   deeej                  ej                  f      deej                  eeej                  ej                  f      f   fdZ xZS )GraniteMoeSharedDecoderLayerr    r   c                    t         |           |j                  | _        t        ||      | _        t        |      | _        t        |j                  |j                        | _	        t        |j                  |j                        | _
        |j                  | _        |j                  dk(  rd | _        y t        |      | _        y )N)r    r   rL   r   )r%   r&   r'   r   	self_attnr   block_sparse_moerF   rms_norm_epsinput_layernormpost_attention_layernormresidual_multiplierr)   r   
shared_mlpr   s      r2   r&   z%GraniteMoeSharedDecoderLayer.__init__  s    !--2&IV 3F ;6v7I7IvObObc(?@R@RX^XkXk(l%#)#=#= "("A"AQ"F$L_`fLgr3   r4   r   r   r   r   r   r   output_router_logitsr   r5   c
                 |   |}| j                  |      } | j                  d||||||||	d|
\  }}}||| j                  z  z   }|}| j                  |      }| j	                  |      \  }}| j
                  |}n|| j                  |      z   }||| j                  z  z   }|f}|r||fz  }|r||fz  }|r||fz  }|S )a  
        Args:
            hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
            attention_mask (`torch.FloatTensor`, *optional*):
                attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1,
                query_sequence_length, key_sequence_length)` if default attention is used.
            output_attentions (`bool`, *optional*):
                Whether or not to return the attentions tensors of all attention layers. See `attentions` under
                returned tensors for more detail.
            use_cache (`bool`, *optional*):
                If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
                (see `past_key_values`).
            past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence
            output_router_logits (`bool`, *optional*):
                Whether or not to return the logits of all the routers. They are useful for computing the router loss, and
                should not be returned during inference.
            position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
                Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
                with `head_dim` being the embedding dimension of each attention head.
            kwargs (`dict`, *optional*):
                Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
                into the model
        )r4   r   r   r   r   r   r   r    )r   r   r  r  r   r  )r0   r4   r   r   r   r   r   r   r  r   r   residualself_attn_weightspresent_key_valuemoe_hidden_statesr   outputss                    r2   r<   z$GraniteMoeSharedDecoderLayer.forward  s   L !,,]; ?Mdnn 
?
')%)/) 3
?
 
