o
    Zh$                     @   s  d dl mZmZmZmZ d dlZd dlZd dlmZ ddlm	Z	 ddl
mZmZ ddlmZ ddlmZ dd	lmZ dd
lmZ ddlmZmZmZmZmZ ddlmZmZ ddlmZm Z  ddl!m"Z" ddl#m$Z$ ddl%m&Z&m'Z'm(Z(m)Z)m*Z* ddl+m,Z, e) rd dl-m.Z. ddl/m0Z0 ddl1m2Z2 e*3e4Z5e2dG dd dej6Z7e$8e7 G dd dej6Z9dd Z:dEddZ;G dd  d ej6Z<d!ej=d"e>d#ej=fd$d%Z?	&dFd'ej6d(ej=d)ej=d*ej=d+eej= d,e@d-e@fd.d/ZAG d0d1 d1ej6ZBG d2d3 d3eZCe'G d4d5 d5e ZDe'G d6d7 d7eDZEG d8d9 d9ee&ZFe'G d:d; d;eDeZGe'd<d=G d>d? d?eDZHe'G d@dA dAeDZIe'G dBdC dCeDZJg dDZKdS )G    )CallableOptionalTupleUnionN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)AttentionMaskConverter)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPastQuestionAnsweringModelOutput SequenceClassifierOutputWithPastTokenClassifierOutput)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)ALL_LAYERNORM_LAYERS)
LossKwargsauto_docstringcan_return_tupleis_torch_flex_attn_availablelogging   )LlamaConfig)	BlockMask)make_flex_block_causal_mask)use_kernel_forward_from_hubZRMSNormc                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	LlamaRMSNormư>c                    s&   t    tt|| _|| _dS )z;
        LlamaRMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	ParametertorchZonesweightvariance_epsilon)selfhidden_sizeeps	__class__ W/var/www/auris/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.pyr'   <   s   

zLlamaRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )N   T)Zkeepdim)	dtypetor)   float32powmeanZrsqrtr+   r*   )r,   hidden_statesZinput_dtypeZvariancer1   r1   r2   forwardD   s
   zLlamaRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler*   shaper+   r,   r1   r1   r2   
extra_reprK   s   zLlamaRMSNorm.extra_repr)r%   )__name__
__module____qualname__r'   r;   r?   __classcell__r1   r1   r/   r2   r$   :   s    r$   c                       s8   e Zd Zddef fddZe edd Z  Z	S )LlamaRotaryEmbeddingNconfigc                    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)r&   r'   hasattrrF   getrG   Zmax_position_embeddingsZmax_seq_len_cachedZoriginal_max_seq_lenrE   r   Zrope_init_fnattention_scalingZregister_bufferrJ   Zoriginal_inv_freq)r,   rE   devicerJ   r/   r1   r2   r'   S   s   
zLlamaRotaryEmbedding.__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   r4   r   ZmpscpuF)device_typeenabledr3   dim)r5   )rJ   floatexpandr=   r6   rO   
isinstancerH   strr)   Zautocast	transposecatcosrN   sinr5   )
r,   xposition_idsZinv_freq_expandedZposition_ids_expandedrQ   ZfreqsZembr[   r\   r1   r1   r2   r;   d   s   0&zLlamaRotaryEmbedding.forwardN)
r@   rA   rB   r    r'   r)   Zno_gradr   r;   rC   r1   r1   r/   r2   rD   R   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..Nr4   r3   rS   )r=   r)   rZ   )r]   x1Zx2r1   r1   r2   rotate_halft   s   ra   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.
    )	unsqueezera   )qkr[   r\   r^   Zunsqueeze_dimZq_embedZk_embedr1   r1   r2   apply_rotary_pos_emb{   s
   

