o
    wi                    @   s  d dl 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 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$m%Z%m&Z&m'Z' ddl(m)Z)m*Z* ddl+m,Z, e* rd dl-m.Z. d dl/m0Z0m1Z1 ndZ.e) rd dl2m3Z3m4Z4 nd\Z4Z3e'5e6Z7G dd deddZ8G dd de
j9Z9G dd dej:Z;d d! Z<d"ej=d#e>d$ej=fd%d&Z?	'dQd(ej:d)ej=d*ej=d+ej=d,eej= d-e@d.e@fd/d0ZAdRd1d2ZBG d3d4 d4ej:ZCG d5d6 d6ejj:ZDd7ej=d8e>fd9d:ZEd;d< ZFd=d> ZGeHe.e3e4fZId?d@ ZJG dAdB dBej:ZKG dCdD dDej:ZLedEG dFdG dGej:ZMG dHdI dIeZNe%G dJdK dKe!ZOe%G dLdM dMeOZPe%G dNdO dOeOeZQg dPZRdS )S    )CallableOptional	TypedDictUnionN)nn)ACT2FN   )Cache)GenerationMixin)use_kernel_forward_from_hub)AttentionMaskConverter)FlashAttentionKwargs)GradientCheckpointingLayer)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.
    cu_seq_lens_qcu_seq_lens_kmax_length_qmax_length_kseq_idxN)	__name__
__module____qualname____doc__torch
LongTensor__annotations__int	IntTensor r2   r2   e/home/ubuntu/sommelier/.venv/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r:   c                        g | ]}t jg g  d qS r<   r-   tensor.0_
batch_sizer:   r2   r3   
<listcomp>        z=HybridMambaAttentionDynamicCache.__init__.<locals>.<listcomp>c                    r=   r>   r?   rA   rD   r2   r3   rF      rG   )super__init__layers_block_typehas_previous_statemamba_d_convmamba_d_stateconv_states
ssm_statestransformer_layersrangenum_hidden_layersr-   zerosmamba_expandhidden_sizemamba_n_groupsmamba_n_headsmamba_d_headr@   append	key_cachevalue_cache)selfr6   rE   r;   r:   conv_kernel_sizessm_state_sizei	__class__rD   r3   rI   i   sD   	
   z)HybridMambaAttentionDynamicCache.__init__)	r)   r*   r+   r,   r-   float16r   rI   __classcell__r2   r2   r`   r3   r5   [   s    "r5   c                       s8   e Zd Zddef fddZe edd Z  Z	S )BambaRotaryEmbeddingNr6   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)rH   rI   hasattrre   getrf   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr6   r   rope_init_fnattention_scalingregister_bufferri   original_inv_freq)r\   r6   r:   ri   r`   r2   r3   rI      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   mpscpuF)device_typeenabledr8   dimr;   )ri   floatexpandshapetor:   
isinstancerg   strr-   autocast	transposecatcosrq   sinr;   )
r\   xposition_idsinv_freq_expandedposition_ids_expandedrw   freqsembr   r   r2   r2   r3   forward   s   0&zBambaRotaryEmbedding.forwardN)
r)   r*   r+   r   rI   r-   no_gradr   r   rc   r2   r2   r`   r3   rd      s
    rd   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..Nrt   r8   ry   )r~   r-   r   )r   x1x2r2   r2   r3   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)r~   r}   reshape)r   r   batchnum_key_value_headsslenhead_dimr2   r2   r3   	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 )Nr8   r   rt   )rz   r;   )ptrainingr   )r   num_key_value_groupsr-   matmulr   r~   r   
functionalsoftmaxfloat32r   r;   r   r   
contiguous)r   r   r   r   r   r   r   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr2   r2   r3   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.
    rt   .Nry   )	unsqueezer~   r   r-   r   )qkr   r   r   unsqueeze_dim
rotary_dimq_rotq_passk_rotk_passq_embedk_embedr2   r2   r3   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' paperr6   	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)rH   rI   r6   r   getattrrU   num_attention_headsr   r   r   r   attention_dropout	is_causalr   Linearattention_biasq_projk_projv_projo_proj)r\   r6   r   r`   r2   r3   rI     s(   
zBambaAttention.__init__Nr   position_embeddingsr   past_key_valuecache_positionr   r   c                 K   s$  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}|\}}t|	|
||\}	}
|d urW|||d}||
|| j	|\}
}t
}| jjdkret| jj }|| |	|
||f| jsqdn| j| jd|\}}|jg |dR   }| |}||fS )Nrt   r   r8   )r   r   r   eagerr   )r   r   )r~   r   r   viewr   r   r   r   updater   r   r6   _attn_implementationr   r   r   r   r   r   r   )r\   r   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   r   r   cache_kwargsattention_interfacer   r   r2   r2   r3   r   #  s8   	

