o
    wir                     @   s`  d dl 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 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$m%Z%m&Z&m'Z'm(Z( ddl)m*Z* e(+e,Z-edG dd dej.Z/G dd dej.Z0dd Z1dBddZ2G dd dej.Z3dej4de5d ej4fd!d"Z6	#dCd$ej.d%ej4d&ej4d'ej4d(eej4 d)e7d*e7fd+d,Z8G d-d. d.ej.Z9G d/d0 d0eZ:e&G d1d2 d2e!Z;e&G d3d4 d4e;Z<G d5d6 d6ee%Z=e&G d7d8 d8e;eZ>e&d9d:G d;d< d<e;Z?e&G d=d> d>e;Z@e&G d?d@ d@e;ZAg dAZBdS )D    )CallableOptionalUnionN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hub)create_causal_mask)FlashAttentionKwargs)GradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPastQuestionAnsweringModelOutput SequenceClassifierOutputWithPastTokenClassifierOutput)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)
LossKwargsauto_docstringcan_return_tuplelogging   )LlamaConfig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torchonesweightvariance_epsilon)selfhidden_sizeeps	__class__ e/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/llama/modeling_llama.pyr#   4   s   

zLlamaRMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )N   T)keepdim)	dtypetor%   float32powmeanrsqrtr(   r'   )r)   hidden_statesinput_dtypevariancer.   r.   r/   forward<   s
   zLlamaRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler'   shaper(   r)   r.   r.   r/   
extra_reprC   s   zLlamaRMSNorm.extra_repr)r!   )__name__
__module____qualname__r#   r<   r@   __classcell__r.   r.   r,   r/   r    2   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#   hasattrrG   getrH   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrF   r   rope_init_fnattention_scalingregister_bufferrK   original_inv_freq)r)   rF   devicerK   r,   r.   r/   r#   H   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   r1   r   mpscpuF)device_typeenabledr0   dim)r3   )rK   floatexpandr>   r4   rV   
isinstancerI   strr%   autocast	transposecatcosrS   sinr3   )
r)   xposition_idsinv_freq_expandedposition_ids_expandedrY   freqsembrd   re   r.   r.   r/   r<   Y   s   0&zLlamaRotaryEmbedding.forwardN)
rA   rB   rC   r   r#   r%   no_gradr   r<   rD   r.   r.   r,   r/   rE   G   s
    rE   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..Nr1   r0   r[   )r>   r%   rc   )rf   x1x2r.   r.   r/   rotate_halfi   s   rp   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.
    )	unsqueezerp   )qkrd   re   rg   unsqueeze_dimq_embedk_embedr.   r.   r/   apply_rotary_pos_embp   s
   

rw   c                       s$   e Zd Z fddZdd Z  ZS )LlamaMLPc                    sx   t    || _|j| _|j| _tj| j| j|jd| _tj| j| j|jd| _	tj| j| j|jd| _
t|j | _d S )Nbias)r"   r#   rF   r*   intermediate_sizer   Linearmlp_bias	gate_projup_proj	down_projr   
hidden_actact_fnr)   rF   r,   r.   r/   r#      s   
zLlamaMLP.__init__c                 C   s$   |  | | || | }|S rl   )r   r   r~   r   )r)   rf   r   r.   r.   r/   r<      s    zLlamaMLP.forward)rA   rB   rC   r#   r<   rD   r.   r.   r,   r/   rx      s    
rx   r9   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)r9   r   batchnum_key_value_headsslenhead_dimr.   r.   r/   	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 )Nr0   r   r1   )r\   r3   )ptrainingr   )r   num_key_value_groupsr%   matmulrb   r>   r   
functionalsoftmaxr5   r4   r3   r   r   
contiguous)r   r   r   r   r   r   r   kwargs
key_statesvalue_statesattn_weightscausal_maskattn_outputr.   r.   r/   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' paperrF   	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      Try   )r"   r#   rF   r   getattrr*   num_attention_headsr   r   r   r   attention_dropout	is_causalr   r|   attention_biasq_projk_projv_projo_projr)   rF   r   r,   r.   r/   r#      s(   
zLlamaAttention.__init__Nr9   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 )Nr1   r   r0   )re   rd   r   eagerr   )r   r   )r>   r   r   viewrb   r   r   rw   updater   r   rF   _attn_implementationr   r   r   r   r   r   r   )r)   r9   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rd   re   cache_kwargsattention_interfacer   r   r.   r.   r/   r<      s8   	

