o
    wÖiÁ  ã                   @   sž  d Z ddlZddlmZ ddlmZmZ ddlZddl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mZmZmZ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!m"Z"m#Z# ddl$m%Z% e# &e'¡Z(da)dd„ Z*dd„ Z+dd„ Z,dd„ Z-G dd„ dej.j/ƒZ0G dd„ dej.j/ƒZ1G dd„ de	j2ƒZ3G dd„ de	j2ƒZ4G d d!„ d!e	j2ƒZ5G d"d#„ d#e	j2ƒZ6G d$d%„ d%e	j2ƒZ7G d&d'„ d'e	j2ƒZ8G d(d)„ d)eƒZ9G d*d+„ d+e	j2ƒZ:G d,d-„ d-e	j2ƒZ;G d.d/„ d/e	j2ƒZ<G d0d1„ d1e	j2ƒZ=e G d2d3„ d3eƒƒZ>e G d4d5„ d5e>ƒƒZ?e G d6d7„ d7e>ƒƒZ@G d8d9„ d9e	j2ƒZAe d:d;G d<d=„ d=e>ƒƒZBe G d>d?„ d?e>ƒƒZCe G d@dA„ dAe>ƒƒZDe G dBdC„ dCe>ƒƒZEg dD¢ZFdS )EzPyTorch YOSO model.é    N)ÚPath)ÚOptionalÚUnion)Únn)ÚBCEWithLogitsLossÚCrossEntropyLossÚMSELossé   )ÚACT2FN)ÚGradientCheckpointingLayer)Ú"BaseModelOutputWithCrossAttentionsÚMaskedLMOutputÚMultipleChoiceModelOutputÚQuestionAnsweringModelOutputÚSequenceClassifierOutputÚTokenClassifierOutput)ÚPreTrainedModel)Úapply_chunking_to_forwardÚ find_pruneable_heads_and_indicesÚprune_linear_layer)Úauto_docstringÚis_ninja_availableÚis_torch_cuda_availableÚloggingé   )Ú
YosoConfigc                  C   s:   ddl m}  dd„ }|g d¢ƒ}| d|dd dd lad S )	Nr   )Úloadc                    s,   t tƒ ¡ jjjd d ‰ ‡ fdd„| D ƒS )NÚkernelsÚyosoc                    s   g | ]}ˆ | ‘qS © r   )Ú.0Úfile©Ú
src_folderr   úc/home/ubuntu/sommelier/.venv/lib/python3.10/site-packages/transformers/models/yoso/modeling_yoso.pyÚ
<listcomp>;   ó    z:load_cuda_kernels.<locals>.append_root.<locals>.<listcomp>)r   Ú__file__ÚresolveÚparent)Úfilesr   r"   r$   Úappend_root9   s   z&load_cuda_kernels.<locals>.append_root)zfast_lsh_cumulation_torch.cppzfast_lsh_cumulation.cuzfast_lsh_cumulation_cuda.cuÚfast_lsh_cumulationT)Úverbose)Útorch.utils.cpp_extensionr   r,   Úlsh_cumulation)r   r+   Ú	src_filesr   r   r$   Úload_cuda_kernels5   s
   r1   c                 C   sJ   t | tƒrg }| D ]}| ¡ s| ¡ }| |¡ q	|S |  ¡ s#|  ¡ } | S ©N)Ú
isinstanceÚlistÚis_contiguousÚ
contiguousÚappend©Úinput_tensorsÚoutÚtensorr   r   r$   Úto_contiguousD   s   
r<   c                 C   sF   t | tƒrg }| D ]}| tjj|ddd¡ q	|S tjj| dddS )Né   éÿÿÿÿ)ÚpÚdim)r3   r4   r7   r   Ú
functionalÚ	normalizer8   r   r   r$   rB   R   s   
rB   c                 C   sü   t |  ¡ ƒdkrtdƒ‚t | ¡ ƒdkrtdƒ‚tj|  d¡|  d¡|| | jd}dtj|| jd }t | |¡ |  d¡|  d¡||¡}t ||¡ | d¡| d¡||¡}|dk 	¡ }|dk 	¡ }	tj
|| dd	}
tj
|	| dd	}
|
 	¡ |
 	¡ fS )
Nr	   zQuery has incorrect size.zKey has incorrect size.r   r=   ©Údevicer   r>   ©r@   )ÚlenÚsizeÚ
ValueErrorÚtorchÚrandnrD   ÚarangeÚmatmulÚreshapeÚintÚsum)ÚqueryÚkeyÚnum_hashÚhash_lenÚrmatÚ	raise_powÚquery_projectionÚkey_projectionÚquery_binaryÚ
key_binaryÚ
query_hashr   r   r$   Úhashing\   s   $$$r[   c                   @   ó$   e Zd Zedd„ ƒZedd„ ƒZdS )ÚYosoCumulationc           
   
