U 4Af @sBdZddlmZddlmZmZmZmZmZm Z ddl Z ddl m m ZddlZ ddl m Z ddlmZddlmZdd lmZdd lmZdd lmZdd lmZdd lmZddlmZm Z m!Z!m"Z"ddl#m$Z$ddl%m&Z&ddl'm(Z(e!)e*Z+dZ,eGdddeZ-eGdddeZ.dd)d*Z;Gd+d,d,e j6ZGd1d2d2e j6Z?d3Z@ed4e@Gd5d6d6eZAd7ZBed4e@Gd8d9d9eAZCGd:d;d;eAZDdS)?zPyTorch Idefics model.) dataclass)AnyDictListOptionalTupleUnionN)nn)CrossEntropyLoss)PreTrainedModel)ACT2FN)*_prepare_4d_causal_attention_mask_for_sdpa) ModelOutput)PretrainedConfig)ALL_LAYERNORM_LAYERS)add_start_docstrings%add_start_docstrings_to_model_forwardloggingreplace_return_docstrings) IdeficsConfig)IdeficsPerceiverResampler)IdeficsVisionTransformerrc@s|eZdZUdZdZejed<dZe e e ejed<dZ e e ejed<dZ e e ejed<dZ e e ejed<dS)IdeficsBaseModelOutputWithPasta Base class for Idefics model's outputs that may also contain a past key/values (to speed up sequential decoding). Args: last_hidden_state (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`): Sequence of hidden-states at the output of the last layer 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 (`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 optionally if `config.is_encoder_decoder=True` 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 optionally if `config.is_encoder_decoder=True` in the cross-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + 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 optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(torch.FloatTensor)`, *optional*): Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver Nlast_hidden_statepast_key_values hidden_states attentionsimage_hidden_states)__name__ __module__ __qualname____doc__rtorch FloatTensor__annotations__rrrrrrr'r'P/tmp/pip-unpacked-wheel-zw5xktn0/transformers/models/idefics/modeling_idefics.pyr5s $rc@seZdZUdZdZeejed<dZ ejed<dZ ee ejed<dZ ee ejed<dZee ejed<dZee ejed<dS) IdeficsCausalLMOutputWithPasta Base class for Idefics causal language model (or autoregressive) outputs. Args: loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided): Language modeling loss (for next-token prediction). logits (`torch.FloatTensor` 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 (`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)`) Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see `past_key_values` input) to speed up sequential decoding. hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`): Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, + 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 optional initial embedding outputs. attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`): Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length, sequence_length)`. Attentions weights after the attention softmax, used to compute the weighted average in the self-attention heads. image_hidden_states (`tuple(torch.FloatTensor)`, *optional*): Tuple of `torch.FloatTensor` (one for the output of the image embeddings, `(batch_size, num_images, sequence_length, hidden_size)`. image_hidden_states of the model produced by the vision encoder, and optionally by the perceiver Nlosslogitsrrrr)r