o
    ei^                     @   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$m%Z% ddl&m'Z'm(Z(m)Z) ddl*m+Z+m,Z, ddl-m.Z. ddl/m0Z0 edG dd dej1Z2G dd dej1Z3G dd dej1Z4dd Z5eddCd d!Z6d"ej7d#e8d$ej7fd%d&Z9	'dDd(ej1d)ej7d*ej7d+ej7d,ej7dB d-e:d.e:d/e%e' fd0d1Z;ee6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   )Qwen3Config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 )Qwen3RMSNormư>epsreturnNc                    s&   t    tt|| _|| _dS )z;
        Qwen3RMSNorm is equivalent to T5LayerNorm
        N)super__init__r   	Parametertorchonesweightvariance_epsilon)selfhidden_sizer&   	__class__ f/home/ubuntu/transcripts/venv/lib/python3.10/site-packages/transformers/models/qwen3/modeling_qwen3.pyr)   3   s   

zQwen3RMSNorm.__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/   r5   input_dtypevariancer3   r3   r4   forward;   s
   zQwen3RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)tupler-   shaper.   )r/   r3   r3   r4   
extra_reprB   s   zQwen3RMSNorm.extra_repr)r%   )
__name__
__module____qualname__floatr)   r+   TensorrA   rD   __classcell__r3   r3   r1   r4   r$   1   s    r$   c                       s$   e Zd Z fddZdd Z  ZS )Qwen3MLPc                    sr   t    || _|j| _|j| _tj| j| jdd| _tj| j| jdd| _tj| j| jdd| _	t
|j | _d S NFbias)r(   r)   configr0   intermediate_sizer   Linear	gate_projup_proj	down_projr   
hidden_actact_fnr/   rO   r1   r3   r4   r)   G   s   
zQwen3MLP.__init__c                 C   s$   |  | | || | }|S N)rT   rV   rR   rS   )r/   xrT   r3   r3   r4   rA   Q   s    zQwen3MLP.forward)rE   rF   rG   r)   rA   rJ   r3   r3   r1   r4   rK   F   s    
rK   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 )Qwen3RotaryEmbeddinginv_freqNrO   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_lenrO   rope_parametersr\   compute_default_rope_parametersr   attention_scalingregister_bufferclone)r/   rO   devicerope_init_fnr[   r1   r3   r4   r)   Y   s   


zQwen3RotaryEmbedding.__init__rh   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   r6   r9   )rh   r9   )	rc   getattrr0   num_attention_headsr+   arangeint64r:   rH   )rO   rh   rj   basedimattention_factorr[   r3   r3   r4   rd   i   s   
&z4Qwen3RotaryEmbedding.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   r7   r!   mpscpuF)device_typeenabledr6   rs   rm   )r[   rH   expandrC   r:   rh   
isinstancetypestrr   	transposer+   catcosre   sinr9   )
r/   rY   position_idsinv_freq_expandedposition_ids_expandedrw   freqsembr   r   r3   r3   r4   rA      s   0&zQwen3RotaryEmbedding.forwardrX   )NNN)rE   rF   rG   r+   rI   __annotations__r"   r)   staticmethodr   intrB   rH   rd   no_gradr   rA   rJ   r3   r3   r1   r4   rZ   V   s&   
 

rZ   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..Nr7   r6   ry   )rC   r+   r   )rY   x1x2r3   r3   r4   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kr   r   unsqueeze_dimq_embedk_embedr3   r3   r4   apply_rotary_pos_emb   s
   

r   r5   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)rC   rz   reshape)r5   r   batchnum_key_value_headsslenrl   r3   r3   r4   	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 )Nr6   r   r7   )rs   r9   )ptrainingr!   )r   num_key_value_groupsr+   matmulr~   r   
functionalsoftmaxr;   r:   r9   r   r   
contiguous)r   r   r   r   r   r   r   r   
key_statesvalue_statesattn_weightsattn_outputr3   r3   r4   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 )Qwen3Attentionz=Multi-headed attention from 'Attention Is All You Need' paperrO   	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
 |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d| _t| j
|jd| _| jdkr|j| _d S d | _d S )Nlayer_typesrl   g      TrM   r&   sliding_attention)r(   r)   hasattrr   
layer_typerO   r   rn   r0   ro   rl   r   r   r   attention_dropout	is_causalr   rQ   attention_biasq_projk_projv_projo_projr$   rms_norm_epsq_normk_normsliding_windowr/   rO   r   r1   r3   r4   r)      s0   
 zQwen3Attention.__init__Nr5   position_embeddingsr   past_key_valuescache_positionr   r'   c                 K   s(  |j d d }g |d| jR }| | ||dd}	| | ||dd}
| ||dd}|\}}t	|	|
||\}	}
|d ur]|||d}|
|
|| j|\}
}t| jjt}|| |	|
||f| jsqdn| j| j| jd|\}}|jg |dR   }| |}||fS )Nr7   r!   r6   )r   r   r   r   )r   r   r   )rC   rl   r   r   viewr~   r   r   r   r   updater   r   get_interfacerO   _attn_implementationr   r   r   r   r   r   r   r   )r/   r5   r   r   r   r   r   input_shapehidden_shapequery_statesr   r   r   r   cache_kwargsattention_interfacer   r   r3   r3   r4   rA      s:   		

