o
    Zh                    @   s  d dl mZ d dlmZmZmZmZmZ d dlZd dlm	Z	 d dl
m  m  m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 ddlmZ ddlmZmZ ddlmZmZ ddl m!Z!m"Z" ddl#m$Z$ ddl%m&Z&m'Z'm(Z( ddl)m*Z*m+Z+ ddl,m-Z- e+ rd dl.m/Z/ d dl0m1Z1m2Z2 ndZ/e* rd dl3m4Z4m5Z5 nd\Z5Z4e(6e7Z8G dd deddZ9G dd dej:Z:G dd de	j;Z<d d! Z=d"ej>d#e?d$ej>fd%d&Z@	'dQd(e	j;d)ej>d*ej>d+ej>d,eej> d-eAd.eAfd/d0ZBdRd1d2ZCG d3d4 d4e	j;ZDG d5d6 d6ej	j;ZEd7ej>d8e?fd9d:ZFd;d< ZGd=d> ZHeIe/e4e5fZJd?d@ ZKG dAdB dBe	j;ZLG dCdD dDe	j;ZMedEG dFdG dGe	j;ZNG dHdI dIe	j;ZOe&G dJdK dKe"ZPe&G dLdM dMePZQe&G dNdO dOePeZRg dPZSdS )S    )partial)CallableOptionalTuple	TypedDictUnionN)nn)ACT2FN   )Cache)GenerationMixin)use_kernel_forward_from_hub)AttentionMaskConverter)FlashAttentionKwargs)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)auto_docstringcan_return_tuplelogging)is_causal_conv1d_availableis_mamba_2_ssm_available   )BambaConfig)selective_state_update)mamba_chunk_scan_combined mamba_split_conv1d_scan_combined)causal_conv1d_fncausal_conv1d_updateNNc                   @   s@   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< dS )BambaFlashAttentionKwargsa  
    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.

    Attributes:
        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.
    Zcu_seq_lens_qZcu_seq_lens_kZmax_length_qZmax_length_kseq_idxN)	__name__
__module____qualname____doc__torch
LongTensor__annotations__int	IntTensor r/   r/   W/var/www/auris/lib/python3.10/site-packages/transformers/models/bamba/modeling_bamba.pyr$   A   s   
 

r$   F)totalc                       s.   e Zd ZdZejdfdef fddZ  ZS ) HybridMambaAttentionDynamicCachea  
    A dynamic cache that can handle both the attention cache (which has a seq_len dimension) and the mamba cache
    (which has a constant shape regardless of seq_len).