   C   sŠ   |d }dt  t  || dd¡¡¡tj  | }||d d …d d …d f  |d d …d d d …f  }t  ||¡}	|  ||||||¡ || _|	S )NÚhash_code_lenr   r>   éþÿÿÿ)rI   ÚacosrL   Ú	transposeÚmathÚpiÚsave_for_backwardÚconfig)
ÚctxÚ
query_maskÚkey_maskrP   rQ   Úvaluere   r^   ÚexpectationÚcumulation_valuer   r   r$   Úforwardp   s   (0zYosoCumulation.forwardc                 C   s”   t |ƒ}| j\}}}}}}| j}|d }	t || dd¡¡| }
t |
|	d | ¡}t |
 dd¡|	d | ¡}t | dd¡|¡}d d |||d fS )Nr^   r>   r_   r=   )r<   Úsaved_tensorsre   rI   rL   ra   )rf   Úgradrg   rh   rj   rP   rQ   ri   re   r^   Úweighted_expÚ
grad_queryÚgrad_keyÚ
grad_valuer   r   r$   Úbackward}   s   zYosoCumulation.backwardN©Ú__name__Ú
__module__Ú__qualname__Ústaticmethodrl   rs   r   r   r   r$   r]   o   s
    
r]   c                   @   r\   )ÚYosoLSHCumulationc              
   C   sV  |  d¡|  d¡krtdƒ‚|  d¡|  d¡krtdƒ‚|  d¡|  d¡kr*tdƒ‚|  d¡|  d¡kr8tdƒ‚|  d¡|  d¡krFtdƒ‚|  d¡|  d¡krTtd	ƒ‚t|||||gƒ\}}}}}|j}|d
 }|d }	td|	 ƒ}
|d r†t ||||||	|d¡\}}n	t||||	ƒ\}}t ||||||
|d¡}|  |||||||¡ || _	|S )Nr   z6Query mask and Key mask differ in sizes in dimension 0z3Query mask and Query differ in sizes in dimension 0z1Query mask and Key differ in sizes in dimension 0z8Query mask and Value mask differ in sizes in dimension 0r   z,Key and Value differ in sizes in dimension 1r=   z,Query and Key differ in sizes in dimension 2rR   r^   Úuse_fast_hash)
rG   rH   r<   Úis_cudarN   r/   Ú	fast_hashr[   rd   re   )rf   rg   rh   rP   rQ   ri   re   Úuse_cudarR   r^   Úhashtable_capacityÚquery_hash_codeÚkey_hash_coderk   r   r   r$   rl      s8   
ÿÿzYosoLSHCumulation.forwardc                 C   sj  t |ƒ}| j\}}}}}}}| j}	|j}
|	d }td| ƒ}|	d rSt |||||||
d¡}t |||||||d | ||
d¡
}t |||||||d | ||
d¡
}nZdt t 	|| 
dd¡¡¡tj  | }||d d …d d …d f  |d d …d d d …f  }t 	|| 
dd¡¡| }t 	||d | ¡}t 	| 
dd¡|d | ¡}t 	| 
dd¡|¡}d d |||d fS )Nr^   r=   Úlsh_backwardr   é   r>   r_   )r<   rm   re   r{   rN   r/   Úlsh_weighted_cumulationrI   r`   rL   ra   rb   rc   )rf   rn   rg   rh   r   r€   rP   rQ   ri   re   r}   r^   r~   rr   rp   rq   rj   ro   r   r   r$   rs   µ   sR   ÿ
ö
ö(0zYosoLSHCumulation.backwardNrt   r   r   r   r$   ry   Ž   s
    
%ry   c                       s*   e Zd ZdZ‡ fdd„Zddd„Z‡  ZS )ÚYosoEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.c                    sÈ   t ƒ  ¡  tj|j|j|jd| _t |jd |j¡| _	t |j
|j¡| _tj|j|jd| _t |j¡| _| jdt |j¡ d¡d dd t|dd	ƒ| _| jd
tj| j ¡ tj| jjddd d S )N)Úpadding_idxr=   ©ÚepsÚposition_ids)r   r>   F)Ú
persistentÚposition_embedding_typeÚabsoluteÚtoken_type_ids©ÚdtyperD   )ÚsuperÚ__init__r   Ú	EmbeddingÚ
vocab_sizeÚhidden_sizeÚpad_token_idÚword_embeddingsÚmax_position_embeddingsÚposition_embeddingsÚtype_vocab_sizeÚtoken_type_embeddingsÚ	LayerNormÚlayer_norm_epsÚDropoutÚhidden_dropout_probÚdropoutÚregister_bufferrI   rK   ÚexpandÚgetattrrŠ   Úzerosrˆ   rG   ÚlongrD   ©Úselfre   ©Ú	__class__r   r$   r   ë   s   
ÿ
ýzYosoEmbeddings.__init__Nc                 C   sô   |d ur	|  ¡ }n|  ¡ d d… }|d }|d u r$| jd d …d |…f }|d u rNt| dƒrC| jd d …d |…f }| |d |¡}|}ntj|tj| jjd}|d u rW|  	|¡}|  
|¡}	||	 }
| jdkrn|  |¡}|
|7 }
|  |
¡}
|  |
¡}
|
S )Nr>   r   rŒ   r   r   r‹   )rG   rˆ   ÚhasattrrŒ   r    rI   r¢   r£   rD   r•   r™   rŠ   r—   rš   rž   )r¥   Ú	input_idsrŒ   rˆ   Úinputs_embedsÚinput_shapeÚ
seq_lengthÚbuffered_token_type_idsÚ buffered_token_type_ids_expandedr™   Ú
embeddingsr—   r   r   r$   rl     s,   







