gq>dZddlZddlZddlmZmZmZmZddlZddlm Z ddl m Z m Z m Z ddlmZddlmZmZmZmZdd lmZdd lmZdd lmZmZmZmZmZmZm Z dd l!m"Z"dd l#m$Z$m%Z%m&Z&m'Z'm(Z(m)Z)m*Z*m+Z+ddl,m-Z-e*j.e/Z0dZ1dZ2Gdde j3Z4Gdde j3Z5Gdde j3Z6Gdde j3Z7Gdde j3Z8Gdde j3Z9Gdde j3Z:Gd d!e j3Z;Gd"d#e j3Z<Gd$d%e"Z=Gd&d'e=Z>d(Z?d)Z@d*ZAe&d+e?Gd,d-e=ZBe&d.e?Gd/d0e=eZCe&d1e?Gd2d3e=ZDe&d4e?Gd5d6e=ZEe&d7e?Gd8d9e=ZFe&d:e?Gd;dcttjt j||_||_dS)ze Construct a layernorm module in the UMT5 style. No bias and no subtraction of mean. N)super__init__r Parametertorchonesweightvariance_epsilon)self hidden_sizeeps __class__s b/var/www/html/ai-engine/env/lib/python3.11/site-packages/transformers/models/umt5/modeling_umt5.pyr*zUMT5LayerNorm.__init__<sD l5:k#:#:;; #ch|tjddd}|tj||jzz}|jjtj tj fvr||jj}|j|zS)NT)keepdim) tor,float32powmeanrsqrtr/r.dtypefloat16bfloat16)r0 hidden_statesvariances r4forwardzUMT5LayerNorm.forwardDs !##EM2266q99>>r4>PP% Ht?T4T(U(UU  ;  ? ? ?),,T[->??M{]**r5)r')__name__ __module__ __qualname__r*rD __classcell__r3s@r4r&r&;sL$$$$$$ + + + + + + +r5r&c*eZdZdeffd ZdZxZS)UMT5DenseActDenseconfigcJttj|j|jd|_tj|j|jd|_tj|j |_ t|j |_ dSNFbias)r)r*rLineard_modeld_ffwiwoDropout dropout_ratedropoutr dense_act_fnactr0rLr3s r4r*zUMT5DenseActDense.__init__Vsx )FNFKeDDD)FKeDDDz&"566 &-.r5c||}||}||}t|jjt jr]|j|jjjkrC|jjjt j kr$| |jjj}||}|SN) rTrZrX isinstancerUr.r,Tensorr?int8r:r0rBs r4rDzUMT5DenseActDense.forward]s .. //  ]33 tw~u| 4 4 C#tw~';;;$ 22),,TW^-ABBM .. r5rErFrGr$r*rDrHrIs@r4rKrKUsS/z//////       r5rKc*eZdZdeffd ZdZxZS)UMT5DenseGatedActDenserLcttj|j|jd|_tj|j|jd|_tj|j|jd|_tj |j |_ t|j |_dSrN)r)r*rrQrRrSwi_0wi_1rUrVrWrXr rYrZr[s r4r*zUMT5DenseGatedActDense.__init__ms IfnfkFFF IfnfkFFF )FKeDDDz&"566 &-.r5c|||}||}||z}||}t |jjtjr]|j |jjj krC|jjj tj kr$| |jjj }||}|Sr]) rZrfrgrXr^rUr.r,r_r?r`r:)r0rB hidden_gelu hidden_linears r4rDzUMT5DenseGatedActDense.forwardushhtyy7788  -00 #m3  ]33 tw~u| 4 4 C#tw~';;;$ 22),,TW^-ABBM .. r5rbrIs@r4rdrdlsS/z//////r5rdc*eZdZdeffd ZdZxZS) UMT5LayerFFrLc$t|jrt||_nt ||_t |j|j|_ tj |j |_ dS)Nr2)r)r* is_gated_actrdDenseReluDenserKr&rRlayer_norm_epsilon layer_normrrVrWrXr[s r4r*zUMT5LayerFF.__init__sx    <"8"@"@D  "3F";";D 'F+B,,,   4<eDDD4<eDDD4<eDDD4>4>FFq!QPQRRr5c~d}|j}|j}|jsC|dz}||dktj|zz }t j|}n(t j|t j| }|dz}||k}t j | |z tj ||z z }|||z z}||tjz}t j|t j ||dz }|t j |||z }|S)a Adapted from Mesh Tensorflow: https://github.com/tensorflow/mesh/blob/0cb87fe07da627bf0b7e60475d59f95ed6b5be3d/mesh_tensorflow/transformer/transformer_layers.py#L593 Translate relative position to a bucket number for relative attention. The relative position is defined as memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for small absolute relative_position and larger buckets for larger absolute relative_positions. All relative positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket. This should allow for more graceful generalization to longer sequences than the model has been trained on Args: relative_position: an int32 Tensor bidirectional: a boolean - whether the attention is bidirectional num_buckets: an integer max_distance: an integer Returns: a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets) rr7r#)r{r|ryr:r,longabsmin zeros_likelogfloatmath full_likewhere) r0relative_positionrelative_buckets num_buckets max_distance max_exactis_small log_ratiorelative_position_if_larges r4_relative_position_bucketz'UMT5Attention._relative_position_bucketsP*9 ;  c A K !2Q!6 : :5: F F T T  % *; < <  !&+NO`>a>a!b!b b  1$ $y0I/5577)CDDtxP\_hPhGiGii y!89 %.ej1I1I%I"%*Y &8RT_bcTc(d(d& & " EK2CE_```r5c||jjj}|,tj|tj|dddf}n |dddf}tj|tj|dddf}||z }||}||} | gdd} | S)z%Compute binned relative position biasN)r?device)r7rr#r) rr.rr,arangerrr unsqueeze) r0 query_length key_lengthrcache_positioncontext_positionmemory_positionrrelative_position_bucketvaluess r4 compute_biaszUMT5Attention.compute_biass >18?F  !