
    i                     z   d dl mZ d dlmZ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 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) ddl*m+Z+m,Z,m-Z-m.Z.m/Z/ ddl0m1Z1m2Z2 ddl3m4Z4 ddl5m6Z6 ddl7m8Z8  e/jr                  e:      Z;d Z< ed      dRd       Z=dej|                  de?dej|                  fdZ@	 dSd ej                  d!ej|                  d"ej|                  d#ej|                  d$ej|                  dz  d%eBd&eBd'e)e+   fd(ZC ee=       G d) d*ej                               ZDd+ej|                  d,e?fd-ZEd. ZFd/ ZGd0 ZH G d1 d2ej                        ZI G d3 d4ej                  j                        ZJ G d5 d6ej                        ZK G d7 d8ej                        ZL G d9 d:ej                        ZM G d; d<ej                        ZN G d= d>ej                        ZO G d? d@edAB      ZP edC       G dD dEej                               ZQ G dF dGe      ZRe, G dH dIe'             ZSe, G dJ dKeS             ZT	 	 	 dTdLej|                  eUej|                     z  dz  dMe?dz  d$ej|                  dz  dej|                  e?z  fdNZVe, G dO dPeSe             ZWg dQZXy)U    )Callable)Optional	TypedDictN)nn)
functional   )initialization)ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hubuse_kernelized_func)lazy_load_kernel)create_causal_mask)GradientCheckpointingLayer)BaseModelOutputWithPastMoeCausalLMOutputWithPastMoeModelOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tupleis_torchdynamo_compilinglogging)maybe_autocastmerge_with_config_defaults)resolve_internal_import)capture_outputs   )GraniteMoeHybridConfigc                     | 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..N   dim)shapetorchcat)xx1x2s      /var/www/vps2.regionflexible.com/Desarrollo/venv/lib/python3.12/site-packages/transformers/models/granitemoehybrid/modeling_granitemoehybrid.pyrotate_halfr3   2   sZ    	
3"!''"+"""	#B	
3q ""	#B99rc2YB''    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.
    )	unsqueezer3   )qkcossinunsqueeze_dimq_embedk_embeds          r2   apply_rotary_pos_embr?   9   sY    & --
&C
--
&C3w;q>C/0G3w;q>C/0GGr4   hidden_statesn_repreturnc                     | 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)r,   expandreshape)r@   rA   batchnum_key_value_headsslenhead_dims         r2   	repeat_kvrJ   S   so    
 2?1D1D.Ehz!!Qa"23::5BUW\^bdlmM  (;e(CT8TTr4   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 )Nr)   r   r(   )r+   dtype)ptrainingr%   )rJ   num_key_value_groupsr-   matmul	transposer   r   softmaxfloat32torT   rQ   rV   
contiguous)rK   rL   rM   rN   rO   rP   rQ   rR   
key_statesvalue_statesattn_weightsattn_outputs               r2   eager_attention_forwardrb   _   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$$r4   c                        e Zd 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
ej                  ej                  f   dz  d
ee   de
ej                  ej                  f   fdZ xZS )GraniteMoeHybridAttentionz=Multi-headed attention from 'Attention Is All You Need' paperconfig	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 )NrI   Tbias)super__init__re   rf   getattrhidden_sizenum_attention_headsrI   rG   rW   attention_multiplierrP   attention_dropout	is_causalr   Linearattention_biasq_projk_projv_projo_projselfre   rf   	__class__s      r2   rk   z"GraniteMoeHybridAttention.__init__|   sJ   "
F4F4F&JdJd4de$*$>$>&B\B\$\!22!'!9!9ii : :T]] JQWQfQf
 ii : :T]] JQWQfQf
 ii : :T]] JQWQfQf
 ii&&68J8JQWQfQf
r4   Nr@   rO   past_key_valuesposition_embeddingsrR   rB   c                    |j                   d d }g |d| j                  }| j                  |      j                  |      j	                  dd      }| j                  |      j                  |      j	                  dd      }	| j                  |      j                  |      j	                  dd      }
||\  }}t        ||	||      \  }}	| |j                  |	|
| j                        \  }	}
t        j                  | j                  j                  t              } || ||	|
|f| j                  sdn| j                   | j"                  d|\  }} |j$                  g |d j'                         }| j)                  |      }||fS )Nr(   r%   r)           )rQ   rP   )r,   rI   rt   viewrY   ru   rv   r?   updaterf   r   get_interfacere   _attn_implementationrb   rV   rp   rP   rE   r]   rw   )ry   r@   rO   r{   r|   rR   input_shapehidden_shapequery_statesr^   r_   r:   r;   attention_interfacera   r`   s                   r2   forwardz!GraniteMoeHybridAttention.forward   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**HC';L*VY[^'_$L*&'6'='=j,X\XfXf'g$J(?(M(MKK,,.E)
 %8	%
  $}}C$2H2HLL	%
 	%
!\ *k));;;;FFHkk+.L((r4   NN)__name__
__module____qualname____doc__r&   intrk   r-   Tensorr   tupler   r   r   __classcell__rz   s   @r2   rd   rd   x   s    G
5 
# 
6 )-HL')||') t+') 	')
 #5<<#=>E') +,') 
u||U\\)	*')r4   rd   input_tensorpad_sizec                     t        | j                        dk(  r
ddddd|ddfnddd|ddf}t        j                  j                  j                  | |dd      S )z
    Padding x tensor with `pad_size` on the seq_len dim (dim=1)

    Assumes that we only have tensors of either size 4 or 3
       r   constant)moderN   )lenr,   r-   r   r   pad)r   r   	pad_shapes      r2   pad_tensor_by_sizer      sf     47|7I7I3Ja3OAq!Q!Q/VWYZ\]_gijlmUnI88""<ST"UUr4   c                    t        | |      } t        | j                        dk(  r.| j                  | j                  d   d|| j                  d         S | j                  | j                  d   d|| j                  d   | j                  d         S )z
    Padding input_tensor with `pad_size` on the seq_len dim (dim=1) and
    simultaneously splitting it into chunk sequences.

