o ZhJ@sdZddlZddlZddlmZmZmZmZddlZddl Zddlm Z ddl m Z m Z mZddlmZddlmZdd lmZmZdd lmZmZmZmZmZmZmZdd lmZdd l m!Z!m"Z"d dl#m$Z$e"%e&Z'dej(de)de)fddZ*Gddde j+Z,Gddde j-Z.Gddde j-Z/Gddde j-Z0Gddde j-Z1Gddde j-Z2e!Gd d!d!eZ3Gd"d#d#e3Z4Gd$d%d%e3Z5e!Gd&d'd'e3Z6e!d(d)Gd*d+d+e3eZ7e!d,d)Gd-d.d.e3Z8e!Gd/d0d0e3Z9Gd1d2d2e3Z:Gd3d4d4e3eZ;gd5Zr?r@rAfloatboolr,r4rBrWrrr9rCr"r"r.r#rD\sP  rDcsdeZdZdeffdd Z d dejdejdejdejd eed e ejeejff d d Z Z S)MvpEncoderLayerconfigcst|j|_t|j|j|jd|_t |j|_ |j |_ t |j |_|j|_t|j|j|_t|j|j|_t |j|_dS)N)rFrGrH)r+r,d_modelrFrDencoder_attention_headsattention_dropout self_attnr LayerNormself_attn_layer_normrHr activation_function activation_fnactivation_dropoutrNZencoder_ffn_dimfc1fc2final_layer_normr-rpr.r"r#r,s  zMvpEncoderLayer.__init__FrXr[r\self_attn_promptr^r_c Cs|}|j|||||d\}}}tjj||j|jd}||}||}|}|||}tjj||j|jd}| |}tjj||j|jd}||}| |}|j t j krwt |set |rwt |j jd} t j|| | d}|f} |r| |f7} | S)a@ Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size `(encoder_attention_heads,)`. self_attn_prompt (`torch.FloatTensor`): prompt of self attention of shape `(2, encoder_attention_heads, pro_len, head_dim)`. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. )rXr[r\r]r^rbi)minmax)rtrrirHrdrvrxrzryr{r|r2r4Zfloat16isinfanyisnanZfinforclamp) r-rXr[r\r~r^residualrlrkZ clamp_valueoutputsr"r"r#r9s:        zMvpEncoderLayer.forward)F) r=r>r?rr,r4 FloatTensorrrnrr9rCr"r"r.r#ros rocseZdZdeffdd Z          ddejdeejd eejd eejd eejd eejd eejdeejdeeejdee dee deej eeej ej fffddZ Z S)MvpDecoderLayerrpcst|j|_t|j|j|jdd|_|j|_t |j |_ |j |_ t |j|_t|j|j|jdd|_t |j|_t |j|j|_t |j|j|_t |j|_dS)NT)rFrGrHrI)rHrI)r+r,rqrFrDdecoder_attention_headsrsrtrHr rwrxryrrurv encoder_attnencoder_attn_layer_normrNZdecoder_ffn_dimrzr{r|r}r.r"r#r,@s,  zMvpDecoderLayer.__init__NFTrXr[encoder_hidden_statesencoder_attention_maskr\cross_attn_layer_head_maskr~cross_attn_promptrZr^ use_cacher_c  Cs^|} | dur | ddnd} |j|| |||| d\}}}tjj||j|jd}| |}||}d}d}|durm|} | durD| ddnd}|j||||||| d\}}}tjj||j|jd}| |}||}||}|} || |}tjj||j |jd}| |}tjj||j|jd}| |}| |}|f}| r|||f7}| r||f7}|S)aC Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. encoder_hidden_states (`torch.FloatTensor`): cross attention input to the layer of shape `(batch, seq_len, embed_dim)` encoder_attention_mask (`torch.FloatTensor`): encoder attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. layer_head_mask (`torch.FloatTensor`): mask for attention heads in a given layer of size `(encoder_attention_heads,)`. cross_attn_layer_head_mask (`torch.FloatTensor`): mask for cross-attention heads in a given layer of size `(decoder_attention_heads,)`. self_attn_prompt (`torch.FloatTensor`): prompt of self attention of shape `(2, decoder_attention_heads, pro_len, head_dim)`. cross_attn_prompt (`torch.FloatTensor`): prompt of cross attention of shape `(2, decoder_attention_heads, pro_len, head_dim)`. past_key_value (`Tuple(torch.FloatTensor)`): cached past key and value projection states