$|L SYZZZ[\[\[\^b[bc  -aaag6 ,zFSSSTXZ[Z[Z[T[\+.>>#'#A#ABS#T#T --.FGG **44Q77 r5rBencoder_hidden_statespast_key_valueattention_masklayer_head_maskrcx|jdd\}}|du} ||} | |d|j|jdd} |0|j|j} | r|j } n|j } | r|n|} | r)|'| r%| j |j}| j |j}n| | }|| }||d|j|jdd}||d|j|jdd}|9| s|nd}| |||jd|i\}}| rd|j|j<t!j| |dd}|||zn|}|jd}|js+t!jd|j||f|j|j}n3||||j| }|dddd| dddf}|$|ddddddd|jdf}||z}|jrUt!j|jd}d |t5|j<|dd|f}n|}||z }t8j|d  |}t8j!||j!|j" }|||z}t!j||}|dd#}|||d}|$|}|||fS) Nr7r8r#rTr )rr?)rrrdim)ptraining)%shaperrrr~ transpose is_updatedgetrwcross_attention_cacheself_attention_cache key_cache value_cacherrupdater,matmulget_seq_lengthrzzerosrr?rrr-listboolr functionalsoftmaxrtype_asrXr contiguousr)r0rBrrrrr batch_size seq_lengthis_cross_attention query_statesrcurr_past_key_valuecurrent_states key_states value_statesscoresreal_seq_lengthr position_bias causal_maskmaskposition_bias_masked attn_weights attn_outputs r4rDzUMT5Attention.forwards"/!4RaR!8 J3$>vvm,, #((RtG^__iijkmnoo  %'266t~FFJ! J&4&J##&4&I#2DW..-  E."<"<,6t~FJ.:4>JLL//J66.11L#RtG^__iijkmnooJ',,ZT\4KbccmmnoqrssL)7I!St+>+E+E dn?OQ_>`,,( L&E@DN-dn=lJ,@,@A,F,FGGKYJd*~'D'D'F'FFFjt%b) / A!KDL*j9&-W]WcMM!--FMR`.M*!!!QQQ aaa*?@M  %(AAAqqq2HJ4DR4H2H)HIK)K7M   1:m1!455D,-Dd'(( )#0DIIKK#@ #0 &&},,V\\^^,DDLLVTT },,\T\TXTa,bb   &'/9Ll<>> !++Aq11<<>> !&&z:rBB ff[)) L.88r5)FN)NNNNNNN)rErFrG__doc__rintr*r,r_rrrrrDrHrIs@r4rvrvs?""XVY]"""""":%, - - - ^    $9=8<152615Y9Y9|Y9 ( 5Y9!u|!45 Y9 !. Y9 "%,/ Y9!.Y9Y9Y9Y9Y9Y9Y9Y9r5rvcBeZdZddeeffd Z ddZxZS)UMT5LayerSelfAttentionNrwctt|d||_t |j|j|_tj |j |_ dS)NTrzrwrn) r)r*rv SelfAttentionr&rRrqrrrrVrWrXr0rLrwr3s r4r*zUMT5LayerSelfAttention.__init___sb *6t_hiii'F >Gr5r]) NNNNNNFFNrrIs@r4rrs~//(3-//////"##'>>>>>>>>r5rcLeZdZdZdeffd ZdejdejfdZxZ S)UMT5ClassificationHeadz-Head for sentence-level classification tasks.rLcttj|j|j|_tj|j|_tj|j|j |_ dS)N)r) r)r*rrQrRdenserVclassifier_dropoutrX num_labelsout_projr[s r4r*zUMT5ClassificationHead.__init__sc Yv~v~>> zF$=>>>  &.&2CDD r5rBrc||}||}tj|}||}||}|Sr])rXrr,tanhrras r4rDzUMT5ClassificationHead.forwards[ ]33  =11  =11  ]33  m44 r5) rErFrGrr$r*r,r_rDrHrIs@r4r r sw77EzEEEEEE U\elr5r cTeZdZdZeZdZdZdZdZ dgZ dgZ e dZ dZdZd S) UMT5PreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. transformerTrrUcvtjt}tjt}|||d}|S)N)decoder_input_ids input_idsdecoder_attention_mask)r,tensorrr)r0r input_mask dummy_inputss r4rz UMT5PreTrainedModel.dummy_inputss=L.. \*-- !*"&0  r5c |jj}t|tr$|jj|dzd St|ttttfr|j jj d|dzt|dr5|jjs)|jjj d|dzt|dr[|jjj d||jjdzz|jjjd Sd St|t(r`t|drN|jjj d|dz|jjjd Sd St|t,r|jjj d||jjdzzt|jdr/|jj#|jjj|jjj d||jjdzzt|jdr1|jj'|jjjd Sd Sd St|t2r|jjj d||jjdzzt|jdr/|jj#|jjj|jjj d||jjdzzt|jdr1|jj'|jjjd Sd Sd St|t:rt|jjj d||jjdzzt|jdr/|jj#|jjj|jjj d||jjdzzt|jdr/|jj#|jjj|jjj d||jjdzzt|jdr1|jj'|jjjd Sd Sd St|t@r|jj}|jj!}|jj"}|j#jj d|||zdzz|j$jj d||dzz|j%jj d||dzz|j&jj d|||zdzz|j'r0|j(jj d||dzzd Sd Sd S) zInitialize the weights?)r=stdlm_head qa_outputs classifierrPN))rLinitializer_factorr^r&r.datafill_ UMT5ModelUMT5ForConditionalGenerationUMT5EncoderModelUMT5ForQuestionAnsweringsharednormal_hasattrtie_word_embeddingsr#r$rRrPzero_UMT5ForTokenClassificationr&r rrrKrTrUrSrdrfrgrvr}rrrrrrzr)r0modulefactorrRr~rs r4 _init_weightsz!UMT5PreTrainedModel._init_weightssY/ fm , ,= o M  $ $Vc\ 2 2 2 2 2  , (    ; o M % - -3FSL - I I Ivy)) O$+2Q O%*22#2NNNv|,, 4!(-553Ft{ObgkNkDl5mmm!&+1133333 4 4 : ; ;* ov|,, 4!(-553FSL5QQQ!&+1133333 4 4 6 7 7& o L  $ , ,#6dkFY^bEb;c , d d dv|V,, /1B1N !&,,... O " ' / /SfI\aeHe>f / g g gv// 2FO4H4T$)//11111 2 24T4T  1 2 2 o I  ! ) )s4;CV[_B_8` ) a a avy&)) ,fin.H #))+++ I  ! ) )s4;CSX\B\8] ) ^ ^ ^vy&)) ,fin.H #))+++++ , ,.H.H  6 7 7 o K  # + +&T[EX]aDa:b + c c cv{F++ . 