zYosoEmbeddings.forward)NNNN©ru   rv   rw   Ú__doc__r   rl   Ú__classcell__r   r   r¦   r$   r„   è   s    r„   c                       ó0   e Zd Zd	‡ fdd„	Zdd„ Zd
dd„Z‡  ZS )ÚYosoSelfAttentionNc              
      s¢  t ƒ  ¡  |j|j dkrt|dƒstd|j› d|j› dƒ‚td u}tƒ rKtƒ rK|sKzt	ƒ  W n t
yJ } zt d|› ¡ W Y d }~nd }~ww |j| _t|j|j ƒ| _| j| j | _t |j| j¡| _t |j| j¡| _t |j| j¡| _t |j¡| _|d ur‡|n|j| _|j| _|j| _|jd u| _|j| _|j| _|j| _| j| j| j| jdœ| _ |jd urÏtj!|j|j|jdf|jd	 dfd
|jd| _"d S d S )Nr   Úembedding_sizezThe hidden size (z6) is not a multiple of the number of attention heads (ú)zGCould not load the custom kernel for multi-scale deformable attention: )r^   rz   rR   r   r   r=   F)Úin_channelsÚout_channelsÚkernel_sizeÚpaddingÚbiasÚgroups)#r   r   r“   Únum_attention_headsr¨   rH   r/   r   r   r1   Ú	ExceptionÚloggerÚwarningrN   Úattention_head_sizeÚall_head_sizer   ÚLinearrP   rQ   ri   rœ   Úattention_probs_dropout_probrž   rŠ   Úuse_expectationr^   Úconv_windowÚuse_convrz   rR   r   Ú
lsh_configÚConv2dÚconv)r¥   re   rŠ   Úkernel_loadedÚer¦   r   r$   r   %  sZ   

ÿÿ
€ÿÿü
úÿzYosoSelfAttention.__init__c                 C   s6   |  ¡ d d… | j| jf }|j|Ž }| dddd¡S )Nr>   r   r=   r   r	   )rG   r½   rÁ   ÚviewÚpermute)r¥   ÚlayerÚnew_layer_shaper   r   r$   Útranspose_for_scoresX  s   
z&YosoSelfAttention.transpose_for_scoresFc                 C   s@  |   |¡}|  |  |¡¡}|  |  |¡¡}|  |¡}| jr.|  ||d d …d d d …d f  ¡}| ¡ \}	}
}}| |	|
 ||¡}| |	|
 ||¡}| |	|
 ||¡}d|d  }| d¡j	|
dd |	|
 |¡ 
¡ }d}| js¨||k r¨|	|
 ||| f}tj|tj||jdgdd}tj|tj||jdgdd}tj|tj||jdgdd}| js®| jr¶t||gƒ\}}| jrÅt |||||| j¡}nt |||||| j¡}| jsä||k rä|d d …d d …d |…f }t|ƒ}| |	|
||¡}| jr÷||7 }| dd	dd
¡ ¡ }| ¡ d d… | jf }|j|Ž }|r||f}|S |f}|S )Nç      ð?g     ˆÃ@r   rE   é    rC   r>   r   r=   r	   r_   )rP   rÑ   rQ   ri   rÇ   rÊ   rG   rM   Ú	unsqueezeÚrepeat_interleaverN   rÅ   rI   Úcatr¢   rD   ÚtrainingrB   r]   ÚapplyrÈ   ry   rÎ   r6   rÂ   rÍ   )r¥   Úhidden_statesÚattention_maskÚoutput_attentionsÚmixed_query_layerÚ	key_layerÚvalue_layerÚquery_layerÚconv_value_layerÚ
batch_sizeÚ	num_headsÚseq_lenÚhead_dimÚgpu_warp_sizeÚpad_sizeÚcontext_layerÚnew_context_layer_shapeÚoutputsr   r   r$   rl   ]  sx   

