o
    wiս                     @   s`  d dl 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	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 ddlmZmZmZmZmZ ddlmZm Z  ddl!m"Z" ddl#m$Z$ ddl%m&Z&m'Z'm(Z(m)Z) ddl*m+Z+ e),e-Z.G dd dej/Z0dd Z1d?ddZ2dej3de4dej3fddZ5dd Z6G d d! d!ej/Z7G d"d# d#e7Z8G d$d% d%e7Z9ed&G d'd( d(ej/Z:e7e8e9d)Z;G d*d+ d+eZ<e'G d,d- d-e"Z=G d.d/ d/ej/Z>e'G d0d1 d1e=Z?G d2d3 d3ee&Z@e'G d4d5 d5e=eZAe'd6d7G d8d9 d9e=ZBe'G d:d; d;e=ZCe'G d<d= d=e=ZDg d>ZEdS )@    N)OptionalUnion)nn   )ACT2FN)CacheDynamicCacheStaticCache)GenerationMixin)use_kernel_forward_from_hub)create_causal_mask)FlashAttentionKwargs_flash_attention_forward!flash_attn_supports_top_left_mask)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPastQuestionAnsweringModelOutput SequenceClassifierOutputWithPastTokenClassifierOutput)ROPE_INIT_FUNCTIONSdynamic_rope_update)PreTrainedModel)Unpack)
LossKwargsauto_docstringcan_return_tuplelogging   )DiffLlamaConfigc                       s$   e Zd Z fddZdd Z  ZS )DiffLlamaMLPc                    sr   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _d S NFbias)super__init__confighidden_sizeintermediate_sizer   Linear	gate_projup_proj	down_projr   
hidden_actact_fnselfr&   	__class__ m/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/diffllama/modeling_diffllama.pyr%   ;   s   
zDiffLlamaMLP.__init__c                 C   s$   |  | | || | }|S N)r,   r.   r*   r+   )r0   xr,   r3   r3   r4   forwardE   s    zDiffLlamaMLP.forward)__name__
__module____qualname__r%   r7   __classcell__r3   r3   r1   r4   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..N   dim)shapetorchcat)r6   x1x2r3   r3   r4   rotate_halfJ   s   rE   c                 C   sD   | |}| |}| | t| |  }|| t||  }||fS )a  Applies Rotary Position Embedding to the query and key tensors.

    Args:
        q (`torch.Tensor`): The query tensor.
        k (`torch.Tensor`): The key tensor.
        cos (`torch.Tensor`): The cosine part of the rotary embedding.
        sin (`torch.Tensor`): The sine part of the rotary embedding.
        position_ids (`torch.Tensor`, *optional*):
            Deprecated and unused.
        unsqueeze_dim (`int`, *optional*, defaults to 1):
            The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
            sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
            that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
            k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
            cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
            the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
    Returns:
        `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
    )	unsqueezerE   )qkcossinposition_idsunsqueeze_dimq_embedk_embedr3   r3   r4   apply_rotary_pos_embQ   s
   

rO   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@   expandreshape)rP   rQ   batchnum_key_value_headsslenhead_dimr3   r3   r4   	repeat_kvl   s
   0rY   c                 C   s   ddt d|    S )Ng?g333333?g333333ӿ)mathexp)	layer_idxr3   r3   r4   lambda_init_fnx   s   r]   c                       s   e Zd ZdZddedee f fddZ						ddej	d	e
ej	ej	f d
eej	 deej dee dededeej de
ej	eej	 ee
ej	  f fddZ  ZS )DiffLlamaAttentionz=Multi-headed attention from 'Attention Is All You Need' paperNr&   r\   c                    s  t    || _|| _|d u rtd| jj d |j| _|j	| _	|j
| _t|d| j	| j | _|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| _tj| j| j | j	|jd| _t|| _ttjd|j| jfd| _ ttjd|j| jfd| _!ttjd|j| jfd| _"ttjd|j| jfd| _#tj$d| j |j%d	d
| _&d S )NzInstantiating z without passing a `layer_idx` is not recommended and will lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` when creating this class.rX   Tr"   r   )sizer=   F)epselementwise_affine)'r$   r%   r&   r\   loggerwarning_oncer2   r8   attention_dropoutr'   num_attention_heads	num_headsgetattrrX   rV   num_key_value_groupsmax_position_embeddings
rope_theta	is_causalr   r)   attention_biasq_projk_projv_projo_projr]   lambda_init	ParameterrA   normallambda_std_dev	lambda_q1	lambda_k1	lambda_q2	lambda_k2RMSNormrms_norm_eps	groupnormr0   r&   r\   r1   r3   r4   r%      s4   