0@0L  %++--- K  # + +&T[EX]aDa:b + c c cv{F++ . 0@0L  %++--- I  ! ) )s4;CSX\B\8] ) ^ ^ ^vy&)) ,fin.H #))+++++ , ,.H.H  . . ok)G!%!1 k+G HO ( (cv'L^B^cgAg7h ( i i i HO ( (cv$7O ( P P P HO ( (cv$7O ( P P P HO ( (cv'L^B^cgAg7h ( i i i1 o.5:BBQW\chl[lQmBnnnnn o o o or5c|jj}|jj}|tdt |rHt j|jdddz|}t j||dddfgd}nD| |j}|dddf |dddf<||d<|td| |d k||S) Nzself.model.config.decoder_start_token_id has to be defined. In UMT5 it is usually set to the pad_token_id. See UMT5 docs for more information.r8)r#.rr#).rz1self.model.config.pad_token_id has to be defined.) rLdecoder_start_token_id pad_token_id ValueErrorrr,fullrcat new_zerosclone masked_fill_)r0rr9r:shifted_input_idss r4 _shift_rightz UMT5PreTrainedModel._shift_rightRs!%!C{/ ! )6  Y ' ' ? % 9?3B3+?$+FH^ _ _  % +Q*RXZ [ [ [   ) 3 3IO D D )238)<)B)B)D)D c122g &(> f %  PQQ Q&&'8D'@,OOO  r5N)rErFrGrr$ config_classbase_model_prefixsupports_gradient_checkpointing_supports_cache_class_supports_static_cache_no_split_modules_keep_in_fp32_modulespropertyrr6rBr5r4rrs L%&*# !$ !F X@o@o@oD!!!!!r5rceZdZdfd ZdZdZ ddZdejdejdejd e d e f d Z e dejd e d e dejdejdejde fdZxZS) UMT5StackNct||_j|_t jfdt jD|_tj j |_ t j j|_d|_|dS)Nc2g|]}t|S)r)r).0irLs r4 z&UMT5Stack.__init__..ss&#e#e#eqIf$B$B$B#e#e#er5rnF)r)r* embed_tokensryrrrange num_layersblockr&rRrqfinal_layer_normrVrWrXgradient_checkpointing post_init)r0rLrSr3s ` r4r*zUMT5Stack.__init__os    ( +]#e#e#e#eERXRcLdLd#e#e#eff -fn&B[ \ \ \z&"566 ',# r5c|jSr]rSr0s r4get_input_embeddingszUMT5Stack.get_input_embeddings{s   r5c||_dSr]r[r0new_embeddingss r4set_input_embeddingszUMT5Stack.set_input_embeddings~s*r5c | | n |jj} | | n |jj} | | n |jj} | | n |jj} |#|!|jrdnd}t d|d|d|1|}|d|d}n@||dd}n!|jrdnd}t d|d|d |j r%|j r| rt d d } |+|j t d | |}|\}}| d ur|jst d|dd }d }|jr| s|t|tr4t|t sd }t!|t#}nzt|t s1d }t dt!j|}n4|(t!t#t#}n |jsd}||nd}| t)j|||z|j} |/t/s!||z}t)j|||j}|jr#|||| ||jnd| }nT|P|ddddddf}||j}d|z t)j|jjz}nd}|jrQ|O|\}}}||f}|t)j||j}||}nd}| ||jj!}| ||jj!}| rdnd}| rdnd}| r |jrdnd}|"|}tG|j$D]\} }!|| }"|| }#| r||fz}|j r,|j r%|%|!j&|||||"|#d| | | }$n|!|||||"|#|| | |  }$|$d}| r|$d}%| r||$dfz }|jr ||$dfz }|'|}|"|}| r||fz}| r|%nd}&|r|j}&|r|(}&| stSd||&|||fDStU||&|||S)Ndecoder_zYou cannot specify both zinput_ids and zinputs_embeds at the same timer8zYou have to specify either z input_ids or inputs_embedszZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...Fzz$UMT5Stack.forward..7s4  =!===  r5)last_hidden_statepast_key_valuesrB attentionscross_attentions)+rLrroutput_hidden_statesuse_return_dictryr;rrrXrrrrSr^r rrfrom_legacy_cacherr,rrr r-_update_causal_maskrr:r?rrinvert_attention_mask get_head_maskrUrX enumeraterV_gradient_checkpointing_funcrDrWto_legacy_cachetupler)'r0rrrrre head_maskcross_attn_head_maskrjrrrm return_dictrerr_msg_prefix input_shaperrreturn_legacy_cachereturn_self_attention_cachepast_key_values_lengthmask_seq_lengthrencoder_batch_sizeencoder_sequence_length_encoder_hidden_shapeencoder_extended_attention_maskall_hidden_statesall_attentionsall_cross_attentionsrBrQ layer_modulerr layer_outputsnext_decoder_cache next_caches' r4rDzUMT5Stack.forwards "+!6IIDK+/?BZZNw>wwwww  "#..**K!r;r?;;II  &',,..ss3KK+/?BZZNu>uuXfuuuvv v  & "4= " "##p"   ( !_``` --i88M!, J   ? j !hT!h!h!hiii$&+# ? #  #_-H/511 V*_Vi:j:j V.2+"5o|~~"V"V1DEE V&*###` #6"G"X"X ("5lnnlnn"U"U ##OETE`!?!?!A!A!Afg  !"\&(>(KTaThN  !*B*D*D !4zAO"Z OML`aaaN ? 228G8S44Y]! KK '(D$)9:K%..}/B.CCK, M''$ & >(=J  ;(88::J   "%"(       9+&+%1     r5r input_tensorrrjrc |jjdkr |d|vr|SdS||nd}t|t}|jjdkr#|s!