    This cache has two sets of lists of tensors: `key_cache` and `value_cache` for attention cache and `conv_states`
    and `ssm_states` for mamba cache. Each of these lists has `num_layers` tensors. The expected shape for each tensor
    For attention layers, `key_cache` and `value_cache` have a shape of `(batch_size, num_heads, seq_len, head_dim)`,
    while `conv_states` and `ssm_states` have a shape of `(batch_size, 0)` (empty tensors).
    For mamba layers, `key_cache` and `value_cache` have a shape of `(batch_size, 0)` (empty tensors),
    while `conv_states` represents the convolution state and has a shape of `(batch_size, d_inner, d_conv)`,
    and `ssm_states` represents the ssm state and has a shape of `(batch_size, d_inner, d_state)`.
    Nconfigc                    sB  t  | | |j| _d| _|j}|j}g | _g | _g | _t	|j
D ]^}| j| dkr\|  jtj |j|j d|j |  ||dg7  _|  jtj |j|j||dg7  _q$|  jtjg g  dg7  _|  jtjg g  dg7  _| j| q$ fddt	|j
D | _ fddt	|j
D | _d S )	NFmamba   devicedtyper7   c                        g | ]}t jg g  d qS r9   r*   tensor.0_
batch_sizer7   r/   r0   
<listcomp>        z=HybridMambaAttentionDynamicCache.__init__.<locals>.<listcomp>c                    r:   r;   r<   r>   rA   r/   r0   rC      rD   )super__init__layers_block_typehas_previous_statemamba_d_convmamba_d_stateconv_states
ssm_statesZtransformer_layersrangenum_hidden_layersr*   Zzerosmamba_expandhidden_sizemamba_n_groupsmamba_n_headsmamba_d_headr=   appendZ	key_cacheZvalue_cache)selfr3   rB   r8   r7   conv_kernel_sizessm_state_sizei	__class__rA   r0   rF   i   sD   	
   z)HybridMambaAttentionDynamicCache.__init__)	r&   r'   r(   r)   r*   Zfloat16r   rF   __classcell__r/   r/   rY   r0   r2   [   s    "r2   c                       s8   e Zd Zddef fddZe edd Z  Z	S )BambaRotaryEmbeddingNr3   c                    s   t    t|dr|jd ur|jd|jd| _nd| _|j| _|j| _|| _	t
| j | _| | j	|\}| _| jd|dd | j| _d S )Nrope_scaling	rope_typetypedefaultinv_freqF)
persistent)rE   rF   hasattrr]   getr^   Zmax_position_embeddingsZmax_seq_len_cachedZoriginal_max_seq_lenr3   r   Zrope_init_fnattention_scalingZregister_bufferra   Zoriginal_inv_freq)rU   r3   r7   ra   rY   r/   r0   rF      s   
zBambaRotaryEmbedding.__init__c           
      C   s   | j d d d d f  |jd dd|j}|d d d d d f  }t|jjtr6|jjdkr6|jjnd}t	j
|dd+ | |  dd}t	j||fdd	}| | j }| | j }	W d    n1 smw   Y  |j|jd
|	j|jd
fS )Nr   r   ZmpscpuF)device_typeenabledr5   dimr8   )ra   floatexpandshapetor7   
isinstancer_   strr*   Zautocast	transposecatcosre   sinr8   )
rU   xposition_idsZinv_freq_expandedZposition_ids_expandedrh   ZfreqsZembru   rv   r/   r/   r0   forward   s   0&zBambaRotaryEmbedding.forwardN)
r&   r'   r(   r   rF   r*   Zno_gradr   ry   r[   r/   r/   rY   r0   r\      s
    r\   c                 C   sH   | dd| j d d f }| d| j d d df }tj| |fddS )z*Rotates half the hidden dims of the input..Nrf   r5   rj   )ro   r*   rt   )rw   x1Zx2r/   r/   r0   rotate_half   s   r|   hidden_statesn_repreturnc                 C   s^   | j \}}}}|dkr| S | dddddddddf |||||} | ||| ||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)ro   rn   reshape)r}   r~   batchnum_key_value_headsslenhead_dimr/   r/   r0   	repeat_kv   s
   0r           modulequerykeyvalueattention_maskscalingdropoutc                 K   s   t || j}t || j}	t||dd| }
|d ur3|d d d d d d d |jd f }|
| }
tjj|
dtj	d
|j}
tjj|
|| jd}
t|
|	}|dd }||
fS )Nr5   r
   rf   )rk   r8   )ptrainingr   )r   num_key_value_groupsr*   matmulrs   ro   r   
functionalZsoftmaxfloat32rp   r8   r   r   
contiguous)r   r   r   r   r   r   r   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr/   r/   r0   eager_attention_forward   s   
&r   c                 C   s   | |}| |}|jd }| dd|f | d|df }}|dd|f |d|df }	}
|| t||  }|	| t|	|  }tj||gdd}tj||
gdd}||fS )a  Applies Rotary Position Embedding to the query and key tensors.

    Removes the interleaving of cos and sin from GLM

    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.
    rf   .Nrj   )	unsqueezero   r|   r*   rt   )qkru   rv   rx   Zunsqueeze_dimZ
rotary_dimZq_rotZq_passZk_rotZk_passZq_embedZk_embedr/   r/   r0   apply_rotary_pos_emb   s   