re   c                       s$   e Zd Z fddZdd Z  ZS )LlamaMLPc                    sx   t    || _|j| _|j| _tj| j| j|jd| _tj| j| j|jd| _	tj| j| j|jd| _
t|j | _d S )Nbias)r&   r'   rE   r-   Zintermediate_sizer   LinearZmlp_bias	gate_projup_proj	down_projr   Z
hidden_actact_fnr,   rE   r/   r1   r2   r'      s   
zLlamaMLP.__init__c                 C   s$   |  | | || | }|S r_   )rl   rm   rj   rk   )r,   r]   rl   r1   r1   r2   r;      s    zLlamaMLP.forward)r@   rA   rB   r'   r;   rC   r1   r1   r/   r2   rf      s    
rf   r:   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=   rV   reshape)r:   ro   batchnum_key_value_headsslenhead_dimr1   r1   r2   	repeat_kv   s
   0rv           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 )Nr3   r   r4   )rT   r5   )ptrainingr   )rv   num_key_value_groupsr)   matmulrY   r=   r   Z
functionalZsoftmaxr7   r6   r5   r~   r   
contiguous)rx   ry   rz   r{   r|   r}   r~   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr1   r1   r2   eager_attention_forward   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 )LlamaAttentionz=Multi-headed attention from 'Attention Is All You Need' paperrE   	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 )Nru   g      Trg   )r&   r'   rE   r   getattrr-   Znum_attention_headsru   rs   r   r}   attention_dropoutZ	is_causalr   ri   Zattention_biasq_projk_projv_projo_projr,   rE   r   r/   r1   r2   r'      s(   
zLlamaAttention.__init__Nr:   position_embeddingsr|   past_key_valuecache_positionr   rp   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 )Nr4   r   r3   )r\   r[   r   eagersdpaoutput_attentionsFz`torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to eager attention. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.rw   )r~   r}   )r=   ru   r   viewrY   r   r   re   updater   r   rE   _attn_implementationrM   loggerwarning_oncer   r   r   r}   rq   r   r   )r,   r:   r   r|   r   r   r   Zinput_shapeZhidden_shapeZquery_statesr   r   r[   r\   Zcache_kwargsZattention_interfacer   r   r1   r1   r2   r;      s@   	

zLlamaAttention.forward)NN)r@   rA   rB   __doc__r    intr'   r)   Tensorr   r   r	   
LongTensorr   r   r;   rC   r1   r1   r/   r2   r      s(    r   c                       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 )LlamaDecoderLayerrE   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rE   r   r.   )r&   r'   r-   r   	self_attnrf   mlpr$   rms_norm_epsinput_layernormpost_attention_layernormr   r/   r1   r2   r'     s   

zLlamaDecoderLayer.__init__NFr:   r|   r^   r   r   	use_cacher   r   r   rp   c	                 K   st   |}
|  |}| jd||||||||d|	\}}|
| }|}
| |}| |}|
| }|f}|r8||f7 }|S )N)r:   r|   r^   r   r   r   r   r   r1   )r   r   r   r   )r,   r:   r|   r^   r   r   r   r   r   r   ZresidualZself_attn_weightsoutputsr1   r1   r2   r;   $  s.   
	



zLlamaDecoderLayer.forward)NNNFFNN)r@   rA   rB   r    r   r'   r)   r   r   r   r	   boolr   r   r   FloatTensorr;   rC   r1   r1   r/   r2   r     s<    	
r   c                   @   sH   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d ZdS )LlamaPreTrainedModelmodelTr   past_key_valuesc                 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 d S )Nrw   )r9   stdg      ?)rE   Zinitializer_rangerW   r   ri   r*   dataZnormal_rh   Zzero_	Embeddingpadding_idxr$   Zfill_)r,   rx   r   r1   r1   r2   _init_weights]  s   


z"LlamaPreTrainedModel._init_weightsN)r@   rA   rB   r    Zconfig_classbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_skip_keys_device_placementZ_supports_flash_attn_2Z_supports_sdpaZ_supports_flex_attnZ_supports_cache_classZ_supports_quantized_cacheZ_supports_static_cacheZ_supports_attention_backendr   r1   r1   r1   r2   r   N  s    r   c                       s  e Zd Zdef fddZdd Zdd Zee									d!d	e	e
j d
e	e
j de	e
j de	e de	e
j de	e de	e de	e de	e
j dee defddZ	d"d
ee
jdf de
jde
jdedef
ddZed
e
jdedede
jde
jdefdd Z  ZS )#
LlamaModelrE   c                    s   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j jd| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r1   )r   ).0r   rE   r1   r2   
<listcomp>t  s    z'LlamaModel.__init__.<locals>.<listcomp>r   r   F)r&   r'   pad_token_idr   
vocab_sizer   r   r-   embed_tokensZ
ModuleListrangenum_hidden_layerslayersr$   r   normrD   
rotary_embgradient_checkpointing	post_initrn   r/   r   r2   r'   m  s   zLlamaModel.__init__c                 C      | j S r_   r   r>   r1   r1   r2   get_input_embeddings}     zLlamaModel.get_input_embeddingsc                 C   
   || _ d S r_   r   r,   r{   r1   r1   r2   set_input_embeddings     
zLlamaModel.set_input_embeddingsN	input_idsr|   r^   r   inputs_embedsr   r   output_hidden_statesr   flash_attn_kwargsrp   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}| |||	||}|}| ||}|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   rO   r1   )r|   r^   r   r   r   r   r   )last_hidden_stater   r:   
attentions)rE   r   r   r   
ValueErrorr   r   r   r   rW   rH   r	   r   r
   get_seq_lengthr)   aranger=   rO   rb   _update_causal_maskr   r   r   r   r   )r,   r   r|   r^   r   r   r   r   r   r   r   past_seen_tokensr   r:   r   Zall_hidden_statesZall_self_attnsZdecoder_layerZlayer_outputsr1   r1   r2   r;     sx   