|st j||||jrdS|j|j } }|j d} |r| } n/t|tj r |j dn|| zdz} ||| | || ||j d} |jjdkrB|@|j jd kr0|s.tj|j} t j| | } | S) Nflash_attention_2r!rsdpa)rer~ is_trainingr#r8)sequence_length target_lengthr?rrrcuda)rL_attn_implementationrr^rr_ignore_causal_mask_sdparr?rrget_max_cache_shaper,r_5_prepare_4d_causal_attention_mask_with_cache_positiontyperr_unmask_unattended)r0rrrrjrpast_seen_tokensusing_static_cacher?rrrr min_dtypes r4rpzUMT5Stack._update_causal_maskKs ; +/B B B)c^.C.C%%4 @O?Z?99;;;`a'EE ; +v 5 5>P 5Yj 5%>*'7 M   t$*L,?v&,Q/  +??AAMMnel;;<$R((%7!; PP +')#)!,Q   K , 6 6*%*f44%5  E**.I0CKQZ[[Kr5rrr?rrc ||dkr|}n+tj|j} tj||f| ||}|dkrtj|d}|tj|||ddkz}|ddddddf|ddd}|| }|j d} |ddddddd| f|ddddddfz} | dk} |ddddddd| f | | |ddddddd| f<|S) a Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. Args: attention_mask (`torch.Tensor`): A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`. sequence_length (`int`): The sequence length being processed. target_length (`int`): The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet. dtype (`torch.dtype`): The dtype to use for the 4D attention mask. device (`torch.device`): The device to plcae the 4D attention mask on. cache_position (`torch.Tensor`): Indices depicting the position of the input sequence tokens in the sequence. batch_size (`torch.Tensor`): Batch size. N) fill_valuer?rr#)diagonalrfr8r) rr,rrr<triurreshapeexpandr?r masked_fill) rrrr?rrrkwargsrr mask_length padding_masks r4rz?UMT5Stack._prepare_4d_causal_attention_mask_with_cache_positionsD  %.*<*<*>*>!*C*C(KK E**.I* -0Ye\bK!###jqAAA 5< fEEEH^H^_acdHeHee eK%dD!!!QQQ&67>>z1bRTUUK))//11 ,226 *111aaaL[L+@ANSTSTSTVZ\`bcbcbcScDdd +q0 5@AAAqqq,;,AV5W5c5c )66 AAAqqq!!!\k\12r5r]) NNNNNNNNNNNNN)rErFrGr*r]rarDr,r_r rrp staticmethodrr?rrrHrIs@r4rMrMnsR      !!!+++ "#!!G G G G T? ?l? ?  ?  ????B5 555{ 5  5  5555\55555r5rMa The UMT5 model was proposed in [Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer](https://arxiv.org/abs/1910.10683) by Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, Peter J. Liu. It's an encoder decoder transformer pre-trained in a text-to-text denoising generative setting. This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads etc.) This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage and behavior. Parameters: config ([`UMT5Config`]): Model configuration class with all the parameters of the model. Initializing with a config file does not load the weights associated with the model, only the configuration. Check out the [`~PreTrainedModel.from_pretrained`] method to load the model weights. a Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. UMT5 is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. [What are input IDs?](../glossary#input-ids) To know more on how to prepare `input_ids` for pretraining take a look a [UMT5 Training](./umt5#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) UMT5 uses the `pad_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). To know more on how to prepare `decoder_input_ids` for pretraining take a look at [UMT5 Training](./umt5#training). decoder_attention_mask (`torch.BoolTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the encoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. decoder_head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. cross_attn_head_mask (`torch.Tensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. encoder_outputs (`tuple(tuple(torch.FloatTensor)`, *optional*): Tuple consists of (`last_hidden_state`, `optional`: *hidden_states*, `optional`: *attentions*) `last_hidden_state` of shape `(batch_size, sequence_length, hidden_size)` is a sequence of hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. past_key_values (`tuple(tuple(torch.FloatTensor))` of length `config.n_layers` with each tuple having 4 tensors of shape `(batch_size, num_heads, sequence_length - 1, embed_size_per_head)`): Contains precomputed key and value hidden states of the attention blocks. Can be used to speed up decoding. If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `decoder_input_ids` of shape `(batch_size, sequence_length)`. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, 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. decoder_inputs_embeds (`torch.FloatTensor` of shape `(batch_size, target_sequence_length, hidden_size)`, *optional*): Optionally, instead of passing `decoder_input_ids` you can choose to directly pass an embedded representation. If `past_key_values` is used, optionally only the last `decoder_inputs_embeds` have to be input (see `past_key_values`). This is useful if you want more control over how to convert `decoder_input_ids` indices into associated vectors than the model's internal embedding lookup matrix. If `decoder_input_ids` and `decoder_inputs_embeds` are both unset, `decoder_inputs_embeds` takes the value of `inputs_embeds`. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): Indices depicting the position of the input sequence tokens in the sequence. It is used to update the cache in the correct position and to infer the complete sequence length. a Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. UMT5 is a model with relative position embeddings so you should be able to pad the inputs on both the right and the left. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for detail. To know more on how to prepare `input_ids` for pretraining take a look a [UMT5 Training](./umt5#training). attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. [What are attention masks?](../glossary#attention-mask) head_mask (`torch.FloatTensor` of shape `(num_heads,)` or `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the self-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`torch.FloatTensor` of shape `(batch_size, 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. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. z^The bare UMT5 Model transformer outputting raw hidden-states without any specific head on top.c'eZdZdZdZeZddgZfdZdZ dZ dZ d Z d Z d Zeeeee  d deejdeejdeejdeejdeejdeejdeejdeeeejdeeeejdeejdeejdeedeedeedeedeejdeeejeff"dZxZS)!r*ao Examples: ```python >>> from transformers import UMT5Model, AutoTokenizer >>> model = UMT5Model.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> noisy_text = "UN Offizier sagt, dass weiter werden muss in Syrien." >>> label = " verhandelt" >>> inputs = tokenizer(inputs, return_tensors="pt") >>> labels = tokenizer(label=label, return_tensors="pt") >>> outputs = model(input_ids=inputs["input_ids"], decoder_input_ids=labels["input_ids"]) >>> hidden_states = outputs.last_hidden_state ```umt5encoder.embed_tokens.weightdecoder.embed_tokens.weightct|tj|j|j|_tj|}d|_ d|_ d|_ t||j|_ tj|}d|_ d|_ |j|_t||j|_|dSNFT)r)r*rr vocab_sizerRr.copydeepcopyryris_encoder_decoderrMencodernum_decoder_layersrUdecoderrYr0rLencoder_configdecoder_configr3s r4r*zUMT5Model.__init__{s    l6#4fnEE v..$)!#( ,1) == v..$(!,1)$*$=! ==  r5c|jSr]r.r\s r4r]zUMT5Model.get_input_embeddings {r5c|||_|j||j|dSr]r.rrarr_s r4razUMT5Model.set_input_embeddings;$  )).999 )).99999r5c|jjrL||jj|j||jj|jdSdSr]rLr1_tie_or_clone_weightsrrSr.rr\s r4 _tie_weightszUMT5Model._tie_weights\ ; * O  & &t|'@$+ N N N  & &t|'@$+ N N N N N O Or5c|jSr]rr\s r4 get_encoderzUMT5Model.get_encoder |r5c|jSr]rr\s r4 get_decoderzUMT5Model.get_decoderrr5c|D]/\}}|jj|j|0dS) 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)itemsrr attention prune_headsr0heads_to_prunerheadss r4 _prune_headszUMT5Model._prune_headssU +0022 C CLE5 L u % / ; ;E B B B B C Cr5 output_typerCNrrrrrwdecoder_head_maskrxencoder_outputsrjredecoder_inputs_embedsrrrmryrrc*| | n |jj} ||n |jj}||||| || ||}ne|rct |t sNt |dt |dkr|dndt |dkr|dnd}|d}|||| | ||||| | ||| }|s||zSt|j |j |j |j |j |j |j |j S) a Returns: Example: ```python >>> from transformers