?
;(*; !=43K3K#KK !55mD+/+@+@+O(=??"-M-0NNM =43K3K#KK ")++G)++G''Gr3   )NNNFFNFN)r=   r>   r?   r   rp   r&   rA   rB   r   r   r   r   r   FloatTensorr<   rC   rD   s   @r2   r   r     s   
h5 
h# 
h 2637*.,1$)59/4KOP||P !.P u//0	P
 !P $D>P D>P !!1!12P 'tnP &eELL%,,,F&GHP 
u  (51B1BEDUDU1U+V"WW	XPr3   r   c                   >    e Zd ZeZdZdZdgZdgZdZ	dZ
dZdZdZd Zy)GraniteMoeSharedPreTrainedModelmodelTr   past_key_valuesFc                 >   t        |t        j                        rm|j                  j                  j                  d| j                  j                         |j                  %|j                  j                  j                          y y t        |t        j                        rz|j                  j                  j                  d| j                  j                         |j                  2|j                  j                  |j                     j                          y y t        |t              r&|j                  j                  j                  d       y t        |t              r<|j                  j                  j                  d| j                  j                         y y )Nr   )rS   stdg      ?)
isinstancer   r,   rJ   datanormal_r    initializer_ranger$   zero_	Embeddingpadding_idxrF   fill_r]   )r0   r   s     r2   _init_weightsz-GraniteMoeSharedPreTrainedModel._init_weights  s   fbii(MM&&CT[[5R5R&S{{&  &&( '-MM&&CT[[5R5R&S!!-""6#5#56<<> . 78MM$$S) ?@MM&&CT[[5R5R&S Ar3   N)r=   r>   r?   r   config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_2_supports_sdpa_supports_cache_class_supports_quantized_cache_supports_static_cacher  r  r3   r2   r  r    sH    )L&*#78#4"5!N  $"Tr3   r  c                   ^     e Zd Zddef fdZ ej                         ed               Z xZ	S )GraniteMoeSharedRotaryEmbeddingr    c                    t         |           t        |d      rG|j                  ;|j                  j	                  d|j                  j	                  d            | _        nd| _        |j                  | _        |j                  | _        || _	        t        | j
                     | _        | j                  | j                  |      \  }| _        | j                  d|d       | j                  | _        y )Nrope_scaling	rope_typetypedefaultinv_freqF)
persistent)r%   r&   hasattrr)  r   r*  max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr    r   rope_init_fnattention_scalingregister_bufferr-  original_inv_freq)r0   r    rx   r-  r1   s       r2   r&   z(GraniteMoeSharedRotaryEmbedding.__init__  s    6>*v/B/B/N#0044[&BUBUBYBYZ`BabDN&DN"("@"@$*$B$B!/?+/+<+<T[[&+Q($(ZeD!%r3   c                 b   | 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        j                  |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   r7   r   mpscpuF)device_typeenabledr"   r8   )rO   )r-  r{   r   rY   rP   rx   r  r+  strrA   autocastr   rh   r   r4  r   rO   )
r0   r   r   inv_freq_expandedposition_ids_expandedr:  freqsembr   r   s
             r2   r<   z'GraniteMoeSharedRotaryEmbedding.forward-  sV    !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.r   )
r=   r>   r?   r   r&   rA   no_gradr   r<   rC   rD   s   @r2   r'  r'    s4    /5 /" U]]_<  <r3   r'  c                   *    e Zd Zdef fdZd Zd Ze	 	 	 	 	 	 	 	 	 	 	 ddee	j                     dee	j                     dee	j                     deeeee	j                     f      d	ee	j                     d
ee   dee   dee   dee   dee   dee	j                     deeef   fd       Z	 ddee	j                  df   de	j                  de	j                  dedef
dZede	j                  dedede	j.                  de	j                  defd       Z xZS )GraniteMoeSharedModelr    c           	      4   t         |   |       |j                  | _        |j                  | _        t        j                  |j                  |j                  | j                        | _        t        j                  t        |j                        D cg c]  }t        ||       c}      | _        t        |j                  |j                        | _        d| _        |j$                  | _        |j                  | _        |j&                  | _        | j                  | j(                  z  | _        |j,                  | _        |j.                  | _        |j0                  | _        | j0                  dk(  rt3        |      nd | _        | j7                          y c c}w )Nr   Frope)r%   r&   pad_token_idr  
vocab_sizer   r  r'   embed_tokens
ModuleListrd   num_hidden_layersr   layersrF   r   normgradient_checkpointingembedding_multiplierr   r   r   r0  
rope_thetaposition_embedding_typer'  
rotary_emb	post_initr   s      r2   r&   zGraniteMoeSharedModel.__init__?  sA    !.. ++LL):):F<N<NPTP`P`ammNSTZTlTlNmn)&)<n
 ,F,>,>FDWDWX	&+#$*$?$?!!--33((DNN:'-'E'E$ ++'-'E'E$EIEaEaekEk9&Aqu 	! os   Fc                     | j                   S r   rI  rZ   s    r2   get_input_embeddingsz*GraniteMoeSharedModel.get_input_embeddingsX  s       r3   c                     || _         y r   rU  r0   r   s     r2   set_input_embeddingsz*GraniteMoeSharedModel.set_input_embeddings[  s
    !r3   	input_idsr   r   r  inputs_embedsr   r   output_hidden_statesr  return_dictr   r5   c                    ||n| j                   j                  }||n| j                   j                  }||n| j                   j                  }|
|
n| j                   j                  }
|d u |d uz  rt        d      | j                  r%| j                  r|rt        j                  d       d}|| j                  |      }|| j                  z  }d}|r<t        |t              s,d}t        j                  |      }t        j                  d       |F||j!                         nd}t#        j$                  |||j&                  d   z   |j(                        }||j+                  d      }| j-                  |||||      }|}d }| j.                  | j/                  ||      }|rd	nd }|rd	nd }|	rd	nd }d }| j0                  D ]B  }|r||fz  } |||||||||	|
	      }|d   }|r	||rdnd   }|r	||d   fz  }|	s:||d   fz  }D | j3                  |      }|r||fz  }|r|nd }|r|j5                         }|
st7        d ||||fD              S t9        |||||      S )Nz:You must specify exactly one of input_ids or inputs_embedszX`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`.FTzWe detected that you are passing `past_key_values` as a tuple and this is deprecated and will be removed in v4.43. Please use an appropriate `Cache` class (https://huggingface.co/docs/transformers/v4.41.3/en/internal/generation_utils#transformers.Cache)r   r   rx   r  )r   r   r   r   r   r   r  r   r"   r7   c              3   &   K   | ]	  }||  y wr   r  ).0vs     r2   	<genexpr>z0GraniteMoeSharedModel.forward.<locals>.<genexpr>  s     tqfgfsts   )last_hidden_stater  r4   
attentionsr   )r    r   r\  r   use_return_dictr   rN  r   r   r   rI  rO  r  r   r   from_legacy_cacheget_seq_lengthrA   arangerY   rx   r   _update_causal_maskrR  rL  rM  to_legacy_cacherX   r   )r0   rZ  r   r   r  r[  r   r   r\  r  r]  r   return_legacy_cachepast_seen_tokensr   r4   r   all_hidden_statesall_self_attnsall_router_logitsnext_decoder_cachedecoder_layerlayer_outputs
next_caches                           r2   r<   zGraniteMoeSharedModel.forward^  s    2C1N-TXT_T_TqTq$8$D $++JjJj 	 "+!6IDKK<Q<Q	%0%<k$++B]B]-t";<YZZ&&4==Yj I  --i8M%(A(AA#Z?"&*<<_MO]
 !CRC^==?de"\\ "2]5H5H5K"KTaThThN )33A6L..M>?L]