"ÿü	þûþûþûÿÿ
þzYosoSelfAttention.forwardr2   ©NF)ru   rv   rw   r   rÑ   rl   r²   r   r   r¦   r$   r´   $  s    3r´   c                       ó8   e Zd Z‡ fdd„Zdejdejdejfdd„Z‡  ZS )ÚYosoSelfOutputc                    sB   t ƒ  ¡  t |j|j¡| _tj|j|jd| _t |j	¡| _
d S ©Nr†   )r   r   r   rÃ   r“   Údenserš   r›   rœ   r   rž   r¤   r¦   r   r$   r   ³  ó   
zYosoSelfOutput.__init__rÙ   Úinput_tensorÚreturnc                 C   ó&   |   |¡}|  |¡}|  || ¡}|S r2   ©rî   rž   rš   ©r¥   rÙ   rð   r   r   r$   rl   ¹  ó   

zYosoSelfOutput.forward©ru   rv   rw   r   rI   ÚTensorrl   r²   r   r   r¦   r$   rì   ²  ó    $rì   c                       r³   )ÚYosoAttentionNc                    s.   t ƒ  ¡  t||d| _t|ƒ| _tƒ | _d S )N)rŠ   )r   r   r´   r¥   rì   ÚoutputÚsetÚpruned_heads)r¥   re   rŠ   r¦   r   r$   r   Á  s   

zYosoAttention.__init__c                 C   s²   t |ƒdkrd S t|| jj| jj| jƒ\}}t| jj|ƒ| j_t| jj|ƒ| j_t| jj	|ƒ| j_	t| j
j|dd| j
_| jjt |ƒ | j_| jj| jj | j_| j |¡| _d S )Nr   r   rE   )rF   r   r¥   r½   rÁ   rü   r   rP   rQ   ri   rú   rî   rÂ   Úunion)r¥   ÚheadsÚindexr   r   r$   Úprune_headsÇ  s   ÿzYosoAttention.prune_headsFc                 C   s4   |   |||¡}|  |d |¡}|f|dd …  }|S )Nr   r   )r¥   rú   )r¥   rÙ   rÚ   rÛ   Úself_outputsÚattention_outputré   r   r   r$   rl   Ù  s   zYosoAttention.forwardr2   rê   )ru   rv   rw   r   r   rl   r²   r   r   r¦   r$   rù   À  s    rù   c                       ó2   e Zd Z‡ fdd„Zdejdejfdd„Z‡  ZS )ÚYosoIntermediatec                    sD   t ƒ  ¡  t |j|j¡| _t|jt	ƒrt
|j | _d S |j| _d S r2   )r   r   r   rÃ   r“   Úintermediate_sizerî   r3   Ú
hidden_actÚstrr
   Úintermediate_act_fnr¤   r¦   r   r$   r   â  s
   
zYosoIntermediate.__init__rÙ   rñ   c                 C   ó   |   |¡}|  |¡}|S r2   )rî   r  ©r¥   rÙ   r   r   r$   rl   ê  ó   

zYosoIntermediate.forwardrö   r   r   r¦   r$   r  á  s    r  c                       rë   )Ú
YosoOutputc                    sB   t ƒ  ¡  t |j|j¡| _tj|j|jd| _t 	|j
¡| _d S rí   )r   r   r   rÃ   r  r“   rî   rš   r›   rœ   r   rž   r¤   r¦   r   r$   r   ò  rï   zYosoOutput.__init__rÙ   rð   rñ   c                 C   rò   r2   ró   rô   r   r   r$   rl   ø  rõ   zYosoOutput.forwardrö   r   r   r¦   r$   r  ñ  rø   r  c                       s.   e Zd Z‡ fdd„Zd	dd„Zdd„ Z‡  ZS )
Ú	YosoLayerc                    sB   t ƒ  ¡  |j| _d| _t|ƒ| _|j| _t|ƒ| _t	|ƒ| _
d S ©Nr   )r   r   Úchunk_size_feed_forwardÚseq_len_dimrù   Ú	attentionÚadd_cross_attentionr  Úintermediater  rú   r¤   r¦   r   r$   r      s   


zYosoLayer.__init__NFc                 C   sF   | j |||d}|d }|dd … }t| j| j| j|ƒ}|f| }|S )N)rÛ   r   r   )r  r   Úfeed_forward_chunkr  r  )r¥   rÙ   rÚ   rÛ   Úself_attention_outputsr  ré   Úlayer_outputr   r   r$   rl   	  s   ÿ
zYosoLayer.forwardc                 C   s   |   |¡}|  ||¡}|S r2   )r  rú   )r¥   r  Úintermediate_outputr  r   r   r$   r    s   
zYosoLayer.feed_forward_chunkrê   )ru   rv   rw   r   rl   r  r²   r   r   r¦   r$   r  ÿ  s    
	r  c                       s0   e Zd Z‡ fdd„Z					ddd„Z‡  ZS )	ÚYosoEncoderc                    s:   t ƒ  ¡  ˆ | _t ‡ fdd„tˆ jƒD ƒ¡| _d| _d S )Nc                    s   g | ]}t ˆ ƒ‘qS r   )r  )r    Ú_©re   r   r$   r%      r&   z(YosoEncoder.__init__.<locals>.<listcomp>F)	r   r   re   r   Ú
ModuleListÚrangeÚnum_hidden_layersrÏ   Úgradient_checkpointingr¤   r¦   r  r$   r     s   
 
