o
    Zh                    @   s  d Z ddlZddlmZmZ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mZ ddlmZ dd	lmZmZ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$Z$ ddl%m&Z& e$'e(Z)G dd de	j*Z+G dd de	j*Z,G dd de	j*Z-G dd de	j*Z.G dd de	j*Z/G dd de	j*Z0G dd de	j*Z1G dd de	j*Z2G dd  d e	j*Z3G d!d" d"e	j*Z4e#G d#d$ d$eZ5e#d%d&G d'd( d(e5Z6e#d)d&G d*d+ d+e5eZ7e#G d,d- d-e5Z8G d.d/ d/e	j*Z9e#d0d&G d1d2 d2e5Z:e#G d3d4 d4e5Z;e#G d5d6 d6e5Z<G d7d8 d8e	j*Z=e#G d9d: d:e5Z>d>d;d<Z?g d=Z@dS )?zPyTorch X-MOD model.    N)ListOptionalTupleUnion)nn)BCEWithLogitsLossCrossEntropyLossMSELoss   )ACT2FNgelu)GenerationMixin))BaseModelOutputWithPastAndCrossAttentions,BaseModelOutputWithPoolingAndCrossAttentions!CausalLMOutputWithCrossAttentionsMaskedLMOutputMultipleChoiceModelOutputQuestionAnsweringModelOutputSequenceClassifierOutputTokenClassifierOutput)PreTrainedModel)apply_chunking_to_forward find_pruneable_heads_and_indicesprune_linear_layer)auto_docstringlogging   )
XmodConfigc                       s4   e Zd ZdZ fddZ	d
ddZdd	 Z  ZS )XmodEmbeddingszV
    Same as BertEmbeddings with a tiny tweak for positional embeddings indexing.
    c                    s   t    tj|j|j|jd| _t|j|j| _	t|j
|j| _tj|j|jd| _t|j| _t|dd| _| jdt|jddd | jd	tj| j tjd
dd |j| _tj|j|j| jd| _	d S )N)padding_idxZepsposition_embedding_typeabsoluteposition_ids)r   F)
persistenttoken_type_idsdtype)super__init__r   	Embedding
vocab_sizehidden_sizeZpad_token_idword_embeddingsmax_position_embeddingsposition_embeddingsZtype_vocab_sizetoken_type_embeddings	LayerNormlayer_norm_epsDropouthidden_dropout_probdropoutgetattrr!   Zregister_buffertorcharangeexpandzerosr#   sizelongr   selfconfig	__class__ U/var/www/auris/lib/python3.10/site-packages/transformers/models/xmod/modeling_xmod.pyr*   5   s"   
zXmodEmbeddings.__init__Nr   c                 C   s   |d u r|d urt || j|}n| |}|d ur| }n| d d }|d }|d u rTt| drI| jd d d |f }||d |}	|	}ntj|tj	| j
jd}|d u r]| |}| |}
||
 }| jdkrt| |}||7 }| |}| |}|S )Nr$   r   r&   r   r(   devicer"   )"create_position_ids_from_input_idsr   &create_position_ids_from_inputs_embedsr<   hasattrr&   r:   r8   r;   r=   r#   rF   r.   r1   r!   r0   r2   r6   )r?   	input_idsr&   r#   inputs_embedspast_key_values_lengthinput_shape
seq_lengthbuffered_token_type_ids buffered_token_type_ids_expandedr1   
embeddingsr0   rC   rC   rD   forwardN   s0   








zXmodEmbeddings.forwardc                 C   sN   |  dd }|d }tj| jd || j d tj|jd}|d|S )z
        We are provided embeddings directly. We cannot infer which are padded so just generate sequential position ids.

        Args:
            inputs_embeds: torch.Tensor

        Returns: torch.Tensor
        Nr$   r   rE   r   )r<   r8   r9   r   r=   rF   Z	unsqueezer:   )r?   rK   rM   Zsequence_lengthr#   rC   rC   rD   rH   v   s   	z5XmodEmbeddings.create_position_ids_from_inputs_embeds)NNNNr   )__name__
__module____qualname____doc__r*   rR   rH   __classcell__rC   rC   rA   rD   r   /   s    
(r   c                       s   e Zd Zd fdd	ZdejdejfddZ						dd	ejd
eej deej deej deej dee	e	ej   dee
 de	ej fddZ  ZS )XmodSelfAttentionNc                    s   t    |j|j dkrt|dstd|j d|j d|j| _t|j|j | _| j| j | _t	
|j| j| _t	
|j| j| _t	
|j| j| _t	|j| _|p\t|dd| _| jdksh| jd	kry|j| _t	d
|j d | j| _|j| _d S )Nr   Zembedding_sizezThe hidden size (z6) is not a multiple of the number of attention heads ()r!   r"   relative_keyrelative_key_query   r   )r)   r*   r-   num_attention_headsrI   
ValueErrorintattention_head_sizeall_head_sizer   Linearquerykeyvaluer4   Zattention_probs_dropout_probr6   r7   r!   r/   r+   distance_embedding
is_decoderr?   r@   r!   rA   rC   rD   r*      s*   