 &"??&"&//-"N #7BD0d"6BD!![[ 	:M#!m%55!)*)."3#-%9$7
M *!,M%28I1q%Q" =#3"55#!mB&7%99!3	:6 		-0  -!11+4'$
#335Jt]J@QSa$bttt%+&+%+
 	
r3   r   input_tensorc           	         | j                   j                  dk(  r||dk(  j                         r|S y | j                   j                  dk(  r't        |t        j
                        rt        |      }|S ||j                         nd}||j                  nd}| j                   j                  dk(  r(|s&|s$t        j                  |||| j                        ry |j                  }|j                  d   }	|r|j                         }
n1t        |t        j
                        r|j                  d	   n||	z   dz   }
| j                  ||	|
|||j                  d   
      }| j                   j                  dk(  rQ|O|j                   j"                  dv r7|s5t	        j$                  |      j&                  }t        j(                  ||      }|S )Nflash_attention_2r   flex_attentionr   Fr   )r[  past_key_values_lengthis_trainingr   r7   )sequence_lengthtarget_lengthrO   r   
batch_size)cudaxpunpu)r    r   anyr  rA   rB   r   rh  is_compileabler   _ignore_causal_mask_sdpar   rO   rY   get_max_cache_shape5_prepare_4d_causal_attention_mask_with_cache_positionrx   r+  finfomin_unmask_unattended)r0   r   ru  r   r  r   rm  using_compilable_cacherO   r{  r|  r   	min_dtypes                r2   rj  z)GraniteMoeSharedModel._update_causal_mask  s    ;;++/BB)~/D.I.I.K%%;;++/??.%,,7!<^!L!!
 @O?Z?99;`aCRC^!?!?di ;;++v5>T]n%>>*'7 MM	 ""&,,Q/!+??AM nell; $$R(%7!;  PP+')#))!, Q 
 KK,,6*%%**.DD%
 E*..I0CCKQZ[Kr3   r{  r|  rO   r}  c                    | | j                         dk(  r| }|S t        j                  |      j                  }t        j                  ||f|||j
                        }|dk7  rt        j                  |d      }|t        j                  ||j
                        |j                  dd      kD  z  }|ddddddf   j                  |ddd      }| |j                         }| j                  d   }	|ddddddd|	f   | ddddddf   j                  |j
                        z   }
|
dk(  }
|ddddddd|	f   j                  |
|      |ddddddd|	f<   |S )	aM  
        Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
        `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.

