o
    eiX                     @   sp  d dl mZ d dlm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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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. ddl/m0Z0 G dd dej1Z2G dd dej1Z3dd Z4eddCddZ5dej6d e7d!ej6fd"d#Z8	$dDd%ej1d&ej6d'ej6d(ej6d)ej6dB d*e9d+e9d,e%e' fd-d.Z:ee5G d/d0 d0ej1Z;ed1G d2d3 d3ej1Z<G d4d5 d5eZ=e(G d6d7 d7e#Z>e(G d8d9 d9e>Z?e(G d:d; d;e>eZ@G d<d= d=ee>ZAG d>d? d?ee>ZBG d@dA dAee>ZCg dBZDdS )E    )Callable)OptionalN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hubuse_kernelized_func)create_causal_mask!create_sliding_window_causal_mask)FlashAttentionKwargs)GenericForQuestionAnswering GenericForSequenceClassificationGenericForTokenClassificationGradientCheckpointingLayer)BaseModelOutputWithPastCausalLMOutputWithPast)ROPE_INIT_FUNCTIONSdynamic_rope_update)ALL_ATTENTION_FUNCTIONSPreTrainedModel)Unpack)TransformersKwargsauto_docstringcan_return_tuple)maybe_autocastmerge_with_config_defaults)capture_outputs   )Qwen2Configc                       s$   e Zd Z fddZdd Z  ZS )Qwen2MLPc                    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__ f/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/qwen2/modeling_qwen2.pyr(   $   s   
zQwen2MLP.__init__c                 C   s$   |  | | || | }|S N)r/   r1   r-   r.   )r3   xr/   r6   r6   r7   forward.   s    zQwen2MLP.forward)__name__
__module____qualname__r(   r:   __classcell__r6   r6   r4   r7   r#   #   s    
r#   c                       s~   e Zd ZU ejed< ddef fddZe			ddedB de	d de
dB d	ed
ef fddZe edd Z  ZS )Qwen2RotaryEmbeddinginv_freqNr)   c                    s   t    |j| _|j| _|| _| jjd | _| j}| jdkr$t	| j }|| j|\}| _
| jd|dd | jd| dd d S )N	rope_typedefaultr@   F)
persistentoriginal_inv_freq)r'   r(   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr)   rope_parametersrA   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r3   r)   devicerope_init_fnr@   r4   r6   r7   r(   6   s   


zQwen2RotaryEmbedding.__init__rM   ztorch.deviceseq_lenreturnztorch.Tensorc                 C   sZ   | j d }t| ddp| j| j }d}d|tjd|dtjdj|tjd|   }||fS )	a  
        Computes the inverse frequencies according to the original RoPE implementation
        Args:
            config ([`~transformers.PreTrainedConfig`]):
                The model configuration.
            device (`torch.device`):
                The device to use for initialization of the inverse frequencies.
            seq_len (`int`, *optional*):
                The current sequence length. Unused for this type of RoPE.
        Returns:
            Tuple of (`torch.Tensor`, `float`), containing the inverse frequencies for the RoPE embeddings and the
            post-processing scaling factor applied to the computed cos/sin (unused in this type of RoPE).
        
rope_thetahead_dimNg      ?r      dtype)rM   rU   )	rH   getattrr*   num_attention_headstorcharangeint64tofloat)r)   rM   rO   basedimattention_factorr@   r6   r6   r7   rI   F   s   
&z4Qwen2RotaryEmbedding.compute_default_rope_parametersc           
      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	|dd+ | |  
dd}tj||fdd	}| | j }| | j }	W d    n1 slw   Y  |j|jd
|	j|jd
fS )Nr   r!   mpscpuF)device_typeenabledrS   r^   rT   )r@   r\   expandshaper[   rM   
isinstancetypestrr   	transposerX   catcosrJ   sinrU   )
r3   r9   position_idsinv_freq_expandedposition_ids_expandedrc   freqsembrm   rn   r6   r6   r7   r:   d   s   0&zQwen2RotaryEmbedding.forwardr8   )NNN)r;   r<   r=   rX   Tensor__annotations__r"   r(   staticmethodr   inttupler\   rI   no_gradr   r:   r>   r6   r6   r4   r7   r?   3   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..Nr`   rS   re   )rg   rX   rl   )r9   x1x2r6   r6   r7   rotate_halft   s   r|   rotary_pos_embc                 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.
        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.
    )	unsqueezer|   )qkrm   rn   unsqueeze_dimq_embedk_embedr6   r6   r7   apply_rotary_pos_emb{   s
   