    Assumes that we only have tensors of either size 4 or 3
    r   r   r(   r)   )r   r   r,   rE   )r   r   
chunk_sizes      r2   reshape_into_chunksr      s     &lH=L
<!###L$6$6q$92z<K]K]^_K`aa ##q!2z<3E3Ea3H,J\J\]^J_
 	
r4   c                 "   | j                  d      } | d   j                  g | j                         | } t        j                  t        j                  ||| j
                  t        j                        d      }| j                  | d      } t        j                  | d      }t        j                  t        j                  ||| j
                  t        j                        d      }|j                  | t        j                         }|S )zo
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    r(   .NdevicerT   )diagonalr   r*   )
sizerD   r-   trilonesr   boolmasked_fillcumsuminf)r   r   masktensor_segsums       r2   segment_sumr      s     ""2&J 2<	*11S<3D3D3FS
SL::ejjZ@S@S[`[e[efqstD++TE15LLL26M ::ejjZ@S@S[`[e[efqrsD!--teeiiZ@Mr4   c                     |N|j                   d   dkD  r<|j                   d   dkD  r*| j                  }| |dddddf   z  j                  |      } | S )zm
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    Nr%   r   )r,   rT   r\   )r@   rO   rT   s      r2   apply_mask_to_padding_statesr      sa    
 !n&:&:1&=&AnFZFZ[\F]`aFa##&1d
)CCGGNr4   c            
       &    e Zd ZdZdedef fdZ	 	 	 ddej                  de	dz  dej                  dz  d	ej                  dz  fd
Z	 	 dde	dz  dej                  dz  fdZ	 	 	 dde	dz  dej                  dz  d	ej                  dz  fdZ xZS )GraniteMoeHybridMambaLayeruP  
    Compute ∆, A, B, C, and D the state space parameters and compute the `contextualized_states`.
    A, D are input independent (see Mamba paper [1] Section 3.5.2 "Interpretation of A" for why A isn't selective)
    ∆, B, C are input-dependent (this is a key difference between Mamba and the linear time invariant S4,
    and is why Mamba is called **selective** state spaces)

