o
    iW                     @   sP  d dl mZmZ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 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$m%Z%m&Z&m'Z' ddl(m)Z) ddl*m+Z+ ddl,m-Z- dd Z.dAddZ/dej0de1dej0fddZ2	dBdej3d ej0d!ej0d"ej0d#eej0 d$e4d%e4d&e#e% fd'd(Z5G d)d* d*ej3Z6ed+G d,d- d-ej3Z7G d.d/ d/ej3Z8G d0d1 d1eZ9e&G d2d3 d3e!Z:G d4d5 d5ej3Z;e&G d6d7 d7e:Z<e&G d8d9 d9e:eZ=G d:d; d;ee:Z>G d<d= d=ee:Z?G d>d? d?ee:Z@g d@ZAdS )C    )CallableOptionalUnionN)nn   )ACT2FN)CacheDynamicCache)GenerationMixin)use_kernel_forward_from_hub)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)deprecate_kwarg)check_model_inputs   )SmolLM3Configc                 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..N   dim)shapetorchcat)xx1x2 r+   `/home/ubuntu/.local/lib/python3.10/site-packages/transformers/models/smollm3/modeling_smollm3.pyrotate_half1   s   r-   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.
    )	unsqueezer-   )qkcossinposition_idsunsqueeze_dimq_embedk_embedr+   r+   r,   apply_rotary_pos_emb8   s
   

r7   hidden_states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%   expandreshape)r8   r9   batchnum_key_value_headsslenhead_dimr+   r+   r,   	repeat_kvS   s
   0rA           modulequerykeyvalueattention_maskscalingdropoutkwargsc                 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 )Nr"   r   r!   )r$   dtype)ptrainingr   )rA   num_key_value_groupsr&   matmul	transposer%   r   
functionalsoftmaxfloat32torL   rI   rN   
contiguous)rC   rD   rE   rF   rG   rH   rI   rJ   
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edddd		
	
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	 f fddZ  ZS )SmolLM3Attentionz=Multi-headed attention from 'Attention Is All You Need' paperconfig	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| _|j| | _|jr||j| dkr||j| _d S d | _d S )Nr@   g      Tbiassliding_attention)super__init__r^   r_   getattrhidden_sizenum_attention_headsr@   r>   rO   rH   attention_dropout	is_causalr   Linearattention_biasq_projk_projv_projo_projno_rope_layersuse_ropeuse_sliding_windowlayer_typessliding_windowselfr^   r_   	__class__r+   r,   rd   |   s8   
zSmolLM3Attention.__init__past_key_valuepast_key_values4.58new_nameversionNr8   position_embeddingsrG   cache_positionrJ   r:   c                 K   s*  |j d d }g |d| jR }| ||dd}	| ||dd}
| ||dd}| jrE|\}}t|	|
||\}	}
|d urXd|i}|	|
|| j
|\}
}t}| jjdkrft| jj }|| |	|
||f| jsrdn| j| j| jd|\}}|jg |dR   }| |}||fS )Nr!   r   r"   r   eagerrB   )rI   rH   rt   )r%   r@   rl   viewrQ   rm   rn   rq   r7   updater_   r\   r^   _attn_implementationr   rN   rh   rH   rt   r<   rV   ro   )rv   r8   r   rG   rz   r   rJ   input_shapehidden_shapequery_statesrW   rX   r1   r2   cache_kwargsattention_interfacer[   rY   r+   r+   r,   forward   s<   
	

zSmolLM3Attention.forward)NN)__name__
__module____qualname____doc__r    intrd   r   r&   Tensortupler   r   
LongTensorr   r   r   __classcell__r+   r+   rw   r,   r]   y   s*    r]   RMSNormc                       s.   e Zd Zd fdd	Zdd Zdd Z  ZS )	SmolLM3RMSNormư>c                    s&   t    tt|| _|| _dS )z=
        SmolLM3RMSNorm is equivalent to T5LayerNorm
        N)rc   rd   r   	Parameterr&   onesweightvariance_epsilon)rv   rf   epsrw   r+   r,   rd      s   

zSmolLM3RMSNorm.__init__c                 C   sJ   |j }|tj}|djddd}|t|| j  }| j|| S )Nr"   r!   T)keepdim)	rL   rU   r&   rT   powmeanrsqrtr   r   )rv   r8   input_dtypevariancer+   r+   r,   r      s
   zSmolLM3RMSNorm.forwardc                 C   s   t | jj d| j S )Nz, eps=)r   r   r%   r   )rv   r+   r+   r,   
extra_repr   s   zSmolLM3RMSNorm.extra_repr)r   )r   r   r   rd   r   r   r   r+   r+   rw   r,   r      s    r   c                       s$   e Zd Z fddZdd Z  ZS )
SmolLM3MLPc                    sx   t    || _|j| _|j| _tj| j| j|jd| _tj| j| j|jd| _	tj| j| j|jd| _
t|j | _d S )Nr`   )rc   rd   r^   rf   intermediate_sizer   rj   mlp_bias	gate_projup_proj	down_projr   
hidden_actact_fnrv   r^   rw   r+   r,   rd      s   
zSmolLM3MLP.__init__c                 C   s$   |  | | || | }|S N)r   r   r   r   )rv   r(   r   r+   r+   r,   r      s    zSmolLM3MLP.forward)r   r   r   rd   r   r   r+   r+   rw   r,   r      s    
r   c                       s   e Zd Zdedef fddZedddd							
				ddejde	ej de	ej
 de	e de	e de	ej
 de	eejejf  dee dejfddZ  ZS )SmolLM3DecoderLayerr^   r_   c                    s^   t    |j| _t||d| _t|| _t|j|jd| _	t|j|jd| _
|j| | _d S )N)r^   r_   r   )rc   rd   rf   r]   	self_attnr   mlpr   rms_norm_epsinput_layernormpost_attention_layernormrs   attention_typeru   rw   r+   r,   rd      s   