output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. Nr))rXrZr[r\r]r^rb)rXrYr[r\r]rZr^) rtrrirHrdrvrrrxrzryr{r|)r-rXr[rrr\rr~rrZr^rrZself_attn_past_key_valueZself_attn_weightsZpresent_key_valueZcross_attn_present_key_valueZcross_attn_weightsZcross_attn_past_key_valuerr"r"r#r9ZsX$         zMvpDecoderLayer.forward) NNNNNNNNFT) r=r>r?rr,r4rBrrrnrr9rCr"r"r.r#r?sJ     rcsHeZdZdZdedededeffdd Zdejd ejfd d Z Z S) MvpClassificationHeadz-Head for sentence-level classification tasks. input_dim inner_dim num_classespooler_dropoutcs8tt|||_tj|d|_t|||_dS)Nrc)r+r,rrNdenseDropoutrHrR)r-rrrrr.r"r#r,s zMvpClassificationHead.__init__rXr_cCs6||}||}t|}||}||}|SN)rHrr4tanhrR)r-rXr"r"r#r9s     zMvpClassificationHead.forward) r=r>r?r@rArmr,r4rBr9rCr"r"r.r#rs rcs:eZdZdZfddZdejdeejfddZZ S) MvpPromptz)Layer-wise prompt for encoder or decoder.c st|j|_||_||_|j||_tj|j d|_ t |j|j|_ t t |j|jtt |j|d|j|_dS)Nrr))r+r, prompt_length num_layersrGrqrLrrrH Embeddingprompt_embeddingZ SequentialrNZprompt_mid_dimZGELU prompt_trans)r-rprrGr.r"r#r,s   zMvpPrompt.__init__ prompt_idsr_cCsN|||}||j|jd|j|j}||}|gd d}|S)Nr))rr)rr ) rrrTrrrGrLrHZpermutesplit)r-rpromptr"r"r#r9s  zMvpPrompt.forward) r=r>r?r@r,r4rBrr9rCr"r"r.r#rs "rc@s,eZdZeZdZdZddZeddZ dS)MvpPreTrainedModelmodelTcCs|jj}t|tjr"|jjjd|d|jdur |jj dSdSt|tj rA|jjjd|d|j durC|jj|j  dSdSdS)NrE)meanstd) rpZinit_std isinstancerrNr7dataZnormal_rJZzero_r padding_idx)r-modulerr"r"r# _init_weightss    z MvpPreTrainedModel._init_weightscCs>|jj}tjgddddd|gg|jd}|||d}|S)N)r r)r r)r3)r[r)rprr4rSr3ne)r-Z pad_tokenr dummy_inputsr"r"r#rs "zMvpPreTrainedModel.dummy_inputsN) r=r>r?rZ config_classZbase_model_prefixZsupports_gradient_checkpointingrpropertyrr"r"r"r#rs rcseZdZdZ ddedeejdeeffdd Z d d Z d d Z       dd ee j dee jdee jdee jdeedeedeedeeeffddZZS) MvpEncodera Transformer encoder consisting of *config.encoder_layers* self attention layers. Each layer is a [`MvpEncoderLayer`]. Args: config: MvpConfig embed_tokens (nn.Embedding): output embedding use_prompt (bool): whether to use prompt NFrp embed_tokens use_promptcstj|_j|_j}j|_j|_ j r!t |nd|_ |dur,||_n tj||j|_tj||_tfddtjD|_t||_||_|rej|_tjj|_d|_| dS)N?cg|]}tqSr")ro.0rkrpr"r# +z'MvpEncoder.__init__..F)!r+r,rHZencoder_layerdrop layerdroprqrrmax_position_embeddingsZmax_source_positionsscale_embeddingmathsqrt embed_scalerrr vocab_sizer%embed_positions ModuleListrangeZencoder_layerslayersrulayernorm_embeddingrrrrrr~gradient_checkpointing post_init)r-rprrrFr.rr#r,s4    zMvpEncoder.__init__cC|jSrrr-r"r"r#get_input_embeddings;zMvpEncoder.get_input_embeddingscC ||_dSrrr-valuer"r"r#set_input_embeddings> zMvpEncoder.set_input_embeddingsrr[ head_mask inputs_embedsr^output_hidden_states return_dictr_c Cs|dur|n|jj}|dur|n|jj}|dur|n|jj}|dur*|dur*td|dur<|}|j} |d| d}n|durT|dd} |dddddf}ntd|durd|||j }| |} || } | | } t j j| |j|jd} |jrt|j|j} || } |durt||j}|rdnd}|rdnd}|dur|dt|jkrtdt|jd |dd t|jD]n\}}|r|| f}d }|jrtg}||jkrd }|rd }nC|jr|jr||j | ||dur||nd|jr | |nd|}n|| ||dur||nd|jr&| |nd|d}|d} |r9||df}q|rB|| f}|sQt!dd| ||fDSt"| ||dS)a~ Args: input_ids (`torch.LongTensor` 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 (`torch.