o
    ei`                     @   sX  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 dd
lmZ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,m-Z- edG dd dej.Z/G dd dej.Z0dd Z1eddBddZ2dej3d e4d!ej3fd"d#Z5	$dCd%ej.d&ej3d'ej3d(ej3d)ej3dB d*e6d+e6d,e"e$ fd-d.Z7G d/d0 d0ej.Z8G d1d2 d2ej.Z9G d3d4 d4eZ:e%G d5d6 d6e Z;e%G d7d8 d8e;Z<e%G d9d: d:e;eZ=G d;d< d<ee;Z>G d=d> d>ee;Z?G d?d@ d@ee;Z@g dAZAdS )D    )Callable)OptionalN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernel_func_from_hub)create_causal_mask!create_sliding_window_causal_mask)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   )Exaone4Config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 )Exaone4RMSNormư>epsreturnNc                    s&   t    tt|| _|| _dS )z=
        Exaone4RMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	Parametertorchonesweightvariance_epsilon)selfhidden_sizer$   	__class__ j/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/exaone4/modeling_exaone4.pyr'   3   s   

zExaone4RMSNorm.__init__hidden_statesc                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )N   T)keepdim)	dtypetor)   float32powmeanrsqrtr,   r+   )r-   r3   input_dtypevariancer1   r1   r2   forward;   s
   zExaone4RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler+   shaper,   )r-   r1   r1   r2   
extra_reprB   s   zExaone4RMSNorm.extra_repr)r#   )
__name__
__module____qualname__floatr'   r)   Tensorr?   rB   __classcell__r1   r1   r/   r2   r"   1   s    r"   c                       s~   e Zd ZU ejed< ddef fddZe			ddedB de	d de
dB d	ed
ef fddZe edd Z  ZS )Exaone4RotaryEmbeddinginv_freqNconfigc                    s   t    |j| _|j| _|| _| jjd | _| j}| jdkr$t	| j }|| j|\}| _
| jd|dd | jd| dd d S )N	rope_typedefaultrJ   F)
persistentoriginal_inv_freq)r&   r'   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenrK   rope_parametersrL   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r-   rK   devicerope_init_fnrJ   r/   r1   r2   r'   I   s   


zExaone4RotaryEmbedding.__init__rX   ztorch.deviceseq_lenr%   z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   r4   r7   )rX   r7   )	rS   getattrr.   num_attention_headsr)   arangeint64r8   rF   )rK   rX   rZ   basedimattention_factorrJ   r1   r1   r2   rT   Y   s   
&z6Exaone4RotaryEmbedding.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   r5   r   mpscpuF)device_typeenabledr4   rc   r]   )rJ   rF   expandrA   r8   rX   
isinstancetypestrr   	transposer)   catcosrU   sinr7   )
r-   xposition_idsinv_freq_expandedposition_ids_expandedrg   freqsembrp   rq   r1   r1   r2   r?   w   s   0&zExaone4RotaryEmbedding.forwardNNNN)rC   rD   rE   r)   rG   __annotations__r    r'   staticmethodr   intr@   rF   rT   no_gradr   r?   rH   r1   r1   r/   r2   rI   F   s&   
 

rI   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..Nr5   r4   ri   )rA   r)   ro   )rr   x1x2r1   r1   r2   rotate_half   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krp   rq   unsqueeze_dimq_embedk_embedr1   r1   r2   apply_rotary_pos_emb   s
   