    The are a few differences between this and Mamba2Mixer:
    - The variable use_precomputed_states is slightly different due to the hybrid cache structure
    - There's a few non-obvious bugs fixed with batching in the slow path that exist in main
    - Some extra variables that our layer doesn't need have been removed
    - We ported most of the refactors in https://github.com/huggingface/transformers/pull/35154, which is (as of Dec 18, 2024) unmerged
    re   rf   c           	         t         |           |j                  | _        |j                  | _        |j
                  | _        |j                  | _        t        |j                  | j                  z        | _        || _        |j                  | _        |j                  | _        t"        |j                     | _        |j&                  | _        |j*                  | _        |j.                  | _        |j2                  | _        |j6                  | _        |j:                  | _        |j<                  | _        |j>                  | _        | j                  d| j0                  z  | j                  z  z   | _         tC        jD                  | j@                  | j@                  |j                  | j                  | j@                  | j                  dz
        | _#        | j                  | j@                  z   | j                  z   }tC        jH                  | j                  || j(                        | _%        tC        jL                  tO        jP                  | j                              | _)        tO        jT                  d| j                  dz         }tC        jL                  tO        jV                  |            | _,        t[        | j                  | j,                        | _.        tC        jL                  tO        jP                  | j                              | _/        tC        jH                  | j                  | j                  | j(                        | _0        tc        d      }te        |dd       a3te        |dd       a4tc        d	      }tk        |d
      a6tk        |d      a7tk        |d      a8ts        tl        tn        tp        th        tf        f      a:tt        stv        jy                  d       y tv        jy                  d       y )Nr)   r%   )in_channelsout_channelsri   kernel_sizegroupspaddingrh   epszcausal-conv1dcausal_conv1d_updatecausal_conv1d_fnz	mamba-ssmz8ops.triton.selective_state_update.selective_state_update)chained_pathz1ops.triton.ssd_combined.mamba_chunk_scan_combinedz8ops.triton.ssd_combined.mamba_split_conv1d_scan_combineda  The fast path is not available because one of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1dzOThe fast path for GraniteMoeHybrid will be used when running the model on a GPU)=rj   rk   mamba_n_heads	num_headsrm   mamba_d_statessm_state_sizemamba_d_convconv_kernel_sizer   mamba_expandintermediate_sizerf   mamba_conv_biasuse_conv_bias
hidden_act
activationr
   actmamba_proj_biasuse_biasrms_norm_epslayer_norm_epsilonmamba_n_groupsn_groupsmamba_d_headrI   mamba_chunk_sizer   time_step_limittime_step_mintime_step_maxconv_dimr   Conv1dconv1drr   in_proj	Parameterr-   r   dt_biasarangelogA_logGraniteMoeHybridRMSNormGatednormDout_projr   rl   r   r   r#   selective_state_updatemamba_chunk_scan_combined mamba_split_conv1d_scan_combinedallis_fast_path_availableloggerwarning_once)ry   re   rf   projection_sizeAcausal_conv1d	mamba_ssmrz   s          r2   rk   z#GraniteMoeHybridMambaLayer.__init__  s   --!--$22 & 3 3!$V%8%84;K;K%K!L"#33 ++&++,.."("5"5--++ 11%55#11#11..T]]1BTEXEX1XXii''--==))A-
 004==@4>>Qyy
 ||EJJt~~$>? LLDNNQ./\\%))A,/
01G1GTMdMde	ejj89		$"8"8$:J:JQUQ^Q^_ )9&}6LdS"=2DdK %[1	!8$^"
 %<$W%
! ,C$^,
(
 "%&)0 $"
 &>  qrr4   Nr@   cache_paramsrO   seq_idxc                    t        ||      }| j                  |      }|j                  \  }}}| j                  | j                  z  }	|d uxr" |j                  | j                        xr |dk(  }
|
r
|j                  d      j                  | j                  | j                  | j                  gd      \  }}}t        ||j                  | j                     j                  | j                  j                   j                  d      | j                  j"                  | j$                        }t'        j                  || j                  |	|	gd      \  }}}t'        j(                  | j*                  j-                                }|d d d df   d d d d d f   j/                  d| j0                  | j                        j3                  t&        j4                        }|d d d d d f   j/                  dd| j0                        }| j6                  d d d df   j/                  d| j0                        }| j8                  d d d df   j/                  d| j0                        }|j;                  || j                  |j                  d   | j                  z        }|j;                  || j                  |j                  d   | j                  z        }|j;                  || j                  | j0                        }t=        |j                  | j                     j>                  ||||||d |d
      }|j;                  || j                  | j0                  z        }| jA                  ||      }| jC                  |      d d d df   }|S t'        j(                  | j*                  j-                                }| jD                  dt-        d	      fk(  ri nd
| jD                  i}| jF                  r|tI        || j                  j                   j                  d      | j                  j"                  | j6                  |f| j8                  | jJ                  || j$                  | j@                  j                   | j@                  jL                  | jB                  j                   | jB                  j"                  | j0                  | j                  ddd|}|S |j                  | j                  | j                  | j                  gd      \  }}}|j|jO                  dd      }tP        jR                  jU                  || jV                  |j                  d   z
  df      }|jY                  || j                        }| j$                  dvrH| j[                  | j                  |jO                  dd            dd |f   jO                  dd            }nqt]        |jO                  dd      | j                  j                   j                  d      | j                  j"                  | j$                  |      jO                  dd      }t        ||      }t'        j                  || j                  |	|	gd      \  }}}t_        |j;                  ||d| j0                        |||j;                  ||| j                  d      |j;                  ||| j                  d      f| jJ                  | j8                  d |d| j6                  dd|\  }}|||ja                  || j                        }|j;                  ||d      }| jA                  ||      }| jC                  |      }|S )Nr%   r(   r*   .rT   T)zr   dt_softplusr~   r   dt_limitF)r   r   r   r   rmsnorm_weightrmsnorm_epsoutproj_weightoutproj_biasheaddimngroupsnorm_before_gatereturn_final_statesr)   r   )siluswish)r/   weightri   r   r   )r   r   r   r   r   r   r   )1r   r   r,   r   r   has_previous_staterf   squeezesplitr   r   r   r   layersconv_statesr   r   ri   r   r-   expr   floatrD   rI   r\   r[   r   r   r   r   recurrent_statesr   r   r   rV   r   r   variance_epsilonrY   r   r   r   r   update_conv_stater   r   r   update_recurrent_state)ry   r@   r   rO   r   projected_states
batch_sizeseq_len_groups_time_state_sizeuse_precomputed_statesgatehidden_states_B_CdtBCr   r   r   hidden_states_reshapedoutdt_limit_kwargshidden_states_B_C_transposedr  scan_output	ssm_states                             r2   cuda_kernels_forwardz/GraniteMoeHybridMambaLayer.cuda_kernels_forwardj  s    5]NS<<6 "/!4!4
GQ!%1D1D!D $i)H)H)Xi]dhi]i 	
 "*:*B*B1*E*K*K''GR +L +'D#R
 !5!##DNN3??""**1-  ! #(++!'')?AWX#M1a 4::++-..A!T3,1d
+222t}}dFYFYZ]]didqdq]rAAq$J&&r2t}}=Bll1dC<077DMMJGq$|$++B>Az4==!''!*2MNAz4==!''!*2MNA%2%7%7
DNNTXTaTa%b"2##DNN3DD& M *..z4>>DMM;YZM IImT:M --.q$|<C| 
w 4::++-..A$($8$8S%,<O$ObV`bfbvbvUwO }}!56$KK&&..q1KK$$LL ff####'99#3#3 $		 : :#'==#7#7!%!3!3 MM MM%*(-#$ &%l 
A /?.D.D++T]]DNNKQS /E /+'  + 4E3N3NqRS3T0"$--"3"34..1M1S1STV1WWYZ[#K #/"@"@dnn"]K??*;;(,$5$?$?1$EFsHWH}U__`acde)% )9+55a;#{{1199!<![[--#'?? ')  i1o & %AARTb$c!&+kk%++-CE[\'#q! *C!&&z7BNFF:wrBFF:wrB*  $ff#(, LL $* &*&Y" (\-E , C CIt~~ ^I)..z7BG"iiT: mmK0
r4   c                    |j                   \  }}}|j                  }t        ||      }| j                  |      }|j	                  | j
                  | j                  | j                  gd      \  }	}
}|
j                  dd      }
|d uxr" |j                  | j                        xr |dk(  }|r|j                  |