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 (`torch.Tensor` of shape `(encoder_layers, encoder_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**. 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. NzDYou cannot specify both input_ids and inputs_embeds at the same timerz5You have to specify either input_ids or inputs_embedsrbr"rz&The head_mask should be specified for  layers, but it is for .FT)NN)r\r~r^rcs|] }|dur|VqdSrr"rvr"r"r# sz%MvpEncoder.forward..last_hidden_staterX attentions)#rpr^ruse_return_dictr!rrTrerrrrrrirHrdrr4r5rrhr3r~r r2lenr enumeraterandrr_gradient_checkpointing_func__call__tupler)r-rr[rrr^rrinput input_shapeZ embed_posrXrr~Zencoder_statesZall_attentionsidxZ encoder_layerZto_dropdropout_probability layer_outputsr"r"r#r9As.            zMvpEncoder.forwardNF)NNNNNNN)r=r>r?r@rrrrrnr,rrr4 LongTensorrBrrrrr9rCr"r"r.r#r sH & rcseZdZdZ ddedeejdeeffdd Z d d Z d d Z            dd ee j dee jdee jdee j dee jdee jdeee jdee jdeedeedeedeedeeeffddZZS) MvpDecoderz Transformer decoder consisting of *config.decoder_layers* layers. Each layer is a [`MvpDecoderLayer`] Args: config: MvpConfig embed_tokens (nn.Embedding): output embedding use_prompt (bool): whether to use prompt NFrprrcstj|_j|_j|_j|_j rt j nd|_ |dur*||_n tjj |j|_tjj |_tfddtjD|_tj |_||_|roj|_tjj|_tjj|_d|_ |!dS)Nrcrr")rrrr"r#rrz'MvpDecoder.__init__..F)"r+r,rHZdecoder_layerdroprrrrZmax_target_positionsrrrrqrrrrrr%rrrZdecoder_layersrrurrrrrr~rrr)r-rprrr.rr#r,s<   zMvpDecoder.__init__cCrrrrr"r"r#rrzMvpDecoder.get_input_embeddingscCrrrrr"r"r#rrzMvpDecoder.set_input_embeddingsrr[rrrcross_attn_head_maskpast_key_valuesrrr^rrr_c !Cs(| dur| n|jj} | dur| n|jj} | dur| n|jj} | dur$| n|jj} |dur4|dur4td|durF|} |j}|d|d}n|dur^|dd}|dddddf} ntd|duro|ddjdnd}|dur}| ||j }t ||||}|dur|durt ||j |dd}|| |}||}||}tjj||j|jd}|jrt|j|j}||}||}|jr|jr| rtd d } | rd nd}| rd nd}| r|durd nd}| rd nd}t||gd d gD]*\}}|dur%|dt |j!kr%td|dt |j!d|ddqt"|j!D]\}}| r7||f7}|jrHt#g}||j$krHq+|durQ||nd}|jr|jr|%|j&|||||durl||nd|durv||nd|jr||nd|jr||ndd| | }n1||||||dur||nd|dur||nd|jr||nd|jr||nd|| | d }|d}| r||| rdndf7}| r||df7}|dur||df7}q+| r||f7}| r|nd} | s t'dd|| |||fDSt(|| |||dS)a  Args: input_ids (`torch.LongTensor` 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 (`torch.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 (`torch.FloatTensor` 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 (`torch.LongTensor` 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 (`torch.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**. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules in the decoder 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(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`): Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`. Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential 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. 