""r   c                       s   e Zd ZdZdedef fddZ		ddejde	ejejf d	e
ej d
e
e de
ej dee de	eje
ej e
e	ej  f fddZ  ZS )BambaAttentionz=Multi-headed attention from 'Attention Is All You Need' paperr3   	layer_idxc                    s   t    || _|| _t|d|j|j | _|j|j | _	| jd | _
|j| _d| _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j| j |j|jd| _d S )Nr   g      Tbias)rE   rF   r3   r   getattrrP   Znum_attention_headsr   r   r   r   attention_dropoutZ	is_causalr   LinearZattention_biasq_projk_projv_projo_proj)rU   r3   r   rY   r/   r0   rF     s(   
zBambaAttention.__init__Nr}   position_embeddingsr   past_key_valuecache_positionr   r   c                 K   sH  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}t|	|
||\}	}
|d urW|||d}||
|| j	|\}
}t
}| jjdkrw| jjdkrq|ddrqtd	 nt| jj }|| |	|
||f| jsd
n| j| jd|\}}|jg |dR   }| |}||fS )Nrf   r   r5   )rv   ru   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   r   )ro   r   r   viewrs   r   r   r   updater   r   r3   _attn_implementationrd   loggerwarning_oncer   r   r   r   r   r   r   )rU   r}   r   r   r   r   r   Zinput_shapeZhidden_shapeZquery_statesr   r   ru   rv   Zcache_kwargsZattention_interfacer   r   r/   r/   r0   ry   #  s@   	

zBambaAttention.forwardr#   )r&   r'   r(   r)   r   r-   rF   r*   Tensorr   r   r   r+   r   r   ry   r[   r/   r/   rY   r0   r   	  s(    r   c                       s(   e Zd Zd fdd	ZdddZ  ZS )	BambaRMSNormGatedư>c                    s&   t    tt|| _|| _d S rz   rE   rF   r   	Parameterr*   onesweightvariance_epsilonrU   rP   epsrY   r/   r0   rF   W  s   

zBambaRMSNormGated.__init__Nc                 C   sj   |j }|tj}|d ur|tj|tj }|djddd}|t	|| j
  }| j|| S Nr5   rf   T)Zkeepdim)r8   rp   r*   r   r   r   silupowmeanrsqrtr   r   )rU   r}   gateinput_dtypevariancer/   r/   r0   ry   \  s   zBambaRMSNormGated.forwardr   rz   r&   r'   r(   rF   ry   r[   r/   r/   rY   r0   r   V  s    r   input_tensorpad_sizec                 C   sH   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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   Zconstant)moder   )lenro   r*   r   r   pad)r   r   Z	pad_shaper/   r/   r0   pad_tensor_by_sizek  s   2r   c                 C   sX   t | |} t| jdkr| | jd d|| jd S | | 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   rf   r5   )r   r   ro   r   )r   r   
chunk_sizer/   r/   r0   reshape_into_chunksv  s   
r   c                 C   s   |  d}| d jg |   |R  } tjtj||| jtjddd}| | d} tj| dd}tjtj||| jtjddd}|| tj	 }|S )zo
    More stable segment sum calculation. Uses cumulative sums and masking instead of direct subtractions.
    rf   .Nr6   Zdiagonalr   r   rj   )
sizern   r*   Ztrilr   r7   boolmasked_fillcumsuminf)r   r   maskZtensor_segsumr/   r/   r0   segment_sum  s   
  r   c                 C   sN   |dur%|j d dkr%|j d dkr%| j}| |dddddf  |} | S )zm
    Tunes out the hidden states for padding tokens, see https://github.com/state-spaces/mamba/issues/66
    Nr   r   )ro   r8   rp   )r}   r   r8   r/   r/   r0   apply_mask_to_padding_states  s   $ r   c                       s   e Zd ZdZdedef fddZ				ddejde	e
 d	e	ej d
e	ej de	ej f
ddZ			dde	e
 d	e	ej d
e	ej fddZ				dde	e
 d	e	ej d
e	ej de	ej fddZ  ZS )
BambaMixeruO  
    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 HybridCache 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
    r3   r   c                    s  t    |j| _|j| _|j| _|j| _t	|j
| j | _|| _|j| _|j| _t|j | _|j| _|j| _|j| _|j| _|j| _dtdf| _d| _d| _ | jd| j | j  | _!t"j#| j!| j!|j| j| j!| jd d| _$| j| j! | j }t"j%| j|| jd| _&t"'t()| j| _*t(+d| jd }t"'t(,|| _-d	| j-_.t/| j| jd
| _0t"'t()| j| _1d	| j1_.t"j%| j| j| jd| _2t3st45d d S t45d d S )Nr   r   gMbP?g?r5   r   )Zin_channelsZout_channelsr   Zkernel_sizegroupspaddingr   Tr   a  The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1dzDThe fast path for Bamba will be used when running the model on a GPU)6rE   rF   rR   	num_headsrP   rJ   rW   rI   rV   r-   rO   intermediate_sizer   Zmamba_conv_biasuse_conv_bias
hidden_act
activationr	   actZmamba_proj_biasZuse_biasrms_norm_epsZlayer_norm_epsilonrQ   n_groupsrS   r   Zmamba_chunk_sizer   rm   time_step_limitZtime_step_minZtime_step_maxconv_dimr   Conv1dconv1dr   in_projr   r*   r   dt_biasarangelogA_logZ_no_weight_decayr   normDout_projis_fast_path_availabler   r   )rU   r3   r   Zprojection_sizeArY   r/   r0   rF     s\   