zYosoEncoder.__init__NFTc                 C   s˜   |rdnd }|r
dnd }t | jƒD ]\}	}
|r||f }|
|||ƒ}|d }|r/||d f }q|r7||f }|sEtdd„ |||fD ƒƒS t|||dS )Nr   r   r   c                 s   s    | ]	}|d ur|V  qd S r2   r   )r    Úvr   r   r$   Ú	<genexpr>=  s   € z&YosoEncoder.forward.<locals>.<genexpr>)Úlast_hidden_staterÙ   Ú
attentions)Ú	enumeraterÏ   Útupler   )r¥   rÙ   rÚ   Ú	head_maskrÛ   Úoutput_hidden_statesÚreturn_dictÚall_hidden_statesÚall_self_attentionsÚiÚlayer_moduleÚlayer_outputsr   r   r$   rl   #  s&   	
€
ýzYosoEncoder.forward)NNFFT)ru   rv   rw   r   rl   r²   r   r   r¦   r$   r    s    	ùr  c                       r  )ÚYosoPredictionHeadTransformc                    sV   t ƒ  ¡  t |j|j¡| _t|jtƒrt	|j | _
n|j| _
tj|j|jd| _d S rí   )r   r   r   rÃ   r“   rî   r3   r  r  r
   Útransform_act_fnrš   r›   r¤   r¦   r   r$   r   G  s   
z$YosoPredictionHeadTransform.__init__rÙ   rñ   c                 C   s"   |   |¡}|  |¡}|  |¡}|S r2   )rî   r.  rš   r
  r   r   r$   rl   P  s   


z#YosoPredictionHeadTransform.forwardrö   r   r   r¦   r$   r-  F  s    	r-  c                       s,   e Zd Z‡ fdd„Zdd„ Zdd„ Z‡  ZS )ÚYosoLMPredictionHeadc                    sL   t ƒ  ¡  t|ƒ| _tj|j|jdd| _t 	t
 |j¡¡| _| j| j_d S )NF)r»   )r   r   r-  Ú	transformr   rÃ   r“   r’   ÚdecoderÚ	ParameterrI   r¢   r»   r¤   r¦   r   r$   r   Y  s
   

zYosoLMPredictionHead.__init__c                 C   s   | j | j_ d S r2   )r»   r1  ©r¥   r   r   r$   Ú_tie_weightsf  s   z!YosoLMPredictionHead._tie_weightsc                 C   r	  r2   )r0  r1  r
  r   r   r$   rl   i  r  zYosoLMPredictionHead.forward)ru   rv   rw   r   r4  rl   r²   r   r   r¦   r$   r/  X  s    r/  c                       r  )ÚYosoOnlyMLMHeadc                    s   t ƒ  ¡  t|ƒ| _d S r2   )r   r   r/  Úpredictionsr¤   r¦   r   r$   r   q  s   
zYosoOnlyMLMHead.__init__Úsequence_outputrñ   c                 C   s   |   |¡}|S r2   )r6  )r¥   r7  Úprediction_scoresr   r   r$   rl   u  s   
zYosoOnlyMLMHead.forwardrö   r   r   r¦   r$   r5  p  s    r5  c                   @   s    e Zd ZeZdZdZdd„ ZdS )ÚYosoPreTrainedModelr   Tc                 C   s´   t |tjƒr |jjjd| jjd |jdur|jj 	¡  dS dS t |tj
ƒrC|jjjd| jjd |jdurA|jj|j  	¡  dS dS t |tjƒrX|jj 	¡  |jj d¡ dS dS )zInitialize the weightsg        )ÚmeanÚstdNrÒ   )r3   r   rÃ   ÚweightÚdataÚnormal_re   Úinitializer_ranger»   Úzero_r‘   r…   rš   Úfill_)r¥   Úmoduler   r   r$   Ú_init_weights€  s   
ÿ
ÿþz!YosoPreTrainedModel._init_weightsN)ru   rv   rw   r   Úconfig_classÚbase_model_prefixÚsupports_gradient_checkpointingrC  r   r   r   r$   r9  z  s
    r9  c                       s¶   e Zd Z‡ fdd„Zdd„ Zdd„ Zdd„ Ze																		dd
ee	j
 dee	j
 dee	j
 dee	j
 dee	j
 dee	j
 dee dee dee deeef fdd„ƒZ‡  ZS )Ú	YosoModelc                    s2   t ƒ  |¡ || _t|ƒ| _t|ƒ| _|  ¡  d S r2   )r   r   re   r„   r¯   r  ÚencoderÚ	post_initr¤   r¦   r   r$   r   “  s
   