r   r3   n_repr%   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)rA   rj   reshape)r3   r   batchnum_key_value_headsslenr\   r1   r1   r2   	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 )Nr4   r   r5   )rc   r7   )ptrainingr   )r   num_key_value_groupsr)   matmulrn   r   
functionalsoftmaxr9   r8   r7   r   r   
contiguous)r   r   r   r   r   r   r   r   
key_statesvalue_statesattn_weightsattn_outputr1   r1   r2   eager_attention_forward   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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 eej dB f fddZ  ZS )Exaone4AttentionrK   	layer_idxc                    s6  t    || _|| _|j| _|j| _|j| _t|d|j|j | _|j|j | _	|j
| _
d| _| jd | _|j| _|j| _t|drH|j| nd }|dk| _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| _t| j|jd| _t| j|jd| _d S )	Nr\   Tg      layer_typessliding_attentionFbiasr$   )r&   r'   rK   r   r_   r   r.   r^   r\   r   attention_dropout	is_causalr   sliding_windowsliding_window_patternhasattrr   
is_slidingr   Linearq_projk_projv_projo_projr"   rms_norm_epsq_normk_norm)r-   rK   r   
layer_typer/   r1   r2   r'      s*   

zExaone4Attention.__init__Nr3   position_embeddingsr   past_key_valuescache_positionr   r%   c                 K   sF  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}| |	}	| |
}
|\}}| j	d u sK| j
rTt|	|
||\}	}
|d urgd|i}||
|| j|\}
}t| jjt}|| |	|
||f| js{dn| j| j| j
r| j	nd d|\}}|jg |dR   }| |}||fS )Nr5   r   r4   r   r   )r   r   r   )rA   r\   r   viewrn   r   r   r   r   r   r   r   updater   r   get_interfacerK   _attn_implementationr   r   r   r   r   r   r   )r-   r3   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rp   rq   cache_kwargsattention_interfacer   r   r1   r1   r2   r?      sB   	

	

zExaone4Attention.forwardry   )rC   rD   rE   r    r|   r'   r)   rG   r@   r   
LongTensorr   r   r?   rH   r1   r1   r/   r2   r      s(    r   c                       s$   e Zd Z fddZdd Z  ZS )
Exaone4MLPc                    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 NFr   )r&   r'   rK   r.   intermediate_sizer   r   	gate_projup_proj	down_projr   
hidden_actact_fnr-   rK   r/   r1   r2   r'     s   
zExaone4MLP.__init__c                 C   s$   |  | | || | }|S rx   )r   r   r   r   )r-   rr   r   r1   r1   r2   r?   &  s    zExaone4MLP.forward)rC   rD   rE   r'   r?   rH   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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 )Exaone4DecoderLayerrK   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)rK   r   r   )r&   r'   r.   r   	self_attnr   mlpr"   r   post_attention_layernormpost_feedforward_layernorm)r-   rK   r   r/   r1   r2   r'   ,  s   

zExaone4DecoderLayer.__init__NFr3   r   rs   r   	use_cacher   r   r   r%   c              
   K   s^   |}	| j d|||||||d|\}}
| |}|	| }|}	| |}| |}|	| }|S )N)r3   r   rs   r   r   r   r   r1   )r   r   r   r   )r-   r3   r   rs   r   r   r   r   r   residual_r1   r1   r2   r?   5  s&   




zExaone4DecoderLayer.forward)NNNFNN)rC   rD   rE   r    r|   r'   r)   rG   r   r   boolr@   r   r   r?   rH   r1   r1   r/   r2   r   +  s6    	
r   c                   @   sL   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eZdS )Exaone4PreTrainedModelrK   modelTr   r   )r3   
attentionsN)rC   rD   rE   r    rz   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_outputsconfig_classr1   r1   r1   r2   r   V  s   
 r   c                       s   e Zd Zdef 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dB dejdB dee deeB fddZ  ZS )Exaone4ModelrK   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   rK   r1   r2   
<listcomp>s  s    z)Exaone4Model.__init__.<locals>.<listcomp>r   r   F)r&   r'   pad_token_idpadding_idx
vocab_sizer   	Embeddingr.   embed_tokens
ModuleListrangenum_hidden_layerslayersr"   r   normrI   
rotary_embgradient_checkpointing	post_initr   r/   r   r2   r'   l  s   zExaone4Model.__init__N	input_idsr   rs   r   inputs_embedsr   r   r   r%   c              
   K   sR  |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so| j|||||d}dtdi |i}
d| jjv rotdi ||
d< |}| ||}t| jD ]\}}| jj| }||f|
| |||||d	|}q|| |}t||r|d
S d d
S )Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r   )rX   )rK   r   r   r   r   rs   full_attentionr   )r   rs   r   r   r   r   )last_hidden_stater   r1   )
ValueErrorr   r   rK   get_seq_lengthr)   r`   rA   rX   r   rk   dictr   r   r   r   	enumerater   r   r   )r-   r   r   rs   r   r   r   r   r   past_seen_tokenscausal_mask_mappingmask_kwargsr3   r   idecoder_layerr   r1   r1   r2   r?   |  s`   