	zBambaMixer.__init__Nr}   cache_paramsr   r   r%   c                 C   s  t ||}| |}|j\}}}	| j| j }
|d uoD|joD|dkoD|j| j jd |j| j jd   ko8|kn  oD|d uoD|d dk}|r)|	dj
| j| j| jgdd\}}}t||j| j | jj	d| jj| j}tj
|| j|
|
gdd\}}}t| j  }|d d d df d d d d d f d| j| jjtjd}|d d d d d f dd| j}| jd d d df d| j}| jd d d df d| j}||| j|jd | j }||| j|jd | j }||| j| j}t|j| j ||||||d |dd
}||| j| j }| ||}|  |d d d df }|S t| j  }| j!d	td
fkr>i nd| j!i}| j"r||d u r|t#|| jj	d| jj| j|f| j| j$|| j| jj| jj%| j j| j j| j| jddd|}|S |j
| j| j| jgdd\}}}|d ur|&dd}t'j()|| j*|jd  df}|j| j +| | jdvr| ,| |&dddd |f &dd}nt-|&dd| jj	d| jj| j|d&dd}t ||}tj
|| j|
|
gdd\}}}t.|||d| j|||||| jd|||| jdf| j$| jd |d| jdd|\}}|d ur:|d ur:|j| j +| |||d}| ||}|  |}|S )Nr   r   rf   rj   .rl   T)zr   dt_softplusr   r   Zdt_limitF)r   r   r%   r   Zrmsnorm_weightZrmsnorm_epsZoutproj_weightZoutproj_biasZheaddimZngroupsZnorm_before_gatereturn_final_statesr5   )r   Zswish)rw   r   r   r   r%   )r   r   r   r%   r   r   r   )/r   r   ro   r   rW   rH   rK   r   rL   squeezesplitr   r   r   r"   r   r   r   r   r*   expr   rm   rn   r   rp   r   r   r   r   r   r   r   r   r   r    r   r   rs   r   r   r   rV   copy_r   r!   r   )rU   r}   r   r   r   r%   projected_statesrB   seq_lenr@   Zgroups_time_state_sizeuse_precomputed_statesr   hidden_states_B_CdtBCr   r   r   Zhidden_states_reshapedoutZdt_limit_kwargshidden_states_B_C_transposedrK   scan_output	ssm_stater/   r/   r0   cuda_kernels_forward  s  
	