zLlamaAttention.forward)NN)rA   rB   rC   __doc__r   intr#   r%   Tensorr=   r   r   
LongTensorr   r   r<   rD   r.   r.   r,   r/   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 )LlamaDecoderLayerrF   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rF   r   r+   )r"   r#   r*   r   	self_attnrx   mlpr    rms_norm_epsinput_layernormpost_attention_layernormr   r,   r.   r/   r#     s   

zLlamaDecoderLayer.__init__NFr9   r   rg   r   output_attentions	use_cacher   r   r   r   c	                 K   st   |}
|  |}| jd||||||||d|	\}}|
| }|}
| |}| |}|
| }|f}|r8||f7 }|S )N)r9   r   rg   r   r   r   r   r   r.   )r   r   r   r   )r)   r9   r   rg   r   r   r   r   r   r   residualself_attn_weightsoutputsr.   r.   r/   r<     s.   
	



zLlamaDecoderLayer.forward)NNNFFNN)rA   rB   rC   r   r   r#   r%   r   r   r   r   boolr=   r   r   FloatTensorr<   rD   r.   r.   r,   r/   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 )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 )Nr   )r7   stdg      ?)rF   initializer_ranger_   r   r|   r'   datanormal_rz   zero_	Embeddingpadding_idxr    fill_)r)   r   r   r.   r.   r/   _init_weightsL  s   


z"LlamaPreTrainedModel._init_weightsN)rA   rB   rC   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   r.   r.   r.   r/   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  ZS )
LlamaModelrF   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 r.   )r   ).0r   rF   r.   r/   
<listcomp>c  s    z'LlamaModel.__init__.<locals>.<listcomp>r   r   F)r"   r#   pad_token_idr   
vocab_sizer   r   r*   embed_tokens
ModuleListrangenum_hidden_layerslayersr    r   normrE   
rotary_embgradient_checkpointing	post_initr   r,   r   r/   r#   \  s   zLlamaModel.__init__c                 C      | j S rl   r   r?   r.   r.   r/   get_input_embeddingsl     zLlamaModel.get_input_embeddingsc                 C   
   || _ d S rl   r   r)   r   r.   r.   r/   set_input_embeddingso     
zLlamaModel.set_input_embeddingsN	input_idsr   rg   r   inputs_embedsr   r   output_hidden_statesr   flash_attn_kwargsr   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   rV   )rF   input_embedsr   r   r   rg   r.   )r   rg   r   r   r   r   r   )last_hidden_stater   r9   
attentions)rF   r   r   r   
ValueErrorr   r   loggerwarning_oncer_   rI   r   r   r	   get_seq_lengthr%   aranger>   rV   rq   r   r   r   r   r   r   )r)   r   r   rg   r   r   r   r   r   r   r   past_seen_tokensr   r9   r   all_hidden_statesall_self_attnsdecoder_layerlayer_outputsr.   r.   r/   r<   r  s   

	
	


zLlamaModel.forward	NNNNNNNNN)rA   rB   rC   r   r#   r   r   r   r   r   r%   r   r   r   r   r   r   r   r   r<   rD   r.   r.   r,   r/   r   Z  sL    	
r   c                   @   s   e Zd ZdS )KwargsForCausalLMN)rA   rB   rC   r.   r.   r.   r/   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_headcolwise_repr9   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S NFry   )
r"   r#   r   r   r   r   r|   r*   r  r   r   r,   r.   r/   r#     s
   
zLlamaForCausalLM.__init__c                 C      | j jS rl   r   r   r?   r.   r.   r/   r        z%LlamaForCausalLM.get_input_embeddingsc                 C      || j _d S rl   r  r   r.   r.   r/   r        z%LlamaForCausalLM.set_input_embeddingsc                 C   r   rl   r  r?   r.   r.   r/   get_output_embeddings  r   z&LlamaForCausalLM.get_output_embeddingsc                 C   r   rl   r  )r)   new_embeddingsr.   r.   r/   set_output_embeddings  r   z&LlamaForCausalLM.set_output_embeddingsc                 C   r   rl   r   )r)   decoderr.   r.   r/   set_decoder  r   zLlamaForCausalLM.set_decoderc                 C   r   rl   r  r?   r.   r.   r/   get_decoder  r   zLlamaForCausalLM.get_decoderNr   r   r   rg   r   r   labelsr   r   r   r   logits_to_keepr   r   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   rg   r   r   r   r   r   r   )r  r#  r   lossr  r   r9   r  r.   )rF   r   r   r   r  r_   r   slicer  loss_functionr   r   r   r9   r  )r)   r   r   rg   r   r   r#  r   r   r   r   r$  r   r   r9   slice_indicesr  r&  r.   r.   r/   r<     s:   '
zLlamaForCausalLM.forward)NNNNNNNNNNr   )rA   rB   rC   _tied_weights_keys_tp_plan_pp_planr#   r   r   r  r  r!  r"  r   r   r   r%   r   r   r   r   r   r   r   r   r  r   r<   rD   r.   r.   r,   r/   r    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).
    )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   r|   r*   scorer   r   r,   r.   r/   r#   W  s
   