zBambaAttention.forwardr"   )r)   r*   r+   r,   r   r0   rI   r-   Tensortupler   r	   r.   r   r   r   rc   r2   r2   r`   r3   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 r   rH   rI   r   	Parameterr-   onesweightvariance_epsilonr\   rU   epsr`   r2   r3   rI   P  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 Nr8   rt   T)keepdim)r;   r   r-   r   r   r   silupowmeanrsqrtr   r   )r\   r   gateinput_dtypevariancer2   r2   r3   r   U  s   zBambaRMSNormGated.forwardr   r   r)   r*   r+   rI   r   rc   r2   r2   r`   r3   r   O  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   constant)moder   )lenr~   r-   r   r   pad)r   r   	pad_shaper2   r2   r3   pad_tensor_by_sized  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   rt   r8   )r   r   r~   r   )r   r   
chunk_sizer2   r2   r3   reshape_into_chunkso  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.
    rt   .Nr9   diagonalr   r   ry   )
sizer}   r-   trilr   r:   boolmasked_fillcumsuminf)r   r   masktensor_segsumr2   r2   r3   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   )r~   r;   r   )r   r   r;   r2   r2   r3   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
    r6   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| _$| j| j! | j }t"j%| j|| jd| _&t"'t()| j| _*t(+d| jd }t"'t(,|| _-d	| j-_.t/| j| jd
| _0t"'t()| j| _1d	| j1_.t"j%| j| j| jd| _2t3st45d d S t45d d S )Nr   r  gMbP?g?r8   r   )in_channelsout_channelsr   kernel_sizegroupspaddingr   Tr   a  The fast path is not available because on of `(selective_state_update, causal_conv1d_fn, causal_conv1d_update)` is None. Falling back to the naive implementation. To install follow https://github.com/state-spaces/mamba/#installation and https://github.com/Dao-AILab/causal-conv1dzDThe fast path for Bamba will be used when running the model on a GPU)6rH   rI   rW   	num_headsrU   rM   r^   rL   r]   r0   rT   intermediate_sizer   mamba_conv_biasuse_conv_bias
hidden_act
activationr   actmamba_proj_biasuse_biasrms_norm_epslayer_norm_epsilonrV   n_groupsrX   r   mamba_chunk_sizer   r|   time_step_limittime_step_mintime_step_maxconv_dimr   Conv1dconv1dr   in_projr   r-   r   dt_biasarangelogA_log_no_weight_decayr   normDout_projis_fast_path_availableloggerwarning_once)r\   r6   r   projection_sizeAr`   r2   r3   rI     s\   

	zBambaMixer.__init__Nr   cache_paramsr   r   r(   c                 C   s  t ||}| |}|j\}}}	| j| j }
|d uoD|joD|dkoD|j| j jd |j| j jd   ko8|kn  oD|d uoD|d dk}|r)|	dj
| j| j| jgdd\}}}t||j| j | jj	d| jj| j}tj
|| j|
|
gdd\}}}t| j  }|d d d df d d d d d f d| j| jjtjd}|d d d d d f dd| j}| jd d d df d| j}| jd d d df d| j}||| j|jd | j }||| j|jd | j }||| j| j}t|j| j ||||||d |dd
}||| j| j }| ||}|  |d d d df }|S t| j  }| j!d	td
fkr>i nd| j!i}| j"r||d u r|t#|| jj	d| jj| j|f| j| j$|| 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   rt   ry   .r{   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_statesr8   )r   swish)r   r   r   r  r(   )r   r)  r1  r(   r;  r#  r2  )/r  r"  r~   r  r^   rK   rN   r   rO   squeezesplitr  r  r  r!   r!  r   r   r  r-   expr&  r|   r}   r   r   r   r#  r)  r   r   r(  r*  r  r   r   r   r   r   r   r   r   r]   copy_r  r    r   )r\   r   r0  r   r   r(   projected_statesrE   seq_lenrC   groups_time_state_sizeuse_precomputed_statesr   hidden_states_B_CdtBCr/  r#  r)  hidden_states_reshapedoutdt_limit_kwargshidden_states_B_C_transposedrN   scan_output	ssm_stater2   r2   r3   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
 |, 6||
}17|1|}2|2S )Nrt   ry   r   r   )shiftsdimsr<   r8   .r   ).NNr{   r9   )rz   output_sizec                    s   g | ]	}t | jqS r2   )r   r   )rB   tr   r\   r2   r3   rF   *  s    z,BambaMixer.torch_forward.<locals>.<listcomp>r   r   r   )r   r   )8r~   r;   r  r"  r>  r  r  r  rK   rN   r   rO   rollr   r:   r!  r   r-   sumr=  r  r   r  r   r   r   r   r]   r@  r  r^   r?  r&  r|   r}   r   r#  softplusclampr  r   r   r   r   bmmr)  repeat_interleaver   r   permuter  r  
zeros_liker   r(  r*  )3r\   input_statesr0  r   r   rE   rB  rC   r;   rA  r   rE  rF  rD  rN   rL  r   rG  rH  r/  cache_devicer#  dAdBdBxrO   ssm_states_reshaped
C_reshapedyr)  
D_residualA_cumsumLG_intermediateGM_intermediateMY_diagdecay_statesB_decaystatesprevious_statesdecay_chunk
new_statesrN  state_decay_outC_times_statesstate_decay_out_permutedY_offrM  contextualized_statesr2   rT  r3   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   r:   rg   rO  NotImplementedErrorr;   r~   r   rx  )r\   r   r0  r   r   r(   r   r;   r2   r2   r3   r   p  s   	$ zBambaMixer.forward)NNNN)NNN)r)   r*   r+   r,   r   r0   rI   r-   r   r   r5   r.   r1   rO  rx  r   rc   r2   r2   r`   r3   r    sV    F
 .
 Sr  c                       s$   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   )rH   rI   r6   rU   r  r   r   mlp_bias	gate_projup_proj	down_projr   r  act_fnr\   r6   r`   r2   r3   rI     s   
zBambaMLP.__init__c                 C   s$   |  | | || | }|S r   )r  r  r}  r~  )r\   r   r  r2   r2   r3   r     s    zBambaMLP.forwardr   r2   r2   r`   r3   r{    s    
r{  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   r`   r2   r3   rI     s   

zBambaRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S r   )	r;   r   r-   r   r   r   r   r   r   )r\   r   r   r   r2   r2   r3   r     s
   zBambaRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)r   r   r~   r   r\   r2   r2   r3   
extra_repr  s   zBambaRMSNorm.extra_reprr   )r)   r*   r+   rI   r   r  rc   r2   r2   r`   r3   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 )BambaDecoderLayerr7   r6   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  r7   )r6   r   	attentionzInvalid layer_type)rH   rI   r{  feed_forwardr  rU   r  input_layernormpre_ff_layernormr  r  r7   r   	self_attn
ValueError)r\   r6   r   r  num_expertsffn_layer_classr`   r2   r3   rI     s   

zBambaDecoderLayer.__init__NFr   r   r   r   output_attentions	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.
        r7   )r   r0  r   r   Nr  )r   r   r   r   r  r  r   r   r2   )r  r  r7   r  r  r  )r\   r   r   r   r   r  r  r   r   r   residualself_attn_weightsoutputsr2   r2   r3   r     sD   "


	



zBambaDecoderLayer.forward)r7   )NNNFFNN)r)   r*   r+   r   r0   r   rI   r-   r   r   r.   r5   r   r   r   r#   FloatTensorr   rc   r2   r2   r`   r3   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   )r6   initializer_ranger   r   r   r   r   datanormal_r   zero_r   r  fill_	Embeddingpadding_idxr  r#  r-   r%  r$  r  r&  r)  )r\   r   r  r2   r2   r3   _init_weights  s$   


z"BambaPreTrainedModel._init_weightsN)r)   r*   r+   r   config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_2_supports_sdpa_supports_cache_class_is_statefulr  r2   r2   r2   r3   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 )"
BambaModelr6   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  )r6   F)rH   rI   pad_token_idr  
vocab_sizer   r  rU   embed_tokensrQ   rR   rY   r  rJ   
ModuleListlayersr   r  r  final_layernormrd   
rotary_embgradient_checkpointing	post_init)r\   r6   decoder_layersr_   r`   r2   r3   rI   ,  s   zBambaModel.__init__c                 C      | j S r   r  r  r2   r2   r3   get_input_embeddings?     zBambaModel.get_input_embeddingsc                 C   
   || _ d S r   r  r\   r   r2   r2   r3   set_input_embeddingsB     
zBambaModel.set_input_embeddingsN	input_idsr   r   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 ]5}|jd	kr|n|}|r||f7 }||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   r<   r   r2   r7   )r   r   r   r  r  r   r   T)last_hidden_stater  r   
attentions)r6   r  r  r  r  r  r   r,  r-  r  r-   r$  r~   r:   r   _update_causal_mask_update_mamba_maskr  r  r  r  rK   r   )r\   r  r   r   r  r  r  r  r  r   r   r   r   
mamba_maskr   all_hidden_statesall_self_attnsdecoder_layer
layer_masklayer_outputs
next_cacher2   r2   r3   r   E  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 )
Nflash_attention_2r   r   sdpa)r  past_key_values_lengthis_trainingr   rt   )sequence_lengthtarget_lengthr;   r   rE   )ry  xpunpu)r6   r   get_seq_lengthr   _ignore_causal_mask_sdpar   r;   r~   r   r-   r   5_prepare_4d_causal_attention_mask_with_cache_positionr:   rg   finfomin_unmask_unattended)r\   r   r   r   r  r  past_seen_tokensr;   r  r  r   	min_dtyper2   r2   r3   r    sF   



