sgxdZddlmZddlZddlmZddlmZmZm Z m Z ddl Z ddl ZddlmZddlmZdd lmZmZmZmZmZmZdd lmZmZmZdd lmZm Z m!Z!m"Z"m#Z#m$Z$d d l%m&Z&e#jNe(Z)dZ*dZ+dZ,d8dZ-d9d:dZ.d;deGd*d+ej`jfZ?eGd,d-ej`jfZ@eGd.d/ej`jfZAe!d0e=Gd1d2e9ZBGd3d4ej`jfZCe!d5e=Gd6d7e9ZDy)=zTF 2.0 LED model.) annotationsN) dataclass)ListOptionalTupleUnion)get_tf_activation)+TFBaseModelOutputWithPastAndCrossAttentions)TFModelInputTypeTFPreTrainedModelget_initializerkeraskeras_serializable unpack_inputs)check_embeddings_within_bounds shape_liststable_softmax) ModelOutputadd_code_sample_docstringsadd_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) LEDConfigzallenai/led-base-16384rgחc tj||j}tj||j}tjt |ddftj ||j}tj ||ddddfgd}tj|dk(tjt |tj ||j|}tjj|tjd|j}tj|g5tj|}ddd|S#1swY|SxYw)Nrridtype) tfcastr fillrconvert_to_tensorconcatwhere debuggingassert_greater_equalconstantcontrol_dependenciesidentity) input_ids pad_token_iddecoder_start_token_id start_tokensshifted_input_ids assert_gte0s Z/var/www/html/venv/lib/python3.12/site-packages/transformers/models/led/modeling_tf_led.pyshift_tokens_rightr3<s877<9LWW%;Y__M77 I q !1%r';'; ::ct|d}||n|}tjd}tj||j}tj |ddddddfdd|df}||z t zS)z_ Expands attention_mask from `[bsz, seq_len]` to `[bsz, 1, tgt_seq_len, src_seq_len]`. rNg?r)rr!r)r"r r=r9)rBrAsrc_lenone_cst expanded_masks r2 _expand_maskrJgsxq!G ,g'Gkk#G 774w}} -DGGDD$!12Q7A4FGM m #~ 55rEc2eZdZdZdfd Zddfd ZxZS)TFLEDLearnedPositionalEmbeddingzN This module learns positional embeddings up to a fixed maximum size. c (t|||fi|yN)super__init__)selfnum_embeddings embedding_dimkwargs __class__s r2rPz(TFLEDLearnedPositionalEmbedding.__init__ys A&ArEc|d}tj|dd}||z }t| tj|tj S)z/Input is expected to be of size [bsz x seqlen].rr:)deltanamer)r!r:rOcallr"int32)rQ input_shaper?seq_len position_idsrUs r2rYz$TFLEDLearnedPositionalEmbedding.call|sGa.xxqw? .. w|BGGLABBrE)rRintrSr^r)r[tf.TensorShaper?r^__name__ __module__ __qualname____doc__rPrY __classcell__rUs@r2rLrLtsBCCrErLceZdZfdZddZ ddZdZedZdZ edZ edZ ed Z ed Z d Zd Zd ZdZxZS)TFLEDEncoderSelfAttentionc t|di|||_|j|jzdk7r%t d|jd|j|j|_t|j|jz |_|j|_ tjj|jt|jd|_tjj|jt|jd|_tjj|jt|jd|_tjj|jt|jd|_tjj|jt|jd |_tjj|jt|jd |_tjj+|j,|_tjj+|j,|_||_|j4|j2}|d zdk(sJd |j2d ||dkDsJd |j2d||d z|_y)NrzThe hidden size (z6) is not a multiple of the number of attention heads (query)kernel_initializerrXkeyvalue query_global key_global value_globalz`attention_window` for layer z has to be an even value. Given z has to be positive. Given )rOrPconfig hidden_sizenum_attention_heads ValueError num_headsr^head_dim embed_dimrlayersDenserinitializer_rangerkrmrnrorprqDropoutattention_probs_dropout_probdropoutglobal_dropoutlayer_idattention_windowone_sided_attn_window_size)rQrtrrTrrUs r2rPz"TFLEDEncoderSelfAttention.__init__s "6"    : : :a ?#F$6$6#78 4457   33F..1K1KKL ++\\'' NN.v/G/GH(  <<%% NN.v/G/GH&  \\'' NN.v/G/GH(  "LL.. NN.v/G/GH/   ,,,, NN.v/G/GH-  "LL.. NN.v/G/GH/  ||++F,O,OP #ll2263V3VW  !224==A q A % m *4==/9YZjYk l m % q  h *4==/9TUeTf g h +;a*?'rEcL|jstjd5|jj |j j fdddtjd5|jj |j j fdddtjd5|jj |j j fddd|jryd|_t|dddtj|jj5|jj dd|j j gdddt|dddtj|jj5|jj dd|j j gdddt|dddtj|jj5|jj dd|j j gdddt|dddtj|jj5|jj dd|j j gdddt|dddtj|jj5|jj dd|j j gdddt|ddetj|jj5|jj dd|j j gdddyy#1swYbxYw#1swY!xYw#1swYxYw#1swYhxYw#1swYxYw#1swYxYw#1swY'>&@A B~. D!!'')@)@(BC D ::  4$ ' 3tzz/ H   $dkk.E.E!FG H 4 % 1txx}}- FdDKK,C,CDE F 4$ ' 3tzz/ H   $dkk.E.E!FG H 4 . :t00556 O!!''tT[[5L5L(MN O 4t , 8t334 M%%tT4;;3J3J&KL M 4 . :t00556 O!!''tT[[5L5L(MN O O ;3 D D B B D D H H F F H H O O M M O Osk1N301O>1O 3O23O'#3O43P3P63P3N=O  OO$'O14O>P PP#c |\}}}}}}|j|} |j|} |j|} t|\} } }tj j ||jd|jd|| tjjt j|j| jz} t j| | | |j|jf} t j| | | |j|jf} |j| | |j }|dk7}t j|| jt"z}|jt j$t|||j }||z }tj j t|| | |j|j dzdzgd| d | d |jd |j dzdzd t| |j'|\}}}}|r|j)|| | |||| }t+|d }|rCt j,|ddddddfdd|j|j dz|zdzf}n?t j,|ddddddfdd|j|j dzdzf}t j.|t j0t||j|}|htj j t||jgd|jdt|t j|d|z}|j3||}t j| | | |j|jf} |r|j5| ||||}n|j7|| |j }tj j t|| | |j|jgdt j|| | |f}|r|j9||||||||| \}}n#t j0| |j|| f}|rCt j,|ddddddfdd|j|j dz|zdzf}n?t j,|ddddddfdd|j|j dzdzf}t j.|t j0t||j|}|||f}|S)a LongformerSelfAttention expects *len(hidden_states)* to be multiple of *attention_window*. Padding to *attention_window* happens in LongformerModel.forward to avoid redoing the padding on each layer. The *attention_mask* is changed in [`LongformerModel.forward`] from 0, 1, 2 to: - -10000: no attention - 0: local attention - +10000: global attention z&hidden_states should have embed_dim = z , but has messagerrrrrzattn_probs should be of size (z, z), but is of size ) attn_scores query_vectors key_vectorsmax_num_global_attn_indicesis_index_global_attn_nonzero"is_local_index_global_attn_nonzero%is_local_index_no_global_attn_nonzerorr6N/Head mask for a single layer should be of size , but is )rrrrtraining) value_vectors attn_probsrrrzUnexpected size) attn_output hidden_statesrlayer_head_maskrrris_index_maskedr)rkrmrnrr!r' assert_equalrzmathsqrtr"ryr r;rx _sliding_chunks_query_key_matmulrr9r8_get_global_attn_indices"_concat_with_global_key_attn_probsrr=r&r<r(_compute_attn_output_with_global_indices'_sliding_chunks_matmul_attn_probs_value'_compute_global_attn_output_from_hidden)rQinputsrrattention_maskrris_index_global_attnis_global_attnrrr batch_sizer\rzr#remove_from_windowed_attention_mask float_mask diagonal_maskrrrrr masked_indexrglobal_attn_probsmasked_global_attn_indexoutputss r2rYzTFLEDEncoderSelfAttention.calls.       =1 hh}-  =1 )3M)B& GY !!  NN ?  ' ( . 1 AA'+',G-I3U6[BK$Kb9 771dD 01At~~t'F'F'JMh'hkl'lmL 771dD 01At~~t'F'F'JQ'NOLXX  HHZ -Z5E5E F   & LL % %?+ Et~~EWX"?346 & O]CjPJ\\*x\@  =:wX\XeXe2fg  GG+%,G-I3U HKFFM4+J+JK !! { #j'4>>4==%Yct " jjz7I.NO  -1-Y-Y'+,G /3U-I6[ /!.Z . *K*!#*dnnFacj)k l  ')ww$Q4%56At~~t'F'F'JMh'hkl'lm( $ (*ww$Q4%56At~~t'F'F'JQ'NO( $XX $ HHZ 89AQAQ R   ,=>rEc t|\}}}}tjj||dzzdd|dzd|tjjt|t|dt|dt|||zdz }tjtj |d ||z||f}tjtj |d ||z||f}|j ||} |j ||} tj| | j } tjd | | } tjddgddgddgddgg} |j| | } tj| d d d d d |d |dzf| d d d d |d d |dzfgd}tjtj||zd||f| j | d d d d |dz d |dzd fgd}tjtj| d|gddgd d d d d |d |ftj||zd||f| j gd}tjtj |dztj" d d d d d f||zd||fdk}tj$|||}tj||gd }tj tj||||d|zdzfd }|j'||}|S)a Matrix multiplication of query and key tensors using with a sliding window attention pattern. This implementation splits the input into overlapping chunks of size 2w (e.g. 512 for pretrained Longformer) with an overlap of size window_overlap rrrz&Sequence length should be multiple of z. Given rz7Shape of query and key should be equal, but got query: z and key: rrrrrr rzbcxd,bcyd->bcxyNrr6r )shiftr7)rr!r'rr; transpose_chunkr"r einsumr$ _pad_and_transpose_last_two_dimsr%r<rollr=r:int64r&_mask_invalid_locations)rQrkrmwindow_overlaprr\rxry chunks_count chunked_query chunked_keychunked_attention_scorespaddings!diagonal_chunked_attention_scoresdiagonal_attn_scores_up_triangdiagonal_attn_scores_low_triang diagonal_attn_scores_first_chunkfirst_chunk_maskdiagonal_attention_scoress r2rz:TFLEDEncoderSelfAttention._sliding_chunks_query_key_matmuls 4>e3D0 GY !! ~) * <^a=O: kk#~6  [5F5FG #%99-> {#[ ''!Q!Q!Q!Q(HI,0,Q,QRjlt,u)*,1!QI]>\]K]I]2]^1!RS./K_^^_M_K_2_`  * &+-)))+QO;AA2!Q.1:L8MPR8RTbefTfTh2hi   + ',.995n-QQ.8 : )+QO;AA  , ( GG):4D$;NOi'NNK    +-((  , ++ '%'II ,.L MTV% ! %'LL JJ)Y^1Ca1GH   % !%)$@$@AZ\j$k!((rEc tjtjjtj||dzfdddg}tj dt |d|z gdt |d|z dz gg}tj||}|tj|ddgz}tj|ddddddft |ddddf}td tj|z}tjtjj|d||}|S)Nrshaperrr6r inf)r!reverselinalg band_partr8r$rpadr=float ones_liker&rgreater) input_tensorr mask_2d_upperpaddingmask_2dmask_4d inf_tensors r2rz1TFLEDEncoderSelfAttention._mask_invalid_locationss4 II  ~~PQ?Q.R SUWYZ [ &&L)!,~= >J|D\]^D_bpDpstDt@u v  &&0BJJwaV<<'''$4"23j6Nq6QSTVWYZ5[\El]R\\,%?? xx ;ZV rEc t|\}}}}tjj||dzzddtjjt|ddt|dddtjjt|dd|zdzd ||zdz }tjtj |d ||z||z|d|zdzf} tjtj |d ||z||f}tj ddg||gddgg} tj|| d } d|z|z} t| d|z| z |z} tjjtj| ||zd f| | }tj|||z|dzd|z|f}tjjt|||z|dzd|z|gd |j| } tjd| |}tj tj|||||fd }|S)z Same as _sliding_chunks_query_key_matmul but for attn_probs and value tensors. Returned tensor will be of the same shape as `attn_probs` rrrz0Seq_len has to be multiple of 2 * window_overlaprNr z:value and attn_probs must have same dims (except head_dim)rz4attn_probs last dim has to be 2 * window_overlap + 1rrconstant_valuesz!Chunked value has the wrong shapezbcwd,bcdh->bcwh) rr!r'rr;rr$rsignalframe_pad_and_diagonalizer)rQrrnrrr\rxryrchunked_attn_probsr padded_value frame_sizeframe_hop_size chunked_valuecontexts r2rzATFLEDEncoderSelfAttention._sliding_chunks_matmul_attn_probs_value2s 4>e3D0 GY !! ~) *A7i "  !! z "2A & u bq !P " !! z "1 %   "J " .014  ZZ LL\ 2Y&>)N"Q&    LL - ) #Wh 7  ''!Q..1QTUWXSY(Z[vveXrB '(2 $\215@:MR^^  JJ|j9&7I8 T !! } % ) #\A%5q>7I8 T7 " "667IJ))-/A=Q,, JJwY J K   rEctj||}t|\}}}}tj|||||f}|S)z)pads rows and then flips rows and columns)r!rrr;)hidden_states_paddedrr chunk_size seq_length hidden_dims r2rz:TFLEDEncoderSelfAttention._pad_and_transpose_last_two_dims{sR "vv ( :DDX9Y6 J J!zz*>ZYceo@pq##rEcFt|\}}}}tjddgddgddgd|dzgg}tj||}tj|||df}|ddddd| f}tj||||||zf}|ddddddddf}|S)aY shift every row 1 step right, converting columns into diagonals. Example: ```python chunked_hidden_states: [ 0.4983, 2.6918, -0.0071, 1.0492, -1.8348, 0.7672, 0.2986, 0.0285, -0.7584, 0.4206, -0.0405, 0.1599, 2.0514, -1.1600, 0.5372, 0.2629, ] window_overlap = num_rows = 4 ``` (pad & diagonalize) => [ 0.4983, 2.6918, -0.0071, 1.0492, 0.0000, 0.0000, 0.0000 0.0000, -1.8348, 0.7672, 0.2986, 0.0285, 0.0000, 0.0000 0.0000, 0.0000, -0.7584, 0.4206, -0.0405, 0.1599, 0.0000 0.0000, 0.0000, 0.0000, 2.0514, -1.1600, 0.5372, 0.2629 ] rrrN)rr!r$rr;)chunked_hidden_statestotal_num_heads num_chunksrrrs r2rz.TFLEDEncoderSelfAttention._pad_and_diagonalizesBCMMbBc?^Z''!Q!Q!Q!^VWEWAX(YZ " !8! !# !OZ#D! !6 q"N?" "! !# ! j..::U V! !6aAssl C$$rEc t|\}}}d|d|zzzdz }||z}d|z}tj||||zf}tjj |||}tj j t||||gd|||gdt|dtj|||d|z|f}|S)zBconvert into overlapping chunks. Chunk size = 2w, overlap size = wrrrz^Make sure chunking is correctly applied. `Chunked hidden states should have output dimension z , but got .r)rr!r;rrr'r) rrrrrnum_output_chunksrrrs r2rz TFLEDEncoderSelfAttention._chunks.8 -F* J N0B!CDqH(*4'  =:zJ?V2WX !#  z> Z !! , - *J 7->?@ :VkKlJmmnp " !# ! *A,> K!  %$rEctjj|d}tj|tjdj }tj |}tj|}tj|tj|dk}tj|}tjtjj|}||||fS)zblhs)rr rrrr)rrrr rrr6)rr! gather_nd scatter_ndrxryrrtupler8r"r tensor_scatter_nd_updater%)rQrrrrrrrrglobal_key_vectorskey_vectors_only_globalattn_probs_from_global_key attn_probs_from_global_key_trans mask_shaperBs r2rzblhd) rr!rrrxryrrr) rQrrrrrrattn_probs_only_globalglobal_value_vectorsvalue_vectors_only_globalattn_output_only_globalattn_probs_without_globalattn_output_without_globals r2rzBTFLEDEncoderSelfAttention._compute_attn_output_with_global_indices-s +A. ",Aq!5Q6Q5Q,Q!R "||M;WX%'MM . +  % !#%)),=?UWp"q%/q!Q8S8T/T$U!&*%Q%Q %}d6U6U& "')CCCrEc v t|dd\} } tj||} tj|| | ||jf} |j | } |j |}|j|}| tjjtj|j| jz} |j| | } |j|| }|j|| }tj| |d}tjj!t|| |j"z|| gd| |j"z|| fdt|d tj$|| |j"|| f}tj&|d }t|d ft)t|d dz}tj*|d z}tj||j}tj,|||}tj&|d }tj.|ddddddfdt|dddf}tj0|d |}tj$|| |j"z|| f}t3|d}|tjj!t||j"gd|j"dt| tj$|dtj$|| |j"|| fz}tj$|| |j"z|| f}|j5|| }tj||}tjj!t|| |j"z||jgd| |j"z||jfdt|d tj$|| |j"||jf}tjtj&|d |}tj$|t|d df}tj,|||}tj$|| |j"|| f}||fS)NrrrrT transpose_bz7global_attn_scores have the wrong size. Size should be rrrrrrrrrr6rrrrrrz=global_attn_output tensor has the wrong size. Size should be )rr!rrrzrorprqrrr"ryr reshape_and_transposematmulr'rrxr;rrr8rr=r&rr)rQrrrrrrrrrrr\global_attn_hidden_states global_query_vectors_only_globalrrglobal_attn_scoresglobal_attn_scores_transrglobal_attn_mask attn_maskglobal_attn_probs_floatrglobal_attn_outputnonzero_global_attn_outputs r2rzATFLEDEncoderSelfAttention._compute_global_attn_output_from_hiddenVs)7; G%'LL@\$]!$&MM . %:DNNK% !,0+<+<=V+W(!__];#00? )BGGLL GGDMM)I)O)O P-  (,0+E+EFfhr+s(!778JJW#99:NPZ[ YY'GI[imn !! ) * $.. (*Ew O$..02MwWXY12316 "  ZZ  )Dg N $&<<0BL#Q  !FGJLu / 0 5P  77:.977#3;S;Y;YZ$&#>#> $ 1 $   \\*BLQGGOAtT1,<=:N`CabcCdfgij?kl XXi;MNZZ  $.. (*Ew O  #11C""M  & LL % %?+ Et~~EWX"?346 & ')jj-&PSUS]S]'*dnnFacj)kT' #')jj'*t~~*EGbdk)l' # !//0GRZ/[ YY'8:NO !! ) * $.. (*Et}} U$..02Mt}}]^_12316 "  ZZ  )Ddmm T  &(\\ LL+\ : .& "&(ZZ & : ;A > C& " 11 57Q JJ  DNN 6))//5 6 4 . :t00556 K!!''tT[[5H5H(IJ K K ; 6 6 K Ks C"%3C."C+.C7rrNrbrcrdrPrYrrfrgs@r2r!r!s ( KrEr!cxeZdZdZ d dfd ZddZ d d dZd dZxZS) TFLEDDecoderAttentionz6Multi-headed attention from "Attention Is All You Needc Pt|di|||_||_tj j ||_||z|_|j|z|jk(sJd|jdz|_ ||_ tj j||d|_ tj j||d|_ tj j||d|_tj j||d|_y) Nz(embed_dim must be divisible by num_headsgk_projr%q_projv_projout_projrs)rOrPrzrxrr{r~rryscaling is_decoderr|r2r3r4r5)rQrzrxrr7biasrTrUs r2rPzTFLEDDecoderAttention.__init__s "6"""||++G4 !Y. }}y(DNN:f zz$++lGS"I:V ZZ J7 zz, ; Z(+yyztL  !! | $ 4>> !7G 46dnn8LgW^7_6`a|,-/ "   % LL % %>*a'*7a'8R7ST">235 & ::lS$..'SZ4[\_a_f_fl&8&8`L::lS4>>5I7T[4\]L%l<  & LL % %?+ Et~~EWX"?346 & ::o}E sDNNGWEIL::lS4>>5I7T[4\]L\\,\B ii L9  !! { # 4>> !7DMM :2CRVR_R_3`2ab{+,. " ll JJ{S$..'4==$Q RT` jjsGY.GH mmK0 "$**\CQXZa;b"c L.88rEc|jryd|_t|ddZtj|jj 5|jj dd|jgdddt|ddZtj|jj 5|jj dd|jgdddt|ddZtj|jj 5|jj dd|jgdddt|dd[tj|jj 5|jj dd|jgdddyy#1swYAxYw#1swYxYw#1swYxYw#1swYyxYw)NTr2r3r4r5) rrr!rr2rXrrzr3r4r5rs r2rzTFLEDDecoderAttention.builds ::  44 ( 4t{{//0 @ !!4t~~">? @ 44 ( 4t{{//0 @ !!4t~~">? @ 44 ( 4t{{//0 @ !!4t~~">? @ 4T * 6t}}112 B ##T4$@A B B 7 @ @ @ @ @ @ B Bs0 )F32)G)G )G3F=G  GG!)r5FT) rzr^rxr^rrr7boolr8rF)r: tf.Tensorr\r^r@r^)NNNNF) rrGr=tf.Tensor | Noner>Tuple[Tuple[tf.Tensor]] | NonerrHrrHreturnz"Tuple[tf.Tensor, tf.Tensor | None]rN) rbrcrdrerPr;rYrrfrgs@r2r0r0s@   VVV V  V  V0m .29=+/,0v9 v9+v97 v9 ) v9 * v9 ,v9pBrEr0cJeZdZdfd Z d ddZddZxZS)TFLEDEncoderLayerc t|d i||j|_t ||d|_t jjdd|_ t jj|j|_ t|j|_t jj|j|_t jj!|j"d|_t jj!|jd|_t jjdd|_||_y) N self_attnrXh㈵>self_attn_layer_normepsilonrXfc1fc2final_layer_normrs)rOrPr'rzr!rNrr{LayerNormalizationrQr~rr activation_function activation_fnactivation_dropoutr|encoder_ffn_dimrTrUrVrtr)s r2rPzTFLEDEncoderLayer.__init__s "6".vxkR$)LL$C$CDWm$C$n!||++FNN; .v/I/IJ"',,"6"6v7P7P"Q<<%%f&<&<5%I<<%%dnn5%A % ? ?Se ? f rEc |}|j||||||g|} | d}tjjt |t |dt |dt ||j ||}||z}|j |}|}|j|j|}|j||}|j|}|j ||}||z}|j|}|f| ddzS)a Args: hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)* attention_mask (`tf.Tensor`): attention mask of size *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values. layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size *(config.encoder_attention_heads,)*. rrz&Self attn modified the shape of query  to rrN) rNr!r'rrrrQrYrTrZrUrV) rQrrrrrrrresidual layer_outputss r2rYzTFLEDEncoderLayer.calls3$! NO_Nbdr s' &a(  !! } % x  N))//tT^^0LM N 4 % 1txx}}- =dDNN;< = 4 % 1txx}}- JdDKK,G,GHI J 4+T 2 >t4499: J%%++T4,HI J J ? + + N N = = J J J Js< H%)H$ )H133H=$)I H!$H.1H:=I I)rtrrr^r) rrGrrGrrGrrGrrGrrFrNr.rgs@r2rLrLsS *+4 +4"+4# +4 # +4 ( +4+4ZJrErLcZeZdZdfd Z d ddZddZxZS)TFLEDDecoderLayerc t|di||j|_t |j|j |j dd|_tjj|j|_ t|j|_tjj|j|_tjj!dd|_t |j|j |j dd|_tjj!dd |_tjj)|j*d |_tjj)|jd |_tjj!dd |_||_y)NrNT)rzrxrrXr7rPrQrR encoder_attn)rrXr7encoder_attn_layer_normrTrOrUrVrs)rOrPr'rzr0decoder_attention_headsattention_dropoutrNrr{r~rr rXrYrZrWrQrdrer|decoder_ffn_dimrTrUrVrtrQrtrTrUs r2rPzTFLEDDecoderLayer.