r   hidden_statesn_reprP   c                 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)rg   rf   reshape)r   r   batchnum_key_value_headsslenrR   r6   r6   r7   	repeat_kv   s
   0r           modulequerykeyvalueattention_maskscalingdropoutkwargsc                 K   s   t || j}t || j}	t||dd| }
|d ur |
| }
tjj|
dtjd	|j
}
tjj|
|| jd}
t|
|	}|dd }||
fS )NrS   r   r`   )r^   rU   )ptrainingr!   )r   num_key_value_groupsrX   matmulrk   r   
functionalsoftmaxfloat32r[   rU   r   r   
contiguous)r   r   r   r   r   r   r   r   
key_statesvalue_statesattn_weightsattn_outputr6   r6   r7   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jdB d
e
dB dejdB dee de	ejejdB f fddZ  ZS )Qwen2Attentionz=Multi-headed attention from 'Attention Is All You Need' paperr)   	layer_idxc                    s   t    t|dr|j| nd | _|| _|| _t|d|j|j	 | _
|j	|j | _| j
d | _|j| _d| _tj|j|j	| j
 dd| _tj|j|j| j
 dd| _tj|j|j| j
 dd| _tj|j	| j
 |jdd| _| jdkrz|j| _d S d | _d S )Nlayer_typesrR   g      Tr%   Fsliding_attention)r'   r(   hasattrr   
layer_typer)   r   rV   r*   rW   rR   r   r   r   attention_dropout	is_causalr   r,   q_projk_projv_projo_projsliding_windowr3   r)   r   r4   r6   r7   r(      s   
 zQwen2Attention.__init__Nr   position_embeddingsr   past_key_valuescache_positionr   rP   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t}|| |	|
||f| jskdn| j| j| jd|\}}|jg |dR   }| |}||fS )Nr`   r!   rS   )rn   rm   r   r   )r   r   r   )rg   rR   r   viewrk   r   r   r   updater   r   get_interfacer)   _attn_implementationr   r   r   r   r   r   r   r   )r3   r   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rm   rn   cache_kwargsattention_interfacer   r   r6   r6   r7   r:      s:   		

zQwen2Attention.forward)NN)r;   r<   r=   __doc__r"   rw   r(   rX   rt   rx   r   
LongTensorr   r   r:   r>   r6   r6   r4   r7   r      s(    r   RMSNormc                       sF   e Zd Zddeddf fddZdejdejfdd	Zd
d Z  Z	S )Qwen2RMSNormư>epsrP   Nc                    s&   t    tt|| _|| _dS )z;
        Qwen2RMSNorm is equivalent to T5LayerNorm
        N)r'   r(   r   	ParameterrX   onesweightvariance_epsilon)r3   r*   r   r4   r6   r7   r(      s   

zQwen2RMSNorm.__init__r   c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )NrS   r`   T)keepdim)	rU   r[   rX   r   powmeanrsqrtr   r   )r3   r   input_dtypevariancer6   r6   r7   r:     s
   zQwen2RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)rx   r   rg   r   )r3   r6   r6   r7   
extra_repr  s   zQwen2RMSNorm.extra_repr)r   )
r;   r<   r=   r\   r(   rX   rt   r:   r   r>   r6   r6   r4   r7   r      s    r   c                       s   e Zd Zdedef fddZ						ddejdejdB d	ejdB d
e	dB de
dB dejdB deejejf dB dee dejfddZ  ZS )Qwen2DecoderLayerr)   r   c                    s^   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
|j| | _d S )N)r)   r   r   )r'   r(   r*   r   	self_attnr#   mlpr   rms_norm_epsinput_layernormpost_attention_layernormr   attention_typer   r4   r6   r7   r(     s   

zQwen2DecoderLayer.__init__NFr   r   ro   r   	use_cacher   r   r   rP   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r   r   ro   r   r   r   r   r6   )r   r   r   r   )r3   r   r   ro   r   r   r   r   r   residual_r6   r6   r7   r:     s&   




zQwen2DecoderLayer.forward)NNNFNN)r;   r<   r=   r"   rw   r(   rX   rt   r   r   boolrx   r   r   r:   r>   r6   r6   r4   r7   r     s6    	
r   c                   @   sH   e Zd ZU eed< dZdZdgZdgZdZ	dZ
dZdZdZeedZdS )Qwen2PreTrainedModelr)   modelTr   r   )r   
attentionsN)r;   r<   r=   r"   ru   base_model_prefixsupports_gradient_checkpointing_no_split_modules_skip_keys_device_placement_supports_flash_attn_supports_sdpa_supports_flex_attn_can_compile_fullgraph_supports_attention_backendr   r   _can_record_outputsr6   r6   r6   r7   r   >  s   
 