zXmodSelfAttention.__init__xreturnc                 C   s6   |  d d | j| jf }||}|ddddS )Nr$   r   r\   r   r
   )r<   r]   r`   viewpermute)r?   ri   Znew_x_shaperC   rC   rD   transpose_for_scores   s   
z&XmodSelfAttention.transpose_for_scoresFhidden_statesattention_mask	head_maskencoder_hidden_statesencoder_attention_maskpast_key_valueoutput_attentionsc                 C   s  |  |}|d u}	|	r|d ur|d }
|d }|}nP|	r/| | |}
| | |}|}n;|d urZ| | |}
| | |}tj|d |
gdd}
tj|d |gdd}n| | |}
| | |}| |}|d u}| jrz|
|f}t||
dd}| j	dks| j	dkr	|j
d |
j
d }}|rtj|d tj|jd	dd}ntj|tj|jd	dd}tj|tj|jd	dd}|| }| || j d }|j|jd
}| j	dkrtd||}|| }n| j	dkr	td||}td|
|}|| | }|t| j }|d ur|| }tjj|dd}| |}|d ur0|| }t||}|dddd }| d d | jf }||}|rX||fn|f}| jrd||f }|S )Nr   r   r\   dimr$   rZ   r[   rE   r'   zbhld,lrd->bhlrzbhrd,lrd->bhlrr
   ) rc   rm   rd   re   r8   catrg   matmulZ	transposer!   shapeZtensorr=   rF   rk   r9   rf   r/   tor(   Zeinsummathsqrtr`   r   Z
functionalZsoftmaxr6   rl   
contiguousr<   ra   )r?   rn   ro   rp   rq   rr   rs   rt   Zmixed_query_layerZis_cross_attentionZ	key_layerZvalue_layerZquery_layer	use_cacheZattention_scoresZquery_lengthZ
key_lengthZposition_ids_lZposition_ids_rZdistanceZpositional_embeddingZrelative_position_scoresZrelative_position_scores_queryZrelative_position_scores_keyZattention_probsZcontext_layerZnew_context_layer_shapeoutputsrC   rC   rD   rR      sn   









zXmodSelfAttention.forwardNNNNNNF)rS   rT   rU   r*   r8   Tensorrm   r   FloatTensorr   boolrR   rW   rC   rC   rA   rD   rX      s4    	rX   c                       s8   e Zd Z fddZdejdejdejfddZ  ZS )XmodSelfOutputc                    sB   t    t|j|j| _tj|j|jd| _t|j	| _
d S Nr    )r)   r*   r   rb   r-   denser2   r3   r4   r5   r6   r>   rA   rC   rD   r*     s   
zXmodSelfOutput.__init__rn   input_tensorrj   c                 C   s    |  |}| |}|| }|S r   )r   r6   )r?   rn   r   rC   rC   rD   rR     s   

zXmodSelfOutput.forwardrS   rT   rU   r*   r8   r   rR   rW   rC   rC   rA   rD   r     s    $r   c                       s   e Zd Zd fdd	Zdd Z						ddejdeej d	eej d
eej deej dee	e	ej   dee
 de	ej fddZ  ZS )XmodAttentionNc                    s6   t    t||d| _t|| _t | _|j| _d S )Nr!   )	r)   r*   rX   r?   r   outputsetpruned_headspre_normrh   rA   rC   rD   r*     s
   