__init__ sa "6".nn44,,  ||++FNN; .v/I/IJ"',,"6"6v7P7P"Q$)LL$C$CDWm$C$n!1 NN  * *,,  (-||'F'FtZs'F't$<<%%f&<&<5%I<<%%dnn5%A % ? ?Se ? f rEc 8|} ||ddnd} |j|| ||\}} } |j||}| |z}|j|}d} d}|S|} ||ddnd}|j|||||\}}} |j||}| |z}|j |}| | z} |} |j |j |}|j||}|j|}|j||}| |z}|j|}|| || fS)a Args: hidden_states (`tf.Tensor`): input to the layer of shape *(batch, seq_len, embed_dim)* attention_mask (`tf.Tensor`): attention mask of size *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values. encoder_hidden_states (`tf.Tensor`): cross attention input to the layer of shape *(batch, seq_len, embed_dim)* encoder_attention_mask (`tf.Tensor`): encoder attention mask of size *(batch, 1, tgt_len, src_len)* where padding elements are indicated by very large negative values. layer_head_mask (`tf.Tensor`): mask for attention heads in a given layer of size *(config.encoder_attention_heads,)*. encoder_layer_head_mask (`tf.Tensor`): mask for encoder attention heads in a given layer of size *(config.encoder_attention_heads,)*. past_key_value (`Tuple(tf.Tensor)`): cached past key and value projection states Nrr)rr>rrrr)rr=rrr>) rNrrQrdrerYrTrZrUrV)rQrrencoder_hidden_statesencoder_attention_maskrencoder_layer_head_maskr>rr^self_attn_past_key_valueself_attn_weightspresent_key_valuecross_attn_present_key_valuecross_attn_weightscross_attn_past_key_values r2rYzTFLEDDecoderLayer.call%s4!:H9S>"1#5Y] >Bnn'3)+ ?M? ; (*;  ]X F  =0 11-@ (,$! ,$H@N?Yrs(;_c %NRN_N_+!65 78 O`O KM-/K!LLLJM$}4M 88GM!24P P !**488M+BC // /Q /  ]X F  =0 --m<        rEcb|jryd|_t|ddMtj|jj 5|jj ddddt|ddZtj|jj 5|jj dd|jgdddt|ddMtj|jj 5|jj ddddt|ddZtj|jj 5|jj dd|jgdddt|ddZtj|jj 5|jj dd|jgdddt|dddtj|jj 5|jj dd|jjgdddt|dd[tj|jj 5|jj dd|jgdddyy#1swYsxYw#1swYxYw#1swYxYw#1swYrxYw#1swYxYw#1swYxYw#1swYyxYw) NTrNrQrdrerTrUrV)rrr!rrNrXrrQrzrdrerTrUrtrhrVrs r2rzTFLEDDecoderLayer.buildus ::  4d + 7t~~223 +$$T* + 4/ 6 Bt88==> N))//tT^^0LM N 4 . :t00556 .!!''- . 42D 9 Et;;@@A Q,,22D$3OP Q 4 % 1txx}}- =dDNN;< = 4 % 1txx}}- JdDKK,G,GHI J 4+T 2 >t4499: J%%++T4,HI J J ?# + + N N . . Q Q = = J J J JsT K%)K% K2&)K? )L 43L%)L%K"%K/2K<?L  LL"%L.rtr)NNNNNNF)rrHrkrHrlrHrrHrmrHr>zTuple[tf.Tensor] | NonerJz?Tuple[tf.Tensor, tf.Tensor, tf.Tensor, Tuple[Tuple[tf.Tensor]]]rNr.rgs@r2rbrbsz>,02637,04826N )N  0 N !1 N * N "2N 0N  IN `JrErbc0eZdZeZdZefdZxZS)TFLEDPreTrainedModelledcnt|}tjdtjd|d<|S)N)NNglobal_attention_maskrO)rOinput_signaturer! TensorSpecrZ)rQsigrUs r2r{z$TFLEDPreTrainedModel.input_signatures/g%')}}\288Ri'j #$ rE) rbrcrdr config_classbase_model_prefixpropertyr{rfrgs@r2rwrws L rErwcJeZdZUdZdZded<dZded<dZded<dZded<y) TFLEDEncoderBaseModelOutputaI Base class for Longformer's outputs, with potential hidden states, local and global attentions. Args: last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the model. hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the model at the output of each layer plus the initial embedding outputs. attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x + attention_window + 1)`, where `x` is the number of tokens with global attention mask. Local attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token in the sequence to every token with global attention (first `x` values) and to every token in the attention window (remaining `attention_window + 1` values). Note that the first `x` values refer to tokens with fixed positions in the text, but the remaining `attention_window + 1` values refer to tokens with relative positions: the attention weight of a token to itself is located at index `x + attention_window / 2` and the `attention_window / 2` preceding (succeeding) values are the attention weights to the `attention_window / 2` preceding (succeeding) tokens. If the attention window contains a token with global attention, the attention weight at the corresponding index is set to 0; the value should be accessed from the first `x` attention weights. If a token has global attention, the attention weights to all other tokens in `attentions` is set to 0, the values should be accessed from `global_attentions`. global_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`, where `x` is the number of tokens with global attention mask. Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence. NrGlast_hidden_stateTuple[tf.Tensor, ...] | Noner attentionsglobal_attentions) rbrcrdrer__annotations__rrrrsrEr2rrs7!F$(y'26M/6/3J,36:3:rErceZdZUdZdZded<dZded<dZded<dZded <dZ ded <dZ d ed <dZ ded <dZ ded<dZ ded<y)TFLEDSeq2SeqModelOutputa Base class for model encoder's outputs that also contains : pre-computed hidden states that can speed up sequential decoding. Args: last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer of the decoder of the model. If `past_key_values` is used only the last hidden-state of the sequences of shape `(batch_size, 1, hidden_size)` is output. past_key_values (`List[tf.Tensor]`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): List of `tf.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size, num_heads, sequence_length, embed_size_per_head)`). Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be used (see `past_key_values` input) to speed up sequential decoding. decoder_hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the decoder at the output of each layer plus the initial embedding outputs. decoder_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads. cross_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads. encoder_last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder of the model. encoder_hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each layer plus the initial embedding outputs. encoder_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads. encoder_global_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`, where `x` is the number of tokens with global attention mask. Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence. NrGrList[tf.Tensor] | Nonepast_key_valuesrdecoder_hidden_statesdecoder_attentionscross_attentionsrHencoder_last_hidden_staterkencoder_attentionsencoder_global_attentions)rbrcrdrerrrrrrrrkrrrsrEr2rrst5n$(y'.2O+2:>7>7;4;592926/6:>7>7;4;>B;BrErceZdZUdZdZded<dZded<dZded<dZd ed <dZ d ed <dZ d ed <dZ ded <dZ d ed<dZ d ed<dZd ed<y)TFLEDSeq2SeqLMOutputap Base class for sequence-to-sequence language models outputs. Args: loss (`tf.Tensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss. logits (`tf.Tensor` of shape `(batch_size, sequence_length, config.vocab_size)`): Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax). past_key_values (`List[tf.Tensor]`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): List of `tf.Tensor` of length `config.n_layers`, with each tensor of shape `(2, batch_size, num_heads, sequence_length, embed_size_per_head)`). Contains pre-computed hidden-states (key and values in the attention blocks) of the decoder that can be used (see `past_key_values` input) to speed up sequential decoding. decoder_hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the decoder at the output of each layer plus the initial embedding outputs. decoder_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the decoder, after the attention softmax, used to compute the weighted average in the self-attention heads. cross_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the decoder's cross-attention layer, after the attention softmax, used to compute the weighted average in the cross-attention heads. encoder_last_hidden_state (`tf.Tensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder of the model. encoder_hidden_states (`tuple(tf.Tensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `tf.Tensor` (one for the output of the embeddings + one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`. Hidden-states of the encoder at the output of each layer plus the initial embedding outputs. encoder_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights of the encoder, after the attention softmax, used to compute the weighted average in the self-attention heads. encoder_global_attentions (`tuple(tf.Tensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `tf.Tensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, x)`, where `x` is the number of tokens with global attention mask. Global attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. Those are the attention weights from every token with global attention to every token in the sequence. NrHlossrGlogitsrrrrrrrrkrr)rbrcrdrerrrrrrrrrkrrrsrEr2rr s~3j"D !FI.2O+2:>7>7;4;592926/6:>7>7;4;>B;BrErat This model inherits from [`TFPreTrainedModel`]. 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 [keras.Model](https://www.tensorflow.org/api_docs/python/tf/keras/Model) subclass. Use it as a regular TF 2.0 Keras Model and refer to the TF 2.0 documentation for all matter related to general usage and behavior. TensorFlow models and layers in `transformers` accept two formats as input: - having all inputs as keyword arguments (like PyTorch models), or - having all inputs as a list, tuple or dict in the first positional argument. The reason the second format is supported is that Keras methods prefer this format when passing inputs to models and layers. Because of this support, when using methods like `model.fit()` things should "just work" for you - just pass your inputs and labels in any format that `model.fit()` supports! If, however, you want to use the second format outside of Keras methods like `fit()` and `predict()`, such as when creating your own layers or models with the Keras `Functional` API, there are three possibilities you can use to gather all the input Tensors in the first positional argument: - a single Tensor with `input_ids` only and nothing else: `model(input_ids)` - a list of varying length with one or several input Tensors IN THE ORDER given in the docstring: `model([input_ids, attention_mask])` or `model([input_ids, attention_mask, token_type_ids])` - a dictionary with one or several input Tensors associated to the input names given in the docstring: `model({"input_ids": input_ids, "token_type_ids": token_type_ids})` Note that when creating models and layers with [subclassing](https://keras.io/guides/making_new_layers_and_models_via_subclassing/) then you don't need to worry about any of this, as you can just pass inputs like you would to any other Python function! Args: config ([`LEDConfig`]): 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 [`~TFPreTrainedModel.from_pretrained`] method to load the model