zDiffLlamaAttention.__init__FrP   position_embeddingsattention_maskrK   past_key_valueoutput_attentions	use_cachecache_positionrR   c	                 K   sj  |  \}
}}|}| |}| |}| |}||
|| j| jdd}||
|| j| jdd}||
|| j| jdd}|\}}t	||||\}}|d urd|||d}|
||| j|\}}t|| j}t|| j}tjtj|ddddd}|dddd}t||ddt| j }|d ur|d d d d d d d |jd f }|| }tjj|dtjd|j}tjj|| j| jd	}ttj | j!| j" dtjd|j}ttj | j#| j$ dtjd|j}|| | j% }t||}tj|ddd\}}|||  }d| j% | &| }|dd' }|(|
|d}| )|}|s1d }||fS )
Nr   r=   rJ   rI   r   r>   r<   r   r?   dtype)ptraining)*r_   rm   rn   ro   viewrf   rX   	transposerV   rO   updater\   rY   rh   rA   rB   chunkrepeatmatmulrZ   sqrtr@   r   
functionalsoftmaxfloat32tor   dropoutrd   r   r[   sumru   rv   rw   rx   rq   r{   
contiguousrT   rp   )r0   rP   r}   r~   rK   r   r   r   r   kwargsbsz
target_len_q_lenquery_states
key_statesvalue_statesrI   rJ   cache_kwargsattn_weightscausal_masklambda_1lambda_2lambda_fullattn_outputattn_output1attn_output2r3   r3   r4   r7      sP   


 &  
zDiffLlamaAttention.forwardr5   NNNFFN)r8   r9   r:   __doc__r   r   intr%   rA   Tensortuple
LongTensorr   boolr7   r;   r3   r3   r1   r4   r^   |   s8    &	r^   c                       s   e Zd ZdZ fddZ						ddejdeejejf deej	 d	eej	 d
ee
 dededeej	 deejeej eeej  f fddZ  ZS )DiffLlamaFlashAttention2aN  
    DiffLlama flash attention module. This module inherits from `DiffLlamaAttention` as the weights of the module stays
    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
    flash attention and deal with padding tokens in case the input contains any of them.
    c                    s   t  j|i | t | _d S r5   )r$   r%   r   _flash_attn_uses_top_left_mask)r0   argsr   r1   r3   r4   r%      s   z!DiffLlamaFlashAttention2.__init__NFrP   r}   r~   rK   r   r   r   r   rR   c	                 C   s&  t |tr	tdd}| \}	}
}| |}| |}| |}||	|
| j| j	
dd}||	|
| j| j	
dd}||	|
| j| j	
dd}|d u r]td | ||\}}n|\}}t||||\}}|d ur|||d}|||| j|\}}|
dd}|
dd}|
dd}| jr| jnd}|j}|jjdkr|jjnd	}|tjkrt rttd
rt|nt }nt| jdr| jj}n| jjj}td| d | |}| |}| |}tj!|ddd\}}|"dddd}|"dddd}t#|||||
||t$| dd | j%| j&d
}t#|||||
||t$| dd | j%| j&d
}tj'||gdd}tj!|ddd\}}t(tj)| j*| j+ dtjd |j}t(tj)| j,| j- dtjd |j}|| | j. }|||  }d| j. | /| }|0|	|
d1 }| 2|}|sd }||fS )Nz`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformersFr   r=   aY  The attention layers in this model are transitioning from computing the RoPE embeddings internally through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed `position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be removed and `position_embeddings` will be mandatory.r           mpscpuget_autocast_dtype_pre_quantization_dtypezThe input hidden states seems to be silently casted in float32, this might be related to the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in .r>   sliding_window)rK   r   r   use_top_left_maskrk   r<   r   )3
isinstancer	   
ValueErrorr_   rm   rn   ro   r   rf   rX   r   rV   rb   rc   
rotary_embrO   r   r\   r   rd   r   devicetyperA   r   is_autocast_enabledhasattrr   get_autocast_gpu_dtyper&   r   weightr   r   r   r   rg   r   rk   rB   r[   r   ru   rv   rw   rx   rq   r{   rT   r   rp   )r0   rP   r}   r~   rK   r   r   r   r   r   r   r   r   r   r   rI   rJ   r   dropout_rateinput_dtypedevice_typetarget_dtypevalue_states1value_states2r   r   r   r   r   r   r   r3   r3   r4   r7      s   