zXmodAttention.__init__c                 C   s   t |dkrd S t|| jj| jj| j\}}t| jj|| j_t| jj|| j_t| jj	|| j_	t| j
j|dd| j
_| jjt | | j_| jj| jj | j_| j|| _d S )Nr   r   ru   )lenr   r?   r]   r`   r   r   rc   rd   re   r   r   ra   union)r?   headsindexrC   rC   rD   prune_heads'  s   zXmodAttention.prune_headsFrn   ro   rp   rq   rr   rs   rt   rj   c              	   C   sd   |}| j r| j|}| |||||||}	| |	d |}
| j s'| j|
}
|
f|	dd   }|S )Nr   r   )r   r   r2   r?   )r?   rn   ro   rp   rq   rr   rs   rt   residualZself_outputsattention_outputr   rC   rC   rD   rR   9  s"   
	zXmodAttention.forwardr   r   )rS   rT   rU   r*   r   r8   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r     s4    	r   c                       2   e Zd Z fddZdejdejfddZ  ZS )XmodIntermediatec                    sD   t    t|j|j| _t|jt	rt
|j | _d S |j| _d S r   )r)   r*   r   rb   r-   intermediate_sizer   
isinstance
hidden_actstrr   intermediate_act_fnr>   rA   rC   rD   r*   X  s
   
zXmodIntermediate.__init__rn   rj   c                 C   s   |  |}| |}|S r   )r   r   r?   rn   rC   rC   rD   rR   `  s   

zXmodIntermediate.forwardr   rC   rC   rA   rD   r   W  s    r   c                       r   )XmodAdapterc                    sd   t    |j|j | _t|j| j| _t| j|j| _t	|j
tr,t|j
 | _d S |j
| _d S r   )r)   r*   r-   Zadapter_reduction_factorZbottleneck_sizer   rb   dense1dense2r   r   r   r   adapter_act_fnr>   rA   rC   rD   r*   g  s   
zXmodAdapter.__init__rn   rj   c                 C   s"   |  |}| |}| |}|S r   )r   r   r   r   rC   rC   rD   rR   q  s   


zXmodAdapter.forwardr   rC   rC   rA   rD   r   f  s    
r   c                       sT   e Zd Z fddZdejdejdejdejfddZdejdejfd	d
Z  ZS )
XmodOutputc                    s   t    t|j|j| _tj|j|jd| _|j	| _	t
|j| _|jr1tj|j|jd| _nd | _|j| _ti | _|jD ]}t|| jt|< qAd S r   )r)   r*   r   rb   r   r-   r   r2   r3   ln_before_adapterr4   r5   r6   adapter_layer_normadapter_reuse_layer_normZ
ModuleDictadapter_modules	languagesr   r   )r?   r@   languagerA   rC   rD   r*   y  s   

zXmodOutput.__init__rn   r   lang_idsrj   c                 C   s,   |  |}| |}|| }| ||}|S r   )r   r6   lang_adapter)r?   rn   r   r   rC   rC   rD   rR     s
   

zXmodOutput.forwardc                 C   s   t j|dd\}}| js|}| jd ur| |}n| jr!| |}| jr&|}t || d}g }tt	||D ]\}\}}	t
| j t|  }
|| j|
 |	 q8t |d}| |}||7 }|S )NT)Zreturn_countsr   )r8   Zunique_consecutiver   r   r   r2   splittolist	enumerateziplistr   keysr_   itemappendrx   r6   )r?   r   rn   Zlang_lengthsr   Zsplit_hidden_statesZlang_wise_outputsiZlang_idZsplit_hidden_statelangrC   rC   rD   r     s$   


zXmodOutput.lang_adapter)	rS   rT   rU   r*   r8   r   rR   r   rW   rC   rC   rA   rD   r   x  s    "r   c                       s   e Zd Z fddZ						ddejdejdeej deej d	eej d
eej deeeej   dee	 deej fddZ
dd Z  ZS )	XmodLayerc                    sz   t    |j| _d| _t|| _|j| _|j| _| jr-| js&t|  dt|dd| _	t
|| _t|| _|j| _d S )Nr   z> should be used as a decoder model if cross attention is addedr"   r   )r)   r*   chunk_size_feed_forwardseq_len_dimr   	attentionrg   add_cross_attentionr^   crossattentionr   intermediater   r   r   r>   rA   rC   rD   r*     s   



zXmodLayer.__init__NFrn   r   ro   rp   rq   rr   rs   rt   rj   c	              	   C   sF  |d ur
|d d nd }	| j |||||	d}
|
d }| jr(|
dd }|
d }n|
dd  }d }| jro|d urot| dsDtd|  d|d urN|d	d  nd }| |||||||}|d }||dd  }|d }|| }|}| jrz| j|}t| j	| j
| j|}| |||}| js| j|}|f| }| jr||f }|S )
Nr\   )rt   rs   r   r   r$   r   z'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`rw   )r   rg   rI   r^   r   r   r   r2   r   feed_forward_chunkr   r   )r?   rn   r   ro   rp   rq   rr   rs   rt   Zself_attn_past_key_valueZself_attention_outputsr   r   Zpresent_key_valueZcross_attn_present_key_valueZcross_attn_past_key_valueZcross_attention_outputsr   Zintermediate_outputZlayer_outputrC   rC   rD   rR     sb   