weights. a Args: input_ids (`tf.Tensor` of shape `({0})`): 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) attention_mask (`tf.Tensor` of shape `({0})`, *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 (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`LedTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) LED uses the `eos_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`). decoder_attention_mask (`tf.Tensor` of shape `(batch_size, target_sequence_length)`, *optional*): will be made by default and ignore pad tokens. It is not recommended to set this for most use cases. head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): Mask to nullify selected heads of the 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 (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the 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 (`tf.Tensor`, *optional*): hidden states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. of shape `(batch_size, sequence_length, hidden_size)` is a sequence of past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers`) 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)`. use_cache (`bool`, *optional*, defaults to `True`): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). Set to `False` during training, `True` during generation output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. 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. This argument can be used only in eager mode, in graph mode the value in the config will be used instead. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. This argument can be used in eager mode, in graph mode the value will always be set to True. training (`bool`, *optional*, defaults to `False`): Whether or not to use the model in training mode (some modules like dropout modules have different behaviors between training and evaluation). ceZdZeZ dd fd ZdZdZe d dZ e jdZ dZ ddZxZS) TFLEDEncoderc t|di|||_tjj |j |_|jdkDrtjdd|_ |j|_ t|jtrO|jdzdk(sJd|jdkDsJd|jg|j z|_nLt#|j|j k(s*Jd|j dt#|j|j|_||_t'|j(|j*d |_t/|j0Dcgc]}t3||d | c}|_tjj5d d |_|j*|_ycc}w)Nr0Layerdrop is currently disabled in TFLED models.r5rrz1`config.attention_window` has to be an even valuez,`config.attention_window` has to be positivezQ`len(config.attention_window)` should equal `config.num_hidden_layers`. Expected z, given embed_positionsrOlayers.rPlayernorm_embeddingrRrs)rOrPrtrr{r~rencoder_layerdroploggerwarning layerdropr- padding_idx isinstancerr^num_hidden_layerslen embed_tokensrLmax_encoder_position_embeddingsr'rr:encoder_layersrLrWrrzrQrtrrTirUs r2rPzTFLEDEncoder.__init__s "6" ||++FNN;  # #a ' NNM N!.. f--s 3**Q.!3 h5h h3**Q. ^0^ ^.'-'>'>&?&BZBZ&ZF #v../63K3KK "445Xc&BYBY>Z=[] K !' 7 7(>  2 2 NN"  RWW]WlWlQmnA(71#Gn #(<<#B#B4Vk#B#l os G%c|jSrNrrQs r2get_embed_tokenszTFLEDEncoder.get_embed_tokenss   rEc||_yrNrrQrs r2set_embed_tokenszTFLEDEncoder.set_embed_tokens (rEc | | td|=t|} t||jj|j|}n|t|dd} n td|t j | d}|'|t j|dz|jz}|j||||j\} }}}t|} t jjt j|t jd} t jjt j|t jd} t jj| }|j!| }||z}|j#|}|j%|| }|#t'|ddd d ddf}|ddddddf}|rd nd}|rd ndx}}|gt j(j+t|d t-|j.d t-|j.d t|d d t1|j.D]\}}|r|j3|| }||fz}t5j6d d}| r||j8krH||||||nd| | |}|d }|se|t j:|ddfz}|t j:|ddfz}|j3|| }|r0| d kDr)t=|Dcgc]}|ddddd| ddfc}n|}|r||fz}|st=d|||fDSt?||||Scc}w)aW Args: input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`tf.Tensor` 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 (`tf.Tensor` of shape `(num_layers, num_heads)`, *optional*): Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. inputs_embeds (`tf.Tensor` 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. NzDYou cannot specify both input_ids and inputs_embeds at the same timerz5You have to specify either input_ids or inputs_embedsrr)r,r inputs_embedsr-rrrs&The head_mask should be specified for  layers, but it is for rr)rrrrrrrrr)rrr rrc3&K|] }|| ywrNrs.0vs r2 z$TFLEDEncoder.call..seqWXWdes)rrrr) rwrrr input_dimr!r#r"r _pad_to_window_sizerrlessint8r reduce_anyrrrrJr'rrr{ enumeratecompute_hidden_statesrandomuniformrrrr)rQr,rrrz head_maskoutput_attentionsoutput_hidden_states return_dictrr[ padding_lenrrr embed_posrencoder_statesall_attentionsall_global_attentionsidx encoder_layerhidden_states_to_adddropout_probabilityr_states r2rYzTFLEDEncoder.callsd  ]%>cd d  "$Y/K *9d6G6G6Q6Q R --i8M  &$]3CR8KTU U  !WW[!4N ! ,+bgg7Lq7PYgYmYm.nnN@D@X@X)')) AYA = Y !0 '',,rww~rww'GK!wwrww~rww/OQRS++,@A((5 % 1 00?  ]X F   %).9!Q1*EN+Aq$,<=N37HdR.  LL % %9%a(DKK >A)+-2;2G #T /%9- M*!,M !/2<< a@PR^3_2a!a)>m\]N^`lAm@o(o%3 p:22=+N  ?.QuQ=[L=!34QR#  +}.>>Ne]NN$See e*+(%3   Rs,M<c(|dkDr |ddd| fS|S)Nrrs)rQrrs r2rz"TFLEDEncoder.compute_hidden_statess#2=/}Q + -.T}TrEcvt|jtr |jnt|j}|dzdk(s Jd|| t |n t |}|dd\}}|||zz |z} | dkDr!t j d|d|| zd|tjddgd| gg} |tj|| |}|G| dkDrBtj|| f|} |j| } tj|| gd }tj|| d }| |||fS) zaA helper function to pad tokens and mask to work with implementation of Longformer selfattention.rrrz2`attention_window` should be an even value. Given Nz(Input ids are automatically padded from r]z0 to be a multiple of `config.attention_window`: rrr6F) rrr^maxrr warning_oncer!r$rr#rr%) rQr,rrr-rr[rr\rrinput_ids_paddinginputs_embeds_paddings r2rz TFLEDEncoder._pad_to_window_sizesz&00E0Es%KD ! !QTUYUjUjQk  !