        Args:
            attention_mask (`torch.Tensor`):
                A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape
                `(batch_size, 1, query_length, key_value_length)`.
            sequence_length (`int`):
                The sequence length being processed.
            target_length (`int`):
                The target length: when generating with static cache, the mask should be as long as the static cache,
                to account for the 0 padding, the part of the cache that is not filled yet.
            dtype (`torch.dtype`):
                The dtype to use for the 4D attention mask.
            cache_position (`torch.Tensor`):
                Indices depicting the position of the input sequence tokens in the sequence.
            batch_size (`torch.Tensor`):
                Batch size.
        N   )
fill_valuerO   rx   r   )diagonalr_  r7   r   )r9   rA   r  r  fullrx   triuri  r   r   clonerY   rP   masked_fill)r   r{  r|  rO   r   r}  r   r   r  mask_lengthpadding_masks              r2   r  zKGraniteMoeSharedModel._prepare_4d_causal_attention_mask_with_cache_position  s   < %.*<*<*>!*C(K* ' E*..I** -0Ye\j\q\qK !##jjqA5<<n>S>STWeWmWmnprsWtttK%dD!Q&67>>z1bRTUK))//1,2226*1aL[L+@ANSTVZ\`bcScDdDgDg&&E    ,q05@Aq,;,AV5W5c5c )6Aq!\k\12 r3   )NNNNNNNNNNN)F)r=   r>   r?   r   r&   rV  rY  r   r   rA   r   rB   r   r   r   r  r   r   r   r<   rj  staticmethodrp   rO   r  rC   rD   s   @r2   rD  rD  =  s   5 2!"  151537KO59$(,0/3/3&*59s
E,,-s
 !.s
 u//0	s

 "%tE4E4E/F(F"GHs
   1 12s
 D>s
 $D>s
 'tns
 'tns
 d^s
 !!1!12s
 
u--	.s
 s
v #(BellK78B llB 	B
 B  BH 444 4 {{	4
 4 4 4r3   rD  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://arxiv.org/abs/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   r8   r7   )r  rX   rx   rA   rh   rP   r   r   r}   r|   one_hotrS   r{   rY   r   r   r   r   )r  r^   rs   r   compute_device
layer_gateconcatenated_gate_logitsrouting_weightsr   selected_expertsexpert_masktokens_per_expertrouter_prob_per_expertr}  r{  rK  expert_attention_mask router_per_expert_attention_maskoverall_losss                      r2   load_balancing_loss_funcr  P  s9   : *[%"@+u%$Q..#(99^i-jPZjmmN.K-jpq#r hh))112JPR1SO**_eDA((%%--.>LK!JJ{'8'8':B "'O!C&4&:&:#
O4::1=*B^_ 4AtT12V&
OUKXYWR,R	 	 "IIk&7&7&9<Q&QWXY\a\e\e!q]
 
 4At+,V&
O[QRWR%R	 	) "'?=]+]cd!ehmhqhq,!i
 "
 99.1G1Q1QRS1TTUL+%%[ .ks   Ic                        e Zd ZdgZdef fdZd Zd Zd Zd Z	d Z
d	 Ze	 	 	 	 	 	 	 	 	 	 	 	 	 dd
eej                     deej                      deej                     deeeeej(                     f      deej(                     deej                     dee   dee   dee   dee   dee   deej                     deeej                   f   deeef   fd       Zed        Z xZS )GraniteMoeSharedForCausalLMzlm_head.weightr    c                 N   t         |   |       t        |      | _        |j                  | _        t        j                  |j                  |j                  d      | _        |j                  | _	        |j                  | _        |j                  | _        | j                          y )NFr#   )r%   r&   rD  r  rH  r   r,   r'   lm_headrouter_aux_loss_coefr   r^   r   rS  r/   s     r2   r&   z$GraniteMoeSharedForCausalLM.__init__  s     *62
 ++yy!3!3V5F5FUS$*$?$?!!33#)#=#=  	r3   c                 .    | j                   j                  S r   r  rI  rZ   s    r2   rV  z0GraniteMoeSharedForCausalLM.get_input_embeddings  s    zz&&&r3   c                 &    || j                   _        y r   r  rX  s     r2   rY  z0GraniteMoeSharedForCausalLM.set_input_embeddings  s    "'

