o
    ei2X                     @   s>  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 dd
l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, G dd dej-Z.G dd dej-Z/dej0de1dej0fddZ2	d>dej-dej0d ej0d!ej0d"ej0dB d#e3d$e3d%e!e# fd&d'Z4d(d) Z5d?d*d+Z6ee6G d,d- d-ej-Z7ed.G d/d0 d0ej-Z8G d1d2 d2eZ9e$G d3d4 d4eZ:e$G d5d6 d6e:Z;e$G d7d8 d8e:eZ<G d9d: d:ee:Z=G d;d< d<ee:Z>g d=Z?dS )@    )Callable)OptionalN   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hubuse_kernelized_func)create_causal_mask) 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   )	GlmConfigc                       s2   e Zd Z fddZdejdejfddZ  ZS )GlmMLPc                    sP   t    || _tj|jd|j dd| _tj|j|jdd| _t	|j
 | _d S )N   Fbias)super__init__confignnLinearhidden_sizeintermediate_sizegate_up_proj	down_projr   
hidden_actactivation_fnselfr$   	__class__ b/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/glm/modeling_glm.pyr#   0   s
   
zGlmMLP.__init__hidden_statesreturnc                 C   s4   |  |}|jddd\}}|| | }| |S )Nr   dim)r)   chunkr,   r*   )r.   r3   	up_statesgater1   r1   r2   forward8   s   

zGlmMLP.forward)__name__
__module____qualname__r#   torchFloatTensorr;   __classcell__r1   r1   r/   r2   r   /   s    r   c                       s~   e Zd ZU ejed< ddef fddZe			ddedB de	d de
dB d	ed
ef fddZe edd Z  ZS )GlmRotaryEmbeddinginv_freqNr$   c                    s   t    |j| _|j| _|| _| jjd | _| j}| jdkr$t	| j }|| j|\}| _
| jd|dd | jd| dd d S )N	rope_typedefaultrC   F)
persistentoriginal_inv_freq)r"   r#   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr$   rope_parametersrD   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r.   r$   devicerope_init_fnrC   r/   r1   r2   r#   D   s   


zGlmRotaryEmbedding.__init__rP   ztorch.deviceseq_lenr4   ztorch.Tensorc           	      C   st   | j d }| j dd}t| ddp| j| j }t|| }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partial_rotary_factorg      ?head_dimNr   r   dtype)rP   rW   )rK   getgetattrr'   num_attention_headsintr?   arangeint64tofloat)	r$   rP   rR   baserT   rU   r7   attention_factorrC   r1   r1   r2   rL   T   s   
&z2GlmRotaryEmbedding.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enabledr   r6   rV   )rC   r_   expandshaper^   rP   
isinstancetypestrr   	transposer?   catcosrM   sinrW   )
r.   xposition_idsinv_freq_expandedposition_ids_expandedrd   freqsembrm   rn   r1   r1   r2   r;   t   s   0&zGlmRotaryEmbedding.forwardN)NNN)r<   r=   r>   r?   Tensor__annotations__r   r#   staticmethodr   r[   tupler_   rL   no_gradr   r;   rA   r1   r1   r/   r2   rB   A   s&   
 

rB   r3   n_repr4   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)r3   r{   batchnum_key_value_headsslenrU   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 )Nr   r   r5   )r7   rW   )ptrainingr   )r   num_key_value_groupsr?   matmulrk   r%   
functionalsoftmaxfloat32r^   rW   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                 C   s>   | ddddf }| ddddf }t j| |fdddS )	z*Rotates half the hidden dims of the input..r   Nr   r   r5   r6   )r?   stackflatten)ro   x1x2r1   r1   r2   rotate_half   s   r   c                 C   s   | |}| |}|dd|jd d f jddd}|dd|jd d f jddd}|jd }| dd|f | d|df }}|dd|f |d|df }}	|| t||  }
|| t||  }tj|
|gdd}
tj||	gdd}|
|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.
    .Nr5   r   r6   )	unsqueezerg   repeat_interleaver   r?   rl   )qkrm   rn   unsqueeze_dim
rotary_dimq_rotq_passk_rotk_passq_embedk_embedr1   r1   r2   apply_rotary_pos_emb   s   

$$
""r   c                       s   e Zd ZdZddededB f fddZ				ddejde	ejejf dB d	ejdB d
e
dB dejdB dee de	ejejf fddZ  ZS )GlmAttentionz=Multi-headed attention from 'Attention Is All You Need' paperNr$   	layer_idxc                    s   t    || _|| _t|d|j|j | _|j|j | _	| jd | _
|j| _d| _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j|j| j |jd| _tj|j| j |jdd| _d S )NrU   g      Tr    F)r"   r#   r$   r   rY   r'   rZ   rU   r~   r   r   attention_dropout	is_causalr%   r&   attention_biasq_projk_projv_projo_projr.   r$   r   r/   r1   r2   r#      s$   
 zGlmAttention.__init__r3   position_embeddingsr   past_key_valuescache_positionr   r4   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d|\}}|jg |dR   }| |}||fS )Nr5   r   r   )rn   rm   r   r   )r   r   )rg   rU   r   viewrk   r   r   r   updater   r   get_interfacer$   _attn_implementationr   r   r   r   r|   r   r   )r.   r3   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   rm   rn   cache_kwargsattention_interfacer   r   r1   r1   r2   r;      s8   	