  
z DiffLlamaFlashAttention2.forwardr   )r8   r9   r:   r   r%   rA   r   r   r   r   r   r   r7   r;   r3   r3   r1   r4   r      s8    	
r   c                   @   s   e Zd ZdZ						ddejdeejejf deej deej dee	 d	e
d
e
deej deejeej eeej  f fddZdS )DiffLlamaSdpaAttentiona   
    DiffLlama attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
    `DiffLlamaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
    SDPA API.
    NFrP   r}   r~   rK   r   r   r   r   rR   c	                 K   sd  |  \}
}}| |}| |}| |}||
|| j| jdd}||
|| j| jdd}||
|| j| jdd}|\}}t	||||\}}|d urb|||d}|
||| j|\}}t|| j}t|| j}tjtj|ddddd}|dddd}|}|d ur|d d d d d d d |jd f }|jjdkr|d ur| }| }| }|d u r|dkrdnd	}tjjj||||| jr| jnd
|d}tj|ddd\}}ttj| j| j dtjd |j!}ttj| j"| j# dtjd |j!}|| | j$ }|||  }d| j$ | %| }|dd }||
|d}| &|}|d fS )Nr   r=   r   r>   r<   r   cudaTFr   )	attn_mask	dropout_prk   r   )'r_   rm   rn   ro   r   rf   rX   r   rV   rO   r   r\   rY   rh   rA   rB   r   r   r@   r   r   r   r   r   scaled_dot_product_attentionr   rd   r[   r   ru   rv   r   r   r   rw   rx   rq   r{   rp   )r0   rP   r}   r~   rK   r   r   r   r   r   r   r   r   r   r   r   rI   rJ   r   r   rk   r   r   r   r   r   r   r3   r3   r4   r7     s\   


&	  
zDiffLlamaSdpaAttention.forwardr   )r8   r9   r:   r   rA   r   r   r   r   r   r   r7   r3   r3   r3   r4   r     s6    	r   ry   c                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	DiffLlamaRMSNormư>c                    s&   t    tt|| _|| _dS )z?
        DiffLlamaRMSNorm is equivalent to T5LayerNorm
        N)r$   r%   r   rr   rA   onesr   variance_epsilon)r0   r'   r`   r1   r3   r4   r%     s   

zDiffLlamaRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )Nr=   r<   T)keepdim)	r   r   rA   r   powmeanrsqrtr   r   )r0   rP   r   variancer3   r3   r4   r7     s
   zDiffLlamaRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)r   r   r@   r   r0   r3   r3   r4   
extra_repr  s   zDiffLlamaRMSNorm.extra_repr)r   )r8   r9   r:   r%   r7   r   r;   r3   r3   r1   r4   r     s    r   )eagerflash_attention_2sdpac                       s   e Zd Zdedef fddZ							ddejdeej d	eej	 d
ee
 dee dee deej	 deeejejf  dee deejeeejejf  f fddZ  ZS )DiffLlamaDecoderLayerr&   r\   c                    sX   t    |j| _t|j ||d| _t|| _t|j|j	d| _
t|j|j	d| _d S )N)r&   r\   r`   )r$   r%   r'   DIFFLLAMA_ATTENTION_CLASSES_attn_implementation	self_attnr    mlpr   rz   input_layernormpost_attention_layernormr|   r1   r3   r4   r%     s   

zDiffLlamaDecoderLayer.__init__NFrP   r~   rK   r   r   r   r   r}   r   rR   c	                 K   st   |}
|  |}| jd||||||||d|	\}}|
| }|}
| |}| |}|
| }|f}|r8||f7 }|S )N)rP   r~   rK   r   r   r   r   r}   r3   )r   r   r   r   )r0   rP   r~   rK   r   r   r   r   r}   r   residualself_attn_weightsoutputsr3   r3   r4   r7     s.   
	