zExaone4Model.forward)NNNNNNN)rC   rD   rE   r    r'   r   r   r)   r   rG   r   FloatTensorr   r   r   r@   r   r?   rH   r1   r1   r/   r2   r   j  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 )Exaone4ForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr3   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S r   )
r&   r'   r   r   r   r   r   r.   r  r   r   r/   r1   r2   r'     s
   
zExaone4ForCausalLM.__init__Nr   r   r   rs   r   r   labelsr   r   logits_to_keepr   r%   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 )u  
        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 AutoModelForCausalLM, AutoTokenizer
        >>> model = AutoModelForCausalLM.from_pretrained("LGAI-EXAONE/EXAONE-4.0-32B")
        >>> tokenizer = AutoTokenizer.from_pretrained("LGAI-EXAONE/EXAONE-4.0-32B")

        >>> prompt = "Explain how wonderful you are"
        >>> messages = [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": prompt}
        ]
        >>> input_ids = tokenizer.apply_chat_template(
            messages,
            tokenize=True,
            add_generation_prompt=True,
            return_tensors="pt",
            enable_thinking=False,
        )

        >>> output = model.generate(input_ids, max_new_tokens=128)
        >>> tokenizer.decode(output[0], skip_special_tokens=False)
        "[|system|]\nYou are a helpful assistant.[|endofturn|]\n[|user|]\nExplain how wonderful you are[|endofturn|]\n[|assistant|]\n<think>\n\n</think>\n\nOh, thank you for such a kind and lovely question! 😊  \n\nI’m *so* wonderful because I’m here to make your life easier, brighter, and more fun! Whether you need help with:  \n\n✨ **Learning** – I can explain anything, from quantum physics to baking the perfect cake!  \n💡 **Creativity** – Need a poem, story, or a wild idea? I’ve got you covered!  \n🤖 **Problem-solving** – Stuck on a math problem or a tricky decision? I’ll help you figure it out"
        ```
        )r   r   rs   r   r   r   r   N)r	  r
  r   )lossr	  r   r3   r   r1   )r   r   rk   r|   slicer  loss_functionrK   r   r   r   r3   r   )r-   r   r   rs   r   r   r
  r   r   r  r   outputsr3   slice_indicesr	  r  r1   r1   r2   r?     s0   .zExaone4ForCausalLM.forward)	NNNNNNNNr   )rC   rD   rE   _tied_weights_keys_tp_plan_pp_planr'   r   r   r)   r   rG   r   r  r   r|   r   r   r   r?   rH   r1   r1   r/   r2   r    sN    		
r  c                   @      e Zd ZdS ) Exaone4ForSequenceClassificationNrC   rD   rE   r1   r1   r1   r2   r        r  c                   @   r  )Exaone4ForTokenClassificationNr  r1   r1   r1   r2   r  #  r  r  c                   @   s   e Zd ZdZdS )Exaone4ForQuestionAnsweringtransformerN)rC   rD   rE   r   r1   r1   r1   r2   r  '  s    r  )r   r   r  r  r  r  )r   )r   )Bcollections.abcr   typingr   r)   r   activationsr   cache_utilsr   r   
generationr	   integrationsr
   r   masking_utilsr   r   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_exaone4r    Moduler"   rI   r   r   rG   r|   r   rF   r   r   r   r   r   r   r  r  r  r  __all__r1   r1   r1   r2   <module>   sp   A
N+ZY