import AutoTokenizer, UMT5Model >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5Model.from_pretrained("google/umt5-small") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> decoder_input_ids = tokenizer("Studies show that", return_tensors="pt").input_ids # Batch size 1 >>> # preprocess: Prepend decoder_input_ids with start token which is pad token for UMT5Model. >>> # This is not needed for torch's UMT5ForConditionalGeneration as it does this internally using labels arg. >>> decoder_input_ids = model._shift_right(decoder_input_ids) >>> # forward pass >>> outputs = model(input_ids=input_ids, decoder_input_ids=decoder_input_ids) >>> last_hidden_states = outputs.last_hidden_state ```Nrrrerwrrmryrr#r7rirBrk rrrerjrrrwrxrrrmryr)rirjdecoder_hidden_statesdecoder_attentionsrlencoder_last_hidden_staterencoder_attentions)rLrrnrr^rlenrrrirjrBrkrl)r0rrrrrwrrxrrjrerrrrmryrrBdecoder_outputss r4rDzUMT5Model.forwardsZ"+!6IIDK_ $D>_ 'tn_ d^!_ "!!12#_ $ uU&');; <%_ _ _ ]\BA_ _ _ _ _ r5r*z2UMT5 Model with a `language modeling` head on top.c)eZdZdZdZgdZfdZdZdZdZ dZ d Z d Z d Z eeeee  d#deejdeejdeejdeejdeejdeejdeejdeeeejdeeeejdeejdeejdeejdeedeedeedeedeejdeeejeff$d Zdejfd!Zed"ZxZ S)$r+a Examples: ```python >>> from transformers import UMT5ForConditionalGeneration, AutoTokenizer >>> model = UMT5ForConditionalGeneration.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien." >>> summary = "Weiter Verhandlung in Syrien." >>> inputs = tokenizer(article, text_target=summary, return_tensors="pt") >>> outputs = model(**inputs) >>> loss = outputs.loss ```r)rrzlm_head.weightc4t||j|_t j|j|j|_tj |}d|_ d|_ d|_ t||j|_tj |}d|_ d|_ |j|_t||j|_t j|j|jd|_|dS)NFTrO)r)r*rR model_dimrrrr.rrryrrrMrrrUrrQr#rYrs r4r*z%UMT5ForConditionalGeneration.__init__(s    l6#4fnEE v..$)!#( ,1) == v..$(!,1)$*$=! == y1BOOO  r5c|jSr]rr\s r4r]z1UMT5ForConditionalGeneration.get_input_embeddings@rr5c|||_|j||j|dSr]rr_s r4raz1UMT5ForConditionalGeneration.set_input_embeddingsDrr5c|jjrL||jj|j||jj|jdSdSr]rr\s r4rz)UMT5ForConditionalGeneration._tie_weightsJrr5c||_dSr]r#r_s r4set_output_embeddingsz2UMT5ForConditionalGeneration.set_output_embeddingsPs % r5c|jSr]rr\s r4get_output_embeddingsz2UMT5ForConditionalGeneration.get_output_embeddingsTrr5c|jSr]rr\s r4rz(UMT5ForConditionalGeneration.get_encoderXrr5c|jSr]rr\s r4rz(UMT5ForConditionalGeneration.get_decoder\rr5rNrrrrrwrrxrrjrerlabelsrrrmryrrc| | n |jj} ||n |jj}||||| ||||}ne|rct |t sNt |dt |dkr|dndt |dkr|dnd}|d}| || || }|||| | ||||| |||| }|d}|jj r ||j dzz}| |}d}| ptd }| |j} ||d |d | d }|s|f|ddz|z}||f|zn|St#|||j|j|j|j|j|j|j S) a labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[-100, 0, ..., config.vocab_size - 1]`. All labels set to `-100` are ignored (masked), the loss is only computed for labels in `[0, ..., config.vocab_size]` Returns: Examples: ```python >>> from transformers import AutoTokenizer, UMT5ForConditionalGeneration >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5ForConditionalGeneration.from_pretrained("google/umt5-small") >>> # training >>> input_ids = tokenizer("The walks in park", return_tensors="pt").input_ids >>> labels = tokenizer(" cute dog the ", return_tensors="pt").input_ids >>> outputs = model(input_ids=input_ids, labels=labels) >>> loss = outputs.loss >>> logits = outputs.logits >>> # inference >>> input_ids = tokenizer("Studies have shown that good for you", return_tensors="pt").input_ids >>> outputs = model.generate(input_ids) >>> tokenizer.decode(outputs[0], skip_special_tokens=True) ```Nrrr#r7rrr%r8 ignore_indexr8 losslogitsrjrrrlrrr)rLrrnrr^rrrBrr1rr#r r:rrrrrjrBrkrlri)r0rrrrrwrrxrrjrerrrrrmryrrBrsequence_output lm_logitsrloss_fctoutputs r4rDz$UMT5ForConditionalGeneration.forward_sod"+!6IIDKUMT5ForConditionalGeneration._reorder_cache..sCnnU_j--aZ=N1O1OPPnnnnnnr5)rv)rjr reordered_past layer_pasts ` r4_reorder_cachez+UMT5ForConditionalGeneration._reorder_cachesQ)  J nnnncmnnnnn NNr5)NNNNNNNNNNNNNNNNN)!rErFrGrrrr*r]rarrrrrrrr"rrrr,rrrr_rrrrDrrr rHrIs@r4r+r+s  Jiii0::: OOO &&&+*+@AA?YYY156:8<=A159=7;@D@D59=A-1$(,0/3&*59%{ { E,-{ !!23{ $E$45 { !))9 : { E-. { $E$56{ 'u|4{ "%el(;"<={ "%el(;"<={   12{  ((9:{ )*{ D>{ $D>{ 'tn!