zDiffLlamaDecoderLayer.forward)NNNFFNN)r8   r9   r:   r   r   r%   rA   r   r   r   r   r   r   r   r   FloatTensorr7   r;   r3   r3   r1   r4   r     s<    	
r   c                   @   sL   e Zd ZeZdZdZdgZdgZdZ	dZ
dZdZdZdZdZdZdd ZdS )	DiffLlamaPreTrainedModelmodelTr   past_key_valuesFc                 C   s  | j j}t|tjr"|jjjd|d |jd ur |jj	  d S d S t|tj
rC|jjjd|d |jd urA|jj|j 	  d S d S t|trQ|jjd d S t|tr|jjd| j j |jjd| j j |jjd| j j |jjd| j j d S d S )Nr   )r   stdg      ?r   )r&   initializer_ranger   r   r)   r   datanormal_r#   zero_	Embeddingpadding_idxr   fill_r^   ru   rt   rv   rw   rx   )r0   moduler   r3   r3   r4   _init_weights8  s&   



z&DiffLlamaPreTrainedModel._init_weightsN)r8   r9   r:   r   config_classbase_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_3_supports_flash_attn_2_supports_sdpa_supports_flex_attn_supports_cache_class_supports_quantized_cache_supports_static_cache_supports_attention_backendr   r3   r3   r3   r4   r   (  s    r   c                       s8   e Zd Zddef fddZe edd Z  Z	S )DiffLlamaRotaryEmbeddingNr&   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_typer   defaultinv_freqF)
persistent)r$   r%   r   r  getr  ri   max_seq_len_cachedoriginal_max_seq_lenr&   r   rope_init_fnattention_scalingregister_bufferr  original_inv_freq)r0   r&   r   r  r1   r3   r4   r%   L  s   
z!DiffLlamaRotaryEmbedding.__init__c           
      C   s   | j d d d d f  |jd dd|j}|d d d d d f  }t|jjtr6|jjdkr6|jjnd}t	j
|dd+ | |  dd}t	j||fdd	}| | j }| | j }	W d    n1 smw   Y  |j|jd
|	j|jd
fS )Nr   r<   r   r   r   F)r   enabledr=   r>   )r   )r  floatrS   r@   r   r   r   r   strrA   autocastr   rB   rI   r  rJ   r   )
r0   r6   rK   inv_freq_expandedposition_ids_expandedr   freqsembrI   rJ   r3   r3   r4   r7   ]  s   0&z DiffLlamaRotaryEmbedding.forwardr5   )
r8   r9   r:   r   r%   rA   no_gradr   r7   r;   r3   r3   r1   r4   r  K  s
    r  c                       s   e Zd Zdef fddZdd Zdd Zee									dd	e	e
j d
e	e
j de	e
j de	e de	e
j de	e de	e de	e de	e
j dee defddZ  ZS )DiffLlamaModelr&   c                    s   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j jd| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r3   )r   ).0r\   r&   r3   r4   
<listcomp>v  s    z+DiffLlamaModel.__init__.<locals>.<listcomp>r   r  F)r$   r%   pad_token_idr   
vocab_sizer   r   r'   embed_tokens
ModuleListrangenum_hidden_layerslayersr   rz   normr  r   gradient_checkpointing	post_initr/   r1   r  r4   r%   o  s   zDiffLlamaModel.__init__c                 C      | j S r5   r  r   r3   r3   r4   get_input_embeddings     z#DiffLlamaModel.get_input_embeddingsc                 C   
   || _ d S r5   r'  r0   valuer3   r3   r4   set_input_embeddings     
z#DiffLlamaModel.set_input_embeddingsN	input_idsr~   rK   r   inputs_embedsr   r   output_hidden_statesr   flash_attn_kwargsrR   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}t	|t
d tfsFtd|d u rO| |}|rX|d u rXt }|	d u rt|d urd| nd}tj|||jd  |jd}	|d u r}|	d}t| j |||	||d}|}| ||}|rd	nd }|rd	nd }| jd | j j D ]&}|r||f7 }||f||||||	|d
|
}|d }|r||d f7 }q| |}|r||f7 }t||r|nd ||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`.FzBThe `past_key_values` should be either a `Cache` object or `None`.r   r   r   )r&   input_embedsr~   r   r   rK   r3   )r~   rK   r   r   r   r   r}   )last_hidden_stater   rP   
attentions)r&   r   r1  r   r   r$  r   rb   rc   r   r   r   r  r   get_seq_lengthrA   aranger@   r   rF   r   r   r"  r!  r#  r   )r0   r/  r~   rK   r   r0  r   r   r1  r   r2  past_seen_tokensr   rP   r}   all_hidden_statesall_self_attnsdecoder_layerlayer_outputsr3   r3   r4   r7     s   