z'LlamaForSequenceClassification.__init__c                 C   r  rl   r  r?   r.   r.   r/   r   `  r  z3LlamaForSequenceClassification.get_input_embeddingsc                 C   r  rl   r  r   r.   r.   r/   r   c  r  z3LlamaForSequenceClassification.set_input_embeddingsNr   r   rg   r   r   r#  r   r   r   r   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   rg   r   r   r   r   r   Nr   r   z=Cannot handle batch sizes > 1 if no padding token is defined.r1   )rV   r3   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_logitsrF   r%  )r   r  r1  r>   rF   r   r  r4   rV   r%   int32r	  argmaxr  r  r-   rA   r(  r   r   r9   r  )r)   r   r   rg   r   r   r#  r   r   r   transformer_outputsr9   r  
batch_sizelast_non_pad_tokennon_pad_masktoken_indicesr4  r&  r.   r.   r/   r<   f  sL   


z&LlamaForSequenceClassification.forwardr  )rA   rB   rC   r#   r   r   r   r   r   r%   r   r   r   r   r   r   r<   rD   r.   r.   r,   r/   r/  H  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 )Nr0   )	r"   r#   r   r=  r   r|   r*   
qa_outputsr   r   r,   r.   r/   r#     s   
z"LlamaForQuestionAnswering.__init__c                 C   r  rl   r=  r   r?   r.   r.   r/   r     r  z.LlamaForQuestionAnswering.get_input_embeddingsc                 C   r  rl   r?  r   r.   r.   r/   r     r  z.LlamaForQuestionAnswering.set_input_embeddingsNr   r   rg   r   r   start_positionsend_positionsr   r   r   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   rg   r   r   r   r   r   r1   r[   )r&  start_logits
end_logitsr9   r  )
r=  r  r>  splitsqueezer   r(  r   r9   r  )r)   r   r   rg   r   r   r@  rA  r   r   r   r   sequence_outputr  rB  rC  r&  r.   r.   r/   r<     s0   

z!LlamaForQuestionAnswering.forwardr  )rA   rB   rC   r   r#   r   r   r   r   r   r%   r   r   r   r   r   r   r<   rD   r.   r.   r,   r/   r<    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#   r0  r   r   r   rH  rI  r   Dropoutr   r|   r*   r1  r   )r)   rF   rH  r,   r.   r/   r#     s   
z$LlamaForTokenClassification.__init__c                 C   r  rl   r  r?   r.   r.   r/   r     r  z0LlamaForTokenClassification.get_input_embeddingsc                 C   r  rl   r  r   r.   r.   r/   r     r  z0LlamaForTokenClassification.set_input_embeddingsNr   r   rg   r   r   r#  r   r   r   r   c
              
   C   sd   | j ||||||||	d}
|
j}| |}| |}d}|dur(| ||| j}t|||
j|
jdS )r2  r3  N)r&  r  r9   r  )	r   r  r   r1  r(  rF   r   r9   r  )r)   r   r   rg   r   r   r#  r   r   r   r   rF  r  r&  r.   r.   r/   r<     s,   


z#LlamaForTokenClassification.forwardr  )rA   rB   rC   r#   r   r   r   r   r   r%   r   r   r   r   r   r   r<   rD   r.   r.   r,   r/   rG    sH    	
rG  )r  r   r   r/  r<  rG  )Nr   )r   )Ctypingr   r   r   r%   torch.utils.checkpointr   activationsr   cache_utilsr   r	   
generationr
   integrationsr   masking_utilsr   modeling_flash_attention_utilsr   modeling_layersr   modeling_outputsr   r   r   r   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   r   configuration_llamar   
get_loggerrA   r  Moduler    rE   rp   rw   rx   r   r   r   r]   r   r   r   r   r   r  r  r/  r<  rG  __all__r.   r.   r.   r/   <module>   sv   
"

F5}lV?F