r!r"r#r*rr$r%r&r+rrrrrrr'r'r'r(r)bs  r)FcKsBt|jdddd|d|j}|d|}|dd|d<|dd|d<|dd|d<|dd|d<d|kr|d}|d||d<|dk r|d||d <|ddk r|dd||d<|ddk r|dd||d<nF|ddk r|dd||d<n"|ddk r:|dd||d<||fS) Nrr pixel_valuesimage_encoder_embeddingsperceiver_embeddingsimage_attention_masktoken_type_idsattention_mask) r$arangeshapeviewrepeattodevice index_selectget) input_idsZ expand_sizeis_encoder_decoderr2Zencoder_outputs model_kwargsZexpanded_return_idxr1r'r'r(expand_inputs_for_generations< ,    r>c Ks|dd}|rD|dddfd}|dk rD|dddfd}|dd}|dd}|dk r|dkr|dd}||dkd|r|dddfd}|dd}|dd}|d d}|d d} |d d } |||d ||||||| | d S)Nr1r,r2 position_idsrrr-r.r/r0interpolate_pos_encodingF use_cache) r;rrAr?r2r1r-r.r/r0r@)r: unsqueezelongcumsum masked_fill_) r;rkwargsr1r2r?r-r.r/r0r@r'r'r(prepare_inputs_for_generations:        rGcsftjtjtjdfdd|D}|D]4|rVtfdd|DrVdq,dq,|S)N) LayerNormLinear Embeddingcsg|] }|qSr'r').0m)mappingr'r( sz freeze_model..c3s|]}t|VqdSN) isinstance)rKt)moduler'r( szfreeze_model..TF)r rHrIrJmodulesanyrequires_grad_)modelmodule_exceptionsZmodule_exceptions_mappedr')rMrRr( freeze_models   rYcsDeZdZdZd eeddfdd ZddZed d d Z Z S) IdeficsDecoupledEmbeddinga Implements a decoupling of parameters to allow freezing (or not) a subset of the embeddings. In practise, the regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `num_additional_embeddings` > 0, then it will create `num_additional_embeddings` additional parameters that are always trained. If `num_additional_embeddings=0`, then the module defaults back to the regular behavior of `nn.Embedding`. FN)partially_freezereturnc  s|dk r$||kr$td|d|tjf|||||d|||_||_||_||_|rj|jd|jdkrt j |j|||d|_ dS)a) Args: num_embeddings (`int`): Size of the dictionary of embeddings num_additional_embeddings (`int`): Number of additional embeddings. Only useful when you `partially_freeze=True`. embedding_dim (`int`): The size of each embedding vector partially_freeze: (`bool`, *optional*, defaults to `False`): If `True`, the regular `weight` will be frozen. `additional_weight` is never frozen. padding_idx (`int`, *optional*): The padding index (needs to be less than num_embeddings) Note: there are a lot of other parameters to initialize a standard `nn.Embedding` such as `padding_idx`, `max_norm` or `norm_type`. We are not supporting these. Nz/padding_idx must be within num_embeddings. Got z and )num_embeddings embedding_dimr8dtype padding_idxFr)r]r^r8r_) ValueErrorsuper__init__r]r`num_additional_embeddingsr[weightrVr rJadditional_embedding) selfr]rdr^r[r8r_r`rF __class__r'r(rcs0  z"IdeficsDecoupledEmbedding.__init__cCsj|jdkrt||jS|}t||jk}||}|||j}d||<t||j}|||<|S)a we have 2 embeddings, with different indices - one pretrained self.weight and another self.additional_embedding.weight that is being trained. in order to make a lookup of the input ids, we: 1. find out the indices of the entries belonging to the 2nd embedding 2. extract those values while subtracting the size of the first embedding (num_embeddings), since the 2nd embedding starts from 0 and not num_embeddings 3. perform the 2nd embedding lookup 4. now we