| j                        }t        j                  || j                  j                  j!                  d      z  d      }
| j"                  r|
| j                  j$                  z   }
| j'                  |
      }
n|Xt(        j*                  j-                  |
| j.                  |
j                   d   z
  df      }|j                  || j                        }| j'                  | j                  |
      dd |f   j                  dd            }
t        |
|      }
t        j                  |
| j
                  | j0                  | j2                  z  | j0                  | j2                  z  gd      \  }}}t        j4                  | j6                  j9                                }|r|j:                  | j                     j<                  j>                  }|d d dd d f   d d d df   }|j                  dd      jA                  ||j                   d   | jB                        }| jD                  d   jA                  | jD                  j                   d   | jB                        }t        j(                  j*                  jG                  ||jI                  |j                        z         }t        jJ                  || jL                  d   | jL                  d         }|d   jA                  | j                  | jB                  | j2                        jI                  t        jN                  	      }t        j4                  |d   |z        jI                  |
      }|jQ                  || j0                  d      dd d d f   }|jA                  || j0                  | j                  | j0                  z  |j                   d         jS                         }|jQ                  |d|j                   d         }|d   |dd d d f   z  }|jQ                  |d| jB                        }||d   z  jI                  |
      }|j:                  | j                     j<                  |z  |z   }|jU                  || j                        }|jQ                  || j0                  d      dd d d f   }|jA                  || j0                  | j                  | j0                  z  |j                   d         jS                         }|jQ                  |d|j                   d         }|jI                  |j>                  |j                        }|jW                  || j                  z  | jB                  | j2                        }|jW                  || j                  z  | j2                  d      }t        jX                  ||      }|jW                  || j                  | jB                        }| jZ                  d   jA                  | jZ                  j                   d   | jB                        }|||z  z   jI                  |j                        }|jQ                  |d      d d d df   }nt(        j*                  jG                  || jD                  z         }t        jJ                  || jL                  d   | jL                  d         }|jQ                  ||d| jB                        j9                         }|jQ                  ||d| j2                        j9                         }|jQ                  ||d| j2                        j9                         }|j]                  | j                  | j0                  z  d| j                        }|j]                  | j                  | j0                  z  d| j                        }| j^                  || j^                  z  z
  | j^                  z  }| jZ                  d   ta        ||      z  }||d   z  }|jI                  |j                        |z  }||||fD cg c]  }tc        ||| j^                         c}\  }}}}|je                  dddd      }t        jf                  |d      }t        j4                  ti        |            } |d d d d d d d d d d d f   |d d d d d d d d d d d f   z  }!|!j                  d      }"|"d   | je                  ddddd      d   z  }#|#j                  d      }$|$d   |d d d d d f   z  j                  d      }%t        j4                  |d d d d d d dd f   |z
        }&||&je                  dddd      d   z  }'|'dd d d f   |d   z  j                  d      }(t        jj                  |(d d d df         })t        jl                  |)|(gd      }(t        j4                  ti        t(        j*                  j-                  |d d d d d d df   d                  }*|*j                  dd      }*|*d   |(d d d d d df   z  j                  d      }+|+d d d df   |+d d df   },}(t        j4                  |      }-|dd d d f   |(d d d d d df   z  }.|-je                  dddd      }/|.j                  d      |/d   z  }0|%|0z   }|jQ                  |d| j                  | jB                        }||z   }|dkD  r|d d d |d d d d f   }|jQ                  ||d      }|,||jU                  |,| j                        },| jo                  ||	      }1| jq                  |1jI                  |            }2|2S c c}w )Nr(   r*   r%   r)   r   .r   ).NNr   r   r   )r+   output_sizer   r   r   )r%   r   )9r,   rT   r   r   r  r   r   r   rY   r   rf   r	  r-   sumr   r   r  r   ri   r   r   r   r   r   r   r   r  r   r  r  r  r   rD   rI   r   softplusr\   clampr   r[   rE   r]   r
  r   bmmr   repeat_interleaver   r   r   permuter   r   
zeros_liker.   r   r   )3ry   input_statesr   rO   r  r  r  rT   r  r  r  r  r  r  r@   r  r  r   cache_devicer   dAdBdBx
ssm_statesssm_states_reshaped
C_reshapedyr   r   
D_residualtA_cumsumLG_intermediateGM_intermediateMY_diagdecay_statesB_decaystatesprevious_statesdecay_chunk
new_statesr  state_decay_outC_times_statesstate_decay_out_permutedY_offr  contextualized_statess3                                                      r2   torch_forwardz(GraniteMoeHybridMambaLayer.torch_forward  s
    ".!3!3
GQ"" 4L.Q<<5&6&<&<''GR '= '
# .77!<!-T!9!~l>]>]^b^l^l>m!~ry}~r~ "&889JDNN[K %		dkk0088;;! !!$58H8H$H! $): ; ' mm//%(=(=@Q@W@WXZ@[([]^'_ +<<[$..Y $5F)GXgX)V)`)`abde)f g89JN[#kk##T]]T5H5H%H$--Z^ZmZmJmn
q! YYtzz'')**!'..t~~>OOVVL Aq!GQc\*Ba#**:rxx|T]]SBll9-44T\\5G5G5JDMMZG$$--b7::bhh3G.GHBR!5!5a!8$:N:Nq:QRB/"))$..$--I\I\]``glgtgt`uA))ByMA-.22,2GB
 		*dmmR8dAFAT]]DNNdmm4SUVU\U\]_U`allnA		*b!''"+6AI3a<0B *11*b$--PMi0044L4IC &,,T^^<MMPRRUXXJ%<<ZXJ 		*dmmR8dAFAT]]DNNdmm4SUVU\U\]_U`allnA		*b!''"+6A $ahhaggFJ",//*t~~2Mt}}^b^q^q"r
T^^ ;T=P=PRSTJ		-z:Az4>>4==AA y!((a$--HA]Q&&**1773A 		*b)!T3,7A ''T\\(9:BR!5!5a!8$:N:Nq:QRB)11*gr4==Y__aM		*gr43F3FGMMOA		*gr43F3FGMMOA##DNNdmm$CX\XfXf#gA##DNNdmm$CX\XfXf#gA'DOO*CCtVH	*-?x-XXJ *ByM9M](()B.A cpqrtuwxay%z\]&9!Xt&W%z"M1a 		!Q1%A||A2.H 		+a.)A q!Qa23a1dAq!8K6LLN""r"*A y\AIIaAq!,DY,OON""r"*A 	l]1a:%>>CCCJF !99XaArsl%;h%FGL,..q"b!<YGGGc4l+mI.FFKKPQKRF $..va!e}=OYY8a@F))K0A0A(1aQRTV;BWY_0`$abK%//15K%o61dC9PPUUZ[U\J *1crc6 2Jq"u4EIF $ii1OT1oq!T30GGN'6'>'>q!Q'J$#''+.Fy.QQE A		*b$..$--HAJA!|a'1a'(		*gr2A $)A(??	4>>Z	ii4(
 !%knnU.C D$$A &{s   o1c                    t         rJd| j                  j                  j                  j                  v rt               s| j                  ||||      S |t        d      |j                  }|B|j                  d   dkD  r0|j                  d   dkD  r||d d d d d f   z  j                  |      }| j                  |||      S )Ncudaz\`seq_idx` support requires fast path support. Please install `mamba_ssm` and `causal_conv1d`r%   r   )r   r   r   r   typer   r  NotImplementedErrorrT   r,   r\   rD  )ry   r@   r   rO   r   rR   rT   s          r2   r   z"GraniteMoeHybridMambaLayer.forward  s     "f0C0C0J0J0O0O&OXpXr,,]L.Zabb%n  ##%.*>*>q*AA*E.J^J^_`JadeJe*^Aq$J-GGKKERM!!-~NNr4   NNNr   )r   r   r   r   r&   r   rk   r-   r   r   	IntTensorr  rD  r   r   r   s   @r2   r   r      s    Zs5 Zs# Zs~ &*.2*._||_ dl_ t+	_
 4'_J &*.2	z% dlz% t+	z%@ &*.2*.O dlO t+	O
 4'Or4   r   c                   (     e Zd Zd fd	ZddZ xZS )r   c                     t         |           t        j                  t	        j
                  |            | _        || _        y Nrj   rk   r   r   r-   r   r   r  ry   rm   r   rz   s      r2   rk   z%GraniteMoeHybridRMSNormGated.__init__  s/    ll5::k#:; #r4   c                    |j                   }|j                  t        j                        }|?|t        j
                  j                  |j                  t        j                              z  }|j                  d      j                  dd      }|t        j                  || j                  z         z  }| j                  |j                  |      z  S Nr)   r(   T)keepdim)rT   r\   r-   r[   r   r   r   powmeanrsqrtr  r   )ry   r@   r  input_dtypevariances        r2   r   z$GraniteMoeHybridRMSNormGated.forward  s    #))%((7)BMM,>,>twwu}}?U,VVM $$Q',,R,>%Ht?T?T4T(UU{{]--k:::r4   gư>rM  )r   r   r   rk   r   r   r   s   @r2   r   r     s    $
	;r4   r   c                   `     e Zd ZdZdef fdZdej                  dej                  fdZ xZ	S )GraniteMoeHybridMLPz~
    MLP layer for shared experts