r   c                       s   e Zd Zdef fddZeee							ddej	dB dej
dB dej	dB dedB d	ejdB d
edB dej	dB dee defddZ  ZS )
Qwen2Modelr)   c                    s   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j jd| _t d| _d| _d| jjv | _|   d S )Nc                    s   g | ]}t  |qS r6   )r   ).0r   r)   r6   r7   
<listcomp>Z  s    z'Qwen2Model.__init__.<locals>.<listcomp>r   r   Fr   )r'   r(   pad_token_idpadding_idx
vocab_sizer   	Embeddingr*   embed_tokens
ModuleListrangenum_hidden_layerslayersr   r   normr?   
rotary_embgradient_checkpointingr)   r   has_sliding_layers	post_initr2   r4   r   r7   r(   S  s   zQwen2Model.__init__N	input_idsr   ro   r   inputs_embedsr   r   r   rP   c              
   K   sF  |d u |d uA rt d|d u r| |}|r!|d u r!t| jd}|d u r=|d ur-| nd}	tj|	|	|jd  |jd}|d u rF|	d}t
| }
tsl| j|||||d}dtdi |i}
| jrltdi ||
d< |}| ||}| jd | jj D ]}||f|
|j |||||d	|}q}| |}t||r|d
S d d
S )Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r!   )rM   )r)   r   r   r   r   ro   full_attentionr   )r   r   ro   r   r   r   )last_hidden_stater   r6   )
ValueErrorr   r   r)   get_seq_lengthrX   rY   rg   rM   r~   rh   dictr   r   r   r   r   r   r   r   r   )r3   r   r   ro   r   r   r   r   r   past_seen_tokenscausal_mask_mappingmask_kwargsr   r   decoder_layerr6   r6   r7   r:   d  s^   



zQwen2Model.forward)NNNNNNN)r;   r<   r=   r"   r(   r   r    r   rX   r   rt   r   FloatTensorr   r   r   r   r:   r>   r6   r6   r4   r7   r   Q  s>    	
r   c                       s   e Zd ZddiZddiZddgdgfiZ fddZee																	
dde	j
d	B de	jd	B de	j
d	B ded	B de	jd	B de	j
d	B ded	B de	j
d	B dee	jB dee defddZ  ZS )Qwen2ForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S r$   )
r'   r(   r   r   r   r   r,   r*   r  r   r2   r4   r6   r7   r(     s
   
zQwen2ForCausalLM.__init__Nr   r   r   ro   r   r   labelsr   r   logits_to_keepr   rP   c
              
   K   s   | j d|||||||d|
}|j}t|	trt|	 dn|	}| |dd|ddf }d}|durB| jd||| jjd|
}t	|||j
|j|jdS )a  
        Example:

        ```python
        >>> from transformers import AutoTokenizer, Qwen2ForCausalLM

        >>> model = Qwen2ForCausalLM.from_pretrained("meta-qwen2/Qwen2-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-qwen2/Qwen2-2-7b-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."
        ```)r   r   ro   r   r   r   r   N)r  r  r   )lossr  r   r   r   r6   )r   r   rh   rw   slicer  loss_functionr)   r   r   r   r   r   )r3   r   r   ro   r   r   r  r   r   r	  r   outputsr   slice_indicesr  r
  r6   r6   r7   r:     s0    zQwen2ForCausalLM.forward)	NNNNNNNNr   )r;   r<   r=   _tied_weights_keys_tp_plan_pp_planr(   r   r   rX   r   rt   r   r  r   rw   r   r   r   r:   r>   r6   r6   r4   r7   r    sN    		
r  c                   @      e Zd ZdS )Qwen2ForSequenceClassificationNr;   r<   r=   r6   r6   r6   r7   r        r  c                   @   r  )Qwen2ForTokenClassificationNr  r6   r6   r6   r7   r    r  r  c                   @   s   e Zd ZdZdS )Qwen2ForQuestionAnsweringtransformerN)r;   r<   r=   r   r6   r6   r6   r7   r    s    r  )r   r   r  r   r  r  r  )r!   )r   )Ecollections.abcr   typingr   rX   r   activationsr   cache_utilsr   r   
generationr	   integrationsr
   r   r   masking_utilsr   r   modeling_flash_attention_utilsr   modeling_layersr   r   r   r   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.genericr   r   utils.output_capturingr    configuration_qwen2r"   Moduler#   r?   r|   r   rt   rw   r   r\   r   r   r   r   r   r   r  r  r  r  __all__r6   r6   r6   r7   <module>   st   A
@.[K