	
	


zDiffLlamaModel.forward	NNNNNNNNN)r8   r9   r:   r   r%   r(  r-  r   r   r   rA   r   r   r   r   r   r   r   r   r7   r;   r3   r3   r1   r4   r  m  sL    	
r  c                   @   s   e Zd ZdS )KwargsForCausalLMN)r8   r9   r:   r3   r3   r3   r4   r?    s    r?  c                       s
  e Zd ZdgZddiZddgdgfiZ fddZdd	 Zd
d Zdd Z	dd Z
dd Zdd Zee											d%deej deej deej dee deej deej dee dee dee deej d eeejf d!ee d"efd#d$Z  ZS )&DiffLlamaForCausalLMzlm_head.weightlm_headcolwise_reprP   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S r!   )
r$   r%   r  r   r  r   r)   r'   rA  r%  r/   r1   r3   r4   r%     s
   
zDiffLlamaForCausalLM.__init__c                 C      | j jS r5   r   r  r   r3   r3   r4   r(       z)DiffLlamaForCausalLM.get_input_embeddingsc                 C      || j _d S r5   rE  r+  r3   r3   r4   r-        z)DiffLlamaForCausalLM.set_input_embeddingsc                 C   r&  r5   rA  r   r3   r3   r4   get_output_embeddings  r)  z*DiffLlamaForCausalLM.get_output_embeddingsc                 C   r*  r5   rI  )r0   new_embeddingsr3   r3   r4   set_output_embeddings  r.  z*DiffLlamaForCausalLM.set_output_embeddingsc                 C   r*  r5   r   )r0   decoderr3   r3   r4   set_decoder	  r.  z DiffLlamaForCausalLM.set_decoderc                 C   r&  r5   rM  r   r3   r3   r4   get_decoder  r)  z DiffLlamaForCausalLM.get_decoderNr   r/  r~   rK   r   r0  labelsr   r   r1  r   logits_to_keepr   rR   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 )a1  
        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, DiffLlamaForCausalLM

        >>> model = DiffLlamaForCausalLM.from_pretrained("google/diffllama-7b")
        >>> tokenizer = AutoTokenizer.from_pretrained("google/diffllama-7b")

        >>> prompt = "What is your favorite condiment?"
        >>> 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]
        "What is your favorite condiment?"
        ```N)	r/  r~   rK   r   r0  r   r   r1  r   )rC  rQ  r  lossrC  r   rP   r6  r3   )r&   r   r1  r   r5  r   r   slicerA  loss_functionr  r   r   rP   r6  )r0   r/  r~   rK   r   r0  rQ  r   r   r1  r   rR  r   r   rP   slice_indicesrC  rT  r3   r3   r4   r7     s:   '
zDiffLlamaForCausalLM.forward)NNNNNNNNNNr   )r8   r9   r:   _tied_weights_keys_tp_plan_pp_planr%   r(  r-  rJ  rL  rO  rP  r   r   r   rA   r   r   r   r   r   r   r   r   r?  r   r7   r;   r3   r3   r1   r4   r@    sf    		
r@  a  
    The DiffLlama Model transformer with a sequence classification head on top (linear layer).

    [`DiffLlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models
    (e.g. GPT-2) do.

    Since it does classification on the last token, it requires to know the position of the last token. If a
    `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If
    no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the
    padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in
    each row of the batch).
    )custom_introc                          e Zd Z 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	j
 dee dee dee defddZ  ZS )"DiffLlamaForSequenceClassificationc                    s@   t  | |j| _t|| _tj|j| jdd| _| 	  d S r!   )
r$   r%   
num_labelsr  r   r   r)   r'   scorer%  r/   r1   r3   r4   r%   j  s
   
z+DiffLlamaForSequenceClassification.__init__c                 C   rD  r5   rE  r   r3   r3   r4   r(  s  rF  z7DiffLlamaForSequenceClassification.get_input_embeddingsc                 C   rG  r5   rE  r+  r3   r3   r4   r-  v  rH  z7DiffLlamaForSequenceClassification.set_input_embeddingsNr/  r~   rK   r   r0  rQ  r   r   r1  rR   c
              