zYosoModel.__init__c                 C   s   | j jS r2   ©r¯   r•   r3  r   r   r$   Úget_input_embeddings  s   zYosoModel.get_input_embeddingsc                 C   s   || j _d S r2   rJ  )r¥   ri   r   r   r$   Úset_input_embeddings   s   zYosoModel.set_input_embeddingsc                 C   s*   |  ¡ D ]\}}| jj| j |¡ qdS )z
        Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base
        class PreTrainedModel
        N)ÚitemsrH  rÏ   r  r   )r¥   Úheads_to_prunerÏ   rþ   r   r   r$   Ú_prune_heads£  s   ÿzYosoModel._prune_headsNr©   rÚ   rŒ   rˆ   r%  rª   rÛ   r&  r'  rñ   c
                 C   s†  |d ur|n| j j}|d ur|n| j j}|	d ur|	n| j j}	|d ur*|d ur*tdƒ‚|d ur9|  ||¡ | ¡ }
n|d urF| ¡ d d… }
ntdƒ‚|
\}}|d urU|jn|j}|d u retj	||f|d}|d u rt
| jdƒr„| jjd d …d |…f }| ||¡}|}n	tj|
tj|d}|  || j j¡}| j||||d}| j||||||	d}|d	 }|	s¸|f|d
d …  S t||j|j|jdS )NzDYou cannot specify both input_ids and inputs_embeds at the same timer>   z5You have to specify either input_ids or inputs_embedsrC   rŒ   r   )r©   rˆ   rŒ   rª   )rÚ   r%  rÛ   r&  r'  r   r   )r!  rÙ   r"  Úcross_attentions)re   rÛ   r&  Úuse_return_dictrH   Ú%warn_if_padding_and_no_attention_maskrG   rD   rI   Úonesr¨   r¯   rŒ   r    r¢   r£   Úget_head_maskr  rH  r   rÙ   r"  rP  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   rÛ   r&  r'  r«   rá   r¬   rD   r­   r®   Úembedding_outputÚencoder_outputsr7  r   r   r$   rl   «  s\   ÿ
üúüzYosoModel.forward)	NNNNNNNNN)ru   rv   rw   r   rK  rL  rO  r   r   rI   r÷   Úboolr   r$  r   rl   r²   r   r   r¦   r$   rG  ‘  sH    
öþýüûúùø	÷
ö
õrG  c                       sÂ   e Zd ZddgZ‡ fdd„Zdd„ Zdd„ Ze																				dd
ee	j
 dee	j
 dee	j
 dee	j
 dee	j
 dee	j
 dee	j
 dee dee dee deeef fdd„ƒZ‡  ZS )ÚYosoForMaskedLMzcls.predictions.decoder.weightzcls.predictions.decoder.biasc                    s,   t ƒ  |¡ t|ƒ| _t|ƒ| _|  ¡  d S r2   )r   r   rG  r   r5  ÚclsrI  r¤   r¦   r   r$   r   ü  s   

zYosoForMaskedLM.__init__c                 C   s
   | j jjS r2   )rY  r6  r1  r3  r   r   r$   Úget_output_embeddings  s   
z%YosoForMaskedLM.get_output_embeddingsc                 C   s   || j j_|j| j j_d S r2   )rY  r6  r1  r»   )r¥   Únew_embeddingsr   r   r$   Úset_output_embeddings  s   
z%YosoForMaskedLM.set_output_embeddingsNr©   rÚ   rŒ   rˆ   r%  rª   ÚlabelsrÛ   r&  r'  rñ   c                 C   s°   |
dur|
n| j j}
| j||||||||	|
d	}|d }|  |¡}d}|dur8tƒ }|| d| j j¡| d¡ƒ}|
sN|f|dd…  }|durL|f| S |S t|||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 be in `[-100, 0, ...,
            config.vocab_size]` (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]`.
        N©rÚ   rŒ   rˆ   r%  rª   rÛ   r&  r'  r   r>   r   ©ÚlossÚlogitsrÙ   r"  )
re   rQ  r   rY  r   rÍ   r’   r   rÙ   r"  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  ré   r7  r8  Úmasked_lm_lossÚloss_fctrú   r   r   r$   rl     s6   ÷
üzYosoForMaskedLM.forward©
NNNNNNNNNN)ru   rv   rw   Ú_tied_weights_keysr   rZ  r\  r   r   rI   r÷   rW  r   r$  r   rl   r²   r   r   r¦   r$   rX  ø  sN    	õþýüûúùø	÷
öõ
ôrX  c                       s(   e Zd ZdZ‡ fdd„Zdd„ Z‡  ZS )ÚYosoClassificationHeadz-Head for sentence-level classification tasks.c                    sF   t ƒ  ¡  t |j|j¡| _t |j¡| _t |j|j	¡| _
|| _d S r2   )r   r   r   rÃ   r“   rî   rœ   r   rž   Ú
num_labelsÚout_projre   r¤   r¦   r   r$   r   E  s
   