zBambaModel._update_causal_maskr  r  r;   rE   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   )
fill_valuer;   r:   r   r   r<   rt   r   )rz   r-   r  r  fullr:   triur$  r   r}   cloner~   r   r  )r   r  r  r;   r   rE   r   r   r  mask_lengthpadding_attention_maskpadding_maskr2   r2   r3   r    s2    $
.$  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)r\   r   r   r  r2   r2   r3   r    s   "zBambaModel._update_mamba_mask)	NNNNNNNNN)r)   r*   r+   r   rI   r  r  r   r   r   r-   r.   r   r5   r  r   r   r#   r   r   r  staticmethodr0   r;   r  r  rc   r2   r2   r`   r3   r  *  s    	
b
<7r  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_headcolwise_repr   logitsc                    sH   t  | t|| _|j| _tj|j|jdd| _|j	| _	| 
  d S )NFr   )rH   rI   r  r  r  r   r   rU   r  z_loss_coefficientr  r  r`   r2   r3   rI   /  s   
zBambaForCausalLM.__init__c                 C   s   | j jS r   r  r  r  r2   r2   r3   r  9  s   z%BambaForCausalLM.get_input_embeddingsc                 C   s   || j _d S r   r  r  r2   r2   r3   r  <  s   z%BambaForCausalLM.set_input_embeddingsc                 C   r  r   r  r  r2   r2   r3   get_output_embeddings?  r  z&BambaForCausalLM.get_output_embeddingsc                 C   r  r   r  )r\   new_embeddingsr2   r2   r3   set_output_embeddingsB  r  z&BambaForCausalLM.set_output_embeddingsc                 C   r  r   r  )r\   decoderr2   r2   r3   set_decoderE  r  zBambaForCausalLM.set_decoderc                 C   r  r   r  r  r2   r2   r3   get_decoderH  r  zBambaForCausalLM.get_decoderNr   r  r   r   r  r  labelsr  r  r  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rt| j	d
||| j j
d|}| jdkrt|jddj|jdd }|| j|  }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)	r  r   r   r  r  r  r  r  r   )r  r  r  r   rt   ry   r{   r8   )lossr  r  r   r  r2   )r6   r  r  r  r  r   r0   slicer  loss_functionr  r  	logsumexpr   r;   r   r   r   r  r   r  )r\   r  r   r   r  r  r  r  r  r  r   r  r   r  r   slice_indicesr  r  z_lossr2   r2   r3   r   K  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 )Nrt   r   r   r<   r  r  )r   r  r  r   r  r   )r~   r5   r6   r;   r:   longr  masked_fill_r   r   num_logits_to_keep)r\   r  r  r   r  r   r   r  r   empty_past_kvmodel_inputsr2   r2   r3   prepare_inputs_for_generation  s:   

z.BambaForCausalLM.prepare_inputs_for_generation)NNNNNNNNNNr   )NNNNNT)r)   r*   r+   _tied_weights_keys_tp_plan_pp_planrI   r  r  r  r  r  r  r   r   r   r-   r.   r   r5   r  r   r   r0   r   r   r   rc   r2   r2   r`   r3   r  )  sp    
	
Pr  )r  r  r  )r   )Nr   )Stypingr   r   r   r   r-   r   (transformers.models.jamba.modeling_jambamodelsjambamodeling_jambatransformers.activationsr   cache_utilsr	   
generationr
   integrationsr   modeling_attn_mask_utilsr   modeling_flash_attention_utilsr   modeling_layersr   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.import_utilsr   r   configuration_bambar   +mamba_ssm.ops.triton.selective_state_updater   !mamba_ssm.ops.triton.ssd_combinedr   r   causal_conv1dr    r!   
get_loggerr)   r,  r#   r5   Modulerd   r   r   r0   r   r|   r   r   r   r   r   r   r  r  r+  r  r  r{  r  r  r  r  r  __all__r2   r2   r2   r3   <module>   s   
6"

(F   d`  ,