r3   c                     | j                   S r   r  rZ   s    r2   get_output_embeddingsz1GraniteMoeSharedForCausalLM.get_output_embeddings  s    ||r3   c                     || _         y r   r  )r0   new_embeddingss     r2   set_output_embeddingsz1GraniteMoeSharedForCausalLM.set_output_embeddings  s	    %r3   c                     || _         y r   r  )r0   decoders     r2   set_decoderz'GraniteMoeSharedForCausalLM.set_decoder  s	    
r3   c                     | j                   S r   r  rZ   s    r2   get_decoderz'GraniteMoeSharedForCausalLM.get_decoder  s    zzr3   rZ  r   r   r  r[  labelsr   r   r\  r  r]  r   logits_to_keepr5   c                    ||n| j                   j                  }|
|
n| j                   j                  }
|	|	n| j                   j                  }	||n| j                   j                  }| j                  ||||||||	|
||      }|d   }t        |t              rt        | d      n|}| j                  |dd|ddf         }|| j                   j                  z  }d}|:|j                         } | j                  ||fd| j                   j                  i|}d}|
r`t        |r|j                  n|d   | j                   | j"                  |      }|+|| j$                  |j'                  |j(                        z  z  }|s|f|dd z   }|
r|f|z   }||f|z   S |S t+        ||||j,                  |j.                  |j0                  |j                        S )ax  
        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, GraniteMoeSharedForCausalLM

        >>> model = GraniteMoeSharedForCausalLM.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)rZ  r   r   r  r[  r   r   r\  r  r]  r   r   rH  r7   r   )lossaux_lossr   r  r4   re  r   )r    r   r  r\  rf  r  r  rp   slicer  logits_scalingr{   loss_functionrH  r  r   r^   r   r  rP   rx   r   r  r4   re  )r0   rZ  r   r   r  r[  r  r   r   r\  r  r]  r   r  r   r  r4   slice_indicesr   r  r  outputs                         r2   r<   z#GraniteMoeSharedForCausalLM.forward  s   P 2C1N-TXT_T_TqTq$8$D $++JjJj 	 %9$D $++JjJj 	 &1%<k$++B]B] **)%+'/!5!5#)  
  
8B>SV8W~ot4]kmA}a,?@A$++444\\^F%4%%  ;;11 	D /)4%%'"+  ((	H !11HKK4LLLY,F#"v-'+'7D7V#CVC(#33!//))!//
 	
r3   c                 J    d}| D ]  }|t        fd|D              fz  } |S )Nr  c              3   t   K   | ]/  }|j                  d j                  |j                               1 yw)r   N)index_selectrP   rx   )ra  
past_statebeam_idxs     r2   rc  z=GraniteMoeSharedForCausalLM._reorder_cache.<locals>.<genexpr>6  s.     nU_j--aZ=N=N1OPns   58)rX   )r  r  reordered_past
layer_pasts    `  r2   _reorder_cachez*GraniteMoeSharedForCausalLM._reorder_cache1  s=    ) 	Jncmnn N	 r3   )NNNNNNNNNNNNr   )r=   r>   r?   _tied_weights_keysr   r&   rV  rY  r  r  r  r  r   r   rA   r   rB   r   r   r   r  r   rp   r   r   r<   r  r  rC   rD   s   @r2   r  r    s   *+5 '(&  151537KO59-1$(,0/3/3&*5934j
E,,-j
 !.j
 u//0	j

 "%tE4E4E/F(F"GHj
   1 12j
 ))*j
 D>j
 $D>j
 'tnj
 'tnj
 d^j
 !!1!12j
 c5<</0j
  
u//	0!j
 j
X  r3   r  )r  rD  r  )Nr   )r   )Nr"   N)Btypingr   r   r   r   r   rA   torch.nn.functionalr   r   rf   activationsr
   cache_utilsr   r   
generationr   modeling_attn_mask_utilsr   modeling_layersr   modeling_outputsr   r   r   modeling_rope_utilsr   r   modeling_utilsr   r   utilsr   r   r   configuration_granitemoesharedr   !torch.nn.attention.flex_attentionr   integrations.flex_attentionr   
get_loggerr=   r   Moduler   rF   r]   rr   r   r   r   rB   rp   r   r{   r   r   r   r  r'  rD  r  r  __all__r  r3   r2   <module>r     sE  , : 9     ! . ) > 9 j j K F J J B  !;J 
		H	%")) 4Jbii J(*bii *Z-S -S`9+")) 9+x(6	UU\\ 	U# 	U%,, 	U& %II%<<% 
% <<	%
 U\\*% % %:Y9		 Y9x]#= ]@ To T T6<bii <D O; O Oh "&
-1	O&u||U5<<%8$>?O&#O& U\\*	O&
 5<<O&dV"A? Vr fr3   