zYosoClassificationHead.__init__c                 K   sR   |d d …dd d …f }|   |¡}|  |¡}t| jj |ƒ}|   |¡}|  |¡}|S )Nr   )rž   rî   r
   re   r  rh  )r¥   ÚfeaturesÚkwargsÚxr   r   r$   rl   M  s   



zYosoClassificationHead.forwardr°   r   r   r¦   r$   rf  B  s    rf  zœ
    YOSO Model transformer with a sequence classification/regression head on top (a linear layer on top of
    the pooled output) e.g. for GLUE tasks.
    )Úcustom_introc                       óª   e Zd Z‡ fdd„Ze										ddeej deej deej deej deej d	eej d
eej dee dee dee de	e
ef fdd„ƒZ‡  ZS )ÚYosoForSequenceClassificationc                    s4   t ƒ  |¡ |j| _t|ƒ| _t|ƒ| _|  ¡  d S r2   )r   r   rg  rG  r   rf  Ú
classifierrI  r¤   r¦   r   r$   r   ^  s
   

z&YosoForSequenceClassification.__init__Nr©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  rñ   c                 C   sh  |
dur|
n| j j}
| j||||||||	|
d	}|d }|  |¡}d}|dur”| j jdu rQ| jdkr7d| j _n| jdkrM|jtjksH|jtj	krMd| j _nd| j _| j jdkrot
ƒ }| jdkri|| ¡ | ¡ ƒ}n+|||ƒ}n%| j jdkr†tƒ }|| d| j¡| d¡ƒ}n| j jdkr”tƒ }|||ƒ}|
sª|f|dd…  }|dur¨|f| S |S t|||j|jd	S )
a  
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the sequence classification/regression loss. Indices should be in `[0, ...,
            config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If
            `config.num_labels > 1` a classification loss is computed (Cross-Entropy).
        Nr^  r   r   Ú
regressionÚsingle_label_classificationÚmulti_label_classificationr>   r_  )re   rQ  r   ro  Úproblem_typerg  rŽ   rI   r£   rN   r   Úsqueezer   rÍ   r   r   rÙ   r"  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  ré   r7  ra  r`  rc  rú   r   r   r$   rl   g  sT   ÷


"


üz%YosoForSequenceClassification.forwardrd  )ru   rv   rw   r   r   r   rI   r÷   rW  r   r$  r   rl   r²   r   r   r¦   r$   rn  W  sH    	õþýüûúùø	÷
öõ
ôrn  c                       rm  )ÚYosoForMultipleChoicec                    sD   t ƒ  |¡ t|ƒ| _t |j|j¡| _t |jd¡| _|  	¡  d S r  )
r   r   rG  r   r   rÃ   r“   Úpre_classifierro  rI  r¤   r¦   r   r$   r   °  s
   
zYosoForMultipleChoice.__init__Nr©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  rñ   c                 C   sŠ  |
dur|
n| j j}
|dur|jd n|jd }|dur%| d| d¡¡nd}|dur4| d| d¡¡nd}|durC| d| d¡¡nd}|durR| d| d¡¡nd}|dure| d| d¡| d¡¡nd}| j||||||||	|
d	}|d }|dd…df }|  |¡}t ¡ |ƒ}|  	|¡}| d|¡}d}|dur¥t
ƒ }|||ƒ}|
s»|f|dd…  }|dur¹|f| S |S t|||j|jdS )a[  
        input_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`):
            Indices of input sequence tokens in the vocabulary.

            Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
            [`PreTrainedTokenizer.__call__`] for details.

            [What are input IDs?](../glossary#input-ids)
        token_type_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`, *optional*):
            Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0,
            1]`:

            - 0 corresponds to a *sentence A* token,
            - 1 corresponds to a *sentence B* token.

            [What are token type IDs?](../glossary#token-type-ids)
        position_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`, *optional*):
            Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
            config.max_position_embeddings - 1]`.