    Args:
        config:
            Configuration object with model hyperparameters.
    re   c                 `   t         |           |j                  | _        |j                  | _        t
        |j                     | _        t        j                  | j                  | j                  dz  d      | _
        t        j                  | j                  | j                  d      | _        y )Nr)   Frh   )rj   rk   rm   
input_sizeshared_intermediate_sizer
   r   r   r   rr   input_linearoutput_linearry   re   rz   s     r2   rk   zGraniteMoeHybridMLP.__init__  s     ,,!:: !2!23IIdoot7G7G!7KRWXYYt'7'7uUr4   r@   rB   c                     | j                  |      }|j                  dd      }| j                  |d         |d   z  }| j                  |      }|S )Nr)   r(   r*   r   r%   )r^  chunkr   r_  )ry   r@   chunked_hidden_statess      r2   r   zGraniteMoeHybridMLP.forward  s^    ))-8 - 3 3A2 3 >(=a(@ADYZ[D\\**=9r4   )
r   r   r   r   r&   rk   r-   r   r   r   r   s   @r2   rZ  rZ    s2    V5 VU\\ ell r4   rZ  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 )GraniteMoeHybridRotaryEmbeddinginv_freqNre   c                    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defaultrf  F)
persistentoriginal_inv_freq)rj   rk   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenre   rope_parametersrh  compute_default_rope_parametersr   attention_scalingregister_bufferclone)ry   re   r   rope_init_fnrf  rz   s        r2   rk   z(GraniteMoeHybridRotaryEmbedding.__init__  s    "("@"@$*$B$B!44[A!%!E!E>>Y&.t~~>L+7V+L($(ZeD0(..2BuUr4   r   ztorch.devicer  rB   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_thetarI   Ng      ?r   r)   r   r   )	ro  rl   rm   rn   r-   r   int64r\   r  )re   r   r  baser+   attention_factorrf  s          r2   rp  z?GraniteMoeHybridRotaryEmbedding.compute_default_rope_parameters  s    & %%l3fj$/c63E3EIcIc3c U\\!S!5;;?BB&X]XcXcBdgjjk
 )))r4   c                 N   | j                   d d d d f   j                         j                  |j                  d   dd      j	                  |j
                        }|d d d d d f   j                         }t        |j
                  j                  t              r/|j
                  j                  dk7  r|j
                  j                  nd}t        |d      5  |j                         |j                         z  j                  dd      }t        j                  ||fd	      }|j                         | j                  z  }|j                         | j                  z  }	d d d        j	                  |j                   
      	j	                  |j                   
      fS # 1 sw Y   AxY w)Nr   r(   r%   mpscpuF)device_typeenabledr)   r*   r   )rf  r  rD   r,   r\   r   
isinstancerG  strr!   rY   r-   r.   r:   rq  r;   rT   )
ry   r/   position_idsinv_freq_expandedposition_ids_expandedr}  freqsembr:   r;   s
             r2   r   z'GraniteMoeHybridRotaryEmbedding.forward<  sR    !MM$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$cff177f&;;;	5 	5s   BFF$rM  rI  )r   r   r   r-   r   __annotations__r&   rk   staticmethodr   r   r   r  rp  no_gradr   r   r   r   s   @r2   re  re    s    llV5 V  04+/"*&-*(* t* 
~u$	%	* *: U]]_<  <r4   re  c                   6     e Zd Zdedededdf fdZd Z xZS )GraniteMoeHybridParallelExpertsnum_expertsr\  r  rB   Nc                     t         |           t        j                  t	        j
                  |||            | _        || _        || _        || _	        y)a  
        Initialize the GraniteMoeHybridParallelExperts 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)
rj   rk   r   r   r-   emptyr   r  r\  r  )ry   r  r\  r  rz   s       r2   rk   z(GraniteMoeHybridParallelExperts.__init__M  sD    " 	ll5;;{K#TU&$&r4   c                     |j                  |d      }g }t        | j                        D ]7  }|j                  t	        j
                  ||   | j                  |                9 t        j                  |d      }|S )a  
        Forward pass of the GraniteMoeHybridParallelExperts module.