	

zXmodLayer.forwardc                 C   s
   |  |S r   )r   )r?   r   rC   rC   rD   r     s   
zXmodLayer.feed_forward_chunkr   )rS   rT   rU   r*   r8   r   r   r   r   r   rR   r   rW   rC   rC   rA   rD   r     s8    	

Kr   c                       s   e Zd Z fddZ									ddejdejdeej d	eej d
eej deej deeeej   dee	 dee	 dee	 dee	 de
eej ef fddZ  ZS )XmodEncoderc                    s\   t     | _t fddt jD | _ j| _	| j	r)tj
 j jd| _
d| _d S )Nc                    s   g | ]}t  qS rC   )r   ).0_r@   rC   rD   
<listcomp>  s    z(XmodEncoder.__init__.<locals>.<listcomp>r    F)r)   r*   r@   r   Z
ModuleListrangenum_hidden_layerslayerr   is_pre_normr2   r-   r3   gradient_checkpointingr>   rA   r   rD   r*   
  s   
 
zXmodEncoder.__init__NFTrn   r   ro   rp   rq   rr   past_key_valuesr   rt   output_hidden_statesreturn_dictrj   c                 C   sr  | j r| jr|rtd d}|
rdnd }|	rdnd }|	r#| jjr#dnd }|r)dnd }t| jD ]`\}}|
r;||f }|d urC|| nd }|d urM|| nd }| j rd| jrd| |j	||||||||		}n|||||||||	}|d }|r|||d f7 }|	r||d f }| jjr||d f }q0| j
r| |}|
r||f }|stdd	 |||||fD S t|||||d
S )NzZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...FrC   r   r$   r   r\   c                 s   s    | ]	}|d ur|V  qd S r   rC   )r   vrC   rC   rD   	<genexpr>Z  s    z&XmodEncoder.forward.<locals>.<genexpr>)last_hidden_stater   rn   
attentionscross_attentions)r   ZtrainingloggerZwarning_oncer@   r   r   r   Z_gradient_checkpointing_func__call__r   r2   tupler   )r?   rn   r   ro   rp   rq   rr   r   r   rt   r   r   Zall_hidden_statesZall_self_attentionsZall_cross_attentionsZnext_decoder_cacher   Zlayer_moduleZlayer_head_maskrs   Zlayer_outputsrC   rC   rD   rR     s   


zXmodEncoder.forward)	NNNNNNFFT)rS   rT   rU   r*   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r   	  sH    	
r   c                       r   )
XmodPoolerc                    s*   t    t|j|j| _t | _d S r   )r)   r*   r   rb   r-   r   ZTanh
activationr>   rA   rC   rD   r*   p  s   
zXmodPooler.__init__rn   rj   c                 C   s(   |d d df }|  |}| |}|S Nr   )r   r   )r?   rn   Zfirst_token_tensorpooled_outputrC   rC   rD   rR   u  s   

zXmodPooler.forwardr   rC   rC   rA   rD   r   o  s    r   c                   @   s6   e Zd ZeZdZdZdd ZdefddZ	dd	 Z
d
S )XmodPreTrainedModelrobertaTc                 C   s   t |tjr |jjjd| jjd |jdur|jj	  dS dS t |tj
rC|jjjd| jjd |jdurA|jj|j 	  dS dS t |tjrX|jj	  |jjd dS t |tre|jj	  dS dS )zInitialize the weightsg        )meanZstdNg      ?)r   r   rb   weightdataZnormal_r@   Zinitializer_rangebiasZzero_r+   r   r2   Zfill_
XmodLMHead)r?   modulerC   rC   rD   _init_weights  s    


z!XmodPreTrainedModel._init_weightsr   c                 C   s8   || j jvrt|  d| dt| j j || j _dS )z
        Set the default language code for the model. This is used when the language is not specified in the input.