	


zLlamaModel.forwardFr!   input_tensorc                 C   s:  | j jdkr|d ur|dk r|S d S | j jdkr&t|tjr$t|}|S |d ur.| nd}|d ur7|jnd}| j jdkrO|sO|sOt	j
|||| jdrOd S |j}|jd }	|r^| }
nt|tjri|jd	 n||	 d }
| j||	|
|||jd d
}| j jdkr|d ur|jjdv r|st|j}t	||}|S )NZflash_attention_2rw   Zflex_attentionr   Fr   )r   Zpast_key_values_lengthZis_trainingr   r4   )sequence_lengthtarget_lengthr5   r   
batch_size)cudaZxpuZnpu)rE   r   anyrW   r)   r   r"   r   Zis_compileabler   Z_ignore_causal_mask_sdpar   r5   r=   Zget_max_cache_shape5_prepare_4d_causal_attention_mask_with_cache_positionrO   rH   finfominZ_unmask_unattended)r,   r|   r   r   r   r   r   Zusing_compilable_cacher5   r   r   r   	min_dtyper1   r1   r2   r     sT   




zLlamaModel._update_causal_maskr   r   r5   r   c                 K   sD  | 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d|	f | ddddddf |j }
|
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.
        N   )Z
fill_valuer5   rO   r   )Zdiagonalr   r4   r   )rT   r)   r   r   fullrO   Ztriur   rq   rV   cloner=   r6   Zmasked_fill)r|   r   r   r5   r   r   r   r   r   Zmask_lengthZpadding_maskr1   r1   r2   r   '  s,    $
6  z@LlamaModel._prepare_4d_causal_attention_mask_with_cache_position	NNNNNNNNN)F)r@   rA   rB   r    r'   r   r   r   r   r   r)   r   r   r	   r   r   r   r   r   r;   r   r   staticmethodr   r5   r   rC   r1   r1   r/   r2   r   k  s    	
d
Dr   c                   @   s   e Zd ZdS )KwargsForCausalLMN)r@   rA   rB   r1   r1   r1   r2   r   _  s    r   c                       s
  e Zd ZdgZddiZddgdgfiZ fddZdd	 Zd
d Zdd Z	dd Z
dd Zdd Zee											d%deej deej deej dee deej deej dee dee dee deej d eeejf d!ee d"efd#d$Z  ZS )&LlamaForCausalLMzlm_head.weightlm_headZcolwise_repr:   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S NFrg   )
r&   r'   r   r   r   r   ri   r-   r   r   rn   r/   r1   r2   r'   h  s
   
zLlamaForCausalLM.__init__c                 C      | j jS r_   r   r   r>   r1   r1   r2   r   q     z%LlamaForCausalLM.get_input_embeddingsc                 C      || j _d S r_   r   r   r1   r1   r2   r   t     z%LlamaForCausalLM.set_input_embeddingsc                 C   r   r_   r   r>   r1   r1   r2   get_output_embeddingsw  r   z&LlamaForCausalLM.get_output_embeddingsc                 C   r   r_   r   )r,   Znew_embeddingsr1   r1   r2   set_output_embeddingsz  r   z&LlamaForCausalLM.set_output_embeddingsc                 C   r   r_   r   )r,   decoderr1   r1   r2   set_decoder}  r   zLlamaForCausalLM.set_decoderc                 C   r   r_   r   r>   r1   r1   r2   get_decoder  r   zLlamaForCausalLM.get_decoderNr   r   r|   r^   r   r   labelsr   r   r   r   logits_to_keepr   rp   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 )at  
        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, LlamaForCausalLM

        >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf")

        >>> 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   lossr   r   r:   r   r1   )rE   r   r   r   r   rW   r   slicer   loss_functionr   r   r   r:   r   )r,   r   r|   r^   r   r   r   r   r   r   r   r   r   r   r:   Zslice_indicesr   r   r1   r1   r2   r;     s:   '
zLlamaForCausalLM.forward)NNNNNNNNNNr   )r@   rA   rB   Z_tied_weights_keysZ_tp_planZ_pp_planr'   r   r   r   r   r   r   r   r   r   r)   r   r   r	   r   r   r   r   r   r   r   r;   rC   r1   r1   r/   r2   r   b  sf    		
r   a  
    The LLaMa Model transformer with a sequence classification head on top (linear layer).