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

        Returns:
            Tensor: Output tensor.
        r   r*   )	r  ranger  appendFlinearr   r-   r.   )ry   inputsexpert_size
input_listoutput_listiresultss          r2   r   z'GraniteMoeHybridParallelExperts.forwardd  sq     \\+1\5
t''( 	HAqxx
1t{{1~FG	H))KQ/r4   r   r   r   r   rk   r   r   r   s   @r2   r  r  L  s)    'C 'S 's 't '.r4   r  c                   2     e Zd Zdededef fdZd Z xZS )GraniteMoeHybridTopKGatingr\  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.
        Frh   N)rj   rk   r  r\  r  r   rr   layer)ry   r\  r  r  rz   s       r2   rk   z#GraniteMoeHybridTopKGating.__init__z  s:     	&$
YYz;UC
r4   c                    | j                  |      j                         }|j                  | j                  d      \  }}t	        j
                  |d      j                  |      }t	        j                  |j                  d      | j                  g|j                  |j                        }|j                  d|d      }|j                         j                  d      }|j                         }|j!                         }	|	j#                  d      \  }
}|j%                  | j                  d      }|j!                         }||   }|||||fS )Nr%   r*   r   rT   r   trunc)rounding_mode)r  r  topkr  r-   rZ   type_aszerosr   r  rT   r   scatterlongr   tolistflattensortdiv)ry   r@   logitstop_k_logitstop_k_indicestop_k_gatesr  gatesr  top_k_expertsr  index_sorted_expertsbatch_indexbatch_gatess                 r2   r   z"GraniteMoeHybridTopKGating.forward  s.   M*002&,kk$**!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Rr4   r  r   s   @r2   r  r  y  s'    D3 DS D D(Sr4   r  c                   .     e Zd ZdZdef fdZd Z xZS )GraniteMoeHybridMoEz
    A Sparsely gated mixture of experts layer with 1-layer Feed-Forward networks as experts.

    Args:
        config:
            Configuration object with model hyperparameters.
    re   c                    t         |           |j                  | _        |j                  | _        t
        |j                     | _        t        |j                  | j                  | j                  dz        | _
        t        |j                  | j                  | j                        | _        t        | j                  |j                  |j                        | _        y )Nr)   )r\  r  r  )rj   rk   rm   r\  r   r
   r   r   r  num_local_expertsr^  r_  r  num_experts_per_tokrouterr`  s     r2   rk   zGraniteMoeHybridMoE.__init__  s     ,,!33 !2!23;$$doot7G7G!7K
 =$$d&6&6
 100,,
r4   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 )Nr(   r)   r*   r   r%   r  )r   rE   r  r^  rb  r   r_  r-   r  r\  rT   r   	index_addr   )ry   layer_inputbszlengthemb_sizer  r  r  r  expert_inputsr@   rc  expert_outputsr  layer_outputs                  r2   r   zGraniteMoeHybridMoE.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r4   )r   r   r   r   r&   rk   r   r   r   s   @r2   r  r    s    
5 
&r4   r  c                       e Zd ZU dZej
                  ed<   ej
                  ed<   eed<   eed<   ej                  ed<   y)GraniteFlashAttentionKwargsaT  
    Keyword arguments for advanced Flash Attention, causal-conv1d, and mamba_ssm kernel usage.
    Use cases include padding-free training and fewer `torch.compile` graph breaks.