        Args:
            language (`str`): The language code, such as `"en_XX"` or `"de_DE"`.
        z does not have an adapter for z. Supported languages: N)r@   r   r^   r   default_language)r?   r   rC   rC   rD   set_default_language  s
   z(XmodPreTrainedModel.set_default_languagec                 C   s|   t d | jj D ]}d|_qt d | jjjD ] }|jj	dur/|jj	 D ]}d|_q)|jj
 D ]}d|_q5qdS )z
        Freeze the embeddings and language adapters of the model. Usually, this is applied before the model is
        fine-tuned on a downstream task.
        zFreezing embeddingsFzFreezing adaptersN)r   infor   rQ   
parametersZrequires_gradencoderr   r   r   r   )r?   Z	parameterr   rC   rC   rD   'freeze_embeddings_and_language_adapters  s   

z;XmodPreTrainedModel.freeze_embeddings_and_language_adaptersN)rS   rT   rU   r   Zconfig_classZbase_model_prefixZsupports_gradient_checkpointingr   r   r   r   rC   rC   rC   rD   r   ~  s    r   a(  
    The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of
    cross-attention is added between the self-attention layers, following the architecture described in *Attention is
    all you need*_ by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz
    Kaiser and Illia Polosukhin.

    To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set
    to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and
    `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass.

    .. _*Attention is all you need*: https://arxiv.org/abs/1706.03762
    )Zcustom_introc                "       s   e Zd Zd 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	j
 dee	j
 dee	j
 deee	j  dee dee dee dee deee	j
 ef fddZ  ZS )	XmodModelTc                    sD   t  | || _t|| _t|| _|rt|nd| _| 	  dS )zv
        add_pooling_layer (bool, *optional*, defaults to `True`):
            Whether to add a pooling layer
        N)
r)   r*   r@   r   rQ   r   r   r   pooler	post_init)r?   r@   add_pooling_layerrA   rC   rD   r*     s   

zXmodModel.__init__c                 C      | j jS r   rQ   r.   r?   rC   rC   rD   get_input_embeddings     zXmodModel.get_input_embeddingsc                 C      || j _d S r   r   )r?   re   rC   rC   rD   set_input_embeddings     zXmodModel.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)itemsr   r   r   r   )r?   Zheads_to_pruner   r   rC   rC   rD   _prune_heads  s   zXmodModel._prune_headsNrJ   r   ro   r&   r#   rp   rK   rq   rr   r   r   rt   r   r   rj   c           "      C   s  |dur|n| j j}|dur|n| j j}|dur|n| j j}| j jr-|dur(|n| j j}nd}|dur;|dur;td|durJ| || | }n|durW| dd }ntd|\}}|durf|j	n|j	}|
durv|
d d j
d nd}|du r| j jdu rtdt| jjd jj }|| j j}|tj||d	 }|du rtj||| f|d	}|du rt| jd
r| jjddd|f }|||}|}n	tj|tj|d}| ||}| j jr|dur| \}}}||f}|	du rtj||d	}	| |	}nd}| || j j}| j|||||d}| j|||||||
||||d}|d } | jdur8| | nd}!|sG| |!f|dd  S t | |!|j!|j"|j#|j$dS )  
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        NFzDYou cannot specify both input_ids and inputs_embeds at the same timer$   z5You have to specify either input_ids or inputs_embedsr   r\   zPInput language unknown. Please call `XmodPreTrainedModel.set_default_language()`)rF   r&   rE   )rJ   r#   r&   rK   rL   )
r   ro   rp   rq   rr   r   r   rt   r   r   r   )r   Zpooler_outputr   rn   r   r   )%r@   rt   r   use_return_dictrg   r   r^   Z%warn_if_padding_and_no_attention_maskr<   rF   rz   r   r   r   r   r   r   r   r   r8   ZonesrI   rQ   r&   r:   r;   r=   Zget_extended_attention_maskZinvert_attention_maskZget_head_maskr   r   r   r   rn   r   r   )"r?   rJ   r   ro   r&   r#   rp   rK   rq   rr   r   r   rt   r   r   rM   Z
batch_sizerN   rF   rL   Zadapter_languagesZdefault_lang_idrO   rP   Zextended_attention_maskZencoder_batch_sizeZencoder_sequence_lengthr   Zencoder_hidden_shapeZencoder_extended_attention_maskZembedding_outputZencoder_outputssequence_outputr   rC   rC   rD   rR     s   
zXmodModel.forward)T)NNNNNNNNNNNNNN)rS   rT   rU   r*   r   r   r   r   r   r8   r   
LongTensorr   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r     sf    	
r   zQ
    X-MOD Model with a `language modeling` head on top for CLM fine-tuning.
    c                $       s  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	j dee	j dee	j
 deeee	j   dee dee dee dee deee	j ef f ddZdd Z  ZS )XmodForCausalLMlm_head.decoder.weightlm_head.decoder.biasc                    s@   t  | |jstd t|dd| _t|| _| 	  d S )NzLIf you want to use `XmodLMHeadModel` as a standalone, add `is_decoder=True.`Fr   
r)   r*   rg   r   warningr   r   r   lm_headr   r>   rA   rC   rD   r*   m  s   