            [What are position IDs?](../glossary#position-ids)
        inputs_embeds (`torch.FloatTensor` of shape `(batch_size, num_choices, sequence_length, hidden_size)`, *optional*):
            Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
            is useful if you want more control over how to convert *input_ids* indices into associated vectors than the
            model's internal embedding lookup matrix.
        labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
            Labels for computing the multiple choice classification loss. Indices should be in `[0, ...,
            num_choices-1]` where `num_choices` is the size of the second dimension of the input tensors. (See
            `input_ids` above)
        Nr   r>   r_   r^  r   r_  )re   rQ  ÚshaperÍ   rG   r   rv  r   ÚReLUro  r   r   rÙ   r"  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  Únum_choicesré   Úhidden_stateÚpooled_outputra  Úreshaped_logitsr`  rc  rú   r   r   r$   rl   º  sP   ,ÿý÷


üzYosoForMultipleChoice.forwardrd  )ru   rv   rw   r   r   r   rI   r÷   rW  r   r$  r   rl   r²   r   r   r¦   r$   ru  ®  sH    
õþýüûúùø	÷
öõ
ôru  c                       rm  )ÚYosoForTokenClassificationc                    sJ   t ƒ  |¡ |j| _t|ƒ| _t |j¡| _t 	|j
|j¡| _|  ¡  d S r2   )r   r   rg  rG  r   r   rœ   r   rž   rÃ   r“   ro  rI  r¤   r¦   r   r$   r     s   
z#YosoForTokenClassification.__init__Nr©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  rñ   c                 C   s
  |
dur|
n| j j}
| j||||||||	|
d	}|d }|  |¡}|  |¡}d}|duretƒ }|durX| d¡dk}| d| j¡}t 	|| d¡t 
|j¡ |¡¡}|||ƒ}n|| d| j¡| d¡ƒ}|
s{|f|dd…  }|dury|f| S |S t|||j|jdS )zÛ
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        Nr^  r   r>   r   r_  )re   rQ  r   rž   ro  r   rÍ   rg  rI   Úwherer;   Úignore_indexÚtype_asr   rÙ   r"  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   r]  rÛ   r&  r'  ré   r7  ra  r`  rc  Úactive_lossÚactive_logitsÚactive_labelsrú   r   r   r$   rl   %  sF   ÷

ÿüz"YosoForTokenClassification.forwardrd  )ru   rv   rw   r   r   r   rI   r÷   rW  r   r$  r   rl   r²   r   r   r¦   r$   r}    sH    õþýüûúùø	÷
öõ
ôr}  c                       s¶   e Zd Z‡ fdd„Ze											ddeej deej deej deej deej d	eej d
eej deej dee dee dee de	e
ef fdd„ƒZ‡  ZS )ÚYosoForQuestionAnsweringc                    sB   t ƒ  |¡ d|_|j| _t|ƒ| _t |j|j¡| _|  	¡  d S )Nr=   )
r   r   rg  rG  r   r   rÃ   r“   Ú
qa_outputsrI  r¤   r¦   r   r$   r   f  s   
z!YosoForQuestionAnswering.__init__Nr©   rÚ   rŒ   rˆ   r%  rª   Ústart_positionsÚend_positionsrÛ   r&  r'  rñ   c                 C   s@  |d ur|n| j j}| j|||||||	|
|d	}|d }|  |¡}|jddd\}}| d¡}| d¡}d }|d ur~|d ur~t| ¡ ƒdkrK| d¡}t| ¡ ƒdkrX| d¡}| d¡}| d|¡}| d|¡}t	|d}|||ƒ}|||ƒ}|| d }|s•||f|dd …  }|d ur“|f| S |S t
||||j|jdS )	Nr^  r   r   r>   rE   )r  r=   )r`  Ústart_logitsÚ
end_logitsrÙ   r"  )re   rQ  r   r…  Úsplitrt  rF   rG   Úclampr   r   rÙ   r"  )r¥   r©   rÚ   rŒ   rˆ   r%  rª   r†  r‡  rÛ   r&  r'  ré   r7  ra  rˆ  r‰  Ú
total_lossÚignored_indexrc  Ú
start_lossÚend_lossrú   r   r   r$   rl   r  sP   ÷








ûz YosoForQuestionAnswering.forward)NNNNNNNNNNN)ru   rv   rw   r   r   r   rI   r÷   rW  r   r$  r   rl   r²   r   r   r¦   r$   r„  d  sN    ôþýüûúùø	÷
öõô
ór„  )rX  ru  r„  rn  r}  r  rG  r9  )Gr±   rb   Úpathlibr   Útypingr   r   rI   Útorch.utils.checkpointr   Útorch.nnr   r   r   Úactivationsr
   Úmodeling_layersr   Úmodeling_outputsr   r   r   r   r   r   Úmodeling_utilsr   Úpytorch_utilsr   r   r   Úutilsr   r   r   r   Úconfiguration_yosor   Ú
get_loggerru   r¿   r/   r1   r<   rB   r[   ÚautogradÚFunctionr]   ry   ÚModuler„   r´   rì   rù   r  r  r  r  r-  r/  r5  r9  rG  rX  rf  rn  ru  r}  r„  Ú__all__r   r   r   r$   Ú<module>   sj    

Z< !*
fIÿQiKO