zQwen3Attention.forward)NN)rE   rF   rG   __doc__r"   r   r)   r+   rI   rB   r   
LongTensorr   r   rA   rJ   r3   r3   r1   r4   r      s(     r   c                       s   e Zd Zdedef fddZ						ddejdejdB d	ejdB d
e	dB de
dB dejdB deejejf dB dee dejfddZ  ZS )Qwen3DecoderLayerrO   r   c                    s^   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
|j| | _d S )N)rO   r   r   )r(   r)   r0   r   	self_attnrK   mlpr$   r   input_layernormpost_attention_layernormr   attention_typer   r1   r3   r4   r)   *  s   

zQwen3DecoderLayer.__init__NFr5   r   r   r   	use_cacher   r   r   r'   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r5   r   r   r   r   r   r   r3   )r   r   r   r   )r/   r5   r   r   r   r   r   r   r   residual_r3   r3   r4   rA   5  s&   




zQwen3DecoderLayer.forward)NNNFNN)rE   rF   rG   r"   r   r)   r+   rI   r   r   boolrB   r   r   rA   rJ   r3   r3   r1   r4   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 )Qwen3PreTrainedModelrO   modelTr   r   )r5   
attentionsN)rE   rF   rG   r"   r   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_outputsr3   r3   r3   r4   r   W  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 )
Qwen3ModelrO   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 r3   )r   ).0r   rO   r3   r4   
<listcomp>s  s    z'Qwen3Model.__init__.<locals>.<listcomp>r   r   Fr   )r(   r)   pad_token_idpadding_idx
vocab_sizer   	Embeddingr0   embed_tokens
ModuleListrangenum_hidden_layerslayersr$   r   normrZ   
rotary_embgradient_checkpointingrO   r   has_sliding_layers	post_initrW   r1   r   r4   r)   l  s   zQwen3Model.__init__N	input_idsr   r   r   inputs_embedsr   r   r   r'   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!   )rh   )rO   r   r   r   r   r   full_attentionr   )r   r   r   r   r   r   )last_hidden_stater   r3   )
ValueErrorr   r   rO   get_seq_lengthr+   rp   rC   rh   r   r{   dictr   r   r   r   r   r   r   r   r   )r/   r   r   r   r   r   r   r   r   past_seen_tokenscausal_mask_mappingmask_kwargsr5   r   decoder_layerr3   r3   r4   rA   }  s^   



zQwen3Model.forward)NNNNNNN)rE   rF   rG   r"   r)   r   r    r   r+   r   rI   r   FloatTensorr   r   r   r   rA   rJ   r3   r3   r1   r4   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 )Qwen3ForCausalLMzlm_head.weightzmodel.embed_tokens.weightlm_headcolwise_gather_outputr5   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S rL   )
r(   r)   r   r   r   r   rQ   r0   r  r   rW   r1   r3   r4   r)     s
   
zQwen3ForCausalLM.__init__Nr   r   r   r   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 )a^  
        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, Qwen3ForCausalLM

        >>> model = Qwen3ForCausalLM.from_pretrained("Qwen/Qwen3-8B")
        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

        >>> 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   r   r   r   r   r   N)r
  r  r   )lossr
  r   r5   r   r3   )r   r   r{   r   slicer  loss_functionrO   r   r   r   r5   r   )r/   r   r   r   r   r   r  r   r   r  r   outputsr5   slice_indicesr
  r  r3   r3   r4   rA     s0   %zQwen3ForCausalLM.forward)	NNNNNNNNr   )rE   rF   rG   _tied_weights_keys_tp_plan_pp_planr)   r   r   r+   r   rI   r   r  r   r   r   r   r   rA   rJ   r3   r3   r1   r4   r    sN    		
r  c                   @      e Zd ZdS )Qwen3ForSequenceClassificationNrE   rF   rG   r3   r3   r3   r4   r        r  c                   @   r  )Qwen3ForTokenClassificationNr  r3   r3   r3   r4   r    r  r  c                   @   s   e Zd ZdZdS )Qwen3ForQuestionAnsweringtransformerN)rE   rF   rG   r   r3   r3   r3   r4   r    s    r  )r  r  r   r   r  r  )r!   )r   )Ecollections.abcr   typingr   r+   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_qwen3r"   Moduler$   rK   rZ   r   r   rI   r   r   rH   r   r   r   r   r   r  r  r  r  __all__r3   r3   r3   r4   <module>   st   A
K.[P