<"
^"V
$




zBambaMixer.cuda_kernels_forwardc           3   
      s  |j \}}}|j}t||}|}	|	jjjjgdd\}
}}|d uoQ|joQ|dkoQ|j	j
 j d |jj
 j d   koE|kn  oQ|d uoQ|d dk}|r|j	j
 jddd|j	j
< |d d dd d f |j	j
 j|j	j
 d d d d df< |j	j
 jjjjd}tj|jjd dd}jr|jj }|}n8|d ur|dd}tj|j|j d  df}|j	j
 | |dddd |f dd}t||}tj|jjj jj gdd\}}}tj !  }|r[|jj
 j}|d d dd d f d d d df }|dd"||j d j#}j$d	 "j$j d j#}tjj%|||j }t&|j'd j'd }|d
 "jj#jjtj(d}t|d	 | j|d}|)|jddd d d f }|"|jjj |j d * }|)|d|j d }|d	 |dd d d f  }|)|dj#}||d	  j|d}|jj
 |jj
 | |  |)|jddd d d f }|"|jjj |j d * }|)|d|j d }|jj
 j|j|jd}|+|j j#j}|+|j jd}t,||}|+|jj#}j-d	 "j-j d j#}|||  |j}|)|dd d d df }ntj%|j$ }t&|j'd j'd }|)||dj#! }|)||dj! }|)||dj! }|j.jj djd}|j.jj djd}j/|j/  j/  j-d	 t0|  }||d	  }||j| } fdd||||fD \}}}}|1dddd}tj2|dd}tt3|} |d d d d d d d d d d d f |d d d d d d d d d d d f  }!|!jdd}"|"d	 | 1dddddd	  }#|#jdd}$|$d	 |d d d d d f  jdd}%t|d d d d d d dd f | }&||&1ddddd	  }'|'dd d d f |d	  jdd}(|r|jj
 d d d df j|(jd})nt4|(d d d df })tj5|)|(gdd}(tt3tj|d d d d d d df d}*|*dd}*|*d
 |(d d d d d df  jdd}+|+d d d df |+d d df }(},t|}-|dd d d f |(d d d d d df  }.|-1dddd}/|.d|/d	  }0|%|0 }|)|djj#}|| } dkrB|d d d |d d d d f }|)||d}|,d ur_|d ur_|jj
 |, d|_6||
}17|1|}2|2S )Nrf   rj   r   r   )Zshiftsdimsr9   r5   .r   ).NNrl   r6   )rk   Zoutput_sizec                    s   g | ]	}t | jqS r/   )r   r   )r?   tr   rU   r/   r0   rC   1  s    z,BambaMixer.torch_forward.<locals>.<listcomp>r
   r   r   )r   r   T)8ro   r8   r   r   r   r   r   r   rH   rK   r   rL   Zrollrp   r7   r   r   r*   sumr   r   r   r   rs   r   r   r   rV   r   r   rW   r   r   rm   rn   r   r   Zsoftplusclampr   r   r   r   r   Zbmmr   Zrepeat_interleaver   r   Zpermuter   r   Z
zeros_likert   r   r   )3rU   Zinput_statesr   r   r   rB   r  r@   r8   r   r   r  r  r  rK   r  r}   r  r  r   Zcache_devicer   ZdAZdBZdBxrL   Zssm_states_reshapedZ
C_reshapedyr   Z
D_residualZA_cumsumLZG_intermediateGZM_intermediateMZY_diagZdecay_statesZB_decayZstatesZprevious_statesZdecay_chunkZ
new_statesr
  Zstate_decay_outZC_times_statesZstate_decay_out_permutedZY_offr	  Zcontextualized_statesr/   r  r0   torch_forward  s   


@,
$"$$$P&*"&0(&
*
 zBambaMixer.torch_forwardc                 K   s   t rd| jjjjv r| |||||S |d urtd|j}|d ur@|jd dkr@|jd dkr@||d d d d d f  	|}| 
||||S )Ncudaz\`seq_idx` support requires fast path support. Please install `mamba_ssm` and `causal_conv1d`r   r   )r   r   r   r7   r_   r  NotImplementedErrorr8   ro   rp   r  )rU   r}   r   r   r   r%   r   r8   r/   r/   r0   ry   x  s   	$ zBambaMixer.forward)NNNN)NNN)r&   r'   r(   r)   r   r-   rF   r*   r   r   r2   r+   r.   r  r  ry   r[   r/   r/   rY   r0   r     sV    F
 .
 Tr   c                       s$   e Zd Z fddZdd Z  ZS )BambaMLPc                    sx   t    || _|j| _|j| _tj| j| j|jd| _tj| j| j|jd| _	tj| j| j|jd| _
t|j | _d S )Nr   )rE   rF   r3   rP   r   r   r   Zmlp_bias	gate_projup_proj	down_projr	   r   act_fnrU   r3   rY   r/   r0   rF     s   
zBambaMLP.__init__c                 C   s$   |  | | || | }|S rz   )r  r  r  r  )rU   rw   r  r/   r/   r0   ry     s    zBambaMLP.forwardr   r/   r/   rY   r0   r    s    
r  ZRMSNormc                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	BambaRMSNormr   c                    s&   t    tt|| _|| _dS )z;
        BambaRMSNorm is equivalent to T5LayerNorm
        Nr   r   rY   r/   r0   rF     s   

zBambaRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S r   )	r8   rp   r*   r   r   r   r   r   r   )rU   r}   r   r   r/   r/   r0   ry     s
   zBambaRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler   ro   r   rU   r/   r/   r0   
extra_repr  s   zBambaRMSNorm.extra_reprr   )r&   r'   r(   rF   ry   r!  r[   r/   r/   rY   r0   r    s    r  c                       s   e Zd Zddededef fd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ejejf  dee deeje	eejejf  f fddZ  ZS )BambaDecoderLayerr4   r3   r   
layer_typec                    s   t    d}|dkrtnd }||| _t|j|jd| _t|j|jd| _|| _	|dkr6t
||d| _d S |dkrBt||| _d S td)Nr   r   r4   )r3   r   	attentionzInvalid layer_type)rE   rF   r  feed_forwardr  rP   r   input_layernormpre_ff_layernormr#  r   r4   r   	self_attn
ValueError)rU   r3   r   r#  Znum_expertsZffn_layer_classrY   r/   r0   rF     s   

zBambaDecoderLayer.__init__NFr}   r   rx   r   r   	use_cacher   r   r   r   c	                 K   s   |}
|  |}| jdkr| jd||||d|	}d}n| jdkr4| jd||||||||d|	\}}|
| }|}
| |}| |}|
| }|f}|rR||f7 }|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, sequence_length)` where padding elements are indicated by 0.
            past_key_value (`HybridMambaAttentionDynamicCache`, *optional*): cached past key and value projection states
            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`).
            cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
                Indices depicting the position of the input sequence tokens in the sequence.
            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. Can be used to provide `BambaFlashAttentionKwargs` for
                padding-free training and/or improve torch.compile performance.
        r4   )r}   r   r   r   Nr$  )r}   r   rx   r   r   r*  r   r   r/   )r&  r#  r4   r(  r'  r%  )rU   r}   r   rx   r   r   r*  r   r   r   ZresidualZself_attn_weightsoutputsr/   r/   r0   ry     sD   "


	



zBambaDecoderLayer.forward)r4   )NNNFFNN)r&   r'   r(   r   r-   rr   rF   r*   r   r   r+   r2   r   r   r   r$   FloatTensorry   r[   r/   r/   rY   r0   r"    s<    	
r"  c                   @   s:   e Zd ZeZdZdZdgZdZdZ	dZ
dZdZdd ZdS )BambaPreTrainedModelmodelTr"  past_key_valuesc                 C   s   | j j}t|tjtjfr%|jjjd|d |j	d ur#|j	j
  d S d S t|ttfr5|jjd d S t|tjrV|jjjd|d |jd urT|jj|j 
  d S d S t|try|jjd ttd|jd |j_|jjd d S d S )Nr   )r   stdg      ?r   )r3   Zinitializer_rangerq   r   r   r   r   dataZnormal_r   Zzero_r   r  Zfill_	Embeddingpadding_idxr   r   r*   r   r   r   r   r   )rU   r   r0  r/   r/   r0   _init_weights   s$   