zSmolLM3DecoderLayer.__init__ry   rz   r{   r|   NFr8   rG   r3   	use_cacher   r   rJ   r:   c              
   K   s^   |}	|  |}| jd|||||||d|\}}
|	| }|}	| |}| |}|	| }|S )N)r8   rG   r3   rz   r   r   r   r+   )r   r   r   r   )rv   r8   rG   r3   rz   r   r   r   rJ   residual_r+   r+   r,   r      s&   




zSmolLM3DecoderLayer.forward)NNNFNN)r   r   r   r    r   rd   r   r&   r   r   r   r   boolr   r   r   r   r   r+   r+   rw   r,   r      s8    	
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 )SmolLM3PreTrainedModelr^   modelTr   rz   )r8   
attentionsN)r   r   r   r    __annotations__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_outputsr+   r+   r+   r,   r     s   
 
r   c                       sD   e Zd ZU ejed< ddef fddZe e	dd Z
  ZS )	SmolLM3RotaryEmbeddinginv_freqNr^   c                    s   t    t|drt|jt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defaultr   F)
persistent)rc   rd   hasattr
isinstancer   dictgetr   max_position_embeddingsmax_seq_len_cachedoriginal_max_seq_lenr^   r   rope_init_fnattention_scalingregister_bufferr   original_inv_freq)rv   r^   devicer   rw   r+   r,   rd   2  s   
zSmolLM3RotaryEmbedding.__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   r!   r   mpscpuF)device_typeenabledr"   r#   )rL   )r   floatr;   r%   rU   r   r   r   strr&   autocastrQ   r'   r1   r   r2   rL   )
rv   r(   r3   inv_freq_expandedposition_ids_expandedr   freqsembr1   r2   r+   r+   r,   r   C  s   0&zSmolLM3RotaryEmbedding.forwardr   )r   r   r   r&   r   r   r    rd   no_gradr   r   r   r+   r+   rw   r,   r   /  s   
 
r   c                       s   e Zd Zdef f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j	 dee defddZ  ZS )SmolLM3Modelr^   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 r+   )r   ).0r_   r^   r+   r,   
<listcomp>\  s    z)SmolLM3Model.__init__.<locals>.<listcomp>r   r   Frb   )rc   rd   pad_token_idpadding_idx
vocab_sizer   	Embeddingrf   embed_tokens
ModuleListrangenum_hidden_layerslayersr   r   normr   
rotary_embgradient_checkpointingr^   rs   has_sliding_layers	post_initr   rw   r   r,   rd   U  s   zSmolLM3Model.__init__N	input_idsrG   r3   rz   inputs_embedsr   r   rJ   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   )r   )r^   input_embedsrG   r   rz   r3   full_attentionrb   )rG   r3   rz   r   r   r   )last_hidden_staterz   r+   )
ValueErrorr   r	   r^   get_seq_lengthr&   aranger%   r   r.   r   r   r   r   r   r   r   r   r   r   r   )rv   r   rG   r3   rz   r   r   r   rJ   past_seen_tokenscausal_mask_mappingmask_kwargsr8   r   decoder_layerr+   r+   r,   r   f  s^   



zSmolLM3Model.forward)NNNNNNN)r   r   r   r    rd   r   r   r   r&   r   r   r   FloatTensorr   r   r   r   r   r   r+   r+   rw   r,   r   S  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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
j deee
jf dee defddZ  ZS )SmolLM3ForCausalLMzlm_head.weightlm_headcolwise_repr8   logitsc                    s@   t  | t|| _|j| _tj|j|jdd| _| 	  d S )NFr`   )
rc   rd   r   r   r   r   rj   rf   r  r   r   rw   r+   r,   rd     s
   
zSmolLM3ForCausalLM.__init__Nr   r   rG   r3   rz   r   labelsr   r   logits_to_keeprJ   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  
        Example:

        ```python
        >>> from transformers import AutoTokenizer, SmolLM3ForCausalLM

        >>> model = SmolLM3ForCausalLM.from_pretrained("meta-smollm3/SmolLM3-2-7b-hf")
        >>> tokenizer = AutoTokenizer.from_pretrained("meta-smollm3/SmolLM3-2-7b-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   rG   r3   rz   r   r   r   N)r  r	  r   )lossr  rz   r8   r   r+   )r   r   r   r   slicer  loss_functionr^   r   r   rz   r8   r   )rv   r   rG   r3   rz   r   r	  r   r   r
  rJ   outputsr8   slice_indicesr  r  r+   r+   r,   r     s0    zSmolLM3ForCausalLM.forward)	NNNNNNNNr   )r   r   r   _tied_weights_keys_tp_plan_pp_planrd   r   r   r   r&   r   r   r   r  r   r   r   r   r   r   r   r   r+   r+   rw   r,   r    sN    		
r  c                   @      e Zd ZdS ) SmolLM3ForSequenceClassificationNr   r   r   r+   r+   r+   r,   r        r  c                   @   r  )SmolLM3ForTokenClassificationNr  r+   r+   r+   r,   r     r  r  c                   @   s   e Zd ZdZdS )SmolLM3ForQuestionAnsweringtransformerN)r   r   r   r   r+   r+   r+   r,   r    s    r  )r   r   r  r  r  r  )Nr   )rB   )Btypingr   r   r   r&   r   activationsr   cache_utilsr   r	   
generationr
   integrationsr   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.deprecationr   utils.genericr   configuration_smollm3r    r-   r7   r   r   rA   Moduler   r\   r]   r   r   r   r   r   r   r  r  r  r  __all__r+   r+   r+   r,   <module>   sn   

O/$\K