#q(q,^_o^p*qq(/8/Dj+*UbJc )"1o G''4D*DDHXX ?   :7)4R]H]G^_..>-?A  ''!Q![1A(BC  y(LQI  $Q$&GGZ,E|$T!(,(9(9:K(L% " =:O*PWY Z %P       rEc|jryd|_t|ddMtj|jj 5|jj ddddt|ddZtj|jj 5|jj dd|jgdddt|ddK|jD];}tj|j 5|j dddd=yy#1swYxYw#1swYnxYw#1swYaxYwNTrrr{) rrr!rrrXrrrzr{rQr[layers r2rzTFLEDEncoder.builds! ::  4*D 1 =t33889 1$$**40 1 4. 5 At77<<= M((..dDNN/KL M 44 ( 4 &]]5::.&KK%&& & 5  1 1 M M&&s$ D/%)D;E/D8;EE rNrtrrz Optional[keras.layers.Embedding]) NNNNNNNNF)rbrcrdrr~rPrrrrYr!functionrrrrfrgs@r2rrsuL(<!)"![ [ z[[UU+ Z &rErcbeZdZeZ ddfd ZdZe ddZddZ xZ S) TFLEDDecoderc $t|d i|||_|j|_||_|j dkDrtjdd|_ t|j|jd|_ t|jDcgc]}t!|d|c}|_t$j"j'dd |_t$j"j+|j,|_ycc}w) Nrrr5rrOrrPrrRrs)rOrPrtr-rrdecoder_layerdroprrrrLmax_decoder_position_embeddingsr'rr:decoder_layersrbr{rrWrr~rrs r2rPzTFLEDDecoder.__init__s "6" !..(  # #a ' NNM N>  2 2 NN"  OTTZTiTiNjk(smDk #(<<#B#B4Vk#B#l ||++FNN; lsD c||_yrNrrs r2rzTFLEDDecoder.set_embed_tokensrrEc | | td| t|}n|t|dd}n td|t|dddnd}|j||}|1t||jj |j |}|}|ddkDrt ||}n.ttj|d|d|zf|d }||ddkDr|t||d z}||t||d }|j||z}|j|| }d }d }d }d }|gtjjt|dt|jd t|jd t|ddt!|jD]w\}}| r||fz }t#j$dd}| r||j&kr6|||nd}||||||||nd|||nd|\}}}}| r||fz }| sl||fz }||fz }y| r||fz }nd}| r|nd}| r|nd}| r|nd}| st)d|||||fDSt+|||||S)aM Args: input_ids (`tf.Tensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide it. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) attention_mask (`tf.Tensor` 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) encoder_hidden_states (`tf.Tensor` of shape `(batch_size, encoder_sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention of the decoder. encoder_attention_mask (`tf.Tensor` of shape `(batch_size, encoder_sequence_length)`, *optional*): Mask to avoid performing cross-attention on padding tokens indices of encoder input_ids. 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 (`tf.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. encoder_head_mask (`tf.Tensor` of shape `(encoder_layers, encoder_attention_heads)`, *optional*): Mask to nullify selected heads of the attention modules in encoder to avoid performing cross-attention on hidden heads. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. past_key_values (`Tuple[Tuple[tf.Tensor]]` of length `config.n_layers` with each tuple having 2 tuples each of which has 2 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 (`tf.Tensor` 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. NzTYou cannot specify both decoder_input_ids and decoder_inputs_embeds at the same timerzEYou have to specify either decoder_input_ids or decoder_inputs_embedsrrrr)r?)rArrsrrrr)rrkrlrrmr>c3$K|]}|| ywrNrsrs r2rz$TFLEDDecoder.call..s=s)rrrrr)rwrrrrrrDrJr!r8rrr'rrr{rrrrrr )rQr,rrrkrlrencoder_head_maskr use_cacherrrrr[r? positionsrcombined_attention_maskall_hidden_statesall_self_attnsall_cross_attentionspresent_key_valuesr decoder_layerrr>layer_self_attnlayer_cross_attnrps r2rYzTFLEDDecoder.callsWN  ]%>st t  "$Y/K  &$]3CR8Kde eIXIdOA,>q,A!B1!Ejk((6LM  *9d6G6G6Q6Q R --i8M%  r?Q &7 \r&s #&2QQ:P)PQR\ghj\k' #  %+b/A*=&= ^epqset@u&u # ,1G1S%12HR]^`Ra%b "001JK  ]X F !  LL % %9%a(DKK >A5D5P_S1VZNR_6&;'=2;2G #TBSB_(9#(>ei-S OM?,<>O"'8&::" ?"44$)9(;;$5 <8 -!1 1  $ +<$7H3d3</$');=NP^`tu  ?"/ 2/)!5  rEc|jryd|_t|ddMtj|jj 5|jj ddddt|dddtj|jj 5|jj dd|jjgdddt|ddK|jD];}tj|j 5|j dddd=yy#1swYxYw#1swYnxYw#1swYaxYwr) rrr!rrrXrrrtr'r{rs r2rzTFLEDDecoder.builds' ::  4*D 1 =t33889 1$$**40 1 4. 5 At77<<= R((..dDKK)FI  ""ll#-&;#+"3%9'!+ OO=X!Y9"1!"4474H14Loa0RV14_1E1I?1-tO Z%G-668O,, 1"1!"4#1''+//!5#'  "_4 4&-??+;;"1"?"?.99,==&5&G&G"1"?"?.99&5&G&G  rEc|jryd|_tj|jjdz|jj zdz5|jj ddddt|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYxYw#1swYqxYw#1swYyxYw)NT/rr) rr!rrrrXrrrrrs r2rzTFLEDMainLayer.build s ::  ]]4;;99C?$++BRBRRUXX Y $ KK  d # $ 4D ) 5t||001 ) ""4( ) 4D ) 5t||001 ) ""4( ) ) 6  $ $ ) ) ) )s$D55EE 5D>E  EruNNNNNNNNNNNNNNNF)r z3Optional[Union[Tuple, TFLEDEncoderBaseModelOutput]]rN) rbrcrdrr~rPrrrrYrrfrgs@r2rrsrL I0 #OS""!#K MK K Z)rErzQThe bare LED Model outputting raw hidden-states without any specific head on top.ceZdZfdZdZdZeeejde e e e  d d dZdZd dZxZS) TFLEDModelcPt||g|i|t|d|_y)NrxrO)rOrPrrxrQrtrrTrUs r2rPzTFLEDModel.__init__+ s( 3&3F3!&u5rEc.|jjSrNrxrrs r2 get_encoderzTFLEDModel.get_encoder0 xxrEc.|jjSrNrxrrs r2 get_decoderzTFLEDModel.get_decoder3 rrEzbatch_size, sequence_length) checkpoint output_typer~c H|j||||||||| | | | | |||}|S)N)r,rr r r rzrr rrrrrrrr)rx)rQr,rr r rr r rzrrrrrrrrrTrs r2rYzTFLEDModel.call6 sN6(()/#9+"7/+'"7/!5#! &rEc Z|jjr"tj|jdnd}|jj rtj |jnd}|jjrtj |jnd}|jjrtj |jnd}|jj rtj |jnd}|jjrtj |jnd}|jjrtj |jnd}t|j|||||j ||| S)Nrr)rtrr!rrrr$rrrrrkrrrrr rQr$pkvdec_hs dec_attns cross_attnsenc_hs enc_attns enc_g_attnss r2serving_outputzTFLEDModel.serving_outputf s>59[[5J5Jbhhv--.q1PTGK{{GgGg%%f&B&BCmqGK{{GdGdB(()B)BCjn GK{{GdGdb**6+B+BCjn GK{{GgGg%%f&B&BCmqGK{{GdGdB(()B)BCjn PTP[P[PmPmb**6+K+KLsw &$66"(((&,&F&F"((&1  rEc|jryd|_t|ddNtj|jj 5|jj ddddyy#1swYyxYw)NTrx)rrr!rrxrXrrs r2rzTFLEDModel.build{ se ::  4 % 1txx}}- %t$ % % 2 % %s A11A:r)"r,TFModelInputType | NonerrHr rHr rHrrHr rHr rHrzrHrrIrrHrrHr bool | Nonerr-rr-rr-rrFrJz*Tuple[tf.Tensor] | TFLEDSeq2SeqModelOutputrN)rbrcrdrPrrrrLED_INPUTS_DOCSTRINGformatr_CHECKPOINT_FOR_DOCr_CONFIG_FOR_DOCrYr*rrfrgs@r2rr& s< 6   *+?