zXmodForCausalLM.__init__c                 C   r   r   r   decoderr   rC   rC   rD   get_output_embeddingsz  r   z%XmodForCausalLM.get_output_embeddingsc                 C   r   r   r   r?   Znew_embeddingsrC   rC   rD   set_output_embeddings~  r   z%XmodForCausalLM.set_output_embeddingsNrJ   r   ro   r&   r#   rp   rK   rq   rr   labelsr   r   rt   r   r   rj   c                 K   s   |dur|n| j j}|
durd}| j|||||||||	|||||d}|d }| |}d}|
dur@| j||
fd| j ji|}|sV|f|dd  }|durT|f| S |S t|||j|j|j	|j
dS )aS  
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the left-to-right language modeling loss (next word prediction). 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]`

        Example:

        ```python
        >>> from transformers import AutoTokenizer, XmodForCausalLM, AutoConfig
        >>> import torch

        >>> tokenizer = AutoTokenizer.from_pretrained("FacebookAI/xlm-roberta-base")
        >>> config = AutoConfig.from_pretrained("facebook/xmod-base")
        >>> config.is_decoder = True
        >>> model = XmodForCausalLM.from_pretrained("facebook/xmod-base", config=config)
        >>> model.set_default_language("en_XX")

        >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt")
        >>> outputs = model(**inputs)

        >>> prediction_logits = outputs.logits
        ```NF)r   ro   r&   r#   rp   rK   rq   rr   r   r   rt   r   r   r   r,   r\   )losslogitsr   rn   r   r   )r@   r   r   r   Zloss_functionr,   r   r   rn   r   r   )r?   rJ   r   ro   r&   r#   rp   rK   rq   rr   r  r   r   rt   r   r   kwargsr   r   prediction_scoresZlm_lossr   rC   rC   rD   rR     sT   .
zXmodForCausalLM.forwardc                    s.   d}|D ]}|t  fdd|D f7 }q|S )NrC   c                 3   s$    | ]}| d  |jV  qdS )r   N)Zindex_selectr{   rF   )r   Z
past_statebeam_idxrC   rD   r     s   " z1XmodForCausalLM._reorder_cache.<locals>.<genexpr>)r   )r?   r   r  Zreordered_pastZ
layer_pastrC   r
  rD   _reorder_cache  s   zXmodForCausalLM._reorder_cache)NNNNNNNNNNNNNNN)rS   rT   rU   _tied_weights_keysr*   r  r  r   r   r8   r   r   r   r   r   r   r   rR   r  rW   rC   rC   rA   rD   r   d  sn    	
\r   c                        s   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	j dee	j dee	j
 dee dee dee deee	j ef fddZ  ZS )XmodForMaskedLMr   r   c                    s@   t  | |jrtd t|dd| _t|| _| 	  d S )NzkIf you want to use `XmodForMaskedLM` make sure `config.is_decoder=False` for bi-directional self-attention.Fr   r   r>   rA   rC   rD   r*     s   
zXmodForMaskedLM.__init__c                 C   r   r   r   r   rC   rC   rD   r    r   z%XmodForMaskedLM.get_output_embeddingsc                 C   r   r   r   r  rC   rC   rD   r     r   z%XmodForMaskedLM.set_output_embeddingsNrJ   r   ro   r&   r#   rp   rK   rq   rr   r  rt   r   r   rj   c                 C   s   |dur|n| j j}| j|||||||||	|||d}|d }| |}d}|
dur;t }||d| j j|
d}|sQ|f|dd  }|durO|f| S |S t|||j|j	dS )a  
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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)r   ro   r&   r#   rp   rK   rq   rr   rt   r   r   r   r$   r\   r  r  rn   r   )
r@   r   r   r   r   rk   r,   r   rn   r   )r?   rJ   r   ro   r&   r#   rp   rK   rq   rr   r  rt   r   r   r   r   r	  Zmasked_lm_lossloss_fctr   rC   rC   rD   rR     s<   
zXmodForMaskedLM.forward)NNNNNNNNNNNNN)rS   rT   rU   r  r*   r  r  r   r   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r    s`    	
r  c                       s0   e Zd ZdZ fddZdd Zdd Z  ZS )r   z*Roberta Head for masked language modeling.c                    sd   t    t|j|j| _tj|j|jd| _t|j|j	| _
tt|j	| _| j| j
_d S r   )r)   r*   r   rb   r-   r   r2   r3   
layer_normr,   r  	Parameterr8   r;   r   r>   rA   rC   rD   r*   E  s   
zXmodLMHead.__init__c                 K   s*   |  |}t|}| |}| |}|S r   )r   r   r  r  r?   featuresr  ri   rC   rC   rD   rR   N  s
   