    [`LlamaForSequenceClassification`] 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).
    )Z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 )LlamaForSequenceClassificationc                    s@   t  | |j| _t|| _tj|j| jdd| _| 	  d S r   )
r&   r'   
num_labelsr   r   r   ri   r-   scorer   rn   r/   r1   r2   r'     s
   
z'LlamaForSequenceClassification.__init__c                 C   r   r_   r   r>   r1   r1   r2   r     r   z3LlamaForSequenceClassification.get_input_embeddingsc                 C   r   r_   r   r   r1   r1   r2   r     r   z3LlamaForSequenceClassification.set_input_embeddingsNr   r|   r^   r   r   r   r   r   r   rp   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|   r^   r   r   r   r   r   Nr   r   z=Cannot handle batch sizes > 1 if no padding token is defined.r4   )rO   r5   z will not detect padding tokens in `inputs_embeds`. Results may be unexpected if using padding tokens in conjunction with `inputs_embeds.`r   )r   r   pooled_logitsrE   r   )r   r   r  r=   rE   r   r   r6   rO   r)   Zint32r   Zargmaxr   r   r0   r@   r   r   r   r:   r   )r,   r   r|   r^   r   r   r   r   r   r   Ztransformer_outputsr:   r   r   Zlast_non_pad_tokenZnon_pad_maskZtoken_indicesr  r   r1   r1   r2   r;     sL   


z&LlamaForSequenceClassification.forwardr   )r@   rA   rB   r'   r   r   r   r   r   r)   r   r   r	   r   r   r   r;   rC   r1   r1   r/   r2   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 )LlamaForQuestionAnsweringtransformerc                    s2   t  | t|| _t|jd| _|   d S )Nr3   )	r&   r'   r   r  r   ri   r-   
qa_outputsr   rn   r/   r1   r2   r'   8  s   
z"LlamaForQuestionAnswering.__init__c                 C   r   r_   r  r   r>   r1   r1   r2   r   @  r   z.LlamaForQuestionAnswering.get_input_embeddingsc                 C   r   r_   r
  r   r1   r1   r2   r   C  r   z.LlamaForQuestionAnswering.set_input_embeddingsNr   r|   r^   r   r   start_positionsend_positionsr   r   rp   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|   r^   r   r   r   r   r   r4   rS   )r   start_logits
end_logitsr:   r   )
r  r   r	  splitZsqueezer   r   r   r:   r   )r,   r   r|   r^   r   r   r  r  r   r   r   r   sequence_outputr   r  r  r   r1   r1   r2   r;   F  s0   

z!LlamaForQuestionAnswering.forwardr   )r@   rA   rB   r   r'   r   r   r   r   r   r)   r   r   r	   r   r   r   r;   rC   r1   r1   r/   r2   r  3  sJ    	
r  c                       r   )LlamaForTokenClassificationc                    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   r   r  r  r   ZDropoutr~   ri   r-   r  r   )r,   rE   r  r/   r1   r2   r'   u  s   
z$LlamaForTokenClassification.__init__c                 C   r   r_   r   r>   r1   r1   r2   r     r   z0LlamaForTokenClassification.get_input_embeddingsc                 C   r   r_   r   r   r1   r1   r2   r     r   z0LlamaForTokenClassification.set_input_embeddingsNr   r|   r^   r   r   r   r   r   r   rp   c
              
   C   sd   | j ||||||||	d}
|
j}| |}| |}d}|dur(| ||| j}t|||
j|
jdS )r  r  N)r   r   r:   r   )	r   r   r~   r  r   rE   r   r:   r   )r,   r   r|   r^   r   r   r   r   r   r   r   r  r   r   r1   r1   r2   r;     s,   


z#LlamaForTokenClassification.forwardr   )r@   rA   rB   r'   r   r   r   r   r   r)   r   r   r	   r   r   r   r;   rC   r1   r1   r/   r2   r  s  sH    	
r  )r   r   r   r  r  r  )Nr   )rw   )Ltypingr   r   r   r   r)   Ztorch.utils.checkpointr   Zactivationsr   Zcache_utilsr	   r
   Z
generationr   Zmodeling_attn_mask_utilsr   Zmodeling_flash_attention_utilsr   Zmodeling_layersr   Zmodeling_outputsr   r   r   r   r   Zmodeling_rope_utilsr   r   Zmodeling_utilsr   r   Zprocessing_utilsr   Zpytorch_utilsr   utilsr   r   r   r   r   Zconfiguration_llamar    Z!torch.nn.attention.flex_attentionr!   Zintegrations.flex_attentionr"   Zintegrationsr#   Z
get_loggerr@   r   Moduler$   appendrD   ra   re   rf   r   r   rv   rU   r   r   r   r   r   r   r   r  r  r  __all__r1   r1   r1   r2   <module>   s   

"

M5 tlV?F