zGlmAttention.forwardru   )NNNN)r<   r=   r>   __doc__r   r[   r#   r?   rv   ry   r   
LongTensorr   r   r;   rA   r1   r1   r/   r2   r      s,    r   RMSNormc                       sF   e Zd Zddeddf fddZdejdejfdd	Zd
d Z  Z	S )
GlmRMSNormư>epsr4   Nc                    s&   t    tt|| _|| _dS )z9
        GlmRMSNorm is equivalent to T5LayerNorm
        N)r"   r#   r%   	Parameterr?   onesweightvariance_epsilon)r.   r'   r   r/   r1   r2   r#     s   

zGlmRMSNorm.__init__r3   c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )Nr   r5   T)keepdim)	rW   r^   r?   r   powmeanrsqrtr   r   )r.   r3   input_dtypevariancer1   r1   r2   r;   '  s
   zGlmRMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)ry   r   rg   r   )r.   r1   r1   r2   
extra_repr.  s   zGlmRMSNorm.extra_repr)r   )
r<   r=   r>   r_   r#   r?   rv   r;   r   rA   r1   r1   r/   r2   r     s    r   c                       s   e Zd Zdedef fddZ						ddejdejdB d	ejdB d
e	dB de
dB dejdB deejejf dB dee dejfddZ  ZS )GlmDecoderLayerr$   r   c                    sR   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
d S )N)r$   r   r   )r"   r#   r'   r   	self_attnr   mlpr   rms_norm_epsinput_layernormpost_attention_layernormr   r/   r1   r2   r#   3  s   

zGlmDecoderLayer.__init__NFr3   r   rp   r   	use_cacher   r   r   r4   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r3   r   rp   r   r   r   r   r1   )r   r   r   r   )r.   r3   r   rp   r   r   r   r   r   residual_r1   r1   r2   r;   =  s&   




zGlmDecoderLayer.forward)NNNFNN)r<   r=   r>   r   r[   r#   r?   rv   r   r   boolry   r   r   r;   rA   r1   r1   r/   r2   r   2  s6    	
r   c                   @   sH   e Zd ZU eed< dZdZdgZdgZdZ	dZ
dZdZdZeedZdS )GlmPreTrainedModelr$   modelTr   r   )r3   
attentionsN)r<   r=   r>   r   rw   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_outputsr1   r1   r1   r2   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j	dB dedB dee defddZ  ZS )GlmModelr$   c                    s   t     j| _ j| _t j j| j| _t	 fddt
 jD | _t j jd| _t d| _d| _|   d S )Nc                    s   g | ]}t  |qS r1   )r   ).0r   r$   r1   r2   
<listcomp>{  s    z%GlmModel.__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   normrB   
rotary_embgradient_checkpointing	post_initr-   r/   r   r2   r#   t  s   zGlmModel.__init__N	input_idsr   rp   r   inputs_embedsr   r   r   r4   c              
   K   s   |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 rE|	d}t
| j|||||d}
|}| j||d}| jd | jj D ]}||f|
|||||d|}qb| |}t||d	S )
Nz:You must specify exactly one of input_ids or inputs_embedsr   r   r   )rP   )r$   r   r   r   r   rp   )rp   )r   r   rp   r   r   r   )last_hidden_stater   )
ValueErrorr   r   r$   get_seq_lengthr?   r\   rg   rP   r   r   r   r   r   r   r   )r.   r   r   rp   r   r   r   r   r   past_seen_tokenscausal_maskr3   r   decoder_layerr1   r1   r2   r;     sP   

	
zGlmModel.forward)NNNNNNN)r<   r=   r>   r   r#   r   r   r   r?   r   rv   r   r@   r   r   r   r   r;   rA   r1   r1   r/   r2   r   r  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 )GlmForCausalLMz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 )NFr    )
r"   r#   r   r   r   r%   r&   r'   r  r   r-   r/   r1   r2   r#     s
   
zGlmForCausalLM.__init__Nr   r   r   rp   r   r   labelsr   r   logits_to_keepr   r4   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, GlmForCausalLM

        >>> model = GlmForCausalLM.from_pretrained("meta-glm/Glm-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-glm/Glm-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   rp   r   r   r   r   N)r  r  r   )lossr  r   r3   r   r1   )r   r   rh   r[   slicer  loss_functionr$   r   r   r   r3   r   )r.   r   r   rp   r   r   r  r   r   r  r   outputsr3   slice_indicesr  r  r1   r1   r2   r;     s0    zGlmForCausalLM.forward)	NNNNNNNNr   )r<   r=   r>   _tied_weights_keys_tp_plan_pp_planr#   r   r   r?   r   rv   r   r@   r   r[   r   r   r   r;   rA   r1   r1   r/   r2   r    sN    		
r  c                   @      e Zd ZdS )GlmForSequenceClassificationNr<   r=   r>   r1   r1   r1   r2   r        r  c                   @   r  )GlmForTokenClassificationNr  r1   r1   r1   r2   r    r  r  )r   r   r  r  r  )r   )r   )@collections.abcr   typingr   r?   torch.nnr%   activationsr   cache_utilsr   r   
generationr   integrationsr	   r
   masking_utilsr   modeling_layersr   r   r   modeling_outputsr   r   modeling_rope_utilsr   r   modeling_utilsr   r   processing_utilsr   utilsr   r   r   utils.genericr   r   utils.output_capturingr   configuration_glmr   Moduler   rB   rv   r[   r   r_   r   r   r   r   r   r   r   r   r  r  r  __all__r1   r1   r1   r2   <module>   sn   C

(D-PK