{ "d^#{ $!!12%{ & uU&'8 9'{ { { ZYBA{ |)EL))))\r5r+zhThe bare UMT5 Model transformer outputting encoder's raw hidden-states without any specific head on top.cleZdZdZdZdgZfdZdZdZdZ dZ d Z e e eee  dd eejd eejdeejdeejdeedeedeedeeejeffdZxZS)r,a Examples: ```python >>> from transformers import UMT5EncoderModel, AutoTokenizer >>> model = UMT5EncoderModel.from_pretrained("google/umt5-small") >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> article = "UN Offizier sagt, dass weiter verhandelt werden muss in Syrien." >>> input_ids = tokenizer(article, return_tensors="pt").input_ids >>> outputs = model(input_ids) >>> hidden_state = outputs.last_hidden_state ```rrc2t|tj|j|j|_tj|}d|_ d|_ t||j|_ | dSNF)r)r*rrrrRr.rrrrrMrrY)r0rLrr3s r4r*zUMT5EncoderModel.__init__s}    l6#4fnEE v..#( ,1) ==  r5c|jSr]rr\s r4r]z%UMT5EncoderModel.get_input_embeddingsrr5cH||_|j|dSr])r.rrar_s r4raz%UMT5EncoderModel.set_input_embeddingss%$  )).99999r5cl|jjr'||jj|jdSdSr])rLr1rrrSr.r\s r4rzUMT5EncoderModel._tie_weightss? ; * O  & &t|'@$+ N N N N N O Or5c|jSr]rr\s r4rzUMT5EncoderModel.get_encoderrr5c|D]:\}}|jj|jdj|;dS)rrN)rrrVrrrrs r4rzUMT5EncoderModel._prune_heads"s] +0022 P PLE5 L u % +A . < H H O O O O P Pr5rNrrrwrerrmryrc ^||n |jj}||||||||}|S)a5 Returns: Example: ```python >>> from transformers import AutoTokenizer, UMT5EncoderModel >>> tokenizer = AutoTokenizer.from_pretrained("google/umt5-small") >>> model = UMT5EncoderModel.from_pretrained("google/umt5-small") >>> input_ids = tokenizer( ... "Studies have been shown that owning a dog is good for you", return_tensors="pt" ... ).input_ids # Batch size 1 >>> outputs = model(input_ids=input_ids) >>> last_hidden_states = outputs.last_hidden_state ```Nr)rLrnr) r0rrrwrerrmryrs r4rDzUMT5EncoderModel.forward*sL:&1% &'tn&d^& uU&'8 9&&&ZYJI&&&&&r5r,z UMT5 model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. c%eZdZdgZddgZdeffd Zeee e e  dde j d ee jd ee j d ee j d ee jd ee jdee jdeee jdee jdee jdee j deedeedeedeedeee ff dZxZS)UMT5ForSequenceClassificationFdecoder.block.0.layer.1.EncDecAttention.relative_attention_bias.weightrrrLct|t||_t ||_|d|_dSr)r)r*r*rr classification_headrYmodel_parallelr[s r4r*z&UMT5ForSequenceClassification.__init__bs[    $V,,#9&#A#A  #r5rNrrrrrwrrxrrerrrrrmryrcp||n |jj}| d} || td|jj|(| &|t d||}|||||||||| | | | ||}|d}||jj  |j }ttj|ddkrt d|j\}}}||ddf|d |ddd ddf}||}d}| | |j } |jjp|jjdkr d |j_nS|jjdkr7| jtjks| jtjkr d |j_n d |j_|jjd kr\t1}|jjdkr1||| }n||| }n|jjd krLt5}||d |jj| d }n*|jjd krt7}||| }|s|f|ddz}||f|zn|St9|||j|j|j|j |j!|j"|j# S)aD labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., config.num_labels - 1]`. If `config.num_labels > 1` a classification loss is computed (Cross-Entropy). Returns: NFz8Passing input embeddings is currently not supported for If no `decoder_input_ids` or `decoder_inputs_embeds` are passed, `input_ids` cannot be `None`. Please pass either `input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`.) rrrrwrrxrrerrrrmryrr#z7All examples must have the same number of tokens.r8 regressionsingle_label_classificationmulti_label_classificationr)$rLrnNotImplementedErrorr3rEr;rBreq eos_token_idr:rrr,unique_consecutivesumrrr problem_typerr?rrr squeezer rrrjrrrlrrr)r0rrrrrwrrxrrerrrrrmryrreos_maskrrr1sentence_representationrrrrs r4rDz%UMT5ForSequenceClassification.forwardls4&1%Kbdd   $)>)F  U !% 1 1) < < "" )/#9/!5+'"7/!5##   "!