z"BambaPreTrainedModel._init_weightsN)r&   r'   r(   r   Zconfig_classZbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_skip_keys_device_placementZ_supports_flash_attn_2Z_supports_sdpaZ_supports_cache_classZ_is_statefulr4  r/   r/   r/   r0   r-    s    r-  c                       s  e Zd Zdef fddZdd Zdd Zee									d!d	e	e
j d
e	e
j de	e
j de	e de	e
j de	e de	e de	e de	e
j dee defddZd
e
jde
jde
jdedef
ddZed
e
jdedede
jde
jdefddZdd  Z  ZS )"
BambaModelr3   c                    s   t  | |j| _|j| _t|j|j| j| _g }t	|j
D ]}|t|||j| d q t|| _|j| _t|j|jd| _t|d| _d| _|   d S )N)r   r#  r   )r3   F)rE   rF   Zpad_token_idr3  
vocab_sizer   r2  rP   embed_tokensrM   rN   rT   r"  rG   Z
ModuleListlayersr   r  r   final_layernormr\   
rotary_embgradient_checkpointing	post_init)rU   r3   Zdecoder_layersrX   rY   r/   r0   rF   4  s   zBambaModel.__init__c                 C      | j S rz   r7  r   r/   r/   r0   get_input_embeddingsG     zBambaModel.get_input_embeddingsc                 C   
   || _ d S rz   r>  rU   r   r/   r/   r0   set_input_embeddingsJ     
zBambaModel.set_input_embeddingsN	input_idsr   rx   r/  inputs_embedsr*  r   output_hidden_statesr   r   r   c
                 K   s  |d ur|n| j j}|d ur|n| j j}|d ur|n| j j}|d u |d uA r*td| jr9| jr9|r9td d}|d u rB| 	|}|}|rO|d u rOtd |	d u r^t
j|jd |jd}	|d u rg|	d}| |||	||}| ||	}| ||}|rdnd }|rdnd }| jD ]P}|jd	kr|n|}|r||f7 }| jr| jr| t|jfi |
|||||||	|	}n||f||||||	|d
|
}|d }|r|d d ur||d f7 }q| |}|r||f7 }|r|jsd|_|sd n|}t||||d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`.FzBamba requires an initialized `HybridMambaAttentionDynamicCache` to return a cache. None was provided, so no cache will be returned.r   r9   r   r/   r4   )r   rx   r   r   r*  r   r   T)last_hidden_stater/  r}   
attentions)r3   r   rG  r*  r)  r;  r   r   r   r7  r*   r   ro   r7   r   _update_causal_mask_update_mamba_maskr:  r8  r#  Z_gradient_checkpointing_funcr   __call__r9  rH   r   )rU   rE  r   rx   r/  rF  r*  r   rG  r   r   r}   r   
mamba_maskr   Zall_hidden_statesZall_self_attnsZdecoder_layerZ
layer_maskZlayer_outputsZ
next_cacher/   r/   r0   ry   M  s   




	


zBambaModel.forwardr   c                 C   s   | j jdkr|d urd|v r|S d S |d ur| nd}| j jdkr0|s0tj|||| jdr0d S |j}|jd }t|t	j
rC|jd n|| d }	| j|||	|||jd d}
| j jdkru|d uru|jjd	v ru|sut	|j}t|
|}
|
S )
NZflash_attention_2r   r   r   )rF  Zpast_key_values_lengthZis_trainingr   rf   )sequence_lengthtarget_lengthr8   r   rB   )r  ZxpuZnpu)r3   r   Zget_seq_lengthr   Z_ignore_causal_mask_sdpar   r8   ro   rq   r*   r   5_prepare_4d_causal_attention_mask_with_cache_positionr7   r_   finfominZ_unmask_unattended)rU   r   r   r   r/  r   Zpast_seen_tokensr8   rN  rO  r   	min_dtyper/   r/   r0   rJ    sF   



zBambaModel._update_causal_maskrN  rO  r8   rB   c                 K   s|  | dur|   dkr| }|S t|j}tj||f|||jd}|dkr+tj|dd}|tj||jd|ddk9 }|ddddddf 	|ddd}| dur|
 }| jd }	| ddddddf | ddddddf kdddd| dddf |}
|ddddddd|	f |
 }|dk}|ddddddd|	f |||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.
        Nr   )Z
fill_valuer8   r7   r   r   r9   rf   r   )rk   r*   rQ  rR  fullr7   Ztriur   r   rn   clonero   rp   r   )r   rN  rO  r8   r   rB   r   r   rS  Zmask_lengthZpadding_attention_maskZpadding_maskr/   r/   r0   rP    s2    $
.$  z@BambaModel._prepare_4d_causal_attention_mask_with_cache_positionc                 C   s.   |}|d dks|durt |dkrd}|S )zv
        No need for zeroing states when
            1. Cached forward
            2. Attending to all inputs
        r   Nr   )r*   all)rU   r   r   rM  r/   r/   r0   rK  2  s   "zBambaModel._update_mamba_mask)	NNNNNNNNN)r&   r'   r(   r   rF   r?  rC  r   r   r   r*   r+   r   r2   r,  r   r   r$   r   ry   rJ  staticmethodr-   r8   rP  rK  r[   r/   r/   rY   r0   r5  2  s    	
o
<7r5  c                       s  e Zd ZdgZddiZddgdgfiZ fddZdd	 Zd
d Zdd Z	dd Z
dd Zdd Zee											d'deej deej deej dee deej deej dee dee dee deej d eeejf d!efd"d#Z						$d(d%d&Z  ZS ))BambaForCausalLMzlm_head.weightlm_headZcolwise_repr}   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S )NFr   )
rE   rF   r5  r.  r6  r   r   rP   rY  r<  r  rY   r/   r0   rF   D  s
   