    cu_seq_lens_q (`torch.LongTensor`):
        Gets cumulative sequence length for query state.
    cu_seq_lens_k (`torch.LongTensor`):
        Gets cumulative sequence length for key state.
    max_length_q (`int`):
        Maximum sequence length for query state.
    max_length_k (`int`):
        Maximum sequence length for key state.
    seq_idx (`torch.IntTensor):
        Index of each packed sequence.
    cu_seq_lens_qcu_seq_lens_kmax_length_qmax_length_kr   N)	r   r   r   r   r-   
LongTensorr  r   rJ   r4   r2   r  r    s7      ######__r4   r  F)total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 )	GraniteMoeHybridRMSNormr   rB   Nc                     t         |           t        j                  t	        j
                  |            | _        || _        y)zF
        GraniteMoeHybridRMSNorm is equivalent to T5LayerNorm
        NrN  rO  s      r2   rk   z GraniteMoeHybridRMSNorm.__init__  s1     	ll5::k#:; #r4   r@   c                 "   |j                   }|j                  t        j                        }|j	                  d      j                  dd      }|t        j                  || j                  z         z  }| j                  |j                  |      z  S rQ  )	rT   r\   r-   r[   rS  rT  rU  r  r   )ry   r@   rV  rW  s       r2   r   zGraniteMoeHybridRMSNorm.forward  sy    #))%((7 $$Q',,R,>%Ht?T?T4T(UU{{]--k:::r4   c                 ^    t        | j                  j                         d| j                   S )Nz, eps=)r   r   r,   r  )ry   s    r2   
extra_reprz"GraniteMoeHybridRMSNorm.extra_repr  s*    ))*+6$2G2G1HIIr4   rX  )
r   r   r   r  rk   r-   r   r   r  r   r   s   @r2   r  r    s7    $ $$ $;U\\ ;ell ;Jr4   r  c                   .    e Zd Zdedef fdZe	 	 	 	 ddej                  dej                  dz  de	dz  de
dz  d	eej                  ej                  f   dz  d
ee   deej                  eej                  ej                  f   dz  f   fd       Z xZS )GraniteMoeHybridDecoderLayerre   rf   c                 8   t         |           |j                  | _        d | _        t	        |j                  |j
                        | _        t	        |j                  |j
                        | _        |j                  dkD  rt        |      nd | _
        |j                  | _        t        |      | _        d | _        |j                  |   dk(  rt!        ||      | _        nt#        ||      | _        |j                  |   | _        t'        |dd      dkD  | _        y )Nr   r   mambar  )rj   rk   rm   	self_attnr  r   input_layernormpost_attention_layernormr  r  block_sparse_moeresidual_multiplierrZ  
shared_mlpr  layers_block_typer   rd   
layer_typerl   has_expertsrx   s      r2   rk   z%GraniteMoeHybridDecoderLayer.__init__  s    !--6v7I7IvObObc(?@R@RX^XkXk(l% @F?W?WZ[?[ 3F ;ae#)#=#= -f5
##I.'93FIFDJ6vyIDN 229= #6+>BQFr4   Nr@   rO   r{   	use_cacher|   rR   rB   c           	         |}| j                  |      }| j                   | j                  d|||d|}n | j                  d|||||d|\  }}||| j                  z  z   }|}| j	                  |      }| j
                  r&| j                  |      }	|	| j                  |      z   }n| j                  |      }||| j                  z  z   }|S )N)r@   r   rO   )r@   rO   r{   r  r|   r  )r  r  r  r  r  r  r  r  )
ry   r@   rO   r{   r  r|   rR   residualr  moe_hidden_statess
             r2   r   z$GraniteMoeHybridDecoderLayer.forward  s    !,,];::!&DJJ +,- 	M  .t~~  +- /#$7   M1 !=43K3K#KK 55mD $ 5 5m D-0NNM OOM:M =43K3K#KKr4   )NNFN)r   r   r   r&   r   rk   r   r-   r   r   r   r   r   r  FloatTensorr   r   r   s   @r2   r  r    s    G5 G# G.  /3(,!&HL(||( t+( 	(
 $;( #5<<#=>E( 45( 
u  %(9(95;L;L(L"MPT"TT	U( (r4   r  c                        e Zd ZU eed<   dZdZdgZdgZdZ	dZ
dZdZdZeedZdZ ej&                          fd       Z xZS )	GraniteMoeHybridPreTrainedModelre   modelTr  r{   F)r@   
attentionsc           
      V   t         |   |       t        |t              r6t	        j
                  |j                  d| j                  j                         t        |t              rt	        j                  |j                         t	        j                  |j                  t        j                  t        j                   d|j"                  dz                      t	        j                  |j$                         y t        |t&              r t	        j                  |j                         y y )Nr~   )rT  stdr%   )rj   _init_weightsr  r  initnormal_r   re   initializer_ranger   ones_r   copy_r   r-   r   r   r   r   r   )ry   rK   rz   s     r2   r  z-GraniteMoeHybridPreTrainedModel._init_weights\  s    f%f=>LLSdkk6S6STf89JJv~~&JJv||UYYu||Av?O?ORS?S/T%UVJJvxx  <=JJv}}% >r4   )r   r   r   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  rd   _can_record_outputs_is_statefulr-   r  r  r   r   s   @r2   r  r  J  ss    ""&*#78#4"5N""&5/ LU]]_	& 	&r4   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e   deez  fd                     Zd Z xZS )GraniteMoeHybridModelre   c           	      N   t         |   |       |j                  | _        |j                  | _        t        j                  |j                  |j                  | j                        | _        t        j                  t        |j                        D cg c]  }t        ||       c}      | _        t        |j                  |j                        | _        |j"                  dk(  rt%        |      nd | _        d| _        |j*                  | _        | j-                          y c c}w )Nr   ropeF)rj   rk   pad_token_idpadding_idx
vocab_sizer   	Embeddingrm   embed_tokens
ModuleListr  num_hidden_layersr  r  r  r   r   position_embedding_typere  
rotary_embgradient_checkpointingembedding_multiplier	post_initrx   s      r2   rk   zGraniteMoeHybridModel.__init__k  s     !.. ++LL):):F<N<NPTP`P`ammNSTZTlTlNmn)&)<n
 ,F,>,>FDWDWX	EKEcEcgmEm9&Asw&+#$*$?$?! 	 os   D"N	input_idsrO   r  r{   inputs_embedsr  rR   rB   c           	         |d u |d uz  rt        d      || j                  |      }|| j                  z  }|r|t        | j                        }|V||j                         nd}t        j                  |j                  d   |j                        |z   }|j                  d      }t        | j                  |||      }	| j                  ||      }
|}d }| j                  | j                  ||      }t        | j                        D ]3  \  }}| j                  j                   |   dk(  r|
n|	} ||f||||d|}5 | j#                  |      }t%        ||	      S )
Nz:You must specify exactly one of input_ids or inputs_embeds)re   r   r%   r  )re   r  rO   r{   r  )rO   r{   r  r|   )last_hidden_stater{   )
ValueErrorr  r  r   re   get_seq_lengthr-   r   r,   r   r7   r   _update_mamba_maskr	  	enumerater  r  r   r   )ry   r  rO   r  r{   r  r  rR   past_seen_tokenscausal_mask
mamba_maskr@   r|   r  decoder_layer
layer_masks                   r2   r   zGraniteMoeHybridModel.forward|  s    -t";<YZZ  --i8M%(A(AA0*$++>OCRC^==?de <<(;(;A(>}G[G[\_ooL'11!4L(;;')+	
 ,,^_M
 &"??&"&//-"N )$++ 6 	A}'+{{'D'DQ'G7'RXcJ)) /#$7 M		 		-0%++
 	
