o ߥi@sdZddlZddlZddlZddlmZddlmZddlmZm Z m Z m Z ddl m Z ddlmZmZddlmZdd lmZdd lmZdd lmZdd lmZd dlmZeZdZGdddejZ GdddejZ!GdddejZ"GdddejZ#GdddejZ$GdddejZ%GdddejZ&GdddejZ'Gd d!d!ejZ(Gd"d#d#eeZ)ej*ej+e j,d$Gd%d&d&e)Z-dS)'z PyTorch MegatronBERT model.N)nn)ACT2FN)PreTrainedModelapply_chunking_to_forward find_pruneable_heads_and_indicesprune_linear_layer)Models)Model TorchModel)MODELS)AttentionBackboneModelOutput)Tasks) get_logger)parse_labels_in_order)MegatronBertConfigrcs<eZdZdZfddZ     d dejfddZZS) MegatronBertEmbeddingszGConstruct the embeddings from word, position and token_type embeddings.csttj|j|j|jd|_t|j|j|_ t|j |j|_ t |j |_|dt|jdt|dd|_dS)N) padding_idx position_ids)rposition_embedding_typeabsolute)super__init__r Embedding vocab_size hidden_size pad_token_idword_embeddingsmax_position_embeddingsposition_embeddingstype_vocab_sizetoken_type_embeddingsDropouthidden_dropout_probdropoutregister_buffertorcharangeexpandgetattrrselfconfig __class__`/home/ubuntu/.local/lib/python3.10/site-packages/modelscope/models/nlp/megatron_bert/backbone.pyr.s*  zMegatronBertEmbeddings.__init__Nrreturnc Cs|dur |}n|dd}|d}|dur&|jdd|||f}|dur5tj|tj|jjd}|dur>||}||}||} |jdkrU| |} | | 7} | | } | S)Nrrdtypedevicer) sizerr'zeroslongr5rr"rr r%) r, input_idstoken_type_idsr inputs_embedspast_key_values_length input_shape seq_lengthr" embeddingsr r0r0r1forwardHs.        zMegatronBertEmbeddings.forward)NNNNr) __name__ __module__ __qualname____doc__rr'Tensorr@ __classcell__r0r0r.r1r+s rcsJeZdZd fdd ZdejdejfddZ      d d d ZZS) MegatronBertSelfAttentionNcst|j|jdkrt|dstd|jd|jd|j|_t|j|j|_|j|j|_t |j|j|_ t |j|j|_ t |j|j|_ t |j|_|p\t|dd|_|jdksh|jd kry|j|_t d |jd |j|_|j|_dS) Nrembedding_sizezThe hidden size (z6) is not a multiple of the number of attention heads ()rr relative_keyrelative_key_queryr)rrrnum_attention_headshasattr ValueErrorintattention_head_size all_head_sizerLinearquerykeyvaluer#attention_probs_dropout_probr%r*rrrdistance_embedding is_decoder)r,r-rr.r0r1rrs8    z"MegatronBertSelfAttention.__init__xr2cCs6|dd|j|jf}||}|ddddS)NrrrLr)r6rMrQviewpermute)r,rZ new_x_shaper0r0r1transpose_for_scoress  z.MegatronBertSelfAttention.transpose_for_scoresFcCs||}|du} | r|dur|d} |d} |}nP| r/|||} |||} |}n;|durZ|||} |||} tj|d| gdd} tj|d| gdd} n|||} |||} ||} |jrv| | f}t| | dd} |j dks|j dkr| d}tj |tj |j d dd}tj |tj |j d dd}||}|||jd}|j| jd }|j dkrtd | |}| |} n|j dkrtd | |}td | |}| ||} | t|j} |dur| |} tjj| dd}||}|dur||}t|| }|dddd }| dd|jf}||}|r9||fn|f}|jrE||f}|S)NrrrLdimrrJrKr3)r4zbhld,lrd->bhlrzbhrd,lrd->bhlrr[)rTr_rUrVr'catrYmatmul transposerr6r(r8r5r\rXrtor4einsummathsqrtrQr functionalsoftmaxr%r] contiguousrR)r, hidden_statesattention_mask head_maskencoder_hidden_statesencoder_attention_maskpast_key_valueoutput_attentionsmixed_query_layeris_cross_attention key_layer value_layer query_layerattention_scoresr>position_ids_lposition_ids_rdistancepositional_embeddingrelative_position_scoresrelative_position_scores_queryrelative_position_scores_keyattention_probs context_layernew_context_layer_shapeoutputsr0r0r1r@s               z!MegatronBertSelfAttention.forwardNNNNNNF) rArBrCrr'rEr_r@rFr0r0r.r1rGps rGc$eZdZfddZddZZS)MegatronBertSelfOutputcs.tt|j|j|_t|j|_dSr) rrrrSrdenser#r$r%r+r.r0r1r zMegatronBertSelfOutput.__init__cC||}||}||Srrr%)r,rmresidualr0r0r1r@ s  zMegatronBertSelfOutput.forwardrArBrCrr@rFr0r0r.r1r rcs:eZdZfddZddZ      d ddZZS) MegatronBertAttentioncs>ttj|j|jd|_t||_t ||_ t |_ dS)Neps) rrr LayerNormrlayer_norm_epslnrGr,routputset pruned_headsr+r.r0r1rs    zMegatronBertAttention.__init__cCst|dkrdSt||jj|jj|j\}}t|jj||j_t|jj||j_t|jj ||j_ t|j j |dd|j _ |jjt||j_|jj|jj|j_ |j ||_dS)Nrrr`)lenrr,rMrQrrrTrUrVrrrRunion)r,headsindexr0r0r1 prune_headss   z!MegatronBertAttention.prune_headsNFc CsF||}||||||||} || d|} | f| dd} | S)Nrr)rr,r) r,rmrnrorprqrrrs ln_outputs self_outputsattention_outputrr0r0r1r@-s   zMegatronBertAttention.forwardr)rArBrCrrr@rFr0r0r.r1rs rcr)MegatronBertIntermediatecsDtt|j|j|_t|jt rt |j|_ dS|j|_ dSr) rrrrSrintermediate_sizer isinstance hidden_actstrrintermediate_act_fnr+r.r0r1rJs   z!MegatronBertIntermediate.__init__cCs||}||}|Sr)rr)r,rmr0r0r1r@Rs  z MegatronBertIntermediate.forwardrr0r0r.r1rHs rcs8eZdZfddZdejdejdejfddZZS)MegatronBertOutputcs.tt|j|j|_t|j|_ dSr) rrrrSrrrr#r$r%r+r.r0r1r[rzMegatronBertOutput.__init__rm input_tensorr2cCrrr)r,rmrr0r0r1r@`s  zMegatronBertOutput.forward)rArBrCrr'rEr@rFr0r0r.r1rYs rcs:eZdZfddZ      d ddZddZZS) MegatronBertLayercst|j|_d|_t||_|j|_|j|_|jr+|js&t|dt||_ t j |j |j d|_t||_t||_dS)Nrz> should be used as a decoder model if cross attention is addedr)rrchunk_size_feed_forward seq_len_dimr attentionrYadd_cross_attention TypeErrorcrossattentionrrrrrr intermediaterrr+r.r0r1rjs    zMegatronBertLayer.__init__NFc Cs|dur |ddnd}|j|||||d} | d} |jr(| dd} | d} n| dd} d} |jro|durot|dsDtd|d|durN|d dnd}|| ||||||}|d} | |dd} |d} | | } t|j|j|j| }|f| } |jr| | f} | S) NrLrsrrrrrrz'If `encoder_hidden_states` are passed, z` has to be instantiated with cross-attention layers by setting `config.add_cross_attention=True`rb) rrYrNAttributeErrorrrfeed_forward_chunkrr)r,rmrnrorprqrrrsself_attn_past_key_valueself_attention_outputsrrpresent_key_valuecross_attn_present_key_valuecross_attn_past_key_valuecross_attention_outputs layer_outputr0r0r1r@{sp        zMegatronBertLayer.forwardcCs$||}||}|||}|Sr)rrr)r,r ln_outputintermediate_outputrr0r0r1rs   z$MegatronBertLayer.feed_forward_chunkr)rArBrCrr@rrFr0r0r.r1rhs  Ercs8eZdZfddZ         dddZZS) MegatronBertEncodercsNt|_tfddtjD|_tjj j d|_ d|_ dS)Ncsg|]}tqSr0)r).0_r-r0r1 sz0MegatronBertEncoder.__init__..rF) rrr-r ModuleListrangenum_hidden_layerslayerrrrrgradient_checkpointingr+r.rr1rs   zMegatronBertEncoder.__init__NFTc  sl| rdnd} r dnd} r|jjrdnd} |rdnd}t|jD]o\}}| r,| |f} |dur4||nd}|dur>||nd|jre|jre|rOtdd}fdd}tj j |||||||}n ||||||}|d}|r|||df7}r| |df} |jjr| |d f} q!| |}| r| |f} | st d d ||| | | fDSt ||| | | d S) Nr0zZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...Fcsfdd}|S)Ncsg|RSrr0)inputs)modulersrrr0r1custom_forwards zRMegatronBertEncoder.forward..create_custom_forward..custom_forwardr0)rrr)rr1create_custom_forwardsz:MegatronBertEncoder.forward..create_custom_forwardrrrrLcss|] }|dur|VqdSrr0)rvr0r0r1 's z.MegatronBertEncoder.forward..)last_hidden_statepast_key_valuesrm attentionscross_attentions)r-r enumeraterrtrainingloggerwarningr'utils checkpointrtupler )r,rmrnrorprqr use_cachersoutput_hidden_states return_dictall_hidden_statesall_self_attentionsall_cross_attentionsnext_decoder_cachei layer_modulelayer_head_maskr layer_outputsr0rr1r@s         zMegatronBertEncoder.forward) NNNNNNFFTrr0r0r.r1rs rcr)MegatronBertPoolercs*tt|j|j|_t|_dSr)rrrrSrrTanh activationr+r.r0r1r:s zMegatronBertPooler.__init__cCs(|dddf}||}||}|S)Nr)rr)r,rmfirst_token_tensor pooled_outputr0r0r1r@?s  zMegatronBertPooler.forwardrr0r0r.r1r8rrcsTeZdZdZeZdZdZdgZfddZ ddZ dd d Z e fd d Z ZS)MegatronBertPreTrainedModelz An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained models. bertTrc s*tj|jfi|tt||dSr)rr name_or_pathr )r,r-kwargsr.r0r1rSsz$MegatronBertPreTrainedModel.__init__cCs~t|tjtjfr|jjjd|jjdnt|tj r(|j j |jj dt|tjr;|j dur=|j j dSdSdS)zInitialize the weightsg)meanstdg?N) rrrSrweightdatanormal_r-initializer_rangerbiaszero_fill_)r,rr0r0r1 _init_weightsWs  z)MegatronBertPreTrainedModel._init_weightsFcCst|tr ||_dSdSr)rrr)r,rrVr0r0r1_set_gradient_checkpointingds  z7MegatronBertPreTrainedModel._set_gradient_checkpointingc sn|dd}|dd}t||fi|}|dur%tdi|}||}n tt|jdd|i|}||_|S)ahInstantiate the model. Args: kwargs: Input args. model_dir: The model dir used to load the checkpoint and the label information. num_labels: An optional arg to tell the model how many classes to initialize. Method will call utils.parse_label_mapping if num_labels not supplied. If num_labels is not found, the model will use the default setting (2 classes). Returns: The loaded model, which is initialized by transformers.PreTrainedModel.from_pretrained model_dirNcfgpretrained_model_name_or_pathr0)poprrrr from_pretrainedr)clsrrr model_argsr-modelr.r0r1 _instantiatehs   z(MegatronBertPreTrainedModel._instantiate)F)rArBrCrDr config_classbase_model_prefixsupports_gradient_checkpointing_keys_to_ignore_on_load_missingrrr classmethodrrFr0r0r.r1rHs  r) group_key module_namecsdeZdZdZdfdd ZddZddZd d Z dd efd dZ Z S)MegatronBertModelaThe bare MegatronBert Model transformer outputting raw hidden-states without any specific head on top.", 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 ([`MegatronBertConfig`]): 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. The model can behave as an encoder (with only self-attention) as well as a decoder, in which case a layer of cross-attention is added between the self-attention layers, following the architecture described in [Attention is all you need](https://arxiv.org/abs/1706.03762) by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin. To behave as an decoder the model needs to be initialized with the `is_decoder` argument of the configuration set to `True`. To be used in a Seq2Seq model, the model needs to initialized with both `is_decoder` argument and `add_cross_attention` set to `True`; an `encoder_hidden_states` is then expected as an input to the forward pass. TcsDt|||_t||_t||_|rt|nd|_| dSr) rrr-rr?rencoderrpooler post_init)r,r-add_pooling_layerr.r0r1rs    zMegatronBertModel.__init__cCs|jjSrr?r)r,r0r0r1get_input_embeddingssz&MegatronBertModel.get_input_embeddingscCs ||j_dSrr )r,rVr0r0r1set_input_embeddingss z&MegatronBertModel.set_input_embeddingscCs*|D]\}}|jj|j|qdS)z Prunes heads of the model. heads_to_prune: dict of {layer_num: list of heads to prune in this layer} See base class PreTrainedModel N)itemsrrrr)r,heads_to_prunerrr0r0r1 _prune_headsszMegatronBertModel._prune_headsNr2c Ks | dur| n|jj} | dur| n|jj} | dur| n|jj} |jjr-| dur(| n|jj} nd} |dur;|dur;td|durD|}n|durQ|dd}ntd|\}}|dur`|jn|j}| durp| ddj dnd}|durt j |||f|d}|durt j |t j |d }||||}|jjr|dur|\}}}||f}|durt j ||d}||}nd}|||jj}|j|||||d }|j|||||| | | | | d }|d}|jdur||nd}| s||f|d dSt|||j|j|j|jd S)a4 Args: input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): Indices of input sequence tokens in the vocabulary. Indices can be obtained using [`BertTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. [What are input IDs?](../glossary#input-ids) 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) token_type_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Segment token indices to indicate first and second portions of the inputs. Indices are selected in `[0, 1]`: - 0 corresponds to a *sentence A* token, - 1 corresponds to a *sentence B* token. [What are token type IDs?](../glossary#token-type-ids) position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, config.max_position_embeddings - 1]`. [What are position IDs?](../glossary#position-ids) 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 [`~file_utils.ModelOutput`] instead of a plain tuple. encoder_hidden_states (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): Sequence of hidden-states at the output of the last layer of the encoder. Used in the cross-attention if the model is configured as a decoder. encoder_attention_mask (`torch.FloatTensor` of shape `(batch_size, sequence_length)`, *optional*): Mask to avoid performing attention on the padding token indices of the encoder input. This mask is used in the cross-attention if the model is configured as a decoder. Mask values selected in `[0, 1]`: - 1 for tokens that are **not masked**, - 0 for tokens that are **masked**. 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)`. 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`). Others (**kwargs) some additional parameters might passed in from upstream pipeline, which not influence the results. NFzDYou cannot specify both input_ids and inputs_embeds at the same timerz5You have to specify either input_ids or inputs_embedsrrL)r5r3)r9rr:r;r<) rnrorprqrrrsrrr)r pooler_outputrrmrr)r-rsruse_return_dictrYrrOr6r5shaper'onesr7r8get_extended_attention_maskinvert_attention_mask get_head_maskrr?rr r rrmrr)r,r9rnr:rror;rprqrrrsrrrr= batch_sizer>r5r<extended_attention_maskencoder_batch_sizeencoder_sequence_lengthrencoder_hidden_shapeencoder_extended_attention_maskembedding_outputencoder_outputssequence_outputrr0r0r1r@sd     zMegatronBertModel.forward)T) NNNNNNNNNNNNN) rArBrCrDrr rrr r@rFr0r0r.r1rs,  r).rDrhr'torch.utils.checkpointrtransformers.activationsrtransformers.modeling_utilsrrrrmodelscope.metainformodelscope.modelsr r modelscope.models.builderr modelscope.outputsr modelscope.utils.constantr modelscope.utils.loggerrmodelscope.utils.nlp.utilsr configurationrr_CONFIG_FOR_DOCModulerrGrrrrrrrrregister_modulebackbone megatron_bertrr0r0r0r1s@         E7_q<