zBambaForCausalLM.__init__c                 C   s   | j jS rz   r.  r7  r   r/   r/   r0   r?  M  s   z%BambaForCausalLM.get_input_embeddingsc                 C   s   || j _d S rz   r[  rB  r/   r/   r0   rC  P  s   z%BambaForCausalLM.set_input_embeddingsc                 C   r=  rz   rY  r   r/   r/   r0   get_output_embeddingsS  r@  z&BambaForCausalLM.get_output_embeddingsc                 C   rA  rz   r\  )rU   Znew_embeddingsr/   r/   r0   set_output_embeddingsV  rD  z&BambaForCausalLM.set_output_embeddingsc                 C   rA  rz   r.  )rU   decoderr/   r/   r0   set_decoderY  rD  zBambaForCausalLM.set_decoderc                 C   r=  rz   r_  r   r/   r/   r0   get_decoder\  r@  zBambaForCausalLM.get_decoderNr   rE  r   rx   r/  rF  labelsr*  r   rG  r   logits_to_keepr   c                 K   s   |dur|n| j j}|	dur|	n| j j}	| jd||||||||	|
d	|}|j}t|tr4t| dn|}| |dd|ddf }d}|durX| j	d||| j j
d|}t|||j|j|jdS )aJ  
        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, BambaForCausalLM

        >>> model = BambaForCausalLM.from_pretrained("...")
        >>> tokenizer = AutoTokenizer.from_pretrained("...")

        >>> 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)	rE  r   rx   r/  rF  r*  r   rG  r   )rZ  rc  r6  )lossrZ  r/  r}   rI  r/   )r3   r   rG  r.  rH  rq   r-   slicerY  Zloss_functionr6  r   r/  r}   rI  )rU   rE  r   rx   r/  rF  rc  r*  r   rG  r   rd  r   r+  r}   Zslice_indicesrZ  re  r/   r/   r0   ry   _  s:   '
zBambaForCausalLM.forwardTc              	   K   s  |d u }	|	s5|d us|d |j d kr"|d d |j d  d f }n!|j d |j d kr4|d d |f }nt| j|j d | j| jd}|d url|d u rl| dd }||dkd |	sl|d d |j d  d f }|d urw|	rwd|i}
nd| i}
|
	||||| jj
|d |
S )Nrf   r   r   r9   rF  rE  )rx   r/  r*  r   rd  r   )ro   r2   r3   r8   r7   longr   Zmasked_fill_r   r   Znum_logits_to_keep)rU   rE  r/  r   rF  r   rx   r*  r   Zempty_past_kvZmodel_inputsr/   r/   r0   prepare_inputs_for_generation  s:   

z.BambaForCausalLM.prepare_inputs_for_generation)NNNNNNNNNNr   )NNNNNT)r&   r'   r(   Z_tied_weights_keysZ_tp_planZ_pp_planrF   r?  rC  r]  r^  ra  rb  r   r   r   r*   r+   r   r2   r,  r   r   r-   r   ry   rh  r[   r/   r/   rY   r0   rX  >  sp    		
LrX  )r5  rX  r-  )r   )Nr   )T	functoolsr   typingr   r   r   r   r   r*   r   Z(transformers.models.jamba.modeling_jambamodelsZjambaZmodeling_jambaZtransformers.activationsr	   Zcache_utilsr   Z
generationr   Zintegrationsr   Zmodeling_attn_mask_utilsr   Zmodeling_flash_attention_utilsr   Zmodeling_outputsr   r   Zmodeling_rope_utilsr   r   Zmodeling_utilsr   r   Zprocessing_utilsr   utilsr   r   r   Zutils.import_utilsr   r   Zconfiguration_bambar   Z+mamba_ssm.ops.triton.selective_state_updater   Z!mamba_ssm.ops.triton.ssd_combinedr   r    Zcausal_conv1dr!   r"   Z
get_loggerr&   r   r$   r2   Moduler\   r|   r   r-   r   rm   r   r   r   r   r   r   r   rV  r   r   r   r  r  r"  r-  r5  rX  __all__r/   r/   r/   r0   <module>   s   
6"

(M   e`   '