   C   s(  | j ||||||||	d}
|
j}| |}|dur|jd }n|jd }| jjdu r2|dkr2td| jjdu r;d}n1|dur`|| jjk|jt	j
}t	j|jd |jt	j
d}|| d}nd}t| jj d |t	j||jd	|f }d}|dur| j|||| jd
}t|||
j|
j|
jdS )  
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        r~   rK   r   r0  r   r   r1  Nr   r   z=Cannot handle batch sizes > 1 if no padding token is defined.r<   )r   r   z will not detect padding tokens in `inputs_embeds`. Results may be unexpected if using padding tokens in conjunction with `inputs_embeds.`r3  )rC  rQ  pooled_logitsr&   rS  )r   r5  r_  r@   r&   r  r   r   r   rA   int32r8  argmaxrb   rc   r2   r8   rV  r   r   rP   r6  )r0   r/  r~   rK   r   r0  rQ  r   r   r1  transformer_outputsrP   rC  
batch_sizelast_non_pad_tokennon_pad_masktoken_indicesrb  rT  r3   r3   r4   r7   y  sL   


z*DiffLlamaForSequenceClassification.forwardr>  )r8   r9   r:   r%   r(  r-  r   r   r   rA   r   r   r   r   r   r   r7   r;   r3   r3   r1   r4   r]  [  sH    		
r]  c                       s   e Zd ZdZ 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
j de	e
j de	e de	e defddZ  ZS )DiffLlamaForQuestionAnsweringtransformerc                    s2   t  | t|| _t|jd| _|   d S )Nr=   )	r$   r%   r  rk  r   r)   r'   
qa_outputsr%  r/   r1   r3   r4   r%     s   
z&DiffLlamaForQuestionAnswering.__init__c                 C   rD  r5   rk  r  r   r3   r3   r4   r(    rF  z2DiffLlamaForQuestionAnswering.get_input_embeddingsc                 C   rG  r5   rm  r+  r3   r3   r4   r-    rH  z2DiffLlamaForQuestionAnswering.set_input_embeddingsNr/  r~   rK   r   r0  start_positionsend_positionsr   r1  rR   c
              	   K   s   | j |||||||	d}|j}| |}|jddd\}}|d }|d }d }|d urA|d urA| j||||fi |
}t||||j|j	dS )N)r~   rK   r   r0  r   r1  r   r<   r>   )rT  start_logits
end_logitsrP   r6  )
rk  r5  rl  splitsqueezer   rV  r   rP   r6  )r0   r/  r~   rK   r   r0  rn  ro  r   r1  r   r   sequence_outputrC  rp  rq  rT  r3   r3   r4   r7     s0   

z%DiffLlamaForQuestionAnswering.forwardr>  )r8   r9   r:   r   r%   r(  r-  r   r   r   rA   r   r   r   r   r   r   r7   r;   r3   r3   r1   r4   rj    sJ    	
rj  c                       r\  )DiffLlamaForTokenClassificationc                    s|   t  | |j| _t|| _t|dd d ur|j}nt|dd d ur'|j}nd}t	|| _
t|j|j| _|   d S )Nclassifier_dropouthidden_dropoutg?)r$   r%   r^  r  r   rg   rv  rw  r   Dropoutr   r)   r'   r_  r%  )r0   r&   rv  r1   r3   r4   r%      s   
z(DiffLlamaForTokenClassification.__init__c                 C   rD  r5   rE  r   r3   r3   r4   r(    rF  z4DiffLlamaForTokenClassification.get_input_embeddingsc                 C   rG  r5   rE  r+  r3   r3   r4   r-    rH  z4DiffLlamaForTokenClassification.set_input_embeddingsNr/  r~   rK   r   r0  rQ  r   r   r1  rR   c
              
   C   sd   | j ||||||||	d}
|
j}| |}| |}d}|dur(| ||| j}t|||
j|
jdS )r`  ra  N)rT  rC  rP   r6  )	r   r5  r   r_  rV  r&   r   rP   r6  )r0   r/  r~   rK   r   r0  rQ  r   r   r1  r   rt  rC  rT  r3   r3   r4   r7     s,   


z'DiffLlamaForTokenClassification.forwardr>  )r8   r9   r:   r%   r(  r-  r   r   r   rA   r   r   r   r   r   r   r7   r;   r3   r3   r1   r4   ru    sH    	
ru  )r   r  r@  r]  rj  ru  )Nr   )FrZ   typingr   r   rA   r   activationsr   cache_utilsr   r   r	   
generationr
   integrationsr   masking_utilsr   modeling_flash_attention_utilsr   r   r   modeling_layersr   modeling_outputsr   r   r   r   r   modeling_rope_utilsr   r   modeling_utilsr   processing_utilsr   utilsr   r   r   r   configuration_diffllamar   
get_loggerr8   rb   Moduler    rE   rO   r   r   rY   r]   r^   r   r   r   r   r   r   r  r  r?  r@  r]  rj  ru  __all__r3   r3   r3   r4   <module>   sf   

j V5""}lV>F