*<< 899<<_=STT u' Q88 9 9A = =VWW W%4%:" A{"1(AAA+">"C"CJPRT_"`"`abababdfhihihiai"j))*ABB  YYv}--F{'/;)Q../;DK,,[+a//V\UZ5O5OSYS_chclSlSl/LDK,,/KDK,{'<77"99;)Q..#8FNN$4$4fnn6F6FGGDD#8FF33DD)-JJJ+--x B 0F G GUWYY)-III,..x// FY,F)-)9TGf$$v E.#3")"?&9$5&-&G")"?&9    r5)NNNNNNNNNNNNNNN)rErFrG"_keys_to_ignore_on_load_unexpectedrr$r*rrr"rrr,rrr_rrrrrrDrHrIs@r4rrVs+s)s&79VW$z$$$$$$+*+@AA+JYhiii'+158<=A,0487;=A59=A-1$(,0/3&*!k k #k !.k $E$45 k !))9 : k EL) k $EL1k 'u|4k "$u'8"9:k   12k  ((9:k )*k D>k $D>k 'tnk d^!k " u55 6#k k k jiBAk k k k k r5rz UMT5 Encoder Model with a token classification head on top (a linear layer on top of the hidden-states output) e.g. for Named-Entity-Recognition (NER) tasks. cneZdZdgZdgZdeffd Zeee e e  dde e jde e jd e e jd e e jd e e jd e ed e ede edeee je ffdZxZS)r3rz'transformer.encoder.embed_tokens.weightrLc6t||j|_t||_t j|j|_t j |j |j|_ | dSr]) r)r*rr,rrrVrrXrQr1r&rYr[s r4r*z#UMT5ForTokenClassification.__init__sz     ++F33z&";<< )F$68IJJ r5rNrrrwrerrrmryrc ||n |jj}||||||||} | d} || } || } d} |Ft } | | d|j|d} |s| | ddf}| | f|zn|St| | | j | j S)z 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]`. Returns: N)rrwrerrmryrr8r7)rrrBrk) rLrnrrXr&r rrrrBrk)r0rrrwrerrrmryrrBrrrrs r4rDz"UMT5ForTokenClassification.forwards&&1%. 'tn. d^.  uU\"$99 :. . . `_BA. . . . . r5r3z UMT5 Model with a span classification head on top for extractive question-answering tasks like SQuAD (linear layers on top of the hidden-states output to compute `span start logits` and `span end logits`). c'peZdZddgZfdZdZdZdZdZdZ e e e e e  dd eejd eejd eejdeejdeejdeejdeejdeeeejdeejdeejdeejdeejdeedeedeedeedeeeje ff"dZxZS)r-rrcHt||j|_t j|j|j|_tj |}d|_ d|_ d|_ t||j|_tj |}d|_ d|_ |j|_t||j|_|j|_t j|j|j|_|dSr)r)r*rRrrrrr.rrryrrrMrrrUrrrQr$rYrs r4r*z!UMT5ForQuestionAnswering.__init__1s    l6#4fnEE v..$)!#( ,1) == v..$(!,1)$*$=! ==  +)FNF4EFF r5c|jSr]rr\s r4r]z-UMT5ForQuestionAnswering.get_input_embeddingsJrr5c|||_|j||j|dSr]rr_s r4raz-UMT5ForQuestionAnswering.set_input_embeddingsNrr5c|jjrL||jj|j||jj|jdSdSr]rr\s r4rz%UMT5ForQuestionAnswering._tie_weightsTrr5c|jSr]rr\s r4rz$UMT5ForQuestionAnswering.get_encoderZrr5c|jSr]rr\s r4rz$UMT5ForQuestionAnswering.get_decoder^rr5rNrrrrrwrrxrstart_positions end_positionsrerrrrmryrc4||n |jj}| | n |jj} | | d} |(| &|td||}| | n |jj} ||n |jj}||||| ||||}ne|rct |tsNt|dt|dkr|dndt|dkr|dnd}|d}| ||| d||||| ||| }|d}| |}| dd \}}| d  }| d  }d}| | t| dkr-| d |j} t| dkr-| d |j} |d}| d|} | d|} t%| }||| }||| }||zdz }|s||f|ddz|z}||f|zn|St'||||j|j|j|j|j|j|j S)a start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the start of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence are not taken into account for computing the loss. end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): Labels for position (index) of the end of the labelled span for computing the token classification loss. Positions are clamped to the length of the sequence (*sequence_length*). Position outside of the sequence are not taken into account for computing the loss. Returns: NFrrrr#r7r) rrrerjrrrwrxrrrmryr8rr) r start_logits end_logitsrjrrrlrrr)rLrnrr;rBrr^rrrr$splitr(rrr:rrr rrjrBrkrlri)r0rrrrrwrrxrr6r7rerrrrmryrBrrrr9r: total_loss ignored_indexr start_lossend_lossrs r4rDz UMT5ForQuestionAnswering.forwardas@&1%)F  U !% 1 1) < < !*!6IIDK} $D>} 'tn!} "d^#} $ uU&')LL M%} } } nmBA} } } } } r5r-)Hrrrtypingrrrrr,rtorch.nnrr r activationsr cache_utilsr rrr generationrmodeling_attn_mask_utilsrmodeling_outputsrrrrrrrmodeling_utilsrutilsrrrrrr r!r"configuration_umt5r$ get_loggerrErr_CHECKPOINT_FOR_DOCModuler&rKrdrlrvrrrr rrMUMT5_START_DOCSTRINGrrr*r+r,rr3r-rKr5r4rNs //////////// AAAAAAAAAA!!!!!!PPPPPPPPPPPP))))))>>>>>>.-----                    +*****  H % %)+++++BI+++4 .RY<")$99999BI999DRY8bi