+F+FGd+ef&+$.2+/.237&*.2,026:>*.26!%)-,0#'#'*')', ' !1 ' $ ','*' 0'8'(' 0''''*' !!'"#'& 4'' g'R *%rErc(eZdZdZfdZdZxZS) BiasLayerz Bias as a layer. It is used for serialization purposes: `keras.Model.save_weights` stores on a per-layer basis, so all weights have to be registered in a layer. c \t|dd|i||j|||||_y)NrXrXr initializer trainablers)rOrP add_weightr8)rQrr6r7rXrTrUs r2rPzBiasLayer.__init__ s3 -d-f-OOU _hOi rEc ||jzSrN)r8)rQxs r2rYzBiasLayer.call s499}rErargs@r2r3r3 s jrEr3zKThe LED Model with a language modeling head. Can be used for summarization.c&eZdZddgZfdZdZdZdZdZdZ d Z e e e eee  d dd Zd Z dd ZddZdZddZxZS)TFLEDForConditionalGenerationzled.encoder.embed_tokens.weightzled.decoder.embed_tokens.weightct||g|i|t|d|_|j|_t dd|j gdd|_d|_y)NrxrOfinal_logits_biasrr<Fr5) rOrPrrxrr3r bias_layersupports_xla_generationrs r2rPz&TFLEDForConditionalGeneration.__init__ sa 3&3F3!&u5))#$Q0A0A,BPWch  (-$rEc.|jjSrNrrs r2rz)TFLEDForConditionalGeneration.get_decoder rrEc.|jjSrNrrs r2rz)TFLEDForConditionalGeneration.get_encoder rrEc2d|jjiS)Nr>)r?r8rs r2get_biasz&TFLEDForConditionalGeneration.get_bias s#T__%9%9::rEc|djd}tdd|gdd|_|jjj |dy)Nr>rrr<Fr5)rr3r?r8assign)rQrnrs r2set_biasz&TFLEDForConditionalGeneration.set_bias sR./55b9 #$Q O\a  ##E*=$>?rEc"|jSrN)rrs r2get_output_embeddingsz3TFLEDForConditionalGeneration.get_output_embeddings s((**rEc&|j|yrN)r)rQrns r2set_output_embeddingsz3TFLEDForConditionalGeneration.set_output_embeddings s !!%(rE)rr~cr|;d} |7| 5t||jj|jj}|j ||||||||| | | | | |||}t j |d|jjjd}|j|}|dn|j||}|s|f|ddz}||f|zS|St|||j|j|j|j|j |j"|j$|j& S) a Returns: Examples: ```python >>> from transformers import AutoTokenizer, TFLEDForConditionalGeneration >>> import tensorflow as tf >>> mname = "allenai/led-base-16384" >>> tokenizer = AutoTokenizer.from_pretrained(mname) >>> TXT = "My friends are but they eat too many carbs." >>> model = TFLEDForConditionalGeneration.from_pretrained(mname) >>> batch = tokenizer([TXT], return_tensors="tf") >>> logits = model(inputs=batch.input_ids).logits >>> probs = tf.nn.softmax(logits[0]) >>> # probs[5] is associated with the mask token ```NF)rr r r rzrr rrrrrrrrrTrr) rrrrrrrrkrr)r3rtr-r.rxr!rrweightsr?hf_compute_lossrrrrrrrkrr)rQr,rr r rr r rzrrrrrrrlabelsrr lm_logitsmasked_lm_lossr$s r2rYz"TFLEDForConditionalGeneration.call s]T  I (-B-J$6DKK44dkk6X6X%!(( )/#9+"7/+'"7/!5#! $IIgaj$((//*A*AtT OOI. !'T5I5I&R[5\\GABK/F3A3M^%. YSY Y##33")"?"?&99$55&-&G&G")"?"?&99&-&G&G  rEc Z|jjr"tj|jdnd}|jj rtj |jnd}|jjrtj |jnd}|jjrtj |jnd}|jj rtj |jnd}|jjrtj |jnd}|jjrtj |jnd}t|j|||||j ||| S)Nr) rrrrrrrkrr)rtrr!rrrr$rrrrrkrrrrrr"s r2r*z,TFLEDForConditionalGeneration.serving_output s<59[[5J5Jbhhv--.q1PTGK{{GgGg%%f&B&BCmqGK{{GdGdB(()B)BCjn GK{{GdGdb**6+B+BCjn GK{{GgGg%%f&B&BCmqGK{{GdGdB(()B)BCjn PTP[P[PmPmb**6+K+KLsw #=="(((&,&F&F"((&1  rEc 2| |ddddf}d|||||||dS)Nr)r,r rr rrr rrs) rQr rrrr rr rTs r2prepare_inputs_for_generationz;TFLEDForConditionalGeneration.prepare_inputs_for_generation/ s?  & 1!RS& 9 ..!2,"!2"  rEclt||jj|jjSrN)r3rtr-r.)rQrOs r2%prepare_decoder_input_ids_from_labelszCTFLEDForConditionalGeneration.prepare_decoder_input_ids_from_labelsI s%!&$++*B*BDKKDfDfggrEc *tjjdtjjj}|j j rtj|d}tj||j j}tjtj|dt|df|}tj||}|||S|tjj||}tj||j jk7|j }||z} tj"| tj"|z } tj| dS)z(CrossEntropyLoss that ignores pad tokensT) from_logits reduction)rrrrr)r)rlossesSparseCategoricalCrossentropy ReductionNONErttf_legacy_lossr!r; not_equalr- boolean_maskrnnrelur"r reduce_sum) rQrOrloss_fn melted_labels active_lossreduced_logits unmasked_loss loss_mask masked_lossreduced_masked_losss r2rNz-TFLEDForConditionalGeneration.hf_compute_lossL s(,,<<Y^YeYeYoYoYtYt<u ;; % %JJvu5M,,}dkk6N6NOK__RZZZPVEWXYEZ@[-\^ijN__]K@F6>2 2  6 2F; GGFdkk&>&>>mFYFYZ #i/  mmK82==;SSzz-t44rEc|jryd|_t|ddMtj|jj 5|jj ddddt|ddNtj|jj 5|jj ddddyy#1swYexYw#1swYyxYw)NTrxr?)rrr!rrxrXrr?rs r2rz#TFLEDForConditionalGeneration.build^ s ::  4 % 1txx}}- %t$ % 4t , 8t334 ,%%d+ , , 9 % % , ,s C %C CC )NNNNNNNNNNNNNNNNF)$r,r,rnp.ndarray | tf.Tensor | Noner rmr rmrrmr rmr z"TFLEDEncoderBaseModelOutput | Nonerzrmrz1Tuple[Tuple[Union[np.ndarray, tf.Tensor]]] | Nonerrmrrmrr-rr-rr-rr-rOrHrrFrJz'Tuple[tf.Tensor] | TFLEDSeq2SeqLMOutput)NNNNNN)rOrGrN)rbrcrd"_keys_to_ignore_on_load_unexpectedrPrrrDrGrIrKrrr.rrr1rYr*rTrVrNrrfrgs@r2r<r< s +**& -  ;@+)*+?@+?o^.28<;?@D37;?>B?CMQ7;?C!%)-,0#'#'%R *R 6R 9 R !> R 1 R 9R <R  =R KR 5R  =R R 'R *R !!R "!#R $%R & 1'R _AR h 0 4h5$ ,rEr<)r,rGr-r^r.r^r_)r>r`r?r^rN)rBrGrAz Optional[int])Ere __future__rr dataclassesrtypingrrrrnumpynp tensorflowr!activations_tfr modeling_tf_outputsr modeling_tf_utilsr r rrrrtf_utilsrrrutilsrrrrrrconfiguration_ledr get_loggerrbrr0r1r9r3rDrJr{rrLLayerrir!r0rLrbrwrrrLED_START_DOCSTRINGr.rrrrr3r<rsrEr2r~sz" !///NSR)   H %.2;$ 6Cell&<&<C$`   2 2`  F$KELL..$KNeBELL..eBPMJ **MJ`EJ **EJP, ';+'; ';T @Ck@C @CF ?C;?C ?CD'RAHJ&5<<%%J&J&Z_&5<<%%_&_&Dv)U\\''v)v)rWW%%W% W%v """QM,$8M, M,rE