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_embedsrr))rjrbzZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...Fr"rrzThe `z` should be specified for rr) r[rrr\rr~rrZr^rr rcsrrr"rr"r"r#rsz%MvpDecoder.forward..)rrrXrcross_attentions))rpr^rrrr!rrTrerrrr r2rrrrirHrdrr4r5rrhr3r~rrloggerZ warning_onceziprrrrrrrrr)!r-rr[rrrrrrrr^rrrrr0Z positionsrXrr~rZall_hidden_statesZall_self_attnsZall_cross_attentionsZnext_decoder_cacheZ attn_maskZ mask_namerZ decoder_layerrrZrZ next_cacher"r"r#r9sP                 zMvpDecoder.forwardr) NNNNNNNNNNNN)r=r>r?r@rrrrrnr,rrr4rrBrrrrrr9rCr"r"r.r#rsf (       rc$s(eZdZdgZddgZdeffdd ZddZd d Zd d Z d dZ ddZ e               d$de ejde ejde ejde ejde ejde ejde ejde eejde eejde ejde ejde ede ede ed e ed!eeeff d"d#ZZS)%MvpModelfinal_logits_biasencoder.embed_tokens.weightdecoder.embed_tokens.weightrpcsdt||j|j}}|j|_t||j||_t ||j|j|_ t ||j|j|_ | dSr)r+r,rrrrrrqsharedrencoderrdecoderr)r-rprrr.r"r#r,s  zMvpModel.__init__cCrr)rrr"r"r#rrzMvpModel.get_input_embeddingscCs||_|j|j_|j|j_dSr)rrrrrr"r"r#rs zMvpModel.set_input_embeddingscCrr)rrr"r"r# get_encoderrzMvpModel.get_encodercCrrrrr"r"r# get_decoderrzMvpModel.get_decodercCsF|jsJd|d|jjd|jjd|jjddS)NzHIf you want to use lightweight tuning, make sure that `use_prompt=True`.FT)rrequires_grad_rr~rrrr"r"r#set_lightweight_tunings  zMvpModel.set_lightweight_tuningNrr[decoder_input_idsdecoder_attention_maskrdecoder_head_maskrencoder_outputsrrdecoder_inputs_embedsrr^rrr_cCsH|dur| dur|durtdt||jj|jj}| dur | n|jj} |dur*|n|jj}| dur4| n|jj} |dur>|n|jj}|durS|j |||| | ||d}n$|rwt |t swt |dt |dkrh|dndt |dkrs|dndd}|j |||d|||| | | | ||d }|s||St|j|j|j|j|j|j|j|jd S) a" decoder_input_ids (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Indices of decoder input sequence tokens in the vocabulary. Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are decoder input IDs?](../glossary#decoder-input-ids) Mvp uses the `eos_token_id` as the starting token for `decoder_input_ids` generation. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). For translation and summarization training, `decoder_input_ids` should be provided. If no `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right for denoising pre-training following the paper. decoder_attention_mask (`torch.LongTensor` of shape `(batch_size, target_sequence_length)`, *optional*): Default behavior: generate a tensor that ignores pad tokens in `decoder_input_ids`. Causal mask will also be used by default. If you want to change padding behavior, you should read [`modeling_mvp._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules in the decoder. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. NzIf no `decoder_input_ids` or `decoder_inputs_embeds` are passed, `input_ids` cannot be `None`. Please pass either `input_ids` or `decoder_input_ids` or `decoder_inputs_embeds`.)rr[rrr^rrrrr)r rr[rrrrrrrr^rr)rrdecoder_hidden_statesdecoder_attentionsrencoder_last_hidden_staterencoder_attentions)r!r$rprrr^rrrrrrrrrrrrXrr)r-rr[rrrrrr rrr rr^rrZdecoder_outputsr"r"r#r9sn2 zMvpModel.forwardNNNNNNNNNNNNNNN)r=r>r?Z"_keys_to_ignore_on_load_unexpected_tied_weights_keysrr,rrrrrrrr4rrBrrrnrrrr9rCr"r"r.r#rst       rze The MVP Model with a language modeling head. Can be used for various text generation tasks. )Z custom_introc&seZdZgdZdeffdd ZddZddZ d/d ed e ed e de j ffdd Z d edd fddZddZddZddZe d0de ejde ejde ejde ejde ejde ejde ejd e eejd!e eejd"e ejd#e ejd$e ejd%e e d&e e d'e e d(e e deeeff"d)d*Zd$ejfd+d,Zed-d.ZZS)1MvpForConditionalGeneration)rrlm_head.weightrpcsXt|t||_|dtd|jjjft j |j |jjjdd|_ | dS)NrrFrK)r+r,rrregister_bufferr4rgrr&rrNrqlm_headrr}r.r"r#r,s   z$MvpForConditionalGeneration.__init__cC |jSr)rrrr"r"r#rrz'MvpForConditionalGeneration.get_encodercCrr)rrrr"r"r#rrz'MvpForConditionalGeneration.get_decoderNTnew_num_tokenspad_to_multiple_of mean_resizingr_cst|||}|||Sr)r+resize_token_embeddings_resize_final_logits_bias)r-rrrnew_embeddingsr.r"r#rs z3MvpForConditionalGeneration.resize_token_embeddingscCsj|jjd}||kr|jddd|f}ntjd||f|jjd}tj|j|gdd}|d|dS)Nrrrr`r)rrr4rgr3rfr)r-rZold_num_tokensZnew_biasZ extra_biasr"r"r#rs z5MvpForConditionalGeneration._resize_final_logits_biascCrrrrr"r"r#get_output_embeddingsrz1MvpForConditionalGeneration.get_output_embeddingscCrrrr-rr"r"r#set_output_embeddingsrz1MvpForConditionalGeneration.set_output_embeddingscC|j|jddSrrrrrrr"r"r#r z2MvpForConditionalGeneration.set_lightweight_tuningrr[rrrrrr rrr labelsrr^rrcCs|dur|n|jj}| dur)| rtdd} |dur)| dur)t| |jj|jj}|j||||||||| | | | |||d}||d|j }d}| dur^t }|| d|jj | d}|st|f|dd}|durr|f|S|St |||j|j|j|j|j|j|jd S) a 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) Mvp 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`). For translation and summarization training, `decoder_input_ids` should be provided. If no `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right for denoising pre-training following the paper. decoder_attention_mask (`torch.LongTensor` 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. If you want to change padding behavior, you should read [`modeling_mvp._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_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**. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (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 of summarization: Fine-tuning a model ```python >>> import torch >>> from transformers import AutoTokenizer, MvpForConditionalGeneration >>> tokenizer = AutoTokenizer.from_pretrained("RUCAIBox/mvp") >>> model = MvpForConditionalGeneration.from_pretrained("RUCAIBox/mvp") >>> inputs = tokenizer( ... "Summarize: You may want to stick it to your boss and leave your job, but don't do it if these are your reasons.", ... return_tensors="pt", ... ) >>> labels = tokenizer("Bad Reasons To Quit Your Job", return_tensors="pt")["input_ids"] >>> loss = model(**inputs, labels=labels).loss >>> loss.backward() ``` Inference after the model fine-tuned ```python >>> with torch.no_grad(): ... generated_ids = model.generate(**inputs) >>> generated_text = tokenizer.batch_decode(generated_ids, skip_special_tokens=True) ``` NzJThe `use_cache` argument is changed to `False` since `labels` is provided.F)r[rr rrrrrrr rr^rrrrr losslogitsrr r rrrr)rprrwarningr$rrrrrrrTrrrr r rrrr)r-rr[rrrrrr rrr r$rr^rrrZ lm_logitsZmasked_lm_lossloss_fctoutputr"r"r#r9sZQ z#MvpForConditionalGeneration.forwardcCst||jj|jjSr)r$rprr)r-r$r"r"r#%prepare_decoder_input_ids_from_labels;szAMvpForConditionalGeneration.prepare_decoder_input_ids_from_labelscsBd}|D]}|tfdd|ddD|ddf7}q|S)Nr"c3$|] }|d|jVqdSr<Z index_selectrhr3rZ past_statebeam_idxr"r#rD"z=MvpForConditionalGeneration._reorder_cache..r)rrr0Zreordered_pastZ layer_pastr"r/r#_reorder_cache>s z*MvpForConditionalGeneration._reorder_cache)NTNNNNNNNNNNNNNNNN) r=r>r?rrr,rrrArrnrrrrrr rrr4rrBrrrrrr9r+ staticmethodr4rCr"r"r.r#rs         rz Mvp model with a sequence classification/head on top (a linear layer on top of the pooled output) e.g. for GLUE tasks. c$seZdZddgZdeffdd ZddZe               dd ee j d ee j d ee j d ee j d ee j dee j dee j dee e j dee j dee j dee j deedeedeedeedeeeff ddZZS)MvpForSequenceClassificationrrrpc sBtj|fi|t||_t|j|j|j|j|_| dSr) r+r,rrrrq num_labelsZclassifier_dropoutclassification_headr)r-rpkwargsr.r"r#r,Ss  z%MvpForSequenceClassification.__init__cCr!r)rrr9rrr"r"r#r`r#z3MvpForSequenceClassification.set_lightweight_tuningNrr[rrrrrr rr r$rr^rrr_cCs.|dur|n|jj}| durd} |dur!| dur!td|jj|j||||||||| | | | ||d}|d}||jj|j }t t | ddkrStd||ddf|dd|ddddddf}||}d}| dur|jjdur|jjdkrd |j_n|jjdkr| jt jks| jt jkrd |j_nd |j_|jjd krt}|jjdkr||| }n,||| }n&|jjd krt}||d|jj| d}n|jjd krt}||| }|s|f|dd}|dur|f|S|St|||j|j|j|j|j |j!|j"d S) a 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) Mvp 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`). For translation and summarization training, `decoder_input_ids` should be provided. If no `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right for denoising pre-training following the paper. decoder_attention_mask (`torch.LongTensor` 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. If you want to change padding behavior, you should read [`modeling_mvp._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_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**. 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). Example of single-label classification: Fine-tuning a model on `num_labels` classes ```python >>> import torch >>> from transformers import AutoTokenizer, MvpForSequenceClassification >>> num_labels = 2 # for example, this is a binary classification task >>> tokenizer = AutoTokenizer.from_pretrained("RUCAIBox/mvp") >>> model = MvpForSequenceClassification.from_pretrained("RUCAIBox/mvp", num_labels=num_labels) >>> inputs = tokenizer("Classify: Hello, my dog is cute", return_tensors="pt") >>> labels = torch.tensor(1) # the real label for inputs >>> loss = model(**inputs, labels=labels).loss >>> loss.backward() ``` Inference after the model fine-tuned ```python >>> with torch.no_grad(): ... logits = model(**inputs).logits >>> predicted_class_id = logits.argmax() ``` NFz8Passing input embeddings is currently not supported for r[rrrrrr rr rr^rrrrz7All examples must have the same number of tokens.rZ regressionZsingle_label_classificationZmulti_label_classificationr%)#rprNotImplementedErrorr/r=reqZ eos_token_idrhr3rr4Zunique_consecutivesumr!rTrer9Z problem_typer8r2r6rAr squeezerrrrr r rrrr)r-rr[rrrrrr rr r$rr^rrrrXZeos_maskZsentence_representationr'r&r)r*r"r"r#r9dsM $    $       z$MvpForSequenceClassification.forwardr)r=r>r?rrr,rrrr4rrBrrrnrrrr9rCr"r"r.r#r7Jsj       r7c&seZdZddgZfddZddZe                ddeej d eej d eej d eej d eej d eej deej dee ej deej deej deej deej dee dee dee dee deeeff"ddZZS)MvpForQuestionAnsweringrrcsBt|d|_|j|_t||_t|j|j|_| dSr() r+r,r8rrrrN hidden_size qa_outputsrr}r.r"r#r,s   z MvpForQuestionAnswering.__init__cCr!r)rrrBrrr"r"r#r r#z.MvpForQuestionAnswering.set_lightweight_tuningNrr[rrrrrr start_positions end_positionsrr rr^rrr_cCsz|dur|n|jj}| dur| durd} |j||||||||| | | |||d}|d}||}|jddd\}}|d}|d}d}| dur| durt| dkr^| d} t| dkrk| d} |d}| d|} | d|} t |d}||| }||| }||d }|s||f|dd}|dur|f|S|St ||||j |j |j|j|j|j|jd S) aX 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) Mvp 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`). For translation and summarization training, `decoder_input_ids` should be provided. If no `decoder_input_ids` is provided, the model will create this tensor by shifting the `input_ids` to the right for denoising pre-training following the paper. decoder_attention_mask (`torch.LongTensor` 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. If you want to change padding behavior, you should read [`modeling_mvp._prepare_decoder_attention_mask`] and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more information on the default strategy. cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_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**. Example: Fine-tuning a model for extrative question answering, and our model also supports generative question answering using `BartForConditionalGeneration` ```python >>> import torch >>> from transformers import AutoTokenizer, MvpForQuestionAnswering >>> tokenizer = AutoTokenizer.from_pretrained("RUCAIBox/mvp") >>> model = MvpForQuestionAnswering.from_pretrained("RUCAIBox/mvp") >>> inputs = tokenizer( ... "Answer the following question: Who was Jim Henson? [SEP] Jim Henson was a nice puppet", ... return_tensors="pt", ... ) >>> target_start_index = torch.tensor([18]) >>> target_end_index = torch.tensor([19]) >>> loss = model(**inputs, start_positions=target_start_index, end_positions=target_end_index).loss >>> loss.backward() ``` Inference after the model fine-tuned ```python >>> with torch.no_grad(): ... outputs = model(**inputs) >>> answer_start_index = outputs.start_logits.argmax() >>> answer_end_index = outputs.end_logits.argmax() >>> predict_answer_tokens = inputs.input_ids[0, answer_start_index : answer_end_index + 1] >>> predict_answer = tokenizer.decode(predict_answer_tokens) ``` NFr;rrrr`)Z ignore_indexr)) r& start_logits end_logitsrr r rrrr)rprrrBrr?rVrrerrrrr r rrrr)r-rr[rrrrrr rCrDrr rr^rrrZsequence_outputr'rErFZ total_lossZ ignored_indexr)Z start_lossZend_lossr*r"r"r#r9spS           zMvpForQuestionAnswering.forwardr5)r=r>r?rr,rrrr4rBrrrrnrrrr9rCr"r"r.r#r@sp        r@cs(eZdZdZfddZddZZS)MvpDecoderWrapperz This wrapper class is a helper class to correctly load pretrained checkpoints when the causal language model is used in combination with the [`EncoderDecoderModel`] framework. cst|t||_dSr)r+r,rrr}r.r"r#r,s zMvpDecoderWrapper.__init__cOs|j|i|Srr)r-argsr:r"r"r#r9szMvpDecoderWrapper.forward)r=r>r?r@r,r9rCr"r"r.r#rGs rGc seZdZdgZfddZddZddZdd Zd d Zd d Z ddZ ddZ e             d%de ejde ejde ejde ejde ejde ejde eejde ejde ejde ede ede ede ed eeeffd!d"Zed#d$ZZS)&MvpForCausalLMrcsNt|}d|_d|_t|t||_tj |j |j dd|_ | dS)NTFrK)copydeepcopyrIZis_encoder_decoderr+r,rGrrrNrArrrr}r.r"r#r,s    zMvpForCausalLM.__init__cCs |jjjSrrrrrr"r"r#rrz#MvpForCausalLM.get_input_embeddingscCs||jj_dSrrLrr"r"r#rsz#MvpForCausalLM.set_input_embeddingscCrrrrr"r"r#rrz$MvpForCausalLM.get_output_embeddingscCrrrrr"r"r#r rz$MvpForCausalLM.set_output_embeddingscCs ||j_dSrrr)r-rr"r"r# set_decoders zMvpForCausalLM.set_decodercCs|jjSrrMrr"r"r#rszMvpForCausalLM.get_decodercCr!rr"rr"r"r#rr#z%MvpForCausalLM.set_lightweight_tuningNrr[rrrrrrr$rr^rrr_cCs| dur| n|jj} | dur| n|jj} | dur| n|jj} |jj||||||||| | | | d }||d}d}| durNt}||d|jj | d}| sd|f|dd}|durb|f|S|St |||j |j |j |jdS)a cross_attn_head_mask (`torch.Tensor` of shape `(decoder_layers, decoder_attention_heads)`, *optional*): Mask to nullify selected heads of the cross-attention modules. Mask values selected in `[0, 1]`: - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. Example: ```python >>> from transformers import AutoTokenizer, MvpForCausalLM >>> tokenizer = AutoTokenizer.from_pretrained("RUCAIBox/mvp") >>> model = MvpForCausalLM.from_pretrained("RUCAIBox/mvp", add_cross_attention=False) >>> inputs = tokenizer("Hello, my dog is cute", return_tensors="pt") >>> outputs = model(**inputs) >>> logits = outputs.logits >>> list(logits.shape) [1, 8, 50267] ```Nr rrr)r&r'rrXrr)rpr^rrrrrrrTrrrrXrr)r-rr[rrrrrrr$rr^rrrr'r&r)r*r"r"r#r9sD,zMvpForCausalLM.forwardcs.d}|D]}|tfdd|Df7}q|S)Nr"c3r,r<r-r.r/r"r#r4r1z0MvpForCausalLM._reorder_cache..r2r3r"r/r#r4/s zMvpForCausalLM._reorder_cache) NNNNNNNNNNNNN)r=r>r?rr,rrrr rNrrrrr4rrBrrrnrrrr9r6r4rCr"r"r.r#rIsn          UrI)rIrr@r7rr)=r@rJrtypingrrrrr4Ztorch.utils.checkpointrZtorch.nnrrr Z activationsr Z generationr Zmodeling_attn_mask_utilsr rZmodeling_outputsrrrrrrrZmodeling_utilsrutilsrrZconfiguration_mvprZ get_loggerr=rrBrAr$rr%ModulerDrorrrrrrrrr7r@rGrI__all__r"r"r"r#sf   $   H~E@-*