r4   c                 f    |}||j                         s|t        j                  |dk(        rd}|S )zv
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        Nr%   )r   r-   r   )ry   rO   r{   r  s       r2   r  z(GraniteMoeHybridModel._update_mamba_mask  s;     $
'O,N,N,P&599^q5H+IJr4   )NNNNNN)r   r   r   r&   rk   r   r"   r$   r-   r  r   r   r  r   r   r  r   r   r   r  r   r   s   @r2   r  r  i  s    5 "  .2.204(,26!%9
##d*9
 t+9
 &&-	9

 9
 ((4/9
 $;9
 459
 
(	(9
    9
vr4   r  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   r*   r(   )r  r   r   r-   r.   r\   r   r   rZ   r  one_hotrT  r  r,   rD   rE   r   r7   )r  r  r  rO   compute_device
layer_gateconcatenated_gate_logitsrouting_weightsr  selected_expertsexpert_masktokens_per_expertrouter_prob_per_expertr  sequence_lengthr  expert_attention_mask router_per_expert_attention_maskoverall_losss                      r2   load_balancing_loss_funcr*    s9   : *[%"@+u%$Q..#(99^i-jPZjmmN.K-jpq#r hh))112JPR1SO**_eDA((%%--.>LK!JJ{'8'8':B "'O!C&4&:&:#
O4::1=*B^_ 4AtT12V&
OUKXYWR,R	 	 "IIk&7&7&9<Q&QWXY\a\e\e!q]
 
 4At+,V&
O[QRWR%R	 	) "'?=]+]cd!ehmhqhq,!i
 "
 99.1G1Q1QRS1TTUL+%%[ .ks   Ic                   D    e Zd ZddiZddiZddgdgfiZdef fdZee		 	 	 	 	 	 	 	 dd
e
j                  d	z  de
j                  d	z  de
j                  d	z  ded	z  de
j                  d	z  de
j                  d	z  ded	z  dee
j                  z  deez  fd              Z xZS )GraniteMoeHybridForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr@   r  re   c                 p   t         |   |       t        |      | _        |j                  | _        t        j                  |j                  |j                  d      | _        |j                  | _	        |j                  | _        |j                  | _        |j                  | _        | j                          y )NFrh   )rj   rk   r  r  r  r   rr   rm   r-  router_aux_loss_coefr  r  r  logits_scalingr  r`  s     r2   rk   z$GraniteMoeHybridForCausalLM.__init__   s     *62
 ++yy!3!3V5F5FUS$*$?$?!!33#)#=#= $33 	r4   Nr  rO   r  r{   r  labelsoutput_router_logitslogits_to_keeprB   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 )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, GraniteMoeHybridForCausalLM

        >>> model = GraniteMoeHybridForCausalLM.from_pretrained("ibm-granite/granite-4.0-h-tiny")
        >>> tokenizer = AutoTokenizer.from_pretrained("ibm-granite/granite-4.0-h-tiny")

        >>> 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)r  rO   r  r{   r  r  )lossaux_lossr  r{   r@   r  router_logitsr  )re   r3  r  r  r  r   slicer-  r1  loss_functionr  r*  r8  r  r  r0  r\   r   r   r{   r@   r  )ry   r  rO   r  r{   r  r2  r3  r4  rR   outputsr@   slice_indicesr  r6  r7  s                   r2   r   z#GraniteMoeHybridForCausalLM.forward-  s|   J %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!//))!//
 	
r4   )NNNNNNNr   )r   r   r   _tied_weights_keys_tp_plan_pp_planr&   rk   r   r   r-   r  r   r   r  r   r   r   r   r   r   r   s   @r2   r,  r,    s    *,GH23H_-z:;H5   .2.204(,26*.,0-.Q
##d*Q
 t+Q
 &&-	Q

 Q
 ((4/Q
   4'Q
 #TkQ
 ell*Q
 
*	*Q
  Q
r4   r,  )r,  r  r  )r%   )r~   )Nr)   N)Ycollections.abcr   typingr   r   r-   r   torch.nnr   r   r	   r  activationsr
   cache_utilsr   r   
generationr   integrationsr   r   r   integrations.hub_kernelsr   masking_utilsr   modeling_layersr   modeling_outputsr   r   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   r   r    utils.genericr!   r"   utils.import_utilsr#   utils.output_capturingr$   configuration_granitemoehybridr&   
get_loggerr   r   r3   r?   r   r   rJ   Moduler  rb   rd   r   r   r   r   r   r   rZ  re  r  r  r  r  r  r  r  r  r   r*  r,  __all__r  r4   r2   <module>rW     s  * % &   $ & ! . ) f f 8 / 9 j j K F & l l G 9 5 B 
		H	%( *+ ,2	UU\\ 	U# 	U%,, 	U& %II%<<% 
% <<	%
 LL4'% % % '(%2 )*A)		 A) +A)NVU\\ VS V
((	\O \O~;588?? ;$")) 4><bii ><B*bii *Z.S .Sb,")) ,^)5 0 Y'Jbii J (J(A#= AH &o & &< [; [ [@ #
*.	O&ell 33d:O&tO& LL4'	O&
 \\CO&d e
"A? e
 e
P fr4   