zXmodLMHead.forwardc                 C   s,   | j jjjdkr| j| j _d S | j j| _d S )Nmeta)r  r   rF   typer   rC   rC   rD   _tie_weightsX  s   zXmodLMHead._tie_weights)rS   rT   rU   rV   r*   rR   r  rW   rC   rC   rA   rD   r   B  s
    	
r   z
    X-MOD Model transformer with a sequence classification/regression head on top (a linear layer on top of the pooled
    output) e.g. for GLUE tasks.
    c                          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j ef fddZ  ZS )XmodForSequenceClassificationc                    s>   t  | |j| _|| _t|dd| _t|| _|   d S NFr   )	r)   r*   
num_labelsr@   r   r   XmodClassificationHead
classifierr   r>   rA   rC   rD   r*   i  s   
z&XmodForSequenceClassification.__init__NrJ   r   ro   r&   r#   rp   rK   r  rt   r   r   rj   c                 C   sj  |dur|n| j j}| j||||||||	|
|d
}|d }| |}d}|dur| j jdu rR| jdkr8d| j _n| jdkrN|jtjksI|jtj	krNd| j _nd| j _| j jdkrpt
 }| jdkrj|| | }n+|||}n%| j jdkrt }||d| j|d}n| j jdkrt }|||}|s|f|d	d  }|dur|f| S |S t|||j|jd
S )a  
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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).
        N	r   ro   r&   r#   rp   rK   rt   r   r   r   r   Z
regressionZsingle_label_classificationZmulti_label_classificationr$   r\   r  )r@   r   r   r  Zproblem_typer  r(   r8   r=   r_   r	   squeezer   rk   r   r   rn   r   r?   rJ   r   ro   r&   r#   rp   rK   r  rt   r   r   r   r   r  r  r  r   rC   rC   rD   rR   t  sV   


"


z%XmodForSequenceClassification.forwardNNNNNNNNNNN)rS   rT   rU   r*   r   r   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r  a  sN    	
r  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j ef fddZ  ZS )XmodForMultipleChoicec                    s@   t  | t|| _t|j| _t|j	d| _
|   d S )Nr   )r)   r*   r   r   r   r4   r5   r6   rb   r-   r  r   r>   rA   rC   rD   r*     s
   
zXmodForMultipleChoice.__init__NrJ   r   r&   ro   r  r#   rp   rK   rt   r   r   rj   c                 C   s  |dur|n| j j}|dur|jd n|jd }|dur%|d|dnd}|dur8||d|d nd}|durG|d|dnd}|durV|d|dnd}|dure|d|dnd}|durx|d|d|dnd}| j||||||||	|
|d
}|d }| |}| |}|d|}d}|durt	 }|||}|s|f|dd  }|dur|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)
        lang_ids (`torch.LongTensor` of shape `(batch_size, num_choices, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        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)
        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)
        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.
        Nr   r$   r   rw   )	r   r#   r&   ro   rp   rK   rt   r   r   r\   r  )r@   r   rz   rk   r<   repeatr   r6   r  r   r   rn   r   )r?   rJ   r   r&   ro   r  r#   rp   rK   rt   r   r   Znum_choicesZflat_input_idsZflat_lang_idsZflat_position_idsZflat_token_type_idsZflat_attention_maskZflat_inputs_embedsr   r   r  Zreshaped_logitsr  r  r   rC   rC   rD   rR     sP   0&


zXmodForMultipleChoice.forwardr!  )rS   rT   rU   r*   r   r   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r"    sN    
	
r"  c                       r  )XmodForTokenClassificationc                    sb   t  | |j| _t|dd| _|jd ur|jn|j}t|| _	t
|j|j| _|   d S r  )r)   r*   r  r   r   classifier_dropoutr5   r   r4   r6   rb   r-   r  r   r?   r@   r%  rA   rC   rD   r*   1  s   z#XmodForTokenClassification.__init__NrJ   r   ro   r&   r#   rp   rK   r  rt   r   r   rj   c                 C   s   |dur|n| j j}| j||||||||	|
|d
}|d }| |}| |}d}|dur=t }||d| j|d}|sS|f|dd  }|durQ|f| S |S t|||j	|j
dS )a  
        lang_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Indices of the language adapters that should be activated for each sample, respectively. Default: the index
            that corresponds to `self.config.default_language`.
        labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
            Labels for computing the token classification loss. Indices should be in `[0, ..., config.num_labels - 1]`.
        Nr  r   r$   r\   r  )r@   r   r   r6   r  r   rk   r  r   rn   r   r   rC   rC   rD   rR   ?  s:   

z"XmodForTokenClassification.forwardr!  )rS   rT   rU   r*   r   r   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r$  .  sN    	
r$  c                       s(   e Zd ZdZ fddZdd Z  ZS )r  z-Head for sentence-level classification tasks.c                    sT   t    t|j|j| _|jd ur|jn|j}t|| _	t|j|j
| _d S r   )r)   r*   r   rb   r-   r   r%  r5   r4   r6   r  out_projr&  rA   rC   rD   r*   ~  s   
zXmodClassificationHead.__init__c                 K   sL   |d d dd d f }|  |}| |}t|}|  |}| |}|S r   )r6   r   r8   tanhr'  r  rC   rC   rD   rR     s   




zXmodClassificationHead.forward)rS   rT   rU   rV   r*   rR   rW   rC   rC   rA   rD   r  {  s    	r  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j dee	 dee	 dee	 de
eej ef fddZ  ZS )XmodForQuestionAnsweringc                    s@   t  | |j| _t|dd| _t|j|j| _| 	  d S r  )
r)   r*   r  r   r   r   rb   r-   
qa_outputsr   r>   rA   rC   rD   r*     s
   z!XmodForQuestionAnswering.__init__NrJ   r   ro   r&   r#   rp   rK   start_positionsend_positionsrt   r   r   rj   c                 C   sJ  |dur|n| j j}| j||||||||
||d
}|d }| |}|jddd\}}|d }|d }d}|dur|	durt| dkrP|d}t|	 dkr]|	d}	|d}|	d|}|		d|}	t
|d}|||}|||	}|| d }|s||f|dd  }|dur|f| S |S t||||j|jd	S )
r   Nr  r   r   r$   ru   )Zignore_indexr\   )r  start_logits
end_logitsrn   r   )r@   r   r   r*  r   r  r~   r   r<   clampr   r   rn   r   )r?   rJ   r   ro   r&   r#   rp   rK   r+  r,  rt   r   r   r   r   r  r-  r.  Z
total_lossZignored_indexr  Z
start_lossZend_lossr   rC   rC   rD   rR     sR   






z XmodForQuestionAnswering.forward)NNNNNNNNNNNN)rS   rT   rU   r*   r   r   r8   r   r   r   r   r   r   r   rR   rW   rC   rC   rA   rD   r)    sT    
	
r)  c                 C   s6   |  | }tj|dd|| | }| | S )a  
    Replace non-padding symbols with their position numbers. Position numbers begin at padding_idx+1. Padding symbols
    are ignored. This is modified from fairseq's `utils.make_positions`.

    Args:
        x: torch.Tensor x:

    Returns: torch.Tensor
    r   ru   )ner_   r8   ZcumsumZtype_asr=   )rJ   r   rL   maskZincremental_indicesrC   rC   rD   rG     s   rG   )r   r  r"  r)  r  r$  r   r   )r   )ArV   r|   typingr   r   r   r   r8   Ztorch.utils.checkpointr   Ztorch.nnr   r   r	   Zactivationsr   r   Z
generationr   Zmodeling_outputsr   r   r   r   r   r   r   r   Zmodeling_utilsr   Zpytorch_utilsr   r   r   utilsr   r   Zconfiguration_xmodr   Z
get_loggerrS   r   Moduler   rX   r   r   r   r   r   r   r   r   r   r   r   r  r   r  r"  r$  r  r)  rG   __all__rC   rC   rC   rD   <module>   sj   (

Z 92_f6 "~ZYmL
V