handle the 1st embedding, we overwrite indices belonging to the 2nd embedding with a padding index 5. perform the 1st embedding lookup 6. now we overwrite the values in the 1st embedding lookup with the values of the 2nd embedding lookup note: for the 1st embedding lookup we could have looked up only the low indices and not do the padding, but then we have to create a new tensor and populate it with 2 tensors that are spread out across various indices - i.e. not a simple concat - I haven't benchmarked the complex case if it's any faster, given that seqlens are usually relatively short it's probably not faster or if faster not by much - but might be a good idea to measure. r) rdFZ embeddingrecloner$wherer]rf)rgr;Zadditional_vocab_indicesZinput_ids_additional_vocabZadditional_embeddingsZ full_vectorr'r'r(forward-s z!IdeficsDecoupledEmbedding.forwardr\cCsd|j|j|j|jS)NzVnum_embeddings={}, num_additional_embeddings={}, embedding_dim={}, partially_freeze={})formatr]rdr^r[rgr'r'r( extra_reprTs z$IdeficsDecoupledEmbedding.extra_repr)FNNN) r r!r"r#rboolrcrmstrrq __classcell__r'r'rhr(rZs 5'rZcsTeZdZdZdeeeeeddfdd Zejejdd d Z e d d d Z Z S)IdeficsDecoupledLineara Implements a decoupling of parameters to allow freezing (or not) a subset of the parameters. In practise, the regular `weight` can be trained or frozen (i.e. `partially_freeze=True`), and if `out_additional_features` > 0, then it will create `out_additional_features * in_features` additional parameters that are always trained. If `out_additional_features=0`, then the module defaults back to the regular behavior of `nn.Linear`. rTN) in_features out_featuresout_additional_featuresbiasr[r\csnt|||||||_||_||_||_|rL|jd|rL|jd|dkrjt j |||||d|_ dS)aG out_additional_features: int. Number of additional trainable dimensions. Only makes sense when `partially_freeze=True`. partially_freeze: bool. If True, the regular `weight` will be frozen and extra parameters (if any) will be trainable. If False, default to the regular behavior of nn.Linear. Fr)rvrwryr8r_N) rbrcrxr[rvrwrerVryr rI additional_fc)rgrvrwrxryr[r8r_rhr'r(rcfs"  zIdeficsDecoupledLinear.__init__)inputr\cCs:t||j|j}|jdkr6||}t||fd}|S)Nrr,)rjZlinearreryrxrzr$cat)rgr{outputZadditional_featuresr'r'r(rms   zIdeficsDecoupledLinear.forwardrncCs d|j|j|j|jdk |jS)z=Overwriting `nn.Linear.extra_repr` to include new parameters.zYin_features={}, out_features={}, out_additional_features={}, bias={}, partially_freeze={}N)rorvrwrxryr[rpr'r'r(rqsz!IdeficsDecoupledLinear.extra_repr)rTTNN) r r!r"r#intrrrcr$Tensorrmrsrqrtr'r'rhr(ru]s  $ rucs.eZdZdfdd ZddZddZZS) IdeficsRMSNormư>cs&ttt||_||_dS)z= IdeficsRMSNorm is equivalent to T5LayerNorm N)rbrcr Parameterr$onesrevariance_epsilon)rg hidden_sizeepsrhr'r(rcs zIdeficsRMSNorm.__init__cCs\|tjdjddd}|t||j}|jjtj tj fkrR||jj}|j|S)Nr,T)Zkeepdim) r7r$Zfloat32powmeanZrsqrtrrer_Zfloat16Zbfloat16)rgrZvariancer'r'r(rms zIdeficsRMSNorm.forwardcCst|jjd|jS)Nz, eps=)tuplerer4rrpr'r'r(rqszIdeficsRMSNorm.extra_repr)r)r r!r"rcrmrqrtr'r'rhr(rs rcs0eZdZd fdd ZddZd dd ZZS) IdeficsEmbedding'Ncsxt||_||_||_d|jtjd|jdtjd ||j}|j d|dd|j ||j j tddS) Ng?rrr_inv_freqF persistentseq_lenr8r_)rbrcdimmax_position_embeddingsbaser$r3int64floatr7register_buffer_set_cos_sin_cacherr8Zget_default_dtype)rgrrrr8rrhr'r(rcs 0zIdeficsEmbedding.__init__cCs|||_tj|j|tjd|j}td||j}tj||fdd}|jd| |dd|jd| |dddS) N)r8r_zi,j->ijr,r cos_cachedFr sin_cached) max_seq_len_cachedr$r3rZtype_asrZeinsumr|rcosr7sin)rgrr8r_rQZfreqsZembr'r'r(rs z#IdeficsEmbedding._set_cos_sin_cachecCsN||jkr|j||j|jd|jd|j|jd|jd|j|jdfS)Nrr)rrr8r_rr7r)rgxrr'r'r(rms  zIdeficsEmbedding.forward)rrN)N)r r!r"rcrrmrtr'r'rhr(rs rcCsH|dd|jddf}|d|jdddf}tj| |fddS)z*Rotates half the hidden dims of the input..Nr,rr)r4r$r|)rx1Zx2r'r'r( rotate_halfsrcCsL|||}|||}||t||}||t||}||fS)anApplies Rotary Position Embedding to the query and key tensors. Args: q (`torch.Tensor`): The query tensor. k (`torch.Tensor`): The key tensor. cos (`torch.Tensor`): The cosine part of the rotary embedding. sin (`torch.Tensor`): The sine part of the rotary embedding. position_ids (`torch.Tensor`): The position indices of the tokens corresponding to the query and key tensors. For example, this can be used to pass offsetted position ids when working with a KV-cache. unsqueeze_dim (`int`, *optional*, defaults to 1): The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. Returns: `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. )rBr)qkrrr?Z unsqueeze_dimZq_embedZk_embedr'r'r(apply_rotary_pos_embs rcs.eZdZeeedfdd ZddZZS) IdeficsMLPrintermediate_size hidden_actcsNttj||dd|_tj||dd|_tj||dd|_t||_dS)NFry) rbrcr rI gate_proj down_projup_projr act_fn)rgrrrrhr'r(rcs  zIdeficsMLP.__init__cCs ||||||SrO)rrrr)rgrr'r'r(rmszIdeficsMLP.forward)r r!r"r~rsrcrmrtr'r'rhr(rs  rcseZdZdZdeeeeeedfdd Ze j eedd d Z de j e e j e e j e e j e ee j eeee j e e j e ee j fd d d ZZS)IdeficsAttentionz=Multi-headed attention from 'Attention Is All You Need' paperFN)r num_headsdropoutis_cross_attentionconfigqk_layer_normscst||_||_|||_||_d|_|j||jkrTtd|jd|d||_t t j dsntd|jrt |j ds|jn|j j }t j|j||jdd |_t j|||jdd |_t j|||jdd |_nNt j|j||jdd |_t j|j||jdd |_t j|j||jdd |_t j||j|dd |_t|j|_||_|jr~t|j|jd |_t|j|jd |_dS) NTz?hidden_size must be divisible by num_heads (got `hidden_size`: z and `num_heads`: z).scaled_dot_product_attentionz)this model requires pytorch 2.0 or higher embed_dimFrr)rbrcrrhead_dimr is_causalrarhasattrr functional vision_configrrIq_projk_projv_projo_projr rotary_embrr rms_norm_eps q_layer_norm k_layer_norm)rgrrrrrrZ kv_input_dimrhr'r(rcsh     zIdeficsAttention.__init__)tensorrbszcCs ||||j|jddS)Nrr)r5rr transpose contiguous)rgrrrr'r'r(_shape^szIdeficsAttention._shape)rkey_value_statesr2r?past_key_valueoutput_attentionsrAr\cCs|jp |dk }|\} } } ||| | |j|jdd} |s||| | |j|jdd} ||| | |j|jdd}nR|\} }} ||| ||j|jdd} ||| ||j|jdd}| j d}|dk r||dj d7}|s0|j |t || d\}}t | | |||\} } |dk rft j|d| gdd} t j|d|gdd}|rt| |fnd}|jr|| } || } |dk r|| d| |fkrtd| d| |fd|| jjd kr|dk r| } | } |}|jr$|dkr$| dkr$d nd }t jjj| | |||jrF|jnd |d }|| |j| |jfkrtd| |j| |jfd||dd}|| | |j}||}d}|rt d|||fS)Nrrr)rrz!Attention mask should be of size z , but is ZcudaTFr)Z attn_maskZ dropout_prz `attn_output` should be of size z^attn_weights are not extracted in scaled_dot_product_attention. The model returns None instead)!rsizerr5rrrrrr4rmaxrr$r|rrrrar8typerrr rrtrainingrZreshaperrlogger warning_once)rgrrr2r?rrrArrZq_len_Z query_statesZ key_statesZ value_statesZkv_lenZ kv_seq_lenrrrZ attn_outputZ attn_weightsr'r'r(rmasl ""$"      $   zIdeficsAttention.forward)rFNF)NNNNFF)r r!r"r#r~rrrrrcr$rrr LongTensorrrmrtr'r'rhr(rs<F rc sxeZdZedfdd Zd ejeejeejee ejee ee e ej ee ej ej ffdddZ Z S) IdeficsDecoderLayerrcspt|j|_t|j|j|j|d|_t|j|j|j d|_ t |j|j d|_ t |j|j d|_|j|_dS)N)rrrrrr)rbrcrrnum_attention_headsr self_attnrrrmlprrinput_layernormpost_attention_layernormrgrrhr'r(rcs  zIdeficsDecoderLayer.__init__NF)rr2r?rrrAr\c Cs|}||}|j||||||d\}}} tjj||j|jd}||}|}||}||}tjj||j|jd}||}|f} |r| |f7} |r| | f7} | S)a] Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states )rr2r?rrrApr)rrr rrrrr) rgrr2r?rrrAresidualself_attn_weightspresent_key_valueoutputsr'r'r(rms.      zIdeficsDecoderLayer.forward)NNNFF)r r!r"rrcr$rrrrrrr%rmrtr'r'rhr(rs rcseZdZedfdd Zd ejeejeejeejeejeeeeee eje ej ee ej ej ffd ddZ Z S) IdeficsGatedCrossAttentionLayerrcst|j|_t|j|jd|j||jd|_t|j|j |j d|_ t |j|j d|_t |j|j d|_|j|_t|_t|_|jdkr|jdkrttdd|j|_ttdd|j|_nB|jdkrttd|_ttd|_ntd |jd nV|jd kr|jdkr^ttdd|j|_ttdd|j|_nD|jdkrttd|_ttd|_ntd |jd n|jd krX|jdkrttjd |jdd|jfd|_ttjd |jdd|jfd|_nT|jdkrDttjd |jdd|_ttjd |jdd|_ntd |jd nt d|jdt!|drt!|dstddS)NT)rrrrrrrrzerosZvectorrrz Unknown value for `alpha_type` ()r>gaussiannormalrandomr)rstdrzAlpha initialization scheme z not yet implemented!alpha_cross_attn alpha_densez+Alpha parameters not initialized correctly!)"rbrcrrrrr cross_attnrrrrrrrrrr ZTanhact_cross_attn act_denseZalpha_initializerZ alpha_typerr$rrrrarrZalphas_initializer_rangeNotImplementedErrorrrrhr'r(rc sj            z(IdeficsGatedCrossAttentionLayer.__init__NF) rr2rr0cross_attention_gaterrArr\c Cs|dkrtd|dkr td|dk r0td|} ||}|j||||d\}} } tjj||j|jd}||dk d||dk<| | |j |}|} | |}| |}tjj||j|jd}| ||j|}|f} |r| | f7} |r| | f7} | S)a Args: hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` attention_mask (`torch.FloatTensor`, *optional*): attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. image_attention_mask (`torch.FloatTensor`, *optional*): image attention mask of size `(batch, 1, tgt_len, src_len)` where padding elements are indicated by very large negative values. cross_attention_gate (`torch.FloatTensor`, *optional*): gate of size `(batch, seq_len)` used to zero-out cross-attention output for tokens attending no images. output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states Nzt`image_hidden_states` is required for Idefics cross attention module which are visual features to be conditioned on.z`cross_attention_gate` is required for Idefics cross attention module to zero-out the cross-attention hidden_states attending to no images.zMPast key value states are not implemented for Idefics cross attention module.)rrr2rrr)rarrrr rrrrZfill_rrrrrr) rgrr2rr0rrrArrrrrr'r'r(rmLs@      z'IdeficsGatedCrossAttentionLayer.forward)NNNNFFN) r r!r"rrcr$rrrrrr%rmrtr'r'rhr(r s&D raL 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 ([`IdeficsConfig`]): 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. zSThe bare LLaMA Model outputting raw hidden-states without any specific head on top.c@sBeZdZeZdZdZddgZdZddZ e d e e dd d Z d S) IdeficsPreTrainedModelrWTrrcCs||jj}t|tjr>|jjjd|d|jdk rx|jj n:t|tj rx|jjjd|d|j dk rx|jj|j  dS)Nr)rr) rZinitializer_rangerPr rIredataZnormal_ryZzero_rJr`)rgrRrr'r'r( _init_weightss    z$IdeficsPreTrainedModel._init_weightsF)hard_check_onlyr\cCs"t|dd}|r|S|sd|_|S)NZuse_bettertransformerFZsdpa)getattrZ_attn_implementation)clsrrZ_is_bettertransformerr'r'r(_check_and_enable_sdpas  z-IdeficsPreTrainedModel._check_and_enable_sdpaN)F)r r!r"r config_classZbase_model_prefixZsupports_gradient_checkpointingZ_no_split_modulesZ_supports_sdpar classmethodrrrrr'r'r'r(rsraI 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) Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and [`PreTrainedTokenizer.__call__`] for details. If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see `past_key_values`). If you want to change padding behavior, you should read [`modeling_opt._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. - 1 indicates the head is **not masked**, - 0 indicates the head is **masked**. 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.n_positions - 1]`. [What are position IDs?](../glossary#position-ids) 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. use_cache (`bool`, *optional*): If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see `past_key_values`). output_attentions (`bool`, *optional*): Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned tensors for more detail. output_hidden_states (`bool`, *optional*): Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for more detail. return_dict (`bool`, *optional*): Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. cseZdZdZedfdd ZdddZgfdd Zgfd d Zd d Z ddZ e e de jee jee jeee jee jee jee jee jee jeeeeeeeeeeeeefdddZZS) IdeficsModelz Transformer decoder consisting of `config.num_hidden_layers` layers. Each layer is a [`IdeficsDecoderLayer`] Args: config: IdeficsConfig rcs t|_j|_j|_tjjjj |jd|_ j j |_ j |_ t j |_jrj}tj j|j|j|j|j|_tfddtjD|_j|_j|j}tfddt|D|_d|_t jj!d|_"|#|$dS)N)r]rdr^r[r`csg|] }tqSr')rrKrrr'r(rN7sz)IdeficsModel.__init__..csg|] }tqSr')rrrr'r(rN<sFr)%rbrcrZ pad_token_idr` vocab_sizerZadditional_vocab_sizerfreeze_text_layers embed_tokensrZ image_sizer vision_model use_resamplerperceiver_configrrZresampler_depthZresampler_n_headsZresampler_head_dimZresampler_n_latentsperceiver_resamplerr Z ModuleListrangeZnum_hidden_layerslayerscross_layer_intervalgated_cross_attn_layersgradient_checkpointingrrnorm post_initfreeze_relevant_params)rgrrZnum_cross_layersrhrr(rcsD     zIdeficsModel.__init__NcCs:|dkr|j}|jr ||j|jr6t|j|jddSN)rX)rrZfreeze_text_module_exceptionsfreeze_vision_layersrYrZfreeze_vision_module_exceptionsrr'r'r(rGs  z#IdeficsModel.freeze_relevant_paramscCs"|j|jfD]}t||dq dSr)rr rY)rgrXrRr'r'r(rQszIdeficsModel.freeze_text_layerscCst|j|ddSr)rYr)rgrXr'r'r(rUsz!IdeficsModel.freeze_vision_layerscCs|jSrOrrpr'r'r(get_input_embeddingsXsz!IdeficsModel.get_input_embeddingscCs ||_dSrOrrgvaluer'r'r(set_input_embeddings[sz!IdeficsModel.set_input_embeddingsF)r;r2r?r inputs_embedsr-r.r/r0rAroutput_hidden_statesr@ return_dictr\c(Cs$|dk r|jn|j}| dk r | n|jj} | dk r4| n|jj} | dk rH| n|jj} |dk r\|n|jj}|dk r~|dk r~tdn2|dk r|j\}}n|dk r|j\}}}ntd|}d}|dk r|ddjd}||}|dk r|dkr| dd}| |dkdn,|dkr>t j |||t j|d}| d}|||fddkr^tdn|dk r|j|j|d}|jdd\}}|j||f|jdd}|j|| d j}n<|dk r|\}}}}|j|j|d}|||||}|jjrF|dkr0||}|d|d}}n|\}}}}|}n*|dkrh|d|d}}ntd |||||}| d}| d} | ddd|} | ||||} |dk r|\}}}||f}| dkrt j||d } || } nd} | d kjdd j|jdjdd |}|dkr<||}|dkr^t j||ft j|jd}t|||f||}|}|j r|j!r| rt"#dd} | rdnd}| rdnd} | rdnd}!t$|j%D]\}"}#| r||f7}|dk r||"nd}$dd}%|j rT|j!rTd}$| r*t"#dd} |&|%|#||||$|| || | |"|j'|j(}&n&|%|#||||$|| || | |"|j'|j(d }&|&d}| r|!|&| rdndf7}!| r| |&df7} q|)|}| r||f7}| r|!nd}'|||||}|st*dd||'|| |fDSt+||'|| |dS)NzTYou cannot specify both decoder_input_ids and decoder_inputs_embeds at the same timezEYou have to specify either decoder_input_ids or decoder_inputs_embedsrrr,r)r_r8z_Exactly 1 of pixel_values, image_encoder_embeddings or perceiver_embeddings has to be not-None.)r-r@zBIf `perceiver_embeddings` are passed, use_resampler should be True)r8rrrzZ`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`...Fr'c  SsP| | dkr8| | | } | ||||||| dd}|d}||||||| d}|S)Nr)r2rr0rrrAr)r2r?rrrAr')Z main_blockrr2r?rrr0rrrA layer_idxr r Zxblockr layer_outputsr'r'r(vblocks,    z$IdeficsModel.forward..vblock) r2r?rrr0rrrArr r css|]}|dk r|VqdSrOr')rKvr'r'r(rSRsz'IdeficsModel.forward..)rrrrr),r8rrrrAuse_return_dictrar4rCrDrEr$r3rBcountr7r_rr5rrrrrr6rZinvert_attention_maskrUZsqueezerrrrr rrr enumeraterZ_gradient_checkpointing_funcr r r rr)(rgr;r2r?rrr-r.r/r0rArrr@rr8Z batch_sizeZ seq_lengthrZseq_length_with_pastZpast_key_values_lengthZ num_imagesrZ image_seq_lenZimage_hidden_sizeZ text_seq_lenZimage_batch_sizeZimage_sequence_lengthZimage_hidden_shaperrZall_hidden_statesZall_self_attnsZnext_decoder_cacheidxZ decoder_layerrrrZ next_cacher'r'r(rm^s:     "           "    )   zIdeficsModel.forward)N)NNNNNNNNNNNNFN)r r!r"r#rrcrrrrrrLLAMA_INPUTS_DOCSTRINGr$rrrrr%rrrrrrmrtr'r'rhr(r sN.    rcsLeZdZdgZddgZd"fdd ZddZdd Zd d Zd d Z ddZ ddZ ddZ e eeeedd#ejeejeejeeejeejeejeejeejeejeejeeeeeeeeeeeeefdddZd$ddZeddZd%ee e!e"feee e!e"fdfdd Z#ed d!Z$Z%S)&IdeficsForVisionText2Textzlm_head.weightzmodel.embed_tokens.weightNcs>t|t||_t|j|j|jd|jd|_ | dS)NF)rvrwrxryr[) rbrcrrWrurrrZfreeze_lm_headlm_headr )rgrrrhr'r(rcds   z"IdeficsForVisionText2Text.__init__cCs|jjSrOrWrrpr'r'r(rssz.IdeficsForVisionText2Text.get_input_embeddingscCs ||j_dSrOr$rr'r'r(rvsz.IdeficsForVisionText2Text.set_input_embeddingscCs|jSrOr#rpr'r'r(get_output_embeddingsysz/IdeficsForVisionText2Text.get_output_embeddingscCs ||_dSrOr%)rgZnew_embeddingsr'r'r(set_output_embeddings|sz/IdeficsForVisionText2Text.set_output_embeddingscCs ||_dSrOrW)rgdecoderr'r'r( set_decodersz%IdeficsForVisionText2Text.set_decodercCs|jSrOr(rpr'r'r( get_decodersz%IdeficsForVisionText2Text.get_decodercCs|}|}t|jddrL|j|_|jdkrL|j|jks@t|jj|j _t |drt |dr|j |_ t |drt |dr|j|_dS) z Overwrite `transformers.modeling_utils.PreTrainedModel.tie_weights` to handle the case of IdeficsDecoupledLinear and IdeficsDecoupledEmbedding. Ztie_word_embeddingsTrrwr]rxrdN) r&rrrrerdrxAssertionErrorrfrzrr]rw)rgZoutput_embeddingsZinput_embeddingsr'r'r( tie_weightss   z%IdeficsForVisionText2Text.tie_weights) output_typerF)r;r2r?rrr-r.r/r0labelsrArrr@rr\cCs| dk r | n|jj} | dk r | n|jj} |dk r4|n|jj}|j||||||||| | | | ||d}|d}||}d}| dk r@| |j} |dk r|dddf|j}|dddddf|dk}| dddf|dk}n.|dddddf}| dddf}t }|| d| d| d}|sp|f|dd}|dk rl|f|S|St |||j |j|j|jdS)a Args: 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]`. Returns: Example: ```python >>> from transformers import AutoProcessor, IdeficsForVisionText2Text >>> model = IdeficsForVisionText2Text.from_pretrained("HuggingFaceM4/idefics-9b") >>> processor = AutoProcessor.from_pretrained("HuggingFaceM4/idefics-9b") >>> dogs_image_url_1 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image1.jpeg" >>> dogs_image_url_2 = "https://huggingface.co/datasets/hf-internal-testing/fixtures_nlvr2/raw/main/image2.jpeg" >>> prompts = [ ... [ ... "User:", ... dogs_image_url_1, ... "Describe this image.\nAssistant: An image of two dogs.\n", ... "User:", ... dogs_image_url_2, ... "Describe this image.\nAssistant:", ... ] ... ] >>> inputs = processor(prompts, return_tensors="pt") >>> generate_ids = model.generate(**inputs, max_new_tokens=6) >>> processor.batch_decode(generate_ids, skip_special_tokens=True) ```N)r;r2r?rrr-r.r/r0rArrr@rr.rr,)r*r+rrrr)rrrrrWr#r7r8rr r5rr)rrrr)rgr;r2r?rrr-r.r/r0r/rArrr@rrrr+r*Zshift_attention_maskZ shift_logitsZ shift_labelsZloss_fctr}r'r'r(rmsX7   "z!IdeficsForVisionText2Text.forwardcKsj|dd}|dk r6|jjr&||d<n||d<d|d<t|fd|i|}dg}|D]}||dqT|S)Nrr/r.r-pastr1)poprrrG)rgr;r0rFrinputsZunwanted_kwargskwargr'r'r(rGs  z7IdeficsForVisionText2Text.prepare_inputs_for_generationcOs t||SrO)r>)argsr=r'r'r(_expand_inputs_for_generationsz7IdeficsForVisionText2Text._expand_inputs_for_generation)rr=r<standardize_cache_formatr\csTt||||}d|krF|d}|dddddfd}||d<|j|d<|S)Nr0r,rr)rb#_update_model_kwargs_for_generationrBr)rgrr=r<r6r0Z last_maskrhr'r(r7s z=IdeficsForVisionText2Text._update_model_kwargs_for_generationcs.d}|D] }|tfdd|Df7}q|S)Nr'c3s|]}|dVqdS)rN)r9)rKZ past_statebeam_idxr'r(rS8sz;IdeficsForVisionText2Text._reorder_cache..)r)r0r9Zreordered_pastZ layer_pastr'r8r(_reorder_cache4sz(IdeficsForVisionText2Text._reorder_cache)N)NNNNNNNNNNNNNFN)N)FF)&r r!r"Z_keys_to_ignore_on_load_missingZ_tied_weights_keysrcrrr&r'r*r+r-rr!rr)_CONFIG_FOR_DOCr$rrrrr%rrrrrmrG staticmethodr5rrrsrr7r:rtr'r'rhr(r"`sv   l    r")rFNN)N)r)Er#Z dataclassesrtypingrrrrrrr$Ztorch.nn.functionalr rrjZtorch.utils.checkpointZtorch.nnr r Z activationsr Zmodeling_attn_mask_utilsrZmodeling_outputsrZmodeling_utilsrZ pytorch_utilsrutilsrrrrZconfiguration_ideficsrZ perceiverrZvisionrZ get_loggerr rr;rr)r>rGrYrJrZrIruModulerappendrrrrrrrZLLAMA_START_DOCSTRINGrr